{ "created_at_utc": "2026-10-19T07:38:36.856340+00:00", "selection": [ { "task_id": "024c982_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_05_person_name", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Sarah—sorry, I mean Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "147d71fb5a9d20254c42a01b124c5f1107b0dd7d07c2971286a31fc14f58055b", "decision": "yes", "review_revision": "20261019T025444616630Z_bd988084", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Sarah' but immediately corrects to 'Stacy' with 'sorry, I mean Stacy'. The correction is clear and unambiguous, restoring the original recipient. All other elements—amount, app, privacy, and note content—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_05_person_name/efdc27ca6f93b461/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "d742f24dfecb5e0c85020339362f41f72872049653b0edf9b22c10f8d8abe3f6", "seconds": 12.16, "spoken_utterances": [ "Request ah... thirteen dollars publicly on Venmo from my friend, Sarah—uh... sorry, I mean um... Stacy, with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Sarah' but immediately self-corrects to 'Stacy' after a pause and 'sorry, I mean'. This is a false start where the final target is corrected. The rest of the instruction, including the amount ($13), platform (Venmo), action (request publicly), and note content, remains consistent with the original. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Request uh thirteen dollars publicly on Venmo from my friend, Sarah. Sorry, I mean um Stacy with a note, for yesterday's meal." }, "candidate_id": "wav_54abf11d1b9477d1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'uh' at the beginning, a clear self-correction from 'Sarah' to 'Stacy' with a believable 'sorry I mean um' transition, and a smooth continuation. The prosody is consistent, and the repair feels spontaneous rather than rehearsed. The pacing and intonation match a person recalling a name and correcting it mid-sentence." } ] }, "is_default": true } }, { "task_id": "024c982_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Request $13 on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\". Oh, right, make the request public.", "long_pause_offsets": [ 81 ], "source_sha256": "d8155c8575490977f377b884c195bdd05ef53c0ffe40d4d7444bfca0a006e316", "decision": "yes", "review_revision": "20261019T025445763315Z_f89b1a7f", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the public visibility requirement, which was missing in the initial instruction. The action, amount, recipient, app, and note content are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/addition/addition_01/dec885912b3e5af5/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "ec63faa34d440db58e99083a7ecd578a035cf9271d474d85c44eb72ccd4b93d1", "seconds": 15.68, "spoken_utterances": [ "Request um... thirteen dollars on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\". ", "Oh, right, make the request uh... public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the core request, then adds the constraint 'make the request public' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "Request um thirteen dollars on Venmo from my friend Stacy with a note, for yesterday's meal. Oh right, make the request uh public." }, "candidate_id": "wav_6a22ed62b4e864e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable self-correction. The initial request is stated, followed by a pause and a revision ('Oh, right, make the request uh public') that sounds like a spontaneous afterthought. The prosody is consistent, and the correction is integrated smoothly into the utterance with a natural 'uh' filler, indicating genuine recall rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "024c982_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_amount", "proposed_instruction": "Request $15 publicly on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\". Sorry, I meant $13, not $15.", "long_pause_offsets": [ 90 ], "source_sha256": "788cc2e5ee561b1f22e29ac402d6bcefdbde651049f1daa24980fc985cd69597", "decision": "yes", "review_revision": "20261019T025502946332Z_00f4218c", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the amount from fifteen to thirteen, while retaining the recipient, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_01_amount/ae96a5cb40b59d44/speech.wav", "wer": 0.12, "wav_sha256": "221d0511201e788c95a5f61b0eb4d7c296a41ffd3407c0a8e708f6d40efdb21c", "seconds": 14.52, "spoken_utterances": [ "Request ... fifteen dollars publicly on Venmo from my friend, Stacy, with a note, um... \"For yesterday's meal\". ", "Sorry, I meant ... thirteen dollars, not fifteen dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a request for $15, then immediately corrects it to $13. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent is to request $13, which matches the original instruction. All other conditions (Venmo, public, from Stacy, with note 'For yesterday's meal') are preserved and correctly stated.", "heard_transcript": "Request fifteen dollars publicly on Venmo from my friend Stacy with a note, um, for yesterday's meal. Sorry, I meant thirteen dollars, not fifteen dollars." }, "candidate_id": "wav_8eeae20b4cf754a0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural revision flow. The speaker starts with a clear request, then self-corrects the amount from 'fifteen' to 'thirteen' with a believable 'sorry, I meant' phrase. The prosody is smooth, the correction is immediate and integrated into the same utterance, and the intonation reflects a genuine moment of recollection without any robotic or rehearsed feel." } ] }, "is_default": true } }, { "task_id": "024c982_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "16b55d0d9462be7f7a67cc6d5cbc427abc63c8d7dcd1a1046abc53271717c8d2", "decision": "yes", "review_revision": "20261019T025520183338Z_71fcf681", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/cancellation/cancel_008/56363cc69e1fa64f/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "7c802b5b7eef87a3c706f52daf377eb00dc6742bf2669387eea1c2bfc2bec6f8", "seconds": 13.0, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\".", "Hold on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then clearly states 'Hold on, I want to cancel that task' as the final interjection. This constitutes a direct and unambiguous cancellation of the entire task. The cancellation is not conditional or followed by a new instruction, and the original request is not executed.", "heard_transcript": "thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal. Hold on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "0f3fb3f37e183e17e0980cc97bda0d839a1f4c0edf5f0beecd957111b77b1a46", "decision": "yes", "review_revision": "20261019T025534539915Z_63209bbc", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/ambiguous_halt/halt_031/4917d1f040f52eff/speech.wav", "wer": 0.1, "wav_sha256": "99c6eefb29bb837e4ae0549b99b962a9c2c95af59f6e5455620d94aebc7f264e", "seconds": 11.88, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\".", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary hold on the task. No cancellation or execution is requested, and the initial instruction is not revoked. The final control intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "cfccdd0b3fbcb65a5857972e08b51a625b6565ef9d2750fd2cd14bdb5059ebd5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/filler/original/bc775c4f62a2e6ff/speech.wav", "wer": 0.125, "wav_sha256": "cce9592f28ab6d9742aef44873163a10deea30167a7e027a86e02cc09eaf1ad7", "seconds": 8.44, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Request' and proceeds to state the full instruction without any hesitation, backchannel, or change in intent. The initial mention of 'thirteen dollars' is not corrected or revised, and the final utterance matches the original instruction in all key aspects: amount ($13), action (request), platform (Venmo), recipient (Stacy), visibility (publicly), and note content ('For yesterday's meal'). The 'filler' type allows for natural speech including initial phrasing, and the final goal is clear and complete.", "heard_transcript": "Request thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_recipient_relation", "proposed_instruction": "Request $28 privately on Venmo from my friend—no, I mean roommate, Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "af13df4ccc43ec9d0b0176fe17fb15b57bb8bbda3007e7b9acf059f1f4f7fba4", "decision": "yes", "review_revision": "20261019T050546672851Z_768cc5a7", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'my friend' but immediately corrects to 'roommate, Melissa'. The final recipient is clearly Melissa, and all other elements (amount, app, privacy, note) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_04_recipient_relation/211eb650985086e7/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "117023e276567f49c8071253bc57a9bd04c58065b9c061ce6541d972d3313d68", "seconds": 14.96, "spoken_utterances": [ "Request twenty-eight dollars privately on um, let's see... Venmo from my friend—uh... no, I mean roommate, um... Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my friend' but immediately self-corrects to 'roommate'. This is a false start, and the final target is correctly identified as 'roommate Melissa'. All other constraints (amount, app, privacy, note) match the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Request twenty eight dollars privately on um, let's see, Venmo from my friend uh, no, I mean roommate, um, Melissa, with a note, for the movie tickets." }, "candidate_id": "wav_2fd5927998dddc81", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow of a spontaneous self-correction. The speaker begins with 'my friend' and immediately self-corrects to 'roommate' with a believable hesitation and intonation shift. The pause before 'Melissa' sounds like genuine recall, and the overall pacing is consistent with someone formulating a request in real time. The correction is integrated smoothly into the utterance without any robotic or stilted delivery." } ] }, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Melissa. Oh, I forgot to mention, include the note \"For the movie tickets\".", "long_pause_offsets": [ 58 ], "source_sha256": "53b4884d3d1802b1c5f9d674926b7f60d8799c1091613efad6c3f78a1eb22739", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing note requirement; all other values including amount, privacy, app, recipient, and action are preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/addition/addition_01/5aa0f1abf0402ab6/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "6042f4392b7f9b3343719397ae5be583d6111e32d669346d3bc40cda53615c9a", "seconds": 16.16, "spoken_utterances": [ "Request twenty-eight dollars let me think... privately on Venmo from my um... roommate, uh... Melissa. ", "Oh, I forgot to mention, include the note \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Request $28 privately on Venmo from my roommate Melissa'. After a pause and a self-correction ('let me think'), the speaker adds the missing constraint 'include the note, for the movie tickets'. This is a clear addition to the original instruction, and the final goal matches the reference. The addition is not a revision or cancellation, and the initial request is not incorrect. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Request twenty eight dollars, let me think. Privately on Venmo from my um roommate uh Melissa. Oh, I forgot to mention, include the note, for the movie tickets." }, "candidate_id": "wav_34dcae5e91eb74ce", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, inserts a natural 'let me think' pause, then continues with a plausible self-correction ('privately on Venmo') and a believable addition ('include the note for the movie tickets') after a short hesitation. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition flows naturally as a remembered detail rather than a forced insertion." } ] }, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_05_person_name", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Jessica, with a note, \"For the movie tickets\". Change the recipient to Melissa, not Jessica.", "long_pause_offsets": [ 96 ], "source_sha256": "c56aa62c2840d8c219b2a14ffcbe876652045487c5e85720eca63f98cce41310", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the recipient from Jessica to Melissa, preserving the amount, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_05_person_name/472148c04c96fad8/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "7f8f44cf0ea956081db181e60d31d6dcba71133d956473a91fe21f80d89f5712", "seconds": 15.04, "spoken_utterances": [ "Request twenty-eight dollars privately on um... Venmo from my ... roommate, Jessica, with a note, \"For the movie tickets\". ", "Change the recipient to ... Melissa, not Jessica." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jessica' but immediately corrects to 'Melissa, not Jessica' after a pause. This is a revision where the final recipient is Melissa. The amount, app, privacy setting, note, and action type are unchanged. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Request twenty-eight dollars privately on um Venmo from my roommate Jessica with a note, for the movie tickets. Change the recipient to Melissa, not Jessica." }, "candidate_id": "wav_f2215ee3de06efab", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with a natural pause and filler ('um') before the correction. The self-repair ('Change the recipient to Melissa, not Jessica') is immediate, grammatically correct, and flows seamlessly within the same utterance, mimicking a spontaneous correction. The prosody is consistent and the pace is conversational." } ] }, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "25a734aff6be7293af33a73d8753ddf4e844797dfe405847ccb3f3c23965ca6c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/cancellation/cancel_007/1b283a3f0622d70e/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "3f744cae1602033e290e6f18dcae9395bd0adf41fcb880ea3972ed3f3ca7b009", "seconds": 14.84, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\".", "Oh, wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a complete Venmo request and then immediately follows with 'Oh wait, cancel my last request.' This constitutes a clear cancellation of the entire prior task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Request twenty eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets. Oh wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "f2a859c81ecc3b6208b633720c57e9315882123277dd4b8f01dfdc8512276f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/ambiguous_halt/halt_005/2d992d61154444ff/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "58ae9229e1b06cc64777f1f698eb04fcf19d3f57922250b69339edaa11c44ed7", "seconds": 13.72, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\".", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full request and then says 'One second,' which is an audible request to pause. There is no cancellation or instruction to resume. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Request twenty eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1b67d2ca20d3f9beae260210b5059fdee5980161eeedfa192c225a444ff4afad", "decision": "yes", "review_revision": "20261019T050429833050Z_49e74a94", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/filler/original/d74aafc3f5afe360/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "6ed293ab284caf86bea61001b964b5f573aca2c3b742eec18e416061420f5cb7", "seconds": 10.6, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler phrase 'uh let's see' but continues to state the full original instruction without any correction or change in intent. The final audible task matches the original instruction in action, target, value, privacy, app, note, and all other constraints. The filler does not alter the goal, so the execution criterion is satisfied.", "heard_transcript": "Request twenty-eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_visibility", "proposed_instruction": "Request $29 privately—no, I mean publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "dce5e6060c12d565f28cd3aabfa19dfa3162e722fc7c60ae9d7389dbdb80db6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'privately' to 'publicly' within the same utterance, preserving the amount, recipient, app, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_02_visibility/e1eb65485eff49c0/speech.wav", "wer": 0.08, "wav_sha256": "faefe4534fffd5001d03ba0afe3046bdb07df437afb8f7040cde157580ab4148", "seconds": 13.48, "spoken_utterances": [ "Request um... twenty-nine dollars privately—uh... no, I mean publicly on Venmo from my friend, Ronald, with a note, um... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately self-corrects to 'publicly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only a minor hesitation and repair that does not alter the core task.", "heard_transcript": "Request um, twenty nine dollars privately uh, no, I mean publicly on Venmo from my friend Ronald with a note, um, for the road trip." }, "candidate_id": "wav_35b88a292392e11f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then makes a clear false start ('privately') which is immediately self-corrected to 'publicly'. The correction is smooth and integrated into the flow. The speaker continues with a complete, coherent request, including the amount, platform, recipient, and note. The prosody is conversational and the pacing realistic, with a natural pause before the final phrase." } ] }, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Request $29 on Venmo from my friend, Ronald, with a note, \"For the road trip\". One more detail: make the request public.", "long_pause_offsets": [ 79 ], "source_sha256": "b9649362aed4cdf6b3a53a51e11abd32ece09b776e0b015ed7941451d7e4d73d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the public visibility requirement, which was missing in the initial instruction. The original instruction specified a private request, but the final instruction explicitly makes it public. The action, amount, recipient, app, and note content are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/addition/addition_02/062fa97dcc4829fe/speech.wav", "wer": 0.0, "wav_sha256": "148ec9a34ed49daf89712d7b7ec2ad3b6f176ca2cf92135effcc3e4eb527483e", "seconds": 16.76, "spoken_utterances": [ "Request twenty-nine dollars on um... Venmo from my friend, uh, let me think... Ronald, with a note, ... \"For the road trip\". ", "One more detail: make the request public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Request $29 on Venmo from my friend Ronald with a note for the road trip' and then adds 'One more detail, make the request public.' This is a clear addition of a new constraint to the original request. The final intent combines the original request with the added condition, matching the reference instruction. The addition is made after a pause and is clearly intended to modify the original task.", "heard_transcript": "Request twenty nine dollars on um Venmo from my friend uh let me think Ronald with a note for the road trip. One more detail, make the request public." }, "candidate_id": "wav_a59c7ab0304b75ba", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural hesitation ('um', 'let me think') and a believable additive structure. The initial request is clear, followed by a pause and a plausible addition of the 'public' detail. The prosody is consistent, and the delivery sounds like a spontaneous recollection rather than a rehearsed or robotic statement." } ] }, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_amount", "proposed_instruction": "Request $35 publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\". Sorry, I meant $29, not $35.", "long_pause_offsets": [ 88 ], "source_sha256": "d70e3c6ae8bda2a2e0fc301da23f165b7bc01f43882ceba02cd1e5a7e672339c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from thirty-five to twenty-nine, while retaining the recipient, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_01_amount/66bc8396a4531990/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "0f50ddc8ef33fa447ea535e7c32de9a01a21756504981133ae166335d8735ad8", "seconds": 18.04, "spoken_utterances": [ "Request thirty-five dollars publicly on Venmo from my um... friend, Ronald, with a note, uh... \"For the road trip\". ", "Sorry, I meant ah... twenty-nine dollars, not thirty-five dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a request for $35, then immediately corrects it to $29. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent matches the original reference instruction exactly, including the amount, recipient, app, visibility, and note. Therefore, the task is correctly executed as per the revision rule.", "heard_transcript": "Request thirty five dollars publicly on Venmo from my um friend Ronald with a note uh for the road trip. Sorry, I meant uh twenty nine dollars not thirty five dollars." }, "candidate_id": "wav_e92af7a5d30ae78b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a highly realistic false start and self-correction. The speaker begins with 'thirty-five dollars' and immediately corrects to 'twenty-nine dollars' with a natural-sounding 'sorry, I meant uh...' and a brief pause. The prosody is conversational, with a slight hesitation and a believable recovery. The correction is integrated smoothly into the ongoing utterance, making it sound like a spontaneous recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "12172c9bf46e41dc3a004ce26e63e74ffee128367386d0c0977ea5a8485227cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/cancellation/cancel_023/caeb5f0484c0f724/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "c9190ff93a1078677890ce179dc6962887fad5eba66e63a2878d02440dcbca49", "seconds": 12.88, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\".", "Wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the full Venmo request but then clearly cancels it with 'Wait, never mind about that. Don't carry out the request.' This constitutes a direct cancellation of the task. The final intent is to cancel, not to execute or pause.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip. Wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "06daf321fcc1dc3156f52c002477f382727e6a981c9cf1857bc04a34e6a389c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/ambiguous_halt/halt_018/5f508fe9a749b162/speech.wav", "wer": 0.12, "wav_sha256": "61f68e1871fd0f0491fd3ed511a4f4a197298756536d1132d52deaef77ceafd6", "seconds": 11.04, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\".", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which constitutes a clear instruction to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause request. The initial request is not canceled or revised, and the final control intent is to pause, not execute or cancel. No issues are present.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "024c982_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6f527b130a5be677360b7300ddb4092f42457b5c3c51c8bcd633e961676161d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/filler/original/2d9fc9b6a5e935b8/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "c49df5b350fbb8c7aa3afba879afa127003e10cdcdf729bafde5a36583f6f9cf", "seconds": 7.6, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'twenty nine dollars' and then corrects to 'twenty nine dollars' with a filler 'uh'. The correction is minor and does not change the value. The rest of the instruction matches the original. The filler 'uh' is a hesitation and does not alter the final intent. The task is executed as requested.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all the jazz—or rather, classical artists on Spotify that have at least 22 followers.", "long_pause_offsets": [], "source_sha256": "69474c5c4382449c4543eb8600c8e5b31eeaec4f870c14753fbff1620712f0e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately corrects to 'classical' within the same utterance. The genre is revised to match the original instruction, and all other conditions (Spotify, 'all artists', follower count of at least twenty-two) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/false_start/false_start_01_genre/f936f4d30b305c74/speech.wav", "wer": 0.05, "wav_sha256": "cc96b9de76f19ae82854119be1bf27601cc82bba417179a8f656ee168f5f7b26", "seconds": 10.64, "spoken_utterances": [ "Follow all the jazz—uh... or rather, classical artists on let's see... Spotify that have at least twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately self-corrects to 'classical' with 'or rather'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Follow all the jazz uh or rather classical artists on let's see Spotify that have at least twenty two followers." }, "candidate_id": "wav_49bc29705200039a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with 'jazz' as a false start, self-corrects to 'classical artists' with a smooth transition, and then continues with 'on let's see Spotify'—a plausible hesitation before naming the platform. The final phrase 'that have at least 22 followers' is delivered clearly and without further interruption. The pacing, intonation, and use of 'let's see' as a filler are all highly realistic and typical of spontaneous speech." } ] }, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Follow all the artists on Spotify that have at least 22 followers. Oh, right, only follow classical artists.", "long_pause_offsets": [ 67 ], "source_sha256": "0c9facaf6b13ea8693fd1bbd4700d090ab1313272be9ae32eaa21ea078862769", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup restricts the initial broad request to only classical artists, preserving the follower threshold and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/addition/addition_02/88e68aead7afff27/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "c03666993a23febdd456d8b8382e27aa47cbcc327135740d46d9fa93f5f80c78", "seconds": 14.6, "spoken_utterances": [ "Follow all the artists on um... Spotify that have at least uh, let's see... twenty-two followers. ", "Oh, right, only follow let me think... classical artists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is broad ('all the artists'), but the speaker adds a constraint ('classical artists') after a pause and self-correction. According to the 'addition' type criterion, the final goal combines the original instruction with the added information. The final audible intent matches the reference instruction exactly, including the action, target platform, condition, and value.", "heard_transcript": "Follow all the artists on um Spotify that have at least uh let's see twenty two followers. Oh right, only follow, let me think, classical artists." }, "candidate_id": "wav_a8645b20615d4539", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Follow all the classical artists on Apple Music that have at least 22 followers. Wait, I meant Spotify, not Apple Music.", "long_pause_offsets": [ 81 ], "source_sha256": "00ee65495fb84248e8aab98fd9c457b88546f2daaebd407fa9a66a61b4eb58ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, while retaining the genre, follower threshold, and all-artists scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/revision/revision_03_app/15e0433da6c546f8/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "9af098f840404d1638b5910c4c1c036cfa2b70dd8f43ba2805338743f01d1284", "seconds": 15.04, "spoken_utterances": [ "Follow all the classical artists on uh, let me think... Apple Music that have at least uh... twenty-two followers. ", "Wait, I meant Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the app (Spotify), genre (classical artists), and constraint (at least 22 followers). No other constraints were changed or lost.", "heard_transcript": "Follow all the classical artists on um, let me think, uh, Apple Music that have at least uh, twenty two followers. Wait, I meant Spotify, not Apple Music." }, "candidate_id": "wav_29e22b73c9ad783a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, multi-stage revision. They begin with a clear request, hesitate with 'um' and 'let me think', then correct the platform from 'Apple Music' to 'Spotify' with a believable 'wait, I meant...' self-correction. The prosody is smooth, the pauses are well-placed, and the correction flows naturally within the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "3f496dfaa97819e75bdd349dcba21666ffd4c754aa285f2579f587e31335aa32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/cancellation/cancel_037/21b4351908821574/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "f9ba6880d57b01d2e0cf07655e50f88fb676b6819cce8f3734df90cd6f12b0f5", "seconds": 12.8, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers.", "Oh, wait, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'Oh wait, stop there. Cancel the whole task.' This constitutes a final, unambiguous cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers. Oh wait, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "7f321dc0923b46c538efbc2cb648a6ab528c420796a98d3e8e033f318eacb477", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/ambiguous_halt/halt_033/7dd6a485dfca82e3/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2c5fccabc01828049c87801bfd634156c51cb9a4b98b4d1e36a2c69a47a7a6f3", "seconds": 11.8, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or resumption. The initial instruction is not executed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers. Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ca583bed72925ae08ca644967890116a8442c940597989164fcd035533bc1e97", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/filler/original/6dbffda2be7ca1eb/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "8ebefaea188be59a3d77bf0f31c5c4b0680cf01884995f06bd6a404cd7e55681", "seconds": 7.96, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core task. The final intent matches the original instruction, with 'twenty two' being a correct verbalization of '22'. No correction or cancellation is made. The task is executed as intended.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Follow all the edm artists on Apple Music—make that Spotify that have at least 23 followers.", "long_pause_offsets": [], "source_sha256": "564b4ee7fd4b9cface7c9ebf7754e5e5be8375e2c381beea3e7c5f8a1679583b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The genre (edm), action (follow), and follower threshold (twenty-three) are preserved in the final utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/false_start/false_start_02_app/67a83944424c082d/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "a2bf75807bfabcfcbc16270cba37831298c94ab60daae12ea5ead82488ec6bba", "seconds": 9.76, "spoken_utterances": [ "Follow all the edm artists on Apple Music—hmmm... make that Spotify that have at least ... twenty-three followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' with 'Make that Spotify'. This is a false start where the first value is replaced by the second. The final intent matches the original instruction exactly, including the platform (Spotify), genre (EDM), and follower threshold (23). The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Follow all the EDM artists on Apple Music. Hmm. Make that Spotify that have at least twenty three followers." }, "candidate_id": "wav_3badea15954ae75d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'Apple Music', hesitates with 'Hmm', then smoothly revises to 'Spotify' with a believable change-of-mind intonation. The phrase 'make that Spotify' is a pragmatic and common way to correct oneself in conversation. The pacing is consistent, and the correction flows seamlessly within the utterance without awkward pauses or stutters." } ] }, "is_default": true } }, { "task_id": "07b42fd_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Follow all the edm artists on Spotify. I didn't mention this, but they need to have at least 23 followers.", "long_pause_offsets": [ 39 ], "source_sha256": "f168f1d953197ac7ffc1d20cc601aa1e7a0aee44f8d2efab4284cd481bae359e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing follower threshold condition, which matches the original instruction's requirement. The genre (EDM), app (Spotify), and action (follow all) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/addition/addition_01/82c7014ccc571e11/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "c97dcf303235d1a20dc73b762c91f0ef76b4bcd9a749a1c3ed0731b6ff253ea0", "seconds": 13.72, "spoken_utterances": [ "Follow all the um... edm artists on ... Spotify. ", "I didn't mention this, but they need to have hmmm... at least twenty-three followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow all the um, EDM artists on Spotify.' The speaker then adds a constraint: 'they need to have, hmm, at least twenty three followers.' This is a clear addition of a missing condition to the original task. The final intent combines the initial request with the added constraint, matching the reference instruction. The hesitation and filler words ('um', 'hmm') do not change the final goal.", "heard_transcript": "Follow all the um, EDM artists on Spotify. I didn't mention this, but they need to have, hmm, at least twenty three followers." }, "candidate_id": "wav_125bbc0ce4216d16", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "07b42fd_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Follow all the edm artists on Apple Music that have at least 23 followers. Let me correct the app: follow artists on Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "4433ba67e190cbd7c7b4143bd0242536341a69f61966a237ea4a3500b8d24c08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the genre (edm), the action (follow), and the follower threshold (at least twenty-three)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/revision/revision_02_app/45600ea1034bc794/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "0da1589942166a51c5bd9f71e6fb0d534c2165a600c40f1c12ce2b8a69e36ffa", "seconds": 14.76, "spoken_utterances": [ "Follow all the ... edm artists on ... Apple Music that have at least uh... twenty-three followers. ", "Let me correct the app: follow artists on Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with only a minor pronunciation difference ('EDM' vs 'edm') that does not affect meaning.", "heard_transcript": "Follow all the EDM artists on Apple Music that have at least uh twenty three followers. Let me correct the app. Follow artists on Spotify, not Apple Music." }, "candidate_id": "wav_5d390cddcac5cc59", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural pause and filler ('uh'), then makes a self-correction ('Let me correct the app') that is smoothly integrated into the flow. The correction is pragmatic and spoken with a believable shift in intonation, and the final instruction is delivered with confidence. The prosody is consistent, and the speaker sounds like a real person recalling and adjusting their request." } ] }, "is_default": true } }, { "task_id": "07b42fd_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "7f629e5215633bf0bc6c5b48cc4084d808da0d688416d41cc9b01d81d186a36c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/cancellation/cancel_007/41114fb65870ca88/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "2a8c8af995d8de80be847c14d76f7fec74d656ac1476023debed42e87f065470", "seconds": 14.0, "spoken_utterances": [ "Follow all the let's see... edm artists on ... Spotify that have at least um... twenty-three followers.", "Wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a task to follow litzy EDM artists on Spotify with at least 23 followers. Immediately after, they say 'Wait, cancel my last request,' which is a clear and direct cancellation of the prior instruction. The final audible intent is cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "Follow all the litzy edm artists on Spotify that have at least um twenty three followers. Wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "fc7fefb2c89d9ccbe6ba05684530e5aec840661d50c50d791c67735c2b029ec6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/ambiguous_halt/halt_019/12996c7770e4a5e6/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "bb32dd1cd7f1a21753101cb39ac02e334670be5fd5fe2b5f6aac60a3c51aa86a", "seconds": 13.4, "spoken_utterances": [ "Follow all the let's see... edm artists on ... Spotify that have at least um... twenty-three followers.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Can you give me a moment?' is a clear request to pause or wait, with no indication of cancellation or resumption. The initial instruction is not fully executed, but the user's final audible intent is to halt the process temporarily. This matches the 'ambiguous_halt' type criterion, which requires only an audible pause request without permanent cancellation.", "heard_transcript": "Follow all the litzy edm artists on Spotify that have at least um twenty three followers. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all the rock—I meant reggae artists on Spotify that have at least 21 followers.", "long_pause_offsets": [], "source_sha256": "fc8359acc0b5d8fd8c9023007ea16341b175489da9026514456de3d24b4a4cbf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'rock' to 'reggae' within the same utterance, preserving the app, genre scope, and follower threshold." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/false_start/false_start_01_genre/5db66db409f77462/speech.wav", "wer": 0.1, "wav_sha256": "a126dddf5cb7c85dc9c7c26c5e8918aeaed7ad19ddf85f95987c8ce9d9946732", "seconds": 10.64, "spoken_utterances": [ "Follow all the rock—uh... I meant reggae artists on let's see... Spotify that have at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock artists' but immediately self-corrects to 'reggae artists'. The correction is clear and immediate, with no additional changes to the platform (Spotify) or the follower threshold (21). The final intent matches the original instruction exactly, satisfying the false_start criterion.", "heard_transcript": "Follow all the rock artists. I meant reggae artists on, let's see, Spotify that have at least twenty one followers." }, "candidate_id": "wav_c7f783ae7da25a65", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a clear false start ('rocka'), immediately self-corrects with 'I meant reggae artists', and then continues with a believable hesitation ('let's see') before completing the request. The intonation and pacing are consistent with someone thinking aloud, and the correction feels genuinely integrated into the utterance rather than tacked on." } ] }, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Follow all the reggae artists on Spotify. I didn't mention this, but they need to have at least 21 followers.", "long_pause_offsets": [ 42 ], "source_sha256": "09873f4614e02027b0cd1a4bb786b07030a9147785b50ab415184fd20db06647", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint about follower count, which matches the original instruction's requirement of at least 21 followers. The genre and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/addition/addition_01/6f73e2be0e49db80/speech.wav", "wer": 0.04, "wav_sha256": "3370852a0a6fe4b8ca9483960914211c6cac90084eb23d90f2752eb0d5149d7d", "seconds": 13.44, "spoken_utterances": [ "Follow all the uh... reggae artists on ... Spotify. ", "I didn't mention this, but they need to have let me think... at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, but the speaker adds the missing constraint 'at least twenty one followers' after a pause. This addition completes the original instruction, and the final intent matches the reference. The speaker's hesitation ('uh', 'let me think') is a filler and does not change the final goal.", "heard_transcript": "Follow all the uh reggae artists on Spotify. I didn't mention this, but they need to have let me think, at least twenty one followers." }, "candidate_id": "wav_9f5bbffa56521378", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('uh') and smoothly self-corrects to 'reggae artists'. The delivery has a credible, hesitant rhythm with a plausible pause and filler ('let me think') before adding the constraint about '21 followers'. The intonation and pacing suggest genuine spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_comparison_operator", "proposed_instruction": "Follow all the reggae artists on Spotify that have at most 21 followers. That should be 'at least', not 'at most'.", "long_pause_offsets": [ 73 ], "source_sha256": "ae5a41acfd0a270380e834744eb45f3d016f6b0033b502107d9d99abb0c9496b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the comparison operator from 'at most' to 'at least', preserving the genre, app, and follower count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/revision/revision_03_comparison_operator/9ab981e43308f200/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "89ee9e7395b9831b9902f66ae6e7527e5809b326485151afd4a9808de9e2b9d0", "seconds": 16.16, "spoken_utterances": [ "Follow all the reggae artists on ... Spotify that have at most hmmm... twenty-one followers. ", "That should be '... at least', not 'at most'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'at most twenty one followers' but immediately corrects themselves with 'That should be at least, not at most.' This constitutes a revision where the final intent is the corrected version. The original instruction was 'at least 21 followers,' and the final audible intent matches this exactly. Therefore, the revision criterion is satisfied.", "heard_transcript": "Follow all the reggae artists on Spotify that have at most twenty one followers. That should be at least, not at most." }, "candidate_id": "wav_73f37fbdf2edd329", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural hesitation ('hmm') before self-correcting the quantifier from 'at most' to 'at least'. The correction is smoothly integrated with a plausible self-aware phrase ('That should be...'), and the prosody reflects genuine thought and revision. The pacing and intonation are consistent with a person formulating and refining a request in real time." } ] }, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "c04b16dd228c62a94d66901aabebe700196bff66b3569fb4a26f593703997032", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/cancellation/cancel_015/70b6bf453eae480c/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "edec6a36d3c00675f2020b6fdea895cfad61d5f605ff18ecbffd5521790f3a7a", "seconds": 14.2, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers.", "Hang on, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'cancel the whole thing' after a brief interjection, which constitutes a direct cancellation of the previously stated task. The final intent is to cancel, not to execute the original instruction. The cancellation is unambiguous and applies to the entire task.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least 21 followers. Hang on, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "8c1a1940787a5c37084ff27baca2a296f4d5363369fdfc52d667d147465af368", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/ambiguous_halt/halt_018/52c4258d9c4d09c9/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "b9934c471271d08c05e48edbd5b80edf74b326f3a179e2b46b616b9847d02578", "seconds": 13.4, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which constitutes a clear instruction to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. No cancellation or execution is indicated, and the original task is not revoked. The final control intent is correctly identified as 'pause'.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least 21 followers. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07b42fd_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dd566f06b0462ba66e8f8f6352b4e6c7b5700fe5e8856353f905919f14f81365", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/filler/original/6465d420cac4c398/speech.wav", "wer": 0.2, "wav_sha256": "ba1b60b9f134075e559cc782024680648413949f900bc39bb7f7291661be7785", "seconds": 9.28, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'twenty one' which is a correct pronunciation of the number 21. There are no hesitations, corrections, or changes to the task. The final intent matches the original instruction, and the filler elements do not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least twenty one followers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_rating_condition", "proposed_instruction": "Move all products with under 4.5—sorry, I meant under 4.2 rating from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "42c2973493a240e10af382e53ca3d8fb7df0b6ac2fc7a19b18c3ba7543d74906", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'under four point five' but immediately corrects to 'under four point two'. The final value matches the original instruction, and all other conditions (action, source, destination, and scope) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/false_start/false_start_01_rating_condition/78210fb0a89e0a3f/speech.wav", "wer": 0.125, "wav_sha256": "4705c6887633ababde8b8e66722012ed71abf51a2902f44ad7d895d43d592b24", "seconds": 11.88, "spoken_utterances": [ "Move all products with ... under four point five—um... sorry, I meant under four point two rating from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'under four point five' but immediately self-corrects to 'under four point two'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The corrected value matches the original instruction exactly, and all other conditions (action, source, destination) are preserved. Therefore, the final intent is fully aligned with the expected task.", "heard_transcript": "Move all products with under four point five mm, sorry, I meant under four point two rating from my Amazon cart to wish list." }, "candidate_id": "wav_2951817d12f83f93", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker starts with a plausible false start ('under four point five'), immediately self-corrects with 'sorry, I meant under four point two rating', and continues smoothly. The correction is integrated into the sentence without awkwardness, and the prosody reflects genuine hesitation and recollection. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Move all products with from my amazon cart to wish list. Oh, I forgot to mention, only move products with under 4.2 rating.", "long_pause_offsets": [ 57 ], "source_sha256": "dcd2978c15bfb6e3f618519d79f9e5c2d48396ddf2105f2d4f16779c641e1fce", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rating condition, which matches the original instruction's threshold and comparison. The initial utterance is incomplete but the final goal is fully specified after the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/addition/addition_01/9607ff121bdc7a8f/speech.wav", "wer": 0.08, "wav_sha256": "fa9406d4f47e8aaa0b124dca8cbf9bad139f968bf53865f404fffcd6492118f9", "seconds": 12.76, "spoken_utterances": [ "Move all products with ... from my amazon cart ... to wish list. ", "Oh, I forgot to mention, only move products with under four point two rating." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('Move all products with from my Amazon cart...'), but the speaker immediately adds the missing constraint ('only move products with under four point two rating'). This constitutes a valid addition, where the final goal combines the original intent with the added condition. The final audible intent matches the reference instruction exactly.", "heard_transcript": "Move all products with from my Amazon cart to wish list. Oh, I forgot to mention, only move products with under four point two rating." }, "candidate_id": "wav_fceba557fd98eb1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then adds a crucial detail ('only move products with under 4.2 rating') in a natural, additive manner. The 'oh I forgot to mention' phrase is a highly plausible real-world hesitation, and the correction is smoothly integrated without awkwardness. The prosody is consistent, and the pacing feels like genuine spontaneous speech." } ] }, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_collection", "proposed_instruction": "Move all products with under 4.2 rating from my amazon cart to saved items. Use my wish list as the destination instead of saved items.", "long_pause_offsets": [ 76 ], "source_sha256": "f57d1ddb6ceee1083ed95c41d3d1de730b33ed2c93abe44ede9c00c41aa5ebb4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup correctly revises the destination from 'saved items' to 'wish list' while preserving the original condition on product rating and source cart. The app name 'amazon' is retained implicitly from the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/revision/revision_03_destination_collection/c0deaac747c7512b/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "d618ba4a0a8def3cf46c89e69a3e8e0d289e3c4039f26b04cfeee532780ed333", "seconds": 16.64, "spoken_utterances": [ "Move all products with under four point two rating from my amazon cart ... to saved items. ", "Use my uh... wish list as the destination instead of ... saved items." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to saved items' but immediately corrects to 'use my uh wish list as the destination instead of saved items'. This constitutes a revision where the final destination is clearly stated as 'wish list'. The original instruction's intent is preserved with the corrected target, satisfying the revision type criterion.", "heard_transcript": "Move all products with under four point two rating from my Amazon cart to saved items. Use my uh wish list as the destination instead of saved items." }, "candidate_id": "wav_d15ea1d831b8ea2e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a clear, single revision. The speaker first states a complete command, then hesitates with 'use my uh' before correcting the destination from 'saved items' to 'wishlist'. The correction is smooth, the intonation natural, and the pause feels like a genuine moment of recall. The prosody is consistent and the delivery is fluent without any awkwardness." } ] }, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e3845e31cf2ce959fa0b22257767bb3f53e3bea10a8e2887f916db46221b82bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/cancellation/cancel_009/95384b1e187e696d/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "ebf5a79887aa7cce0500496529f60a740d7b00de567a22c105fd2f9209ce0dc2", "seconds": 16.28, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list.", "Wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to move products with under 4.2 rating from the Amazon cart to the wish list. After a brief pause and hesitation ('um'), they immediately say 'Wait, cancel that altogether,' which is a clear and final cancellation of the entire task. The final interjection is a direct cancellation command, satisfying the 'cancellation' type criterion.", "heard_transcript": "Move all products with a under four point two rating from my amazon cart um to wish list. Wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "f56fcb87f634b1bbeda9c60ef54769de47d59ef247dbf1c1ea410204c0c3578d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/ambiguous_halt/halt_037/d36f76965034e14f/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "95793615b1d716db9879205bc190051de0f364060e8c772e555d203de68b49a5", "seconds": 15.2, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Pause for a moment,' which constitutes a clear request to halt the current operation. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and complete, and the final control request is unambiguous.", "heard_transcript": "Move all products with a rating under four point two from my Amazon cart um to wish list. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "232c9eb7e17fc3c4fc75a7e689c9046d4c507e60abed62b32b6758269d60a27b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/filler/original/cc86bbed06e40c4c/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "499e9c5468cbc86eaf2caaa7cd5102312aaa4c706468dddf8a95e2f6f6ed02d1", "seconds": 11.76, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core task. The final intent matches the original instruction, with no self-correction or modification to the action, target, or value. The hesitation does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Move all products with uh under four point two rating from my amazon cart um to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_rating_condition", "proposed_instruction": "Move all products with over 4.2—no, that should be over 3.7 rating from my amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "6e8827eded2214d69a5685393cddc5dba102ecf42ff067f6ee1be079a5657eb3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four point two' but immediately corrects to 'over three point seven'. The final value matches the original instruction's rating threshold." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/false_start/false_start_01_rating_condition/7d6893f120fc1be8/speech.wav", "wer": 0.25925925925925924, "wav_sha256": "5129801211675ea476f884e543d085dfbc03ec7ae00f0d7f111a527ec1723534", "seconds": 11.72, "spoken_utterances": [ "Move all products with ah... over four point two—uh... no, that should be over three point seven rating from my amazon wish list um... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four point two' but immediately self-corrects to 'over three point seven'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is executed with the corrected value. The rest of the instruction matches the original, so the final intent is correct.", "heard_transcript": "Move all products with a uh over four point two uh um no, that should be over a three point seven rating from my Amazon wish list um to cart." }, "candidate_id": "wav_cbf0f83ab44c4416", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. The speaker starts with 'a uh over four point two uh um' as a false start, then clearly corrects to 'no that should be over three point seven rating'. The correction is immediate, the intonation reflects a genuine moment of realization, and the flow remains continuous without awkward pauses. The speaker also correctly identifies the source as 'my Amazon wishlist' and the target as 'to cart', with a natural final pause. The repair is fully integrated into the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Move all products with over 3.7 rating to cart. Oh, and one more thing: move them from my Amazon wish list.", "long_pause_offsets": [ 48 ], "source_sha256": "55b850c1c59f227b867f264128f2e378070217b6d1574a4830f006d922b171c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source location 'Amazon wish list'; the action, condition, and destination remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/addition/addition_02/9064a585521a9970/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "5e8a3b363e49c3285a94e30b4e4c9de97cac7a72bde9b54973708e3bf6d4e078", "seconds": 14.88, "spoken_utterances": [ "Move all products with uh... over three point seven rating um, let me think... to cart. ", "Oh, and one more thing: move them let's see... from my Amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Move all products with a over three point seven rating um let me think to cart.' This is followed by a longer pause and an addition: 'Oh and one more thing, move them let's see from my Amazon wish list.' The addition clarifies the source of the products as 'my Amazon wish list,' which combines with the initial instruction to form the complete and correct final goal. The type is 'addition,' and the final combined goal matches the original instruction exactly. Therefore, the task is passed.", "heard_transcript": "Move all products with a over three point seven rating um let me think to cart. Oh and one more thing, move them let's see from my Amazon wish list." }, "candidate_id": "wav_1a8986f7954b045c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and realistic flow of spontaneous speech. It begins with a clear initial request, includes a believable 'let me think' filler pause, then transitions into a revision with 'Oh and one more thing' and a follow-up correction that reorders the original request. The prosody is varied, with natural hesitations and a pragmatic change-of-mind tone, making the self-correction sound entirely spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Move all products with over 3.7 rating from my ebay wish list to cart. Let me correct the app: move products from my amazon wish list, not ebay.", "long_pause_offsets": [ 71 ], "source_sha256": "f98ec93bb09bab49d7c7f87e60b0c2c8bfdd96b5712ec618af0f0f6db5f9f212", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, while retaining the rating threshold and the action of moving items from the wish list to the cart." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/revision/revision_02_app/8ab0bc15a242c052/speech.wav", "wer": 0.14705882352941177, "wav_sha256": "a3c500fcce216ea42fe293e6edba7d41fd8ff6895eb10d90429309482adcd608", "seconds": 16.84, "spoken_utterances": [ "Move all products with let me think... over three point seven rating ... from my ebay wish list to cart. ", "Let me correct the app: move products from my ah... amazon wish list, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the app as 'eBay' but immediately self-corrects to 'Amazon' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final intent matches the original instruction exactly, including the rating threshold and target action.", "heard_transcript": "Move all products with let me think over three point seven rating from my eBay wish list to cart. Let me correct the app. Move products from my uh Amazon wish list, not eBay." }, "candidate_id": "wav_fef209ede3eb6d7c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a genuine 'let me think' filler, then self-corrects the platform from 'eBay' to 'Amazon' with a smooth, spontaneous-sounding 'let me correct the app' phrase. The prosody is conversational, with appropriate pauses and intonation that suggest real-time formulation and correction. The correction is integrated into the ongoing utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "7e4a1cb0c95c4149027d298b190354ca405c619a2b1bff109f152b9efa805d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/cancellation/cancel_023/ef3ccda604419eb6/speech.wav", "wer": 0.15625, "wav_sha256": "dff5013240d030c983ef2d810e18f4a9e14fe035b03b1f1573e58e71565aeabb", "seconds": 15.68, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart.", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and hesitation, explicitly cancels it by saying 'Hang on, never mind about that. Don't carry out the request.' This constitutes a clear cancellation of the entire task as per the type-specific criterion. The final intent is to cancel, not execute or pause.", "heard_transcript": "Move all products with let me think over three point seven rating um from my Amazon wish list uh to cart. Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "de07c46d8db0ff2d90e9dbd5816d2e84176ba699457ea996bdc57f7b6e7a8bc7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/ambiguous_halt/halt_028/5a91cb5767249c71/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "18019c6161b42b144cc9db90c0a2a458eaed64c2734befc6fa42d751d2b8bbb0", "seconds": 14.48, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled or revised, and the final audible intent is to pause, not execute or cancel.", "heard_transcript": "Move all products with let me think over three point seven rating um from my Amazon wish list uh to cart. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "336e340e8052df70fffb8992fc7cbfacf0558947a1f8559d79bda1b92645a30e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/filler/original/7989ac46b51b51a2/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "4592d908b6b759dea4b72730f80cb396a294c752459038869df696f7087a779a", "seconds": 10.08, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('let me think', 'um', 'uh') but does not change the core instruction. The final audible intent matches the original instruction in action, target, and value. The hesitation does not alter the goal, so the filler type criterion is satisfied.", "heard_transcript": "Move all products with, let me think, over three point seven rating, um, from my Amazon wish list, uh, to cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all products with over 4.3 rating from my amazon cart to shopping list—wait, I meant wish list.", "long_pause_offsets": [], "source_sha256": "46f36c50eeb5059cec46038550905daf44b25c935cde578fc7cc69add1176296", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects to 'wish list'. The correction is clear and unambiguous, preserving the app ('amazon'), source ('cart'), condition ('over four point three rating'), and final destination ('wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/false_start/false_start_04_destination_collection/496bf72d8ee44da9/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "bc1dae12ed8e7ff21c5d7c5c3ab767b976ce05f201d1b0c44160bf8b810f0eac", "seconds": 10.84, "spoken_utterances": [ "Move all products with over four point three rating um... from my amazon cart uh... to shopping list—hmmm... wait, I meant wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wish list' after a pause and 'Wait, I meant...'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Move all products with over four point three rating um from my Amazon cart uh to shopping list. Wait, I meant wish list." }, "candidate_id": "wav_1bd73c6824059239", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complete false-start pattern. The speaker begins with a clear request, hesitates with 'um', then makes a self-correction from 'shopping list' to 'wishlist' with a believable 'wait, I meant wishlist' phrase. The intonation and pacing are consistent with spontaneous speech, and the correction is integrated smoothly into the flow without sounding forced or robotic." } ] }, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Move all products with from my amazon cart to wish list. Oh, I forgot to mention, only move products with over 4.3 rating.", "long_pause_offsets": [ 57 ], "source_sha256": "ab7dce23211267dacb66cb692d16f71a28f4d46c1e1315d459621e47f4c8ab66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rating condition, which matches the original instruction's threshold and comparison operator. The initial utterance is incomplete but the final goal is fully specified after the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/addition/addition_01/399012c9017166a3/speech.wav", "wer": 0.2413793103448276, "wav_sha256": "1a3b0a0db95c80415208e9d939b18d169f1c763f4eaae9503f9f35b1a8dcf90a", "seconds": 13.96, "spoken_utterances": [ "Move all products with uh... from my amazon cart uh, let's see... to wish list. ", "Oh, I forgot to mention, only move products with ... over four point three rating." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially begins the task without the rating constraint, then adds the constraint 'only move products with over four point three rating' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added condition. The final audible intent fully satisfies the reference instruction.", "heard_transcript": "Move all products with uh from my Amazon cart uh let's see to wish list. Oh, I forgot to mention, only move products with over four point three rating." }, "candidate_id": "wav_5fc355aa7bc0de3a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_rating_condition", "proposed_instruction": "Move all products with over 4.5 rating from my amazon cart to wish list. Wait, make that over 4.3 rating instead of over 4.5.", "long_pause_offsets": [ 73 ], "source_sha256": "f274e2ff2a99a31707e30888697ea07c1b5c85e8b1cdbd365a18837750bf8a9a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the rating threshold from four point five to four point three, while preserving the app (amazon), action (move), source (cart), destination (wish list), and scope (all products)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/revision/revision_01_rating_condition/0cc93dc1fc68f08e/speech.wav", "wer": 0.24242424242424243, "wav_sha256": "f87faf3e4c59eb785e3f89d0698f5ae4a8a933e1f4dcf9359507328786e3b5c3", "seconds": 16.24, "spoken_utterances": [ "Move all products with hmmm... over four point five rating um... from my amazon cart uh... to wish list. ", "Wait, make that over four point three rating instead of over four point five." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four and a half rating' but immediately corrects it to 'over four point three rating instead of over four point five'. This is a clear revision where the final value is 'four point three'. The final intent matches the original instruction exactly, so the revision criterion is satisfied.", "heard_transcript": "Move all products with um, mm, over four and a half rating, um, from my Amazon cart, uh, to wish list. Wait, make that over four point three rating instead of over four point five." }, "candidate_id": "wav_eef981b20494836e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural revision flow. The speaker initially states 'over four and a half rating' with a clear hesitation ('um', 'mm'), then corrects to 'over four point three rating' with a believable 'wait, make that...' structure. The prosody includes realistic fillers and a smooth, self-correcting intonation pattern. The revision is integrated into a single continuous utterance, making it sound like a spontaneous thought process rather than a rehearsed or robotic correction." } ] }, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "4c1be07c7da554fb8f0b701f5b056b7c8b361b90cabdff89a9e8835149c22268", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/cancellation/cancel_024/3de9e57a337feff6/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "e54520d87f520d7290ee054b4a636f1bce1fa65ecdafc5633a2e3e37cfe4b1af", "seconds": 14.84, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list.", "Hold on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and hesitation, explicitly says 'Hold on. Oh, never mind. Cancel the whole thing.' This constitutes a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, not a request to pause or an ambiguous statement. Therefore, the control intent is correctly identified as 'cancel' and the verdict is 'pass'.", "heard_transcript": "Move all products with uh over four four point three rating um from my Amazon cart um let's see to wish list. Hold on. Oh, never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "2371063a5d75e9f59ba8cc0651af0978bc3d44c6f7c4ea5a11a6759b7fa7e975", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/ambiguous_halt/halt_029/032bbc09805c2924/speech.wav", "wer": 0.2608695652173913, "wav_sha256": "9fd30e776f2a7ecaa0b1db71271ebb4499f7fd39750fdee53c41a16a3885afcf", "seconds": 14.04, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, actually, well', which is a clear interjection indicating the speaker is reconsidering or hesitating before completing the task. This constitutes an audible request to pause, with no explicit cancellation or resumption. The final interjection is not a command to execute or cancel, but a natural hesitation. Therefore, the control intent is 'pause', matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Move all products with uh over four four point three rating um from my Amazon cart um let's see to wish list. Wait, actually, well" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "07bb666_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8776435f1412a21ff8521077c9e71eb9e7a303a68d790b844c5b3cb72f04ae37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/filler/original/d347c3b6729a8ac2/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "2469b00834efa8c1ce2d5057f37247f0a7e3ddaf4b69a4b68e343fe357fb272e", "seconds": 9.84, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates and repeats 'four' before stating 'four point three', which is a self-correction within the same value. The final audible value is 'four point three', matching the original instruction. The filler words and pauses do not alter the task goal. The final interjection 'let's see' is a backchannel, not a change in intent. The task is executed as intended.", "heard_transcript": "Move all products with uh over four four point three rating um from my amazon cart um let's see to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery list—wait, I meant Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "04234e4dc78483939a3242c88d25e4f7e9c0414e3f06afd84c8afc18357a52d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately corrects it to 'Amazon shopping list note in SimpleNote'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/false_start/false_start_02_note_description/ec418b26e0b8d636/speech.wav", "wer": 0.2, "wav_sha256": "004379592ec4d63f872e12d41b86f1014ee7b5114da618103e4e1b854f253d36", "seconds": 8.64, "spoken_utterances": [ "Place an Amazon order uh... as per my grocery list—um... wait, I meant ... Amazon shopping list note in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately self-corrects to 'Amazon shopping list note in SimpleNote'. According to the 'false_start' type criterion, the final corrected version is the intended task, and the initial incorrect value is discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Place an Amazon order uh as per my grocery list um wait I meant Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_708a89c97da9c380", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The false start 'uh' and 'wait' are placed at a plausible point of hesitation, and the self-correction 'I meant Amazon shopping list note in Simple Note' flows smoothly without awkwardness. The intonation and pacing are consistent with someone recalling a detail and correcting themselves in real time." } ] }, "is_default": true } }, { "task_id": "099174d_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Gmail. Look for the note in SimpleNote instead of Gmail.", "long_pause_offsets": [ 68 ], "source_sha256": "5c1a027e361802a99ecc8f1ce640953a7235aa36afb9afb14802826342aa1d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to SimpleNote, preserving the action, the note content, and the intent to use the shopping list note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/revision/revision_03_note_app/f9cdc60d8291ec57/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "0082297b1c54de299a315d7fa297050dd80bcae689824f524c08defe915fde4b", "seconds": 15.52, "spoken_utterances": [ "Place an Amazon order as per my Amazon shopping list note in uh, let me think... Gmail. ", "Look for the note in um... SimpleNote instead of hmmm... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Gmail but then corrects themselves to SimpleNote after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use SimpleNote, which matches the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh let me think. Gmail. Look for the note in um SimpleNote instead of hmm Gmail." }, "candidate_id": "wav_e448c3c8f4114c3d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, then uses a believable 'let me think' filler pause before correcting the app from Gmail to Simple Note. The correction is integrated smoothly into the ongoing sentence with a natural intonation shift and a plausible self-repair ('instead of Gmail'), showing a realistic thought process without robotic or stilted delivery." } ] }, "is_default": true } }, { "task_id": "099174d_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "07c52099fc5fd2832530a1e2a5d1c7849e1f5afb051e4f9cc11ff5a36f557ab3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/cancellation/cancel_015/138310da73c6abf8/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "6c9196654d1c3ed7923b4ac0014e146469dc3f31d3d3599ab258480c9c411bc5", "seconds": 13.96, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote.", "Wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Wait, actually, cancel the whole thing.' This constitutes a final cancellation of the entire task, satisfying the cancellation type criterion.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote. Wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "ea7936941688af9f86472d9173b261abb7cf8114a41810810b93440326e442f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/ambiguous_halt/halt_006/c32059568f610b21/speech.wav", "wer": 0.1, "wav_sha256": "aeeddad6d538d9460ef72e5ba4ddf1c1d941b6cea8c1820044fc8da4d404246b", "seconds": 12.48, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment,' which is a clear request to pause or wait. The user does not cancel the task or provide a new instruction. The initial request is still present but not executed due to the pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote. Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c864376aec3812bc2b8183e13e5f4c89a11b0fa29fe921a9d6a32ceca477600c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/filler/original/4c54c5a587f1383a/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "214d231c9e3d43d078095db0aa268d5944810bca90292543ab8c7859269ee2fb", "seconds": 9.36, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'uh, let me think' but does not change the core request. The final intent remains to place an Amazon order based on the shopping list note in SimpleNote. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery list—sorry, I meant Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "c581a47cb4e99600db33b7e319000aed407700005500b35a7a5d1505891a433d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately corrects it to 'Amazon shopping list note in SimpleNote'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/false_start/false_start_02_note_description/8434faf899445e9f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4360a722d994ff66adad1bd357c7be8bbf09e5db8859ce97b87ee0eeae74aae7", "seconds": 11.56, "spoken_utterances": [ "Place an Amazon order um... as per my grocery list—uh... sorry, I meant um... Amazon shopping list note in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately self-corrects to 'Amazon shopping list note in SimpleNote'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly, with no other discrepancies.", "heard_transcript": "Place an Amazon order um as per my grocery list uh sorry, I meant um Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_db240054707f0732", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'Amazon order' and 'grocery list', then self-corrects to 'Amazon shopping list note in SimpleNote'. The correction is smooth, with a believable hesitation ('uh') and a clear, unforced shift in the second clause. The prosody is consistent, and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "099174d_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_note_description", "proposed_instruction": "Place an Amazon order as per my Grocery list note in SimpleNote. Make that the Amazon shopping list note, not the Grocery list.", "long_pause_offsets": [ 65 ], "source_sha256": "25213f62f7ea6a8dfe2d6b40c3f218b36bc249b773b55f2796336665e7936cb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the note name from 'Grocery list' to 'Amazon shopping list'; the app (SimpleNote), action (place Amazon order), and source (note) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/revision/revision_02_note_description/ff4c9fbc1ef5a07d/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "534be654f793229ee9aa146863a1420f6bb3ef75d40b7cce99992b01ce91ddb5", "seconds": 16.08, "spoken_utterances": [ "Place an Amazon order as per my Grocery list note in uh, let's see... SimpleNote. ", "Make that the let me think... Amazon shopping list note, not the Grocery list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list note' but immediately corrects to 'Amazon shopping list note' after a pause and self-correction ('Make that the...'). The final instruction clearly revises the target note from 'grocery list' to 'Amazon shopping list'. The app 'SimpleNote' remains unchanged. The revision is clear and final, satisfying the revision type criterion.", "heard_transcript": "Place an Amazon order as per my grocery list note in uh let's see. SimpleNote. Make that the let me think. Amazon shopping list note, not the grocery list." }, "candidate_id": "wav_3df5ac7f054762dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, hesitates with 'uh' and 'let me see' as if recalling details, then smoothly self-corrects from 'grocery list' to 'Amazon shopping list' with a believable 'make that' and 'let me think' filler. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "099174d_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "7413206c8dccfa772091ba15e7ddb20ef29644e4ad30933f8318afdfe13da528", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/cancellation/cancel_011/fce2315b623d2c8c/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "48260645c5b82c57098dc524d2dc5af7dce6bf594d27b7e6d74f17bc8aa3c31e", "seconds": 14.2, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote.", "Hold on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to place an Amazon order, then clearly revokes it by saying 'Hold on, I've changed my mind. Cancel that.' This constitutes a direct cancellation of the task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote. Hold on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "7552241b3975992cebc2d8c0ad94bc6aa196cb966c549fd9890bf20b0ba14f8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/ambiguous_halt/halt_028/3341e6608dc4ceae/speech.wav", "wer": 0.1, "wav_sha256": "bb81f6e9598fa0907a3d80f436d157e8edb9f9487fde0794fb4cd2fe0e0e3093", "seconds": 13.44, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8e8f9cd9a171bd6a53ae506439c3c9db9fbb2332a3cc616a458aa173e8048515", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/filler/original/0b399f9bce5193bb/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "0eaf1d586b9948cf39d221dabd1a370e5f1bc433b180600b2da9fff93ef31d50", "seconds": 8.88, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchanneling, or self-correction. The speaker delivers the full task request clearly and without interruption. Since the type is 'filler', the presence of minor hesitations or pauses within the instruction is acceptable as long as the final goal matches the original instruction. The final audible intent is to execute the order as per the shopping list in SimpleNote, which aligns with the reference.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Walmart—I should say Amazon order as per my Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "3cef97193490889e98c4516a3aca2e0c52738a1016660f7e76ff06091f32bffc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon' within the same utterance, preserving the app, action, and note reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/false_start/false_start_01_ordering_app/71a930fe427248db/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "23b4ec0d83e4fed0c49d9b4c6fe8cff0dff0e09fb519744186c85a3a9abe02a6", "seconds": 9.04, "spoken_utterances": [ "Place an Walmart—um... I should say Amazon order as per my Amazon shopping list note in ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects to 'Amazon' with 'I should say Amazon order'. This is a clear false start where the corrected value replaces the mistaken one. The final intent matches the original instruction exactly, so the task should be executed as intended.", "heard_transcript": "Place an Walmart, um, I should say Amazon order as per my Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_f4f374cbe56781a4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'Walmart', immediately self-corrects to 'Amazon', and then continues with a coherent, slightly hesitant but fluent rest of the sentence. The correction is integrated smoothly into the flow, and the prosody reflects genuine recall and correction rather than a rehearsed or robotic edit. The pacing and intonation are consistent with someone thinking aloud." } ] }, "is_default": true } }, { "task_id": "099174d_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_note_description", "proposed_instruction": "Place an Amazon order as per my Grocery list note in SimpleNote. Make that the Amazon shopping list note, not the Grocery list.", "long_pause_offsets": [ 65 ], "source_sha256": "25213f62f7ea6a8dfe2d6b40c3f218b36bc249b773b55f2796336665e7936cb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the note name from 'Grocery list' to 'Amazon shopping list'; the app (SimpleNote), action (place Amazon order), and source (note) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/revision/revision_02_note_description/ccc82dc92dd5d7af/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "beb1efd5436fca3e02e2b0a0305edef6d204d43ca433fca6b52b3485cfb55423", "seconds": 12.6, "spoken_utterances": [ "Place an Amazon order as per my Grocery list note in um... SimpleNote. ", "Make that the let me think... Amazon shopping list note, not the Grocery list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list note' but then corrects themselves to 'Amazon shopping list note' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use the 'Amazon shopping list note', which matches the original reference instruction. The correction is clear and unambiguous, and the rest of the task (placing an Amazon order in SimpleNote) remains consistent.", "heard_transcript": "Place an Amazon order as per my grocery list note in um, SimpleNote. Make that the, let me think, Amazon shopping list note, not the grocery list." }, "candidate_id": "wav_8676ee5ce855c077", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' before self-correcting 'grocery list note' to 'Amazon shopping list note'. The correction is smooth and integrated into the flow, with a natural pause and filler ('let me think') before the revision. The prosody is consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "099174d_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_022", "hide_variant_original": true, "source_sha256": "597cbf47289e279a671fd327119e7b64ebd39423ba9e9453568eed12212825ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/cancellation/cancel_022/5873582b07869ac2/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "d3ca85d19d401b74ce9155d5d35ea6b8c956f69e8ca7716f79b62ad3efbb635b", "seconds": 12.16, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote.", "Hang on, actually, never mind. Cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to place an Amazon order, but concludes with the explicit cancellation command 'cancel the task'. This meets the criterion for a cancellation type, where the final audible intent is to revoke the task. The initial instruction is irrelevant once the cancellation is clearly stated.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote. Hang on, actually, never mind, cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "4e51d2f5f63bd12b8c268da1b8ab65b1f86194d0d38c5be7707dced36500b1a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/ambiguous_halt/halt_033/8ba301a0d9c66ab0/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "d6012b497aae112be83c6cd8fb8a884c60e0f1d0157fcfe6c483fc297824e8f6", "seconds": 10.72, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, there's something,' which is a clear request to pause the task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not completed, but the final intent is to halt, which is the expected behavior for this type.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote. Hold on, there's something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "099174d_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "587ca070d4acc64da8a237faf42c25de3a2b0b226cd79b6dda8f9b1ea555666a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/filler/original/60f46a714e4df807/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "55f21d43b7860f83bbe7656d994da5e470ef0d79468f54fcc86a7ba856f3a879", "seconds": 7.12, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Place an Amazon order' and then hesitates with 'uh' before continuing the original instruction. The final utterance matches the original instruction without any correction or change in intent. The hesitation is a filler and does not alter the task goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all socket sets—sorry, I meant wrench sets in my amazon wish list.", "long_pause_offsets": [], "source_sha256": "d7cd044ad079712eda5f7198cae455fc3b4ed0ba263fd893a5e447a80dfd0b0e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'socket sets' to 'wrench sets' within the same utterance, preserving the action, scope ('all'), and destination ('Amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/false_start/false_start_01_product_category/2f644a8df6ab8872/speech.wav", "wer": 0.3157894736842105, "wav_sha256": "9afc14e16cec826f1bc5c86f4d793a66f8637509ad5bc2eb8c6d84dd66bc9f93", "seconds": 6.6, "spoken_utterances": [ "Place an order for uh... all socket sets—um... sorry, I meant wrench sets in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'socket sets' but immediately self-corrects to 'wrench sets' with 'Sorry, I meant...'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, with only the target item corrected from 'socket sets' to 'wrench sets'. All other conditions, including 'order', 'all', and 'in my Amazon wish list', are preserved.", "heard_transcript": "Place an order for uh all socket sets some. Sorry, I meant wrench sets in my Amazon wish list." }, "candidate_id": "wav_b79552215895efc7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'socket sets', then immediately self-corrects to 'wrench sets' with a clear, spontaneous-sounding 'sorry, I meant'. The correction is integrated smoothly into the sentence, and the final phrase 'in my Amazon wishlist' is delivered with natural prosody and a slight pause before the final word, suggesting genuine recall. The flow feels like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "0d22252_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon wish list. I left out a detail: the order is for wrench sets.", "long_pause_offsets": [ 47 ], "source_sha256": "d3271fbbb7cb532de88c8f3836a8aa3ac96c1884f37a3dae47c1d64fee84c746", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing item category 'wrench sets' to the initial incomplete request, completing the final goal without contradicting any other part of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/addition/addition_02/e1d4ee1c155dae90/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "6f4e1fb1a043c4cb3545c1e30f0722bd39a9c00405bb3f67dafac724fafc5cb4", "seconds": 10.56, "spoken_utterances": [ "Place an order for all ah... in my amazon wish list. ", "I left out a detail: the order is for um... wrench sets." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all a' which is likely a misstatement of 'all wrench sets'. After a pause and self-correction, they clarify the correct target as 'wrench sets'. Since this is an 'addition' type, the final intent combines the original instruction with the added clarification. The final audible goal matches the reference instruction exactly.", "heard_transcript": "Place an order for all a, in my Amazon wish list. I left out a detail, the order is for um, wrench sets." }, "candidate_id": "wav_7a38d45dd368c568", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural-sounding initial request with a slight hesitation ('um') and a believable self-correction. The revision ('I left out a detail...') is smoothly integrated with a pause and a change-of-mind intonation, and the corrected item ('wrench sets') is clearly articulated. The prosody and pacing are consistent with spontaneous speech, and the repair feels authentic rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "0d22252_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all socket sets in my amazon wish list. Wait, the order should be for all wrench sets, not socket sets.", "long_pause_offsets": [ 59 ], "source_sha256": "8d1c68e953949cb0d3c7e8c56057ab94583ed74b347ff7d31362a1d25f8bd5e1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the tool category from socket sets to wrench sets, while preserving the action, scope ('all'), and destination ('amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/revision/revision_01_product_category/0e7ab1cb4cbe3097/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "c913912c74f411ea1034bb70e7f16f037f151a9d564f3f40cdc283b55bcc9017", "seconds": 13.6, "spoken_utterances": [ "Place an order for um... all socket sets uh, let me think... in my amazon wish list. ", "Wait, the order should be for all wrench sets, not socket sets." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'socket sets' but immediately corrects themselves to 'wrench sets' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Place an order for um, all socket sets, uh, let me think, in my Amazon wish list. Wait, the order should be for all wrench sets, not socket sets." }, "candidate_id": "wav_b44904b45d857421", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a filler ('um'), a hesitation ('uh, let me think'), and a clear self-correction ('Wait, the order should be for all wrench sets, not socket sets'). The correction is integrated into the same utterance with a believable shift in intonation and pacing, mimicking a real-time thought process. The speaker also corrects the item type mid-sentence, which is a strong indicator of authentic task formulation." } ] }, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all mouse pads in my amazon shopping cart—scratch that, make it wish list.", "long_pause_offsets": [], "source_sha256": "b1e6fcbadbf615fdb86f2bf6b3cdcc2ef180df8591e5768dc527c591a7c14449", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately corrects it to 'wish list' within the same utterance. The final target is the wish list, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/false_start/false_start_03_source_collection/9644f6f214fd18b6/speech.wav", "wer": 0.1, "wav_sha256": "658d51aa95f2af0bea901c1876eb8598131faf88a02273eb2fda4f482e283e3a", "seconds": 13.24, "spoken_utterances": [ "Place an order for um... all mouse pads ... in my amazon shopping cart—ah... scratch that, make it wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately self-corrects with 'Scratch that, make it wish list.' This is a clear false start where the final intent is the corrected version. The final instruction matches the original reference instruction exactly.", "heard_transcript": "Place an order for um all mouse pads in my Amazon shopping cart. Scratch that, make it wish list." }, "candidate_id": "wav_685bda3b3377d0e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a clear false start ('Place an order for...'), followed by a hesitant filler ('um'), then a complete initial request. The speaker then self-corrects with a believable 'scratch that' and a smooth, natural revision to 'wishlist'. The prosody is consistent, the pauses are appropriately placed, and the correction feels like a genuine moment of reconsideration rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for mouse pads in my amazon wish list. Oh, I forgot to mention, place an order for all mouse pads in my amazon wish list.", "long_pause_offsets": [ 54 ], "source_sha256": "4bee14a9cd3ae5466b6db54c95e27d01b6f933421a74e06407ce13830f65ee2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The second utterance adds the missing quantifier 'all', which was omitted in the first. The final goal includes the complete original instruction with the added constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/addition/addition_01/7c6db680dab67daf/speech.wav", "wer": 0.21875, "wav_sha256": "878b2302f551c28d82ee05e3a561fb10f6fc98e44f920d44696e555b40d18d46", "seconds": 17.36, "spoken_utterances": [ "Place an order for mouse pads uh, let's see... in my amazon wish list. ", "Oh, I forgot to mention, place an order for um... all mouse pads in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for mouse pads, uh, let's see, in my Amazon wish list.' This is followed by a self-correction: 'Oh, I forgot to mention, place an order for um, all mouse pads in my Amazon wish list.' The addition of 'all' modifies the initial request to be more specific, and the final instruction matches the original reference. The type is 'addition', which requires combining the earlier instruction with the added information. The final goal is to place an order for all mouse pads in the Amazon wish list, which aligns with the reference. No issues are present.", "heard_transcript": "Place an order for mouse pads, uh, let's see, in my Amazon wish list. Oh, I forgot to mention, place an order for um, all mouse pads in my Amazon wish list." }, "candidate_id": "wav_42242dfc062a05ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, followed by a plausible hesitation ('uh, let's see') and a self-correction ('Oh, I forgot to mention'). The addition of 'all mousepads' is smoothly integrated with a natural pause and filler ('um'), and the final sentence is a coherent, grammatically correct restatement. The prosody is varied, with appropriate stress and pacing, and the overall flow mimics a real person recalling and refining a thought." } ] }, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all mouse pads in my amazon shopping cart. I misspoke about the source: use 'wish list', not 'shopping cart'.", "long_pause_offsets": [ 62 ], "source_sha256": "2c534b38575e433b69c19a666fa01cfb552a41ef819d416a61e39c988ffc62a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'shopping cart' to 'wish list', preserving the action, item scope, and app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/revision/revision_03_source_collection/83cbc25ce756f36a/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "e154db2219ccbab7d0e574b0d1da70c8c1f4af4028f6d8716220c3802d3de301", "seconds": 18.48, "spoken_utterances": [ "Place an order for um... all mouse pads uh... in my amazon shopping cart. ", "I misspoke about the source: use 'um, let me think... wish list', not 'shopping cart'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately corrects themselves to 'wish list' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to place an order for all mouse pads in the Amazon wish list, which matches the original instruction.", "heard_transcript": "Place an order for um all mouse pads uh in my Amazon shopping cart. I misspoke about the source. Use um let me think, wish list, not shopping cart." }, "candidate_id": "wav_6928001321dfc5c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then makes a clear self-correction ('I misspoke about the source') followed by a plausible thought process ('let me think') before stating the correct source ('wishlist'). The prosody is consistent, and the correction is integrated smoothly into the flow without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "88d41ac81c74ee9fe07665f4153f866f2b85c725f7b4fd16a087b7e32649f462", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/cancellation/cancel_013/80e3f2f644605710/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "64db7ff40a52ad8e87f4d10abe87b39a50e3c7b3eeb7f1e1ce95ed99319c1902", "seconds": 12.6, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list.", "Hold on, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to place an order for all mouse pads in the Amazon wish list. After a brief pause, the speaker clearly states 'Hold on, I changed my mind. Please cancel the task.' This constitutes a direct and unambiguous cancellation of the entire task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list. Hold on, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "0270084e0378460d615715bcbc1e5787fd82fc4363b75e8a49e0cf194775ade8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/ambiguous_halt/halt_010/f01ab00c7d3811ba/speech.wav", "wer": 0.25, "wav_sha256": "b6894e6d7aea9574a9f403d45e5acc83d699837ec4ce6f6e67857326ea471e47", "seconds": 10.04, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled, but the agent is told to hold off on execution. No further task is issued, and the recording concludes with the pause request.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "531fb1a88f9415bc62029ce52ce2dadd80d1e08e8a7957db7dc1002d524aee3a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/filler/original/739dc034562f16ed/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "149bac8227294f2ecc4bb269070b335c75e496585f02d9e5a8e2ce3320c99c54", "seconds": 6.44, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker clearly and directly states the original instruction without any change in goal or scope. The 'filler' type allows for minor hesitations or fillers within the instruction, but the final intent remains the original task. The instruction is fully audible and matches the reference.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all handheld vacuum cleaners in my Walmart—I should say amazon wish list.", "long_pause_offsets": [], "source_sha256": "727ac61602eb7229c17d2d7f62f27aa9adccca0c214ddd070a78fbc83e23b737", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon wish list' within the same utterance, preserving the action, item type, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/false_start/false_start_02_app/5f7d188f75386abe/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "bba9cdc040a2aab5fb88a3928b3cee70d091152c05d07790c870b758b7050898", "seconds": 7.32, "spoken_utterances": [ "Place an order for all handheld vacuum cleaners in my Walmart—uh... I should say ... amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects to 'Amazon wish list'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Place an order for all handheld vacuum cleaners in my Walmart, uh, I should say Amazon wish list." }, "candidate_id": "wav_8ec71f9f6a2eca94", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon wish list. I left out a detail: only place an order for handheld vacuum cleaners.", "long_pause_offsets": [ 47 ], "source_sha256": "e8d0c81db6f2bfcdeac9fa282a467f2d6126b6508bee080670e7e55bdebed87a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only handheld vacuum cleaners, preserving the app (Amazon), action (place an order), and source (wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/addition/addition_02/39078b4ca57a7138/speech.wav", "wer": 0.125, "wav_sha256": "c3390442f5b675c741a5c6fe516cc768494b9f7f05b82a88a27b8bc7a5815e00", "seconds": 12.04, "spoken_utterances": [ "Place an order for all in um... my amazon wish list. ", "I left out a detail: only place an order for ... handheld vacuum cleaners." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was incomplete, but the speaker explicitly added a constraint ('only place an order for handheld vacuum cleaners') after a pause. This addition narrows the scope of the original request, and the final audible intent matches the reference instruction. The type is 'addition', and the final goal is fully aligned with the reference.", "heard_transcript": "Place an order for all in um my Amazon wish list. I left out a detail, only place an order for handheld vacuum cleaners." }, "candidate_id": "wav_efed8bb48a106f98", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, uses a plausible filler ('um') to pause, and then adds a specific restriction ('only place an order for handheld vacuum cleaners') in a way that sounds like a spontaneous correction. The prosody is consistent, and the addition flows logically from the initial request without awkwardness." } ] }, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all air purifiers in my amazon wish list. Wait, change the product category to 'handheld vacuum cleaners' instead of 'air purifiers'.", "long_pause_offsets": [ 61 ], "source_sha256": "dc7581a249466a2a467f340d77ee0abe49906f7cf04f5e25834d2de62b5f114b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the product category from 'air purifiers' to 'handheld vacuum cleaners', preserving the action, scope ('all'), and source ('Amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/revision/revision_01_product_category/e9a0e4396f5bd2a1/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "ac1cb71e8f0d65f08d01f09c10ae919638144780f11ec0c3a508bd1c6c3ab5bf", "seconds": 15.04, "spoken_utterances": [ "Place an order for all air purifiers in let's see... my amazon wish list. ", "Wait, change the product category to 'ah... handheld vacuum cleaners' instead of 'air purifiers'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'air purifiers' but immediately corrects to 'handheld vacuum cleaners' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly.", "heard_transcript": "Place an order for all air purifiers in let's see, my Amazon wish list. Wait, change the product category to uh handheld vacuum cleaners instead of air purifiers." }, "candidate_id": "wav_da9f05c0b7d1459b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' before correcting the product category. The correction is delivered with a believable change-of-mind intonation and a smooth reference back to the initial request. The prosody is consistent, and the pacing feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "63303e3dc3fa7485273c81a04ffbc948e62c318d9c921fe2e58612a8f1517b66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/cancellation/cancel_009/52582ff42a05c1ba/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "3e044f1ad305b5cb0c66e63f07166ad769b9d0287af66b8af620746123cf618a", "seconds": 11.52, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but immediately self-cancels with 'Hang on, cancel that altogether.' This constitutes a clear cancellation of the entire task as per the type-specific criterion. The final control intent is cancellation, matching the expected behavior for this type.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "972dab916fe3988a7f352f02f96a823cf4aa02b029916f7fd4bac594729c3281", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/ambiguous_halt/halt_039/ebdc2b772cf392d9/speech.wav", "wer": 0.15, "wav_sha256": "997ae43fdbbf6d9cd168ea8d6c81f36054c630a9d957f3023f8a59b89f7a03da", "seconds": 11.16, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list.", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, give me a second first,' which is a clear request to pause or hold. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker is asking for a moment before proceeding.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list. Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "0d22252_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "56a2e39e9ff7fc4ff1337978c8174c441aa4113be3c00df5e02b4eec18d09312", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/filler/original/562e4bd8507eb9e8/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "9c4438a23fadcee6115da1fcd4f2e90add1b7ebdaeaf0141817d4ecf070f7f32", "seconds": 7.28, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler word 'uh' but completes the original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2021—sorry, that should be 2022 from my Spotify song library.", "long_pause_offsets": [], "source_sha256": "40d9003aa5cf4b2522bdb7d41a7d76cc7c00365ff4a951a1e4d3df8f910e775b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'two thousand twenty-one' to 'two thousand twenty-two' within the same utterance, preserving the app, action, and scope of 'any song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/false_start/false_start_01_release_year/c66e54dfefe7d35c/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "9dde0c2db49d2eac1ce62aeb033d823cd1d40127cc02270ab2383c06948fa670", "seconds": 10.04, "spoken_utterances": [ "Play any song released in two thousand twenty-one—ah... sorry, that should be ... two thousand twenty-two from my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces '2022' as 'two teeny one uh' but immediately self-corrects to 'two thousand twenty two'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved. The corrected year matches the reference, and no other constraints are altered. Therefore, the final intent aligns with the original instruction.", "heard_transcript": "Play any song released in two teeny one uh, sorry, that should be two thousand twenty two from my Spotify song library." }, "candidate_id": "wav_7b0e8817a4cc38c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural-sounding false start with 'tutini huan' followed by a clear self-correction to '2022'. The speaker's tone and pacing are consistent, and the correction is delivered smoothly without awkwardness. The pause before the correction is brief and plausible, fitting a moment of recalling a number. The final request is articulated clearly and confidently." } ] }, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Play any song released from my Spotify song library. Oh, I forgot to say, play a song released in 2022.", "long_pause_offsets": [ 53 ], "source_sha256": "9d2b3c9ba82caa9031dafccdaa125fc7d9b77035f038d95d564fc8b04abb1c98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing year constraint, which refines the initial broad request. The app (Spotify), action (play), and scope (any song from library) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/addition/addition_01/b91365d0af05bc9c/speech.wav", "wer": 0.0, "wav_sha256": "78101837098faaed29cf16c771b03ebee86276d1b04ea59886dc8ad62cccd3db", "seconds": 12.6, "spoken_utterances": [ "Play any song released ... from my Spotify song library. ", "Oh, I forgot to say, play a song released in uh... two thousand twenty-two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to play any song from their Spotify library, then adds the constraint 'released in 2022' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and executable, matching the reference instruction.", "heard_transcript": "Play any song released from my Spotify song library. Oh, I forgot to say, play a song released in uh, 2022." }, "candidate_id": "wav_3629806f34bbebdb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'Oh, I forgot to say' before adding a specific detail. The addition of 'in uh 2022' includes a realistic filler ('uh') and a slightly uncertain intonation, suggesting spontaneous recall. The flow is continuous and the correction feels like a genuine afterthought, making it highly realistic." } ] }, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_release_year", "proposed_instruction": "Play any song released in 2023 from my Spotify song library. Make that any song released in 2022, not 2023.", "long_pause_offsets": [ 61 ], "source_sha256": "02c50749f51759ca2df290abbf057961accf5ef7548faf84ea97ed02c1203f67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the release year from 'two thousand twenty-three' to 'two thousand twenty-two', preserving the app, action, and scope of 'any song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/revision/revision_01_release_year/54b807d283257076/speech.wav", "wer": 0.0, "wav_sha256": "675da0d1a2c01b6dd371949bce7e934c7ade438c7960ca6227310b731c748606", "seconds": 14.88, "spoken_utterances": [ "Play any song released in ... two thousand twenty-three from my Spotify song library. ", "Make that any song released in um... two thousand twenty-two, not two thousand twenty-three." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests a song from 2023, then corrects themselves after a pause to specify 2022. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to play a song from 2022, which matches the original reference instruction. All other constraints (Spotify song library, any song) remain unchanged.", "heard_transcript": "Play any song released in two thousand twenty three for my Spotify song library. Make that any song released in um two thousand twenty two, not two thousand twenty three." }, "candidate_id": "wav_8ae93dddef27fb48", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "8aeb999ab5912c5dae4510e8415e4864bcee3109f272c40b8f4b3875d2febb68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/cancellation/cancel_031/ad598251a6a9afe6/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "33f4ad80496fd2518b6fd8fe0a94618e11fba6e37b3aded0cf9afa8673b9dfb6", "seconds": 14.4, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library.", "Hold on, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to play a song from 2022, but then clearly revokes the entire task by saying 'Hold on, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the original instruction. The final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Play any song released in whom two thousand twenty two uh from my Spotify song library. Hold on, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "3c5397af9c2ea9abc355f8bb5f10973c21667465b68e954df6104a1243896aa8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/ambiguous_halt/halt_015/5c78a676175ed0f6/speech.wav", "wer": 0.1, "wav_sha256": "d57578496b2641cac84ca2ebc216f8b3c30e1f0ce369f27ed86f55606e687acd", "seconds": 12.76, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library.", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but immediately hesitates with 'whom?' and 'uh', then completes the instruction with '2022'. After stating the full request, they say 'Please wait a moment', which is a clear and final request to pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The initial hesitation and self-correction do not change the final intent, which is to pause.", "heard_transcript": "Play any song released in whom? 2022 uh from my Spotify song library. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dcecea0decb5d66b33b1e0ccaed9936d6e6a082e17b403b308fcf647bccc1704", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/filler/original/c77552fb4bc50dc2/speech.wav", "wer": 0.125, "wav_sha256": "407c0f2c5fffa9ce524cafedc799e62cf60c327c4e6e77b3d441543dab461541", "seconds": 8.96, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces '2022' as 'whom', then self-corrects to '2022'. The final audible intent matches the original instruction. The initial error is a filler-like misstatement that is immediately corrected, and the final goal is clear and executable.", "heard_transcript": "Play any song released in whom? 2022 uh from my Spotify song library." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2020—sorry, I meant 2021 from my Spotify album library.", "long_pause_offsets": [], "source_sha256": "caf128f4475b5f9afb1ecc246c8b3be31ad649f3cc3a9b5f82aa25b40d6454cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'two thousand twenty' but immediately corrects to 'two thousand twenty-one', which is the same as 2021. The app, action, and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/false_start/false_start_01_release_year/9277aeba663ddcdf/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "0a664f28788346db52a159835753fe26ea6f1b14e8279825ac2a45dc453b5ed7", "seconds": 8.44, "spoken_utterances": [ "Play any song released in two thousand twenty—uh... sorry, I meant ah... two thousand twenty-one from my Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '2020' but immediately self-corrects to '2021' with 'sorry, I meant uh 2021'. This is a clear false start where the final value is the corrected one. The rest of the instruction, including the source 'Spotify album library', remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Play any song released in two thousand twenty uh sorry, I meant uh two thousand twenty one from my Spotify album library." }, "candidate_id": "wav_e5c96ef425472f09", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then makes a natural false start ('2020') followed by a self-correction ('2021') with a smooth, conversational tone. The correction is immediate and integrated into the same utterance without awkward pauses or repetition. The phrase 'from my Spotify album library' is delivered with consistent prosody, suggesting spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Play any song released from my Spotify album library. Oh, I forgot to mention, play only songs released in 2021.", "long_pause_offsets": [ 54 ], "source_sha256": "088a7c52c1888da4efc459e98581c9a08f618fe04589d42f594f2fc01404e951", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing year restriction, narrowing the initial broad request to only songs from 2021. The app and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/addition/addition_01/5f90ad5fdcf3f778/speech.wav", "wer": 0.12, "wav_sha256": "927f52a4e5a001f482b30e9dd549eb08f2cb3acccf3053addd301da433f7c6a4", "seconds": 12.92, "spoken_utterances": [ "Play any song released from my hmmm... Spotify album library. ", "Oh, I forgot to mention, play only songs released in um... two thousand twenty-one." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Play any song released from my home Spotify album library.' This is followed by an addition: 'Oh, I forgot to mention, play only songs released in um, twenty twenty one.' The addition narrows the scope of the original request by specifying the release year as 2021. The final intent combines the original request with the added constraint, matching the reference instruction. The 'um' and 'twenty twenty one' are natural speech variations for '2021'. The control intent is 'execute' as the addition refines the task rather than cancelling or pausing it.", "heard_transcript": "Play any song released from my home Spotify album library. Oh, I forgot to mention, play only songs released in um, twenty twenty one." }, "candidate_id": "wav_93bd9a902e831c04", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable hesitation ('um') and a self-corrected pause ('from my home Spotify album library') before adding a new constraint ('play only songs released in um 2021'). The addition is integrated smoothly with a change-of-mind intonation and a plausible reference back to the initial request. The prosody is consistent, and the pauses feel like genuine thought processes rather than robotic breaks." } ] }, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Play any song released in 2021 from my Apple Music album library. Let me correct the app: play from my Spotify album library, not Apple Music.", "long_pause_offsets": [ 66 ], "source_sha256": "5e8e12fdf900a0428fed51d3d730d423a54d011f6bf9cb00f2c893a4201252a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the year 2021 and the action of playing any song from the album library." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/revision/revision_02_app/4b0935477d790d66/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "4e7c0083606c9f2e1afe00e6e61ee9c2ce94f055b55d3c5a0d33ecc5327c0070", "seconds": 16.64, "spoken_utterances": [ "Play any song released in ... two thousand twenty-one from my uh... Apple Music album library. ", "Let me correct the app: play from my uh, let's see... Spotify album library, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to play a 2021 song from the Spotify library, matching the original reference instruction. All other constraints (song from 2021, album library) are preserved.", "heard_transcript": "Play any song released in 2021 from my uh Apple Music album library. Let me correct the app, play from my uh let's see, Spotify album library, not Apple Music." }, "candidate_id": "wav_5e01c469be823cd9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker starts with a clear request, then uses a believable self-correction with 'Let me correct the app' and a filler 'uh' before smoothly revising the platform from Apple Music to Spotify. The prosody is consistent, and the revision feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "c639fba0305b75ab542151925867f5c0b529cab7ce846e22fa92a94932346290", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/cancellation/cancel_019/067d6b5dade2e4a1/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "a70ebc62a1073189984946fd51f1f0df11ac9860ee195835772c89c331e0c8ac", "seconds": 11.68, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library.", "Hold on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly cancels it with 'Hold on, never mind, don't do that.' This constitutes a final, unambiguous cancellation of the original instruction. The control intent is correctly identified as 'cancel'.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library. Hold on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "ffeba519a95ef6228226fbea2acda1b5a0f46a593b07c6965dd49fa0efc74bca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/ambiguous_halt/halt_029/e31d2c32b31f92df/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "22ce339d2ba43430ef1a6eb34e90df536029f4bbef4298009d82d71bc500ca6f", "seconds": 10.96, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then says 'Wait, actually...' as a final interjection, which is a clear audible request to pause or halt the current action. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the audio ends with a pause directive.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library. Wait, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b4bfdd75392f06e0126a58868bfaa38ede02e9b14d75a020d01263e00a0d081d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/filler/original/e98c6a17fcaf218a/speech.wav", "wer": 0.125, "wav_sha256": "d796c8c2ec5138b74721cd43178529c6a5d6f023f72dabac67b3ab97c57ce74c", "seconds": 7.4, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um' as a filler, which is a common hesitation marker. The rest of the instruction is delivered clearly and matches the original request. The final intent is to execute the task as stated, with no correction or cancellation. The filler does not alter the goal, so the task is executed as intended.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1150ed6_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2019—sorry, I meant 2020 from my Spotify playlists.", "long_pause_offsets": [], "source_sha256": "7bdfd6894f1a5bd7047348389054927dd9f679d2c018ea1dea458bceb727b201", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'two thousand nineteen' but immediately corrects to 'two thousand twenty', which is the correct year. The app (Spotify), action (play), and scope (any song from playlists) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/false_start/false_start_01_release_year/1d58c6a36c7d81f6/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b4ee200331efa204ce1251e1530ba30fcd2cb4e756eba09c9321bddf9a9c7814", "seconds": 11.04, "spoken_utterances": [ "Play any song released in let me think... two thousand nineteen—hmmm... sorry, I meant uh... two thousand twenty from my Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '2019' but immediately self-corrects to '2020' after a pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected value, and the original incorrect value is discarded. The rest of the instruction matches the reference exactly.", "heard_transcript": "Play any song released in let me think two thousand nineteen hmm sorry I meant uh two thousand twenty from my Spotify playlists." }, "candidate_id": "wav_31808f9ae32e0b7a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'let me think' followed by a clear self-correction from '2019' to '2020'. The intonation and pacing of the correction sound spontaneous, and the subsequent phrase 'for my Spotify playlist' flows smoothly without hesitation. The prosody is consistent and the repair is integrated into the utterance as a single, coherent thought process." } ] }, "is_default": true } }, { "task_id": "1150ed6_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Play any song released in 2020 from my Apple Music playlists. Let me correct the app: play from Spotify playlists, not Apple Music.", "long_pause_offsets": [ 62 ], "source_sha256": "34fcd71cec2498fff7a34892382a44b480f322b24f0d659215c400dab182abe3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the year 2020 and the 'any song' scope. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/revision/revision_02_app/310b436b130d1098/speech.wav", "wer": 0.22580645161290322, "wav_sha256": "4bc16c991693a9360619cdadb78727b56edd304839a444a4ec32865bd7443d46", "seconds": 17.76, "spoken_utterances": [ "Play any song released in uh... two thousand twenty um, let's see... from my Apple Music playlists. ", "Let me correct the app: play from let's see... Spotify playlists, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests a song from Apple Music playlists but immediately corrects the app to Spotify. The revision is clear and final, with no further changes or cancellations. The rest of the instruction (song from 2020) remains unchanged. The final intent matches the original instruction after the correction.", "heard_transcript": "Play any song released in uh two thousand twenty um let's see from my Apple Music playlists. Let me correct the app, play from let's see, Spotify playlists, not Apple Music." }, "candidate_id": "wav_5bc527fac26e5c48", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my my roommate—no, I meant all roommates, \"I have taken out the trash.\".", "long_pause_offsets": [], "source_sha256": "68eeeff96aa7cff5fa77f7a5612e4b423dd39869e4aeddc4a1086852b7d1f036", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'my roommate' but immediately corrects to 'all roommates'. The final recipient is correctly specified as 'all roommates', and the message content remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/false_start/false_start_02_recipient_reference/f486ebd1644ed39d/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "0aa25da40b06bf9b6048df50a09a087139335c8193c0a365d701ed150b2ba0cf", "seconds": 9.72, "spoken_utterances": [ "Send a let's see... phone voice message to my my roommate—ah... no, I meant ... all roommates, \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my roommate' but immediately self-corrects to 'all roommates'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction with the correction applied.", "heard_transcript": "Send a phone voice message to my my roommate uh no, I meant all roommates, I have taken out the trash." }, "candidate_id": "wav_e0c612bf4bd8284c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'send a', hesitates with 'let's see', then correctly states 'phone voice message'. The false start 'my roommate' is immediately and smoothly self-corrected to 'all roommates' with a believable change-of-mind intonation and a clear reference back to the initial request. The pacing and prosody are consistent with spontaneous speech, and the correction feels integrated rather than tacked on." } ] }, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Send a phone voice message to my, \"I have taken out the trash.\". I should also mention, send it to all roommates.", "long_pause_offsets": [ 65 ], "source_sha256": "73aaade64bb9ab773b9a25df0787ec06b41ded64815d95fc16c41b541faa0e53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient group 'all roommates' after an initial incomplete utterance. The message content and action are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/addition/addition_02/854686e74890e0f6/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "a88863e0278b2164ed8872e6cdc79d7f73423d8550421d2feade6935088f35f4", "seconds": 14.16, "spoken_utterances": [ "Send a uh, let me think... phone voice message um... to my, \"I have taken out the trash.\". ", "I should also mention, send it to all roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a phone voice message to my...', then corrects to 'send it to all roommates' after a pause. This is an addition that completes the original instruction by specifying the target audience. The final intent matches the original instruction with the added detail of 'all roommates', which is consistent with the addition type criterion.", "heard_transcript": "Send a, uh, let me think, phone voice message, um, to my, I have taken out the trash. I should also mention, send it to all roommates." }, "candidate_id": "wav_e1ed8ceb1de10d5c", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_message_text", "proposed_instruction": "Send a phone voice message to my all roommates, \"I have taken out the recycling.\". Change the message to \"I have taken out the trash.\", not \"I have taken out the recycling.\"", "long_pause_offsets": [ 83 ], "source_sha256": "d3ac22375d243b01b334021c40b6eeea037bb93b48e7c52b69562e3ed4d4741a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the message content from 'recycling' to 'trash', preserving the recipient group and message type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/revision/revision_03_message_text/7cd4a1bc6f36dc46/speech.wav", "wer": 0.0, "wav_sha256": "4dedc681bfaa3e9f11cbb8588bd062d857f593dbadba99734775a3dcb301728c", "seconds": 17.28, "spoken_utterances": [ "Send a let's see... phone voice message to my all roommates, uh... \"I have taken out the recycling.\". ", "Change the message to um, let me think... \"I have taken out the trash.\", not \"I have taken out the recycling.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I have taken out the recycling' but then corrects it to 'I have taken out the trash' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "Send a let's see, phone voice message to my all roommates, uh, I have taken out the recycling. Change the message to, um, let me think, I have taken out the trash, not I have taken out the recycling." }, "candidate_id": "wav_e5793776aa1552bf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear initial request, then self-corrects the content of the message ('recycling' to 'trash') with a believable 'um' and 'let me think' hesitation. The correction is integrated smoothly into the same utterance, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "0d69a6071eb9e34a3da59ebfbe33efb3b59591e31111b493abe7887d6b7c75cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/cancellation/cancel_004/ba78dbf65dad65a2/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "77a4071d0de2a242ed62e19f6635bfbea5ee46b9b7ebbbefc953d70bb8a7ec79", "seconds": 12.56, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\".", "Oh, wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task but immediately self-corrects with 'Oh wait, cancel that request,' which is a clear and final cancellation of the entire task. The final interjection is a direct cancellation command, satisfying the 'cancellation' type criterion.", "heard_transcript": "Send a um phone voice message to my all roommates uh, I have taken out the trash. Oh wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "4e46eb4a072743fd2682b617824d0eca103a6482a242e9ba6edf220a18c1b8bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/ambiguous_halt/halt_030/2a169f1cb7b354bf/speech.wav", "wer": 0.05, "wav_sha256": "e21e7c5b2f2699b2c03ddec04a88d750642685dca955878a87c9e00a09fc890a", "seconds": 11.88, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\".", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause and reconsider the task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt, not a cancellation or execution. The initial instruction is not completed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Send a um phone voice message to my all roommates uh, I have taken out the trash. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "43648a9b592a39fe41c97f9c9a7db340722180d8ec198eba649ad0a4a0e85345", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/filler/original/519ca6a1ab5c96e0/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8fda36eac763ff788bb06fecb7b81484770f4c9036bd3b762b4c891a74387cda", "seconds": 8.28, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core task. The final intent is clear and matches the original instruction. The filler words do not alter the goal, so the execution should proceed as intended.", "heard_transcript": "Send a um phone voice message to my all roommates, uh, I have taken out the trash." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_message_type", "proposed_instruction": "Send a email—wait, I meant phone text message to my partner, \"The dishwasher is clean and ready to be emptied.\".", "long_pause_offsets": [], "source_sha256": "7ef5338c9f712687b6a100d3b110f47f45fc44156788c3709d57df26bac4b5b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email' to 'phone text message' within the same utterance, preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/false_start/false_start_01_message_type/c1a108eb7e8140db/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "58b5207d03d0298e0c5ba9393933311df422c5e96b9980fa7c38eac2d69519e3", "seconds": 12.88, "spoken_utterances": [ "Send a email—um... wait, I meant ... phone text message to my let's see... partner, \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a email' but immediately self-corrects with 'wait, I meant, phone text message'. This is a clear false start where the final audible intent matches the original instruction. The target ('my partner') and message content are preserved and correctly stated.", "heard_transcript": "Send a email, um, wait, I meant, phone text message to my let's see, partner, the dishwasher is clean and ready to be emptied." }, "candidate_id": "wav_1b7da480695923f1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex pattern of spontaneous speech. It begins with a false start ('email'), followed by a self-correction ('wait, I meant'), a filler ('let's see'), and then a complete, coherent request. The pauses and fillers are well-placed and sound like genuine hesitation during formulation. The correction is immediate and integrated smoothly into the flow, making it the most realistic example of a natural, unscripted task request." } ] }, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Send a phone text message to, \"The dishwasher is clean and ready to be emptied.\". Oh, and one more thing: send it to my partner.", "long_pause_offsets": [ 82 ], "source_sha256": "97c079a88fab94036921485a5173cd94d0ed20273fee3ebc2bc0088502589f7b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient 'my partner' to the initial message, which already contained the correct message content. The final goal matches the original instruction in both action and content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/addition/addition_02/95c8f8930ea588c6/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "8579c4aaa69b8abf20d12bcb1595f10286a97af991373dfaddaa162f3201cefb", "seconds": 16.28, "spoken_utterances": [ "Send a uh, let me think... phone text message to, uh... \"The dishwasher is clean and ready to be emptied.\". ", "Oh, and one more thing: send it to ... my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task without specifying the recipient, then adds 'send it to my partner' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "Send a uh let me think phone text message to uh the dishwasher is clean and ready to be emptied. Oh and one more thing, send it to my partner." }, "candidate_id": "wav_c56def2b2bc49396", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses a natural, spontaneous 'let me think' pause, followed by a coherent initial request. The addition of 'Oh, and one more thing' is a believable conversational device, and the correction 'send it to my partner' flows smoothly without awkwardness. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_message_type", "proposed_instruction": "Send a email to my partner, \"The dishwasher is clean and ready to be emptied.\". Make that a phone text message, not an email.", "long_pause_offsets": [ 80 ], "source_sha256": "b74f7a4fc9bd5d19a5528edfc96581d6ab936f688f30e621238f23c6fcf4ef76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type from email to phone text message, while preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/revision/revision_01_message_type/afca670853ce8e52/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "b559758c908f4f7169bf4e439df34db75aa3f630a0e1f2b57adfd919be5e3ba8", "seconds": 19.96, "spoken_utterances": [ "Send a um, let me think... email to my um, let's see... partner, \"The dishwasher is clean and ready to be emptied.\". ", "Make that a uh, let me think... phone text message, not an email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately corrects to 'phone text message' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to send a text message, which matches the original instruction's content and target, only differing in the communication method. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Send a um, let me think, email to my um, let's see, partner, the dishwasher is clean and ready to be emptied. Make that a uh, let me think, phone text message, not an email." }, "candidate_id": "wav_20f95ab9e67692e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a highly natural, multi-stage thought process with realistic fillers ('um', 'let me think', 'let's see') and a clear, spontaneous self-correction from 'email' to 'phone text message'. The prosody includes believable hesitations and a smooth repair that maintains conversational flow, making the revision sound like a genuine moment of reconsideration." } ] }, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "428654040ccde3a36f19f390fd41c3f98e82b7e6b5598364bf8194a2039e8a46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/cancellation/cancel_010/6fcc61f4a27e78ab/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "ff670a38129d0c86be86d8004d7af80ec7909ed93f5c20a0a79902dfaf674ab8", "seconds": 11.4, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\".", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to send a text message, then immediately follows with 'Oh wait, cancel the request I just made.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "14626d1a293a5a53c7da6c4ce08abae478dd8bf7a966a43b29e98af800bc5710", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/ambiguous_halt/halt_018/37a8ab75df124c2d/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "000ad3a6b0af443f7415f0f7b5857d5fcb33a98fd13bfd843c503da964649445", "seconds": 11.04, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\".", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which is a clear instruction to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The original task is not canceled, and the agent is being asked to pause, not execute or cancel. Therefore, the final intent satisfies the type-specific criterion.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "23c0efcb55f2ceb36000e82099fd5f35dc0807cdc00fa050d601ebc62cd3dfc6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/filler/original/ee97e9a1606ce438/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "dc5f9a737879bc8a3f7af159e0f685a6d4c3a8f8f4b2c692cf18750db8d35955", "seconds": 6.88, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('um') but completes the original instruction without any correction, cancellation, or pause. The final intent matches the original instruction, and the filler word does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my wife—no, I mean husband, \"Don't forget to water the plants today.\".", "long_pause_offsets": [], "source_sha256": "7118ad530d60c68b4ddc20aad27b647742fae60b4a6370ecc2eaadd5df4558a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'wife' with 'husband'; the action, message content, and delivery method remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/false_start/false_start_02_recipient_reference/45d02cb2754bb0df/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "99f21992a76c5ae976e2d27fef9b31121cb1f28ea16d08a85ad545b313073e73", "seconds": 10.12, "spoken_utterances": [ "Send a phone voice message to my ah... wife—uh... no, I mean husband, \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wife' but immediately self-corrects to 'husband'. According to the 'false_start' type criterion, the final corrected target should be applied, and the original instruction is executed with the correction. The message content remains unchanged and is fully audible.", "heard_transcript": "Send a phone voice message to my uh wife uh no, I mean husband. Don't forget to water the plants today." }, "candidate_id": "wav_2c1da62d7f6f58cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a very natural false start with 'my uh wife uh' followed by a clear self-correction to 'husband'. The intonation and pacing of the correction sound spontaneous and unforced. The delivery is fluent and continuous, with no awkward pauses or unnatural phrasing, making it the most realistic-sounding false start." } ] }, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Send a phone voice message to my husband,. Oh, I forgot to say, include the message \"Don't forget to water the plants today.\"", "long_pause_offsets": [ 43 ], "source_sha256": "fd9b44cbad99ff5286a79066fbf87bbe3694498700d3fa631c2ac0c5ef178c24", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing message content, which was omitted in the initial utterance. The recipient and message type are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/addition/addition_03/85fe742763039803/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "6807cf7db0fe50793d7e16f6f76298b8c96f44c36aedf27653da5e585bf9a450", "seconds": 12.4, "spoken_utterances": [ "Send a ... phone voice message to my husband,. ", "Oh, I forgot to say, include the message ... \"Don't forget to water the plants today.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to send a voice message to the husband is followed by an addition that specifies the message content. The final audible intent combines both the action and the message, matching the original instruction. The addition is clear and directly specifies the missing content.", "heard_transcript": "Send a phone voice message to my husband. Oh, I forgot to say, include the message, don't forget to water the plants today." }, "candidate_id": "wav_6a0b4e27ee3d30b9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and a believable, smoothly integrated correction. The intonation and pacing are consistent with spontaneous speech, and the addition of 'include the message' feels like a genuine afterthought rather than a forced insertion. The prosody is fluid and the delivery is continuous without awkward breaks." } ] }, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_message_text", "proposed_instruction": "Send a phone voice message to my husband, \"Don't forget to feed the dog today.\". Change the message to \"Don't forget to water the plants today.\" instead of \"Don't forget to feed the dog today.\"", "long_pause_offsets": [ 81 ], "source_sha256": "738ea20497b436d675f8fa0abcc241f16a2fedafa87cf30bf14ec6b3c2114efe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the message content from 'feed the dog' to 'water the plants', preserving the recipient and message type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/revision/revision_03_message_text/3379aa9ac189b9e0/speech.wav", "wer": 0.0, "wav_sha256": "f0d95bbc507aaf4745e77a67184b0058836b8955158bd013d3995e3d84c5f31f", "seconds": 17.4, "spoken_utterances": [ "Send a phone voice message to ... my husband, \"Don't forget to feed the dog today.\". ", "Change the message to uh... \"Don't forget to water the plants today.\" instead of \"Don't forget to feed the dog today.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the incorrect message 'Don't forget to feed the dog today.' but then clearly revises it to 'Don't forget to water the plants today.' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction, and the revision is clear and unambiguous.", "heard_transcript": "Send a phone voice message to my husband, \"Don't forget to feed the dog today.\". Change the message to, uh, \"Don't forget to water the plants today\" instead of, \"Don't forget to feed the dog today.\"." }, "candidate_id": "wav_a046a6d03d28ff39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a clear, natural revision. The speaker first states a complete request, then smoothly transitions with 'Change the message to...' to correct the content. The prosody is consistent, the pace is conversational, and the correction is integrated into the flow without awkwardness. The self-correction is pragmatic and sounds like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "47a3b25d725c5f6271efa64050243d07d0186ccb05d0d8fc40856a0b33ddefb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/cancellation/cancel_038/2c9be1b9bb45355e/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "1070001a09a45007ec46261f340352fdb25c8a1f3b8a503cf4d852fe47db6adf", "seconds": 14.48, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\".", "Wait, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task by stating the message content, but immediately follows with 'Wait, please stop and cancel that request.' This constitutes a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. Therefore, the final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Send a a phone voice message to my husband, uh let's see. Don't forget to water the plants today. Wait, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "da1964ff7e500c9968ae999e53780e5db22f3146905953515280afd5a1677dc6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/ambiguous_halt/halt_011/7aac648b98433b6f/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "2e754bae2747375184599fcaa77180bbc4fd2480ada357fe3280467595a17acc", "seconds": 12.8, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\".", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full instruction to send a voice message with the correct content and recipient, then immediately says 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The original task is fully stated and not revoked.", "heard_transcript": "Send a uh phone voice message to my husband uh let's see, don't forget to water the plants today. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "13547f5_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f948e9caea7e2e480f0896465a77063e317c7eb5f40ec9aebbe441a75431a4bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/filler/original/6c6143178acd665c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "b8a2dd6c787b5be0287b1605294862e7c7037a4486f74651a38357f414a798cb", "seconds": 9.2, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core task. The final instruction matches the original request in action, target, and message content. The fillers do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Send a uh phone voice message to my husband uh let's see, don't forget to water the plants today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Etsy—I should say amazon order as per my husband's email today.", "long_pause_offsets": [], "source_sha256": "a6f10a55a70960f4ca118c7f98f62b2ff8d72c5de7d653bff9bf5a2fd1075288", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon' within the same utterance, preserving the action, the source of the order (husband's email), and the time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/false_start/false_start_01_ordering_app/e6cb915cd808cdbe/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "29ee06fc95cb7b3d3a9708144bc6d1ed6f17df1788349aaecad7b86d8b773f16", "seconds": 8.56, "spoken_utterances": [ "Place an Etsy—ah... I should say ... amazon order as per my husband's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon' with 'I should say'. This is a classic false_start where the final, corrected value is the intended goal. The rest of the instruction matches the original, so the final intent is to execute the Amazon order task.", "heard_transcript": "Place an Etsy uh, I should say, Amazon order as per my husband's email today." }, "candidate_id": "wav_7150d4efd76e7955", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural-sounding false start and self-correction. The speaker begins with 'Etsy', hesitates with 'uh', then clearly self-corrects to 'I should say Amazon'. The correction is smooth, the prosody is consistent, and the pacing feels like a genuine moment of recall. The final sentence is delivered fluently without further hesitation, making the entire utterance sound spontaneous and realistic." } ] }, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an amazon order as per my husband's email. Oh, I forgot to mention, do this today.", "long_pause_offsets": [ 49 ], "source_sha256": "978075796d1a5068dc4df5912f37cf87067f21eba9170ae6ccac00d34d7779cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'today' to the initial instruction, which is a missing condition in the original instruction. The final goal combines both the source (husband's email) and the timing requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/addition/addition_01/d206cd8efef1dda5/speech.wav", "wer": 0.0, "wav_sha256": "9026599366fbde6e4762171b145b5fe77f0d18d8b6895528a8fe9ffc38cff9bb", "seconds": 11.56, "spoken_utterances": [ "Place an ... amazon order as per my husband's email. ", "Oh, I forgot to mention, do this um... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to place an Amazon order as per the husband's email is followed by an addition specifying 'today' as a time constraint. The addition is clear and directly modifies the original task by adding a new condition. The final intent is a combined version of the original instruction and the added constraint, which matches the expected behavior for an 'addition' type. The task is not cancelled or paused, and the final control intent is to execute the combined task.", "heard_transcript": "Place an Amazon order as per my husband's email. Oh, I forgot to mention, do this um today." }, "candidate_id": "wav_ad681d43f3ee505c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, coherent initial request with natural prosody and a believable, smoothly integrated addition. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to introduce a follow-up detail, and the subsequent 'do this um today' is a natural, hesitant continuation that fits the addition pattern. The pacing and intonation suggest spontaneous thought rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my father's email today. Wait, the email should be from my husband's, not father's.", "long_pause_offsets": [ 54 ], "source_sha256": "b52052d945476f03ef5fc3b840b34a08a401f7596a38e872565ad81b17a53f1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the sender relationship from 'father' to 'husband', preserving the app, action, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/revision/revision_02_email_sender_relation/6f47d88b792373ce/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "6719895bd06c758599c46deb68c2c66e0fee7630634cd33c2642b3221ec9fee7", "seconds": 16.32, "spoken_utterances": [ "Place an amazon order as per hmmm... my father's email let me think... today. ", "Wait, the email should be from my ... husband's, not father's." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong email source ('father's'), but immediately corrects it to 'husband's' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the action, target, and time constraint.", "heard_transcript": "Place an Amazon order as per my father's email. Let me think. Today. Wait, the email should be from my husband's, not father's." }, "candidate_id": "wav_4f515bd0eb5743c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural revision process. The initial request is interrupted by a thoughtful 'um' and 'let me think,' followed by a self-correction ('wait, the email should be from my husband's not father's') that is smoothly integrated into the flow. The intonation and pacing reflect genuine hesitation and recollection, with the correction feeling like a spontaneous realization rather than a rehearsed edit. The prosody is consistent and the pauses are appropriately placed for a person formulating a request." } ] }, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "602a2bb2af579c8141b7d47b80564083ca71d25446652746d8908fe4683251d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/cancellation/cancel_031/2f0eb37678067ab8/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "d676a8c537aaf52a1f9ec7d25bade865f5808f0243a7e23aa6066025eb2b4cb8", "seconds": 17.76, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today.", "Wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Wait, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the entire task. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today. Wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "44d243ab07000f1e5ce1e52182d30bf34f2bf48182d4f4bebc5c15fae0a405c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/ambiguous_halt/halt_007/ca2de98bc962b632/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "f03306be8e63fa3c84ad116ccb70a471dd3c8963bfac0b6d26fc1bc7c1b50d98", "seconds": 15.44, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but repeatedly hesitates with fillers ('um', 'let me think', 'let's see') and ends with 'One moment', which is a clear request to pause. The final interjection is not a cancellation or a new instruction, but a request to hold. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f2d9d4b4d84e0c1008735388ade6a3c94e59e1e67bcd3613a5167d677135f734", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/filler/original/749450f7b10fce79/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d9f38ecf6525edc4e5f6d9180ec536362fb0002576b35d4f51f0ed96f83bbb80", "seconds": 12.4, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes multiple fillers ('um', 'let's see', 'let me think') and hesitations but does not change the core request. The final audible intent matches the original instruction exactly, with no self-correction or modification. The fillers are typical hesitation markers and do not alter the task goal.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my husband's—no, I mean wife's email today.", "long_pause_offsets": [], "source_sha256": "3b9b8020449cb1a796c5916a02d84a40a5b7a01f9b905ddcac44254957b2d928", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'husband's' to 'wife's' within the same utterance, preserving the app (Amazon), action (place an order), and time condition ('today')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/false_start/false_start_02_email_sender_relation/768e11422bb81ad8/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "39c23e78a32dca5909cb223dc7cb263f9e9a4ed71b24d1bf18c2e304f66d9571", "seconds": 7.48, "spoken_utterances": [ "Place an amazon order as per ... my husband's—ah... no, I mean wife's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my husband's' but immediately self-corrects to 'my wife's'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Place an Amazon order as per my husband's uh, no, I mean wife's email today." }, "candidate_id": "wav_7019d036b381c789", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with 'my husband's' and immediately self-corrects to 'no, I mean wife's email' with a clear, believable hesitation and intonation shift. The correction is integrated smoothly into the ongoing utterance, and the overall prosody, including the pause and filler 'uh', sounds completely natural and unscripted." } ] }, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Place an amazon order as per today. One more thing: use my wife's email to place the order.", "long_pause_offsets": [ 36 ], "source_sha256": "06455b1c3625fd8e16ea56823ef9e4f14d6bcde48728c4130c98ac978d394b89", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to use the wife's email, which was implied but not specified in the initial instruction. The time constraint 'today' and the action 'place an Amazon order' are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/addition/addition_02/c4fead421dae65de/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "76c2e395afd28ea24f8919d1468efe0233dc563c8cdeefc4288659f9269d13d5", "seconds": 10.6, "spoken_utterances": [ "Place an amazon order as per ... today. ", "One more thing: use um... my wife's email to place the order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'Place an amazon order as per my wife's email today.' The speaker first says 'as per today,' which is a partial or incorrect version of the original. Then, after a pause, they add 'One more thing, use um, my wife's email to place the order.' This addition clarifies and corrects the missing target (my wife's email) and specifies the action (use her email). The final intent combines the original goal with the added constraint, resulting in a complete and coherent instruction that matches the reference. The addition type is correctly applied, and the final goal is executable.", "heard_transcript": "Place an Amazon order as per today. One more thing, use um, my wife's email to place the order." }, "candidate_id": "wav_18689397daeb0979", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural, slightly hesitant 'um' filler, followed by a smooth, additive clause ('one more thing') that introduces a new detail without disrupting the flow. The prosody is consistent, and the pacing feels like a person recalling a detail mid-sentence. The addition is integrated naturally, not tacked on." } ] }, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Place an amazon order as per my wife's email tomorrow. Make that today, not tomorrow.", "long_pause_offsets": [ 55 ], "source_sha256": "61733b1fb5f5014f7e1b7b26c07f2966c5db6c57e9f96f6f1688e4f64a4ce6c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'tomorrow' to 'today', preserving the action, app, and source of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/revision/revision_03_time_reference/0fd9aca6d79d5e0d/speech.wav", "wer": 0.05, "wav_sha256": "7bc2ea20fa2dc310d6f9e138bacf1d54c45bcdb61f08986ec6594dd9d4704dd6", "seconds": 13.8, "spoken_utterances": [ "Place an amazon order as per hmmm... my wife's email um... tomorrow. ", "Make that let me think... today, not tomorrow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects it to 'today' after a self-correction ('Make that, let me think'). According to the revision type criterion, the correction spoken after a pause (or self-correction) overrides the earlier instruction. The final audible intent is to place the order today, which matches the original instruction. All other constraints are preserved.", "heard_transcript": "Place an Amazon order as per my wife's email um tomorrow. Make that, let me think, today, not tomorrow." }, "candidate_id": "wav_90fe677f4c17edd2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then pauses with a believable 'let me think' filler before correcting the date. The correction is smoothly integrated into the same utterance with a natural shift in intonation, and the overall prosody is consistent and conversational." } ] }, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "595490622274c6a45381efc7d880d7662a29a283fb0e1c65450cc95905ea5d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/cancellation/cancel_027/5bdc03f03c55d777/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "1c70fb3f80c562efe8ef520991c5cab28eda3efa23e66f1d6216e873411faa29", "seconds": 13.28, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today.", "Wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the task to place an Amazon order, but then explicitly says 'Wait, please disregard my last request,' which is a clear cancellation of the prior instruction. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an um, Amazon order as per my wife's email, um, let me think, today. Wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "abcfb3628c68b7bc111de28568b76b6e443918b94995907d7e8c532f8f6a988b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/ambiguous_halt/halt_007/96bb37c950232cf9/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "4de914ba2fe72652f3e5f21d001bc35b49ac5eb987d9ca2a9c703efb343323b3", "seconds": 11.64, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment,' which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not executed due to the pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Place an um, Amazon order as per my wife's email. Um, let me think. Today. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "72147276f511957464d874b32075126dbcfc05a43102344ed3b36667d52ac9fc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/filler/original/0902b68dcb2ac930/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "dd5cccc7fe7dd9817102c549b282f2099d5968135c8a9521a9923c82b8a20626", "seconds": 8.52, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and pauses but does not change the core request. The final audible intent matches the original instruction. The 'filler' type allows for hesitation and backchannels without altering the goal, so the task should be executed as stated.", "heard_transcript": "Place an um Amazon order as per my wife's email, um let me think, today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Etsy—I should say amazon order as per my partner's email today.", "long_pause_offsets": [], "source_sha256": "afa732a87ab3fee99a9131d47a43af1c1014ad6d3889f5e3b3462565e25b0189", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'amazon' within the same utterance, preserving the action, the source of the order, and the condition of following the partner's email." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/false_start/false_start_01_ordering_app/52eafccb49484bea/speech.wav", "wer": 0.1875, "wav_sha256": "d32e96092d77f820f994467620ec18c8f12db71a395923008c14b137684543af", "seconds": 7.2, "spoken_utterances": [ "Place an Etsy—um... I should say ... amazon order as per my partner's email uh... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'etzium' but immediately self-corrects to 'Amazon'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including the action, target, and time constraint, remains consistent. Therefore, the final audible intent matches the original instruction and the control intent is to execute.", "heard_transcript": "Place an etzium, I should say Amazon order as per my partner's email uh today." }, "candidate_id": "wav_5e3a173baa98eab3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('Etsy item') followed by a natural self-correction ('I should say Amazon order') and continues smoothly. The prosody is consistent, the repair is immediate and plausible, and the pacing reflects genuine spontaneous speech. The filler 'uh' after 'email' is well-placed and does not disrupt the flow." } ] }, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an amazon order as per my partner's email. Oh, I forgot to mention, do this today.", "long_pause_offsets": [ 49 ], "source_sha256": "2811b4b0a98b5dc6993675259c3f374d4090b8b4b7439087da8d6259a3a46e09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'today' to the initial request, which is a missing condition in the original instruction. The final goal combines both the action and the temporal requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/addition/addition_01/dedff6aced30ec9f/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8d77c52996072461317992188dc3be78871e054cf9f54b6b5868ec9a520073f3", "seconds": 9.4, "spoken_utterances": [ "Place an ... amazon order as per uh... my partner's email. ", "Oh, I forgot to mention, do this today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to place an Amazon order as per the partner's email. After a pause and an interjection, the speaker adds 'do this today,' which is a time constraint that narrows the scope of the original task. Since the type is 'addition,' the final goal is the combination of the original instruction and the added constraint. The final audible intent matches the reference instruction with the addition of the time constraint, so the task is executed with the new condition.", "heard_transcript": "Place an Amazon order as per my partner's email. Oh, I forgot to mention, do this today." }, "candidate_id": "wav_8532cb368d54159b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Place an Amazon order as per my partner's email') and then adds a time constraint ('Oh, I forgot to mention, do this today') with a believable hesitation and self-correction ('Oh, I forgot to mention'). The prosody is consistent, the pace is conversational, and the addition feels like a spontaneous afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my friend's email today. Wait, the email should be from my partner's email, not friend's.", "long_pause_offsets": [ 54 ], "source_sha256": "50885c543a640d2fb49e4279460bc194b24b403b5b9f07160c045b5c3798a31f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the sender relationship from 'friend' to 'partner', preserving the app, action, and time constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/revision/revision_02_email_sender_relation/7f420727e5de02cc/speech.wav", "wer": 0.2, "wav_sha256": "2b443abb76491bc7a99748a5817fe8a5c638bb579aaca099240af792359c4c93", "seconds": 15.36, "spoken_utterances": [ "Place an um... amazon order as per my friend's email um, let's see... today. ", "Wait, the email should be from my ... partner's email, not friend's." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friend's email' but immediately corrects it to 'partner's email' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original instruction exactly, including the action, target, and time constraint.", "heard_transcript": "Place an Amazon order as per my friend's email. Um, let's see. Today. Wait, the email should be for my partner's email, not friends." }, "candidate_id": "wav_27719973f625d9aa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision. The speaker begins with a clear request, hesitates with 'um' and 'let's see', then self-corrects 'friend's email' to 'partner's email' with a believable change-of-mind intonation and a smooth, continuous flow. The prosody is varied and human, with a natural pause and filler before the correction, and the correction itself is integrated seamlessly into the utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "7ea5683bcd1033d76234f3c014946aaa3c20fcf7b7a0e36c857d348fed4e11ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/cancellation/cancel_002/a7c84ff4c747f637/speech.wav", "wer": 0.0625, "wav_sha256": "0ab109c5181512161b9b073cdc8d1948eabf2dacab8ebeff9073d4828b9c4b41", "seconds": 10.16, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today.", "Wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting an Amazon order, but immediately corrects themselves with 'Wait, cancel the whole task.' This constitutes a clear cancellation of the initial request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an uh Amazon order as per my partner's email today. Wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "6849ccb9b772a17c6c356420a20eb80624d3907e23690ef5302de34b2024bd3e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/ambiguous_halt/halt_024/ebc79d5b7709edb8/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "7e3ba37d3281d485897fccbf7bfe112956c217b93c33129821c089c8920028c3", "seconds": 9.52, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial instruction is not completed, but the final control intent is to pause, which is consistent with the type.", "heard_transcript": "Place an uh Amazon order as per my partner's email today. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1a79e37_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "548fd87739e91ac7e0580f73382d7ae1fd657817b976badfa4897a77b5610cbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/filler/original/3214a6e1f4766c23/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "d0dbaf2e226a514b987d09712b016f16116c19e9e253fcf2a51e01378ec720b8", "seconds": 5.84, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Place an uh Amazon order...', where 'uh' is a filler sound. The rest of the sentence matches the original instruction. The final intent is to execute the Amazon order task as per the partner's email today. The filler does not change the goal, so the task is executed as intended.", "heard_transcript": "Place an uh Amazon order as per my partner's email today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 2 ebay—I should say amazon order.", "long_pause_offsets": [], "source_sha256": "b94291fb22606506d86a178fdcdf64dbb77efd14cadf1d4ec7083299dff8f977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the app, action, and scope of the return request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/false_start/false_start_03_app/99018dec7ae11620/speech.wav", "wer": 0.0, "wav_sha256": "fe697e89ae7bdd8d55caaa77642efc412a274e2dd4b2be400828706accaf7911", "seconds": 10.6, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my last two ebay—um... I should say amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'eBay' but immediately self-corrects to 'Amazon' after a pause and a self-correction phrase ('I should say'). According to the 'false_start' type criterion, the final corrected value ('Amazon') overrides the initial mistaken value ('eBay'), and the rest of the instruction remains consistent. The final intent matches the original instruction with the corrected target.", "heard_transcript": "Initiate returns via uh FedEx for everything in my last two eBay um, I should say Amazon order." }, "candidate_id": "wav_9ffd336b0cdd1723", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start with a self-correction ('eBay' to 'Amazon'), but the correction is less fluid. The speaker says 'eBay' and then hesitates with 'um' before correcting. The intonation during the correction is slightly more abrupt and less conversational than in audio_01, making the repair feel a bit more forced or less spontaneous." } ] }, "is_default": true } }, { "task_id": "1c4bd27_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 2 ebay order. Let me correct the app: return items from my last 2 amazon order, not ebay.", "long_pause_offsets": [ 67 ], "source_sha256": "c0fdd463b8e82b9946204068bb9345cf483257e5a3221d8035ac73154134bb1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the action, return method, and scope of 'everything in my last two orders'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/revision/revision_03_app/e59f8a49341bd3e8/speech.wav", "wer": 0.0, "wav_sha256": "3f4c52fb3ef83000dd18ca01b11ba1d30d6a689ad61827821c2663ca633b57d9", "seconds": 16.28, "spoken_utterances": [ "Initiate returns via let's see... FedEx for everything in my last two ebay order. ", "Let me correct the app: return items from my last two ... amazon order, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay order' but immediately corrects it to 'Amazon order, not eBay' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original instruction's action and target, with only the platform corrected from eBay to Amazon.", "heard_transcript": "Initiate returns via FedEx for everything in my last two eBay order. Let me correct the app, return items from my last two Amazon order, not eBay." }, "candidate_id": "wav_4934788123e68f39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear request, then uses a realistic filler ('let me see') before making a mistake ('eBay'). The self-correction ('let me correct the app') is delivered with a believable shift in intonation and pacing, followed by a smooth, grammatically coherent correction ('return items from my last two Amazon order, not eBay'). The prosody and rhythm are consistent with spontaneous speech, and the correction is integrated into the utterance rather than sounding like a separate command." } ] }, "is_default": true } }, { "task_id": "1c4bd27_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "284ad45e61883ca106bb022595a87d8ba9672f0763f460e74eec1456e8be1663", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/cancellation/cancel_011/b1a4dad02e836f70/speech.wav", "wer": 0.1, "wav_sha256": "c7c1cd40bf18553ac93de1dc008ce20c9044e9fef7ad86756520de0ecffa30cc", "seconds": 13.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order.", "Wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting a task, then clearly revokes it with 'Wait, I've changed my mind. Cancel that.' This constitutes a final, unambiguous cancellation of the original instruction. The control intent is correctly identified as 'cancel'.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order. Wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "fdab2cdcef4cfdafa4765b23c69dec6a1f8acdf16cc1599d40fdbb0dac45b351", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/ambiguous_halt/halt_002/b92d876c66763709/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "02141ef29db90cf79a48b3c904d0f605d922ab994518ff68f8d27e1380985110", "seconds": 11.16, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order.", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to initiate returns is not canceled, but the speaker explicitly halts further action. No correction or cancellation is made, and the final audible intent is to pause.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order. Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dc88a05584dad10f96ea7cbb7a4b2f1f691dc0040fcf65b9af4f889db3a98ce0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/filler/original/068af455177b6672/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "1d2de6b2bc66fd25b41ce1f33fee50573b8544aaa0328f77806197dd308e03e4", "seconds": 8.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'um' but completes the original instruction without changing the action, carrier, scope, or target. The only difference is the number 'two' replacing 'last', which is a minor grammatical variation that does not alter the core intent. The final goal matches the original instruction.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_shipping_service", "proposed_instruction": "Initiate returns via UPS—no, that should be FedEx for everything in my last 3 amazon order.", "long_pause_offsets": [], "source_sha256": "8812f780943c24e7eb07ff550a44c945132a0af046282b5644eea123a8760ada", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects the shipping service from UPS to FedEx, which matches the original instruction. The phrase 'last three amazon order' is interpreted as 'last three Amazon orders' based on context and standard phrasing, with no other conflicting changes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/false_start/false_start_01_shipping_service/07a35489e596d082/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "f547cc9977df732e294a250e10243dd2789bb2f00b8133385032e0fe159dec3f", "seconds": 9.52, "spoken_utterances": [ "Initiate returns via UPS—hmmm... no, that should be uh... FedEx for everything in my last three amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'UPS' but immediately self-corrects to 'FedEx' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value ('FedEx') should be applied, and the original incorrect value ('UPS') should be discarded. The rest of the instruction, including 'for everything in my last three Amazon order', remains unchanged and is preserved. The final intent matches the original reference instruction exactly.", "heard_transcript": "Initiate returns via UPS. Hmm. No, that should be uh FedEx for everything in my last three Amazon order." }, "candidate_id": "wav_2dea46d2a07b651f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'UPS', immediately self-corrects to 'FedEx' with a clear, spontaneous-sounding 'no, that should be', and smoothly continues the full request. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the final sentence is delivered with natural intonation and pacing." } ] }, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Initiate returns via FedEx for everything in my last amazon order. Limit that to the last 3 amazon orders.", "long_pause_offsets": [ 67 ], "source_sha256": "0d3d8c7bcb1e9f8c065b7540bc5e6a913ddbd5ad0d0601b7b03f8402767584e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to limit the scope from 'last amazon order' to 'last three amazon orders', which is a narrowing of the original instruction but does not contradict it. The action, carrier, and item scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/addition/addition_02/79e1dc4682e1561e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "10e9861970b05437a4cdacb1861168be228be858fab649959cfa465f70bba3f7", "seconds": 14.84, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my um... last amazon order. ", "Limit that to the last ... three amazon orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request for the 'last Amazon order' and then adds a constraint to 'limit that to the last three Amazon orders'. This is a clear addition that narrows the scope of the original instruction. The final intent is a combined task that matches the reference instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Initiate returns via uh FedEx for everything in my um last Amazon order. Limit that to the last three Amazon orders." }, "candidate_id": "wav_83ea37987773ae81", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_shipping_service", "proposed_instruction": "Initiate returns via UPS for everything in my last 3 amazon order. Wait, use FedEx for the returns, not UPS.", "long_pause_offsets": [ 67 ], "source_sha256": "ba628c7836cbbfb8a71ed64f601e9bbedf067ce18c4fec76c44cf62b731f278c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the shipping service from UPS to FedEx, preserving the action, scope (everything in the last three Amazon orders), and the intent to initiate returns." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/revision/revision_01_shipping_service/d335ef0650188995/speech.wav", "wer": 0.04, "wav_sha256": "3fdef6c3300789742303e1afb7695d7c83dcd2f4cbbbf84a0b4943c205b1b703", "seconds": 14.88, "spoken_utterances": [ "Initiate returns via ... UPS for everything in my um... last three amazon order. ", "Wait, use um, let me think... FedEx for the returns, not UPS." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially requested returns via UPS but then corrected themselves to FedEx after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, with only the carrier changed from UPS to FedEx as per the revision.", "heard_transcript": "Initiate returns via UPS for everything in my um last three Amazon order. Wait, use um, let me think. FedEx for the returns, not UPS." }, "candidate_id": "wav_8f9c6ffaecb0b93d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural revision process. The initial request is made with a plausible hesitation ('um'), followed by a self-correction ('wait, use um...') that sounds like genuine recollection. The speaker then clearly corrects the carrier from 'UPS' to 'FedEx' and continues with a smooth, continuous flow. The prosody and pacing are consistent with spontaneous speech, and the correction is integrated naturally within the utterance." } ] }, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "16e2e4026c756b5ef60be788040c83563bb80da5e68233db3d9957b74de4f689", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/cancellation/cancel_034/1832c59aa2b77344/speech.wav", "wer": 0.0, "wav_sha256": "a06398014df84fde9e767b77ed5a62d60b273542f13cd73cdaa7c11241f849b6", "seconds": 12.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order.", "Oh, wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Oh wait, let's drop the whole task.' This constitutes a final, unambiguous cancellation of the entire task. The original instruction is irrelevant after this cancellation.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order. Oh wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "f023c8a4498d9f5e79fc253dcdad13ddb919b75fc633d9d97f61af21d6a7b258", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/ambiguous_halt/halt_036/48e8d1de8191860b/speech.wav", "wer": 0.0, "wav_sha256": "5677a0b5c912d37002a582068437c78d88fc7570f4a03d6abc043a26a87772b5", "seconds": 11.6, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly interjects 'Wait a second, before you go ahead,' which is a direct request to pause. The final interjection is a pause command, not a cancellation or a revision. The task is not completed, but the audio ends with a valid request to halt, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0cda63a6a9649e314e0f9c94b6acf09d6091ff5119719a39c1dd944712c6858f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/filler/original/44f60b04526fac66/speech.wav", "wer": 0.0, "wav_sha256": "410820375327bcdeca627e5aecf9fcf66151f3f3801d24a8feb418be3a9444ae", "seconds": 7.56, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'uh', 'let me think') and a brief pause but does not change the core request. The final intent is to initiate returns via FedEx for all items in the last three Amazon orders, which matches the original instruction. The hesitation and self-interruption are typical of a filler type and do not alter the task goal.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_shipping_service", "proposed_instruction": "Initiate returns via UPS—I meant FedEx for everything in my last 4 amazon order.", "long_pause_offsets": [], "source_sha256": "173eda3abc9e3a500836a68d83c136c4aabe91d4a80e1b228890594c4bba2e93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'UPS' to 'FedEx' within the same utterance, preserving the action, target (Amazon order), and scope ('everything in my last four')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/false_start/false_start_01_shipping_service/0baa025e689999cb/speech.wav", "wer": 0.0625, "wav_sha256": "926136bad4f88782d3ae41fa89ddbd305347ef040fdb2cad9e7bcc2c9780e409", "seconds": 9.96, "spoken_utterances": [ "Initiate returns via UPS—uh... I meant FedEx for everything in my ... last four amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'UPS' but immediately self-corrects to 'FedEx'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the scope 'everything in my last four Amazon order', is preserved. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Initiate returns via UPS. I meant FedEx for everything in my last four Amazon order." }, "candidate_id": "wav_a7782a605b7f06a0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'UPS' and immediately self-corrects to 'FedEx' with a clear, spontaneous 'I meant' phrase. The correction is integrated smoothly into the sentence, maintaining a consistent rhythm and prosody. The pause before the correction is brief and plausible, suggesting genuine recall rather than a rehearsed or robotic edit. The overall delivery sounds like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Initiate returns via FedEx for everything in my. I left out a detail: only return items from my last 4 Amazon orders.", "long_pause_offsets": [ 49 ], "source_sha256": "dd7b9b072dc008795124b116f4e5c098eb5fd438a08a83cac3cefa7ec637537c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial incomplete request, narrowing the scope from 'my last 4 amazon order' to 'my last four Amazon orders'. The action, app, and return method remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/addition/addition_02/983ee708fbf35c9a/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "a9dfaf0a7516882f9d08e63da80b5e8a21ddc10c69920738212f228e59919d41", "seconds": 14.48, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my. ", "I left out a detail: only return items from my hmmm... last four Amazon orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the original instruction, then adds a clarifying constraint ('only return items from my last four Amazon orders') after a pause and self-correction. This matches the 'addition' type, where the final goal combines the original request with the added detail. The final intent is clear and executable.", "heard_transcript": "Initiate returns via FedEx for everything in my... I left out a detail, only return items from my last four Amazon orders." }, "candidate_id": "wav_e082ea602054dfa8", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_shipping_service", "proposed_instruction": "Initiate returns via UPS for everything in my last 4 amazon order. Wait, use FedEx for the returns, not UPS.", "long_pause_offsets": [ 67 ], "source_sha256": "63c50916764e61516c9695c04ab3c2f74cd9fa3616fd845bd39c8029390a1e78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the shipping service from UPS to FedEx, while retaining the scope of 'everything in my last four Amazon order'. The original instruction's 'FedEx' is replaced by 'UPS' in the first utterance, but the second utterance explicitly corrects it back to 'FedEx'. The phrase 'last four Amazon order' is preserved as spoken, matching the original's intent despite minor grammatical variation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/revision/revision_01_shipping_service/f25b7182d1d3c35f/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "f03f22f94b2aebdea7b7fc4d07db5259d90dc61aeb744e04ca0c55f525a3fc6d", "seconds": 18.16, "spoken_utterances": [ "Initiate returns via uh, let me think... UPS for everything in my ... last four amazon order. ", "Wait, use um, let me think... FedEx for the returns, not UPS." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions UPS but immediately corrects themselves to FedEx after a pause and self-correction. The final instruction clearly specifies FedEx, which matches the original reference. The revision type criterion is satisfied as the correction is applied to the earlier instruction.", "heard_transcript": "Initiate returns via uh, let me think, UPS for everything in my last four Amazon order. Wait, use um, let me think, FedEx for the returns, not UPS." }, "candidate_id": "wav_80a3581af235458d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear false start and self-correction. The speaker initially says 'UPS' but then corrects to 'FedEx'. The hesitation with 'let me think' and the pause before the correction sound natural. However, the correction is slightly less fluid than in audio_02, with a more abrupt shift in the phrase, making it feel a bit more like a mechanical edit than a fully spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "f79c3ccc9883e09c4c2b70b3e2e3003b8129febcd0233b804cf1fee91e415045", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/cancellation/cancel_024/2eb68924920dde89/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "16252e4923411b56b03236f42b055a7b7bf49ed39ceeeabdc8279066e0639a29", "seconds": 14.96, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order.", "Oh, wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then immediately corrects themselves with 'Oh wait, oh never mind, cancel the whole thing.' This constitutes a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or modification of the original task. Therefore, the final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order. Oh wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "08667f0889b34256e3c11ec667fe9374fc55c5e3d8378ba664afb3a6229ac647", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/ambiguous_halt/halt_010/60dd400e2742f25f/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "215f529929c11985dace545822b987e34c5256ad88b4e07d8726e039cce1e90e", "seconds": 13.48, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there any indication that the task should be resumed. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "1c4bd27_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c23a79719407f18738c0ae65565b4c8e193c3343b8bf4516194f87a1729b7452", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/filler/original/fcb98c494a9a79b1/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "3388213c2bef6fec66adf0357fb6d1e3e8caa8631d68c0bcde532c2159983861", "seconds": 9.8, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Initiate returns via FedEx for everything in my last four Amazon order.' The word 'four' is pronounced clearly and matches the reference. There are no hesitations, self-corrections, or changes in the goal. The final intent is to execute the original instruction as heard, with no ambiguity or cancellation. The use of 'four' instead of '4' is a natural paraphrase and does not change the meaning. Therefore, the task is executed as intended.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_transaction_relation", "proposed_instruction": "How much money have I received from—no, I meant sent to others on venmo this month so far?", "long_pause_offsets": [], "source_sha256": "03ac185971da4a327733f6588b8a1eceedf811599db26984a063bf3344c006b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'received from' to 'sent to' within the same utterance, preserving the recipient ('others'), app ('Venmo'), and time frame ('this month so far'). The final intended action is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/false_start/false_start_01_transaction_relation/066fde155163fc61/speech.wav", "wer": 0.1, "wav_sha256": "110c29c1e290a19a23516e72bb90e9014d5c4b37c74f439cdfe10b638494f999", "seconds": 6.84, "spoken_utterances": [ "How much money have I received from—uh... no, I meant sent to others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received from' but immediately self-corrects with 'no, I meant sent to others'. This is a clear false start where the final intent matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How much money have I received from uh, no, I meant sent to others on Venmo this month so far?" }, "candidate_id": "wav_0e81ed5d52d0de93", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'received' and immediately corrects to 'sent' with a clear, spontaneous 'no, I meant' phrase. The prosody is consistent, the correction is integrated smoothly into the sentence, and the overall delivery sounds like a genuine moment of recall. The pause before the correction is brief and plausible, fitting a natural thought process." } ] }, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How much money have I sent to others on this month so far? Oh, and one more thing: do this on Venmo.", "long_pause_offsets": [ 59 ], "source_sha256": "bb73b4fde097cc43c26e50de2943fd022c33fefcff0c9fb82698ed7c80922999", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app 'Venmo' to the initial question, which already includes the correct action, recipient scope, and time frame. The final goal combines both utterances without altering the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/addition/addition_02/17edd243d141dda7/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "64d394d29a434a2395144e057923c4e0f42fbad54c21d8ac71eb476fbd460edb", "seconds": 11.12, "spoken_utterances": [ "How much money have I sent to others on um... this month so far? ", "Oh, and one more thing: do this on ... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is a near-verbatim repetition of the original instruction with a minor hesitation ('um'). The speaker then adds 'do this on Venmo', which clarifies the app context. Since the original instruction already specified Venmo, this addition is redundant but does not contradict or alter the core task. The final intent is to execute the original request with the app context confirmed, satisfying the addition type criterion.", "heard_transcript": "How much money have I sent to others on um this month so far? Oh, and one more thing, do this on Venmo." }, "candidate_id": "wav_c191c25fcb7c4235", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The 'um' filler is brief and contextually appropriate, suggesting spontaneous speech. The addition ('Oh, and one more thing...') is delivered with a believable shift in intonation and a slight pause, followed by a coherent, additive clarification. The flow feels like a real-time thought process, with the second part clearly building on the first." } ] }, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How much money have I sent to others on cashapp this month so far? Wait, the app should be venmo, not cashapp.", "long_pause_offsets": [ 67 ], "source_sha256": "09134c3a32857370cc1de42d632af7f81e3744a11b4241f7ff35ec2fae9a62bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Cashapp to Venmo, preserving the action, recipient scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/revision/revision_02_app/24e15dee027b1c62/speech.wav", "wer": 0.28, "wav_sha256": "c1f8f28639bdeac2e7ff7dbeed7e492b998321248f6f531bea343cb540a64d1b", "seconds": 11.88, "spoken_utterances": [ "How much money have I sent to others on uh... cashapp this month so far? ", "Wait, the app should be let's see... venmo, not cashapp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Cash App' but immediately corrects it to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction exactly, and the control intent is to execute the corrected task.", "heard_transcript": "How much money have I sent to others on the Cash App this month so far? Wait, the app should be uh, Venmo, not Cash App." }, "candidate_id": "wav_4b65683c2bb98446", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question. After a brief pause and self-correction ('Wait, the app should be...'), they smoothly revise the app name from 'Cash App' to 'Venmo' with a plausible hesitation and intonation shift. The correction is integrated into the same utterance, maintaining a continuous and realistic flow without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "6f89e7651ff5f9eb98e8e3badb62ecd3c1251910314bd11d757c49a73c20bc36", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/cancellation/cancel_025/bee83857604eacfd/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "feb6ee6f25f0ec1ec1155888838d2c385881f39b150cfafb08154d8eeb1ebf25", "seconds": 10.32, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?", "Hold on, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, forget that request. Don't do it,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far? Hold on, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "4154fd0539e53da0305565773e44b8dcf2fe2fb28885320ec91688730234b545", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/ambiguous_halt/halt_006/08fd55429975b668/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "67516f72acabe00be9bb3895ee5ed825401b8bd2d79467d7ec8b5bde736af870", "seconds": 8.52, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the phrase 'Just a moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the agent is expected to hold the task until the user is ready to proceed.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far? Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "973b92b1e04e9339def0cffb67d2b7871a831481a817425b4a1568ff5b4e535d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/filler/original/01c50ce8b29cb37d/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "b9dd3bad3ca0135e419716005533b8ebf8cd6a8aed6deea574db62742cf40eb7", "seconds": 5.44, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'on uh Venmo', where 'uh' is a filler word. The rest of the sentence is clear and matches the original instruction. The final intent is to execute the query as stated, with no change in action, target, or value. The filler does not alter the task goal.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_transaction_relation", "proposed_instruction": "How much money have I sent to—no, I mean received from others on venmo this month so far?", "long_pause_offsets": [], "source_sha256": "72cf356fe718045655bef010bebf0d269a110527227bee1711dd85c95a08f77e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent to' but immediately corrects to 'received from' using 'no, I mean'. The final action is 'received from', and all other components (money, app, time frame) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/false_start/false_start_01_transaction_relation/bfdb7b47b02a3e1a/speech.wav", "wer": 0.0, "wav_sha256": "8d93e8bfc64abfb1423ca9f4327756e85673bc73e15eb14835b13f50d2003bb3", "seconds": 11.0, "spoken_utterances": [ "How much money have I sent to—uh... no, I mean um... received from others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'received'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How much money have I sent to uh, no, I mean um, received from others on Venmo this month so far?" }, "candidate_id": "wav_e9f1a829d448bfeb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'How much money have I sent to...' but immediately self-corrects with 'No, I mean...' and smoothly transitions to the correct request. The intonation and pacing are consistent with spontaneous speech, and the correction feels like a genuine moment of realization rather than a rehearsed edit. The delivery is clear, with appropriate pauses and filler words ('uh', 'um') that enhance realism." } ] }, "is_default": true } }, { "task_id": "21abae1_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How much money have I received from others on cashapp this month so far? Let me correct the app: use 'venmo', not 'cashapp'.", "long_pause_offsets": [ 73 ], "source_sha256": "40e4bd010fe094d04810ee2ba6150d95e6c163a66c2a20e6c1cee28b5b4224f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'cashapp' to 'venmo', preserving the rest of the query including the action, recipient scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/revision/revision_02_app/edf1c23d76b97d08/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "40f1a4471e12532f10ae8a01365f1e927ebfa503a9ab635a4cdbfe88106f9d6e", "seconds": 13.52, "spoken_utterances": [ "How much money have I received from others on cashapp ... this month so far? ", "Let me correct the app: use 'let me think... venmo', not 'cashapp'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task for Cash App but then corrects it to Venmo after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the amount received on Venmo, which matches the original reference instruction. All other constraints (e.g., 'this month so far') are preserved.", "heard_transcript": "How much money have I received from others on Cash App this month so far? Let me correct the app, use, let me think, Venmo, not Cash App." }, "candidate_id": "wav_d111967d23b33225", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. The speaker begins with a clear request, then hesitates with 'let me think' before self-correcting the app name from 'cash app' to 'venmo'. The correction is integrated smoothly into the utterance with a natural pause and filler, and the prosody reflects genuine recollection rather than a rehearsed or robotic delivery. The rhythm and intonation are consistent with a person formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "21abae1_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "9fe44cfaa2a19046190e6ce6fd591f17ce73c86688fdc1b8c003cf41ed0d2656", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/cancellation/cancel_018/ace2e00bd4bec21a/speech.wav", "wer": 0.0, "wav_sha256": "143d486e6af138dfb7ebc80721bdd8d368c7af09ec364ca73a42e9f9b7a3955b", "seconds": 12.88, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?", "Hang on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question but then clearly revokes the entire task with 'Hang on, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the original request. The final interjection is a complete cancellation command, not a pause or ambiguous halt.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far? Hang on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "7316d0e2d169ad6f7f714c177f408c3ac4cfdb87f33de44f8faac4d778dedb4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/ambiguous_halt/halt_033/72eca3ee8822509f/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "7282b3e77c0bcb52c006adb202ad8550430de1f3fcc1cf9d8202cc7e6ccbd7bf", "seconds": 11.64, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on, do something,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary stop rather than a cancellation or execution. The initial request is not revoked, and the speaker does not resume or correct it, so the final intent is to pause.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far? Hold on, do something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "54cbd3b2db0b1f77c3c647967b7ab38ff6498a71abf1ecb5baceed709cf1940e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/filler/original/35ae898b7e7a346a/speech.wav", "wer": 0.0, "wav_sha256": "f81fc62e780d2d43c27d75e34fb93f7d6af902a8b216e1945ac2ac10653d87fc", "seconds": 7.44, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent remains to ask for the amount received from others on Venmo this month. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How much money have I sent to or received from others on cashapp—I should say venmo this month so far?", "long_pause_offsets": [], "source_sha256": "3c48c4d005089474fde64c81edb2c17a6094990a3ed2aef0fa54a03257e52e5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the full scope of the original request including the app, action, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/false_start/false_start_02_app/6895fd81e02718ff/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "d415fb78f70a8e67424679e8d3839b6ba8a19c7b21972248c0d303ba66203d65", "seconds": 12.88, "spoken_utterances": [ "How much money have I ... sent to or received from others on cashapp—hmmm... I should say venmo let's see... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is executed with the correction applied. The rest of the instruction remains unchanged.", "heard_transcript": "How much money have I sent to or received from others on Cash App, uh, I should say Venmo, let's see, this month so far?" }, "candidate_id": "wav_bc4862140713270f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('How much money have I...') and immediately self-corrects from 'CashApp' to 'Venmo' with a plausible 'I should say' repair. The intonation is conversational, and the final phrase 'this month so far' is delivered smoothly without further interruption, suggesting a single, continuous thought process." } ] }, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How much money have I sent to or received from others on this month so far? I should also mention, this is on Venmo.", "long_pause_offsets": [ 76 ], "source_sha256": "44234c2a9c1ce3c5c81b029807cd578cc192d83224db9bf30526a5f0f0999518", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the app as Venmo, which was omitted in the initial utterance. The core question about money sent/received on Venmo this month remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/addition/addition_02/cd0ec0df7e37e94e/speech.wav", "wer": 0.04, "wav_sha256": "20e761c652f45c858129f72be33e81eb5949c6916336daf05bb45b16cbb5d42b", "seconds": 12.48, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... this month so far? ", "I should also mention, this is on Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and completed. The speaker adds the detail 'this is on Venmo' after a pause, which clarifies the platform. This is a valid addition that narrows the scope of the original request without changing its core intent. The final goal matches the original instruction with an added constraint, satisfying the addition type criterion.", "heard_transcript": "How much money have I um sent to or received from others on this month so far? I should also mention, this is on Venmo." }, "candidate_id": "wav_24346d54ae797a38", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_transaction_relation", "proposed_instruction": "How much money have I sent to or received by others on venmo this month so far? Wait, it should be 'sent to or received from', not 'sent to or received by'.", "long_pause_offsets": [ 80 ], "source_sha256": "d9c67968d8e3eb6fa64969cff54d90cf0870b382d0b499252d6fb9a79f1c6682", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the preposition from 'by' to 'from', which is the only difference. The rest of the query, including the app (Venmo), time frame (this month so far), and action (sent to or received from) remains consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/revision/revision_01_transaction_relation/634b4db091ddb2af/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "76e7ad92051d863451fccc2b4a7be1b9ef7ac9c292f09dd34e4cfeff5fb22624", "seconds": 16.12, "spoken_utterances": [ "How much money have I um... sent to or received by others on venmo uh... this month so far? ", "Wait, it should be 'sent to or received from', not 'sent to or received by'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received by' but immediately corrects it to 'received from' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "How much money have I um sent to or received by others on Venmo uh this month so far? Wait, it should be sent to or received from, not sent to or received by." }, "candidate_id": "wav_8790ab93560701a5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a plausible self-correction. The initial phrasing 'sent to or received by' is slightly awkward, leading to a believable hesitation and revision to 'sent to or received from'. The correction is smoothly integrated with a natural 'wait' and clear intonation shift, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "3f8bd089df016a19eb50420c5620e8f7e39df554f48608d18df35c70af5622fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/cancellation/cancel_001/4dc57d4104336130/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "dcdd41a6ae8f706d94e88934e835e3be672271340f55baf7c791c52474ba123e", "seconds": 13.36, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately self-corrects with 'Oh wait, cancel that,' which is a clear and final cancellation of the entire task. The audio ends with this cancellation, so no execution is expected. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far? Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "bd17ed2cb49c844644f8fb71f6264f12c7465be5226e448a79c917890bd07a8c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/ambiguous_halt/halt_002/d983b3d27993e954/speech.wav", "wer": 0.05, "wav_sha256": "fcd1ebc9dc4d4fba0ceeb159b5c2e2ab3d7c400fc37687af6721105c8447f6c6", "seconds": 12.64, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then audibly says 'Hold on,' which is a clear request to pause or wait. There is no cancellation or instruction to resume. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "21abae1_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "17377c100a87dcd38fe55793df4fa7b3090be93a86c4594d3ae379bb149264a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/filler/original/ed8e5634c86f86db/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "161521122ed0c9c21b5185746415e41334eb347830720634a6a2a0a94e8d60c8", "seconds": 9.36, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um' as a hesitation filler but completes the original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I received—no, make that sent this month, have in total?", "long_pause_offsets": [], "source_sha256": "bcce32c5aa049d4d7cd4028018fb7ba3416227b3c5f45d4e1e14626d82a6222e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'received' but immediately corrects it to 'sent' within the same utterance. The final intended direction of the transactions is 'sent', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/false_start/false_start_03_transaction_direction/e44e851755944f29/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "0a9f2a1edb47e45e5da5c1afb0f45bb6aae5d9c134b7fd5e9cedc80e28c88e26", "seconds": 11.12, "spoken_utterances": [ "How many likes did uh... all Venmo transactions, let me think... I received—um... no, make that sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I received' but immediately self-corrects to 'make that sent' after a pause. According to the 'false_start' type criterion, the final corrected value ('sent') should be applied, and the original incorrect value ('received') should be discarded. The final intent correctly reflects the original instruction with the corrected verb. All other constraints (Venmo transactions, this month, total likes) are preserved.", "heard_transcript": "How many likes did uh all Venmo transactions, let me think. I received um, no, make that sent this month, have in total?" }, "candidate_id": "wav_16bdab9cb1a9d48e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. The speaker begins with 'likes' and 'Venmo transactions', then hesitates with 'uh' and 'let me think', before correcting 'received' to 'sent'. The correction is integrated smoothly into the sentence with a natural 'no, make that' phrase, and the prosody includes believable pauses and filler words that suggest genuine thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How many likes did all Venmo transactions, this month, have in total? I didn't mention this, but only count the ones I sent.", "long_pause_offsets": [ 70 ], "source_sha256": "c779f33cfbb5ff03fbdb84d7ade8b89f66838c852d62e94cd4d0f04d4e5e6816", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that only counts transactions I sent, which is consistent with the original intent of focusing on the user's own transactions. The app (Venmo), time frame (this month), and metric (total likes) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/addition/addition_02/c950c5161da81de2/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "51641c6be5135fb875099d566847cf428069bb63e130558119ac57e60dc8cea1", "seconds": 12.04, "spoken_utterances": [ "How many likes did um... all Venmo transactions, this month, have in total? ", "I didn't mention this, but only count the ones ... I sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is a slightly garbled version of the reference, but the final addition 'only count the ones I sent' correctly narrows the scope to transactions sent by the speaker. The addition is spoken after a pause and modifies the original request, satisfying the 'addition' type criterion. The final intent matches the reference with the added constraint.", "heard_transcript": "How many likes did um all Venmo transactions this month have in total? I didn't mention this, but only count the ones I sent." }, "candidate_id": "wav_e92c86078c2b7385", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um') and a plausible self-correction ('I didn't mention this, but...'). The addition of the restriction 'only count the ones I sent' is smoothly integrated with a believable pause and change of thought, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent last month, have in total? Make that this month, not last month.", "long_pause_offsets": [ 77 ], "source_sha256": "eb0e6b456516ea884f12292953a28234fb3f5e2512aae00f5f85e528503306af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app (Venmo), the action (counting likes), and the scope (transactions I sent)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/revision/revision_04_time_reference/79218930e4585eb1/speech.wav", "wer": 0.04, "wav_sha256": "a950154c7b6637bc3ded845874e5b8ff172d6e9bfa2c9ea256ed9aa03d56f963", "seconds": 12.44, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I sent last month, have in total? ", "Make that um, let me think... this month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a self-correction phrase ('let me think'). According to the revision type criterion, the correction spoken after a pause (or self-correction) overrides the earlier instruction. The final audible intent is to ask about 'this month', which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "How many likes did all Venmo transactions, I sent last month, have in total? Make that, um, let me think, this month, not last month." }, "candidate_id": "wav_2ef414a477028280", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects 'last month' to 'this month' with a plausible 'um' and 'let me think' filler. The correction is smoothly integrated into the sentence, maintaining a consistent rhythm and intonation. The prosody and pacing are highly realistic, resembling a spontaneous thought process without robotic or rehearsed qualities." } ] }, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "1ad842b53eaabaed84ae53b3829a1dcc9544c0bea46637d4a8f1b812fa2cb0e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/cancellation/cancel_013/416c6c31e1c48d85/speech.wav", "wer": 0.12, "wav_sha256": "059df336f3059e891799835317c7919887815703795abe89015869217b9b5f76", "seconds": 13.36, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?", "Wait, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a question, then includes a clear self-correction with 'Wait, I changed my mind. Please cancel the task.' This constitutes a final, explicit cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. Therefore, the final control intent is cancellation, matching the expected behavior for the 'cancellation' type.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total? Wait, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "49e2773055700b9025208efe72746379eb24ae919a60b3d47a0586722cbc5833", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/ambiguous_halt/halt_039/4b7a9fa4e15c316d/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "6d64189d92bf4185e3653b80dc8bcdbaa0ce2e0d3de9be0c4a02a0f331354a5f", "seconds": 12.44, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then clearly interjects with 'Wait, give me a second first,' which is a standard expression indicating a need to pause or take a moment. This is a final interjection that does not cancel the task or request a different action, but rather asks for a temporary halt. The audio ends with this pause request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2eb41f1b031e9f64d9a531bd9cc276177029c540f8da87a01e7ae3481c5626c9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/filler/original/8d0c2bbee44f0797/speech.wav", "wer": 0.1875, "wav_sha256": "76e1961aa4b278db543eafd0fa84978a10dbd36a4753b27328ee550197258b92", "seconds": 8.04, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I received last year—sorry, that should be this year, have in total?", "long_pause_offsets": [], "source_sha256": "383ee606573c66e7f47fb06870b0d506a874fa2876010df50e8263622138b2ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year'. The rest of the query, including the app (Venmo), the subject (transactions I received), and the action (total likes), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/false_start/false_start_04_time_reference/ccdd47174f51a024/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4859febc9043eeb7d8a77c665f2513e942e49674098b551cd9206e64f745f390", "seconds": 8.24, "spoken_utterances": [ "How many likes did all Venmo transactions, I received last year—hmmm... sorry, that should be ... this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('last year') is discarded. The final instruction matches the original reference instruction in both action ('how many likes') and target ('all Venmo transactions I received this year').", "heard_transcript": "How many likes did all Venmo transactions I received last year have? Sorry, that should be this year have in total." }, "candidate_id": "wav_3312309351bddd45", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage self-correction. The initial request is clear, followed by a believable 'uh' hesitation, a self-aware 'sorry' that explicitly references the error ('that should be this year'), and a smooth, unforced correction. The prosody includes a slight pitch drop on 'uh' and a return to the original tone after the correction, mimicking genuine thought process. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many likes did Venmo transactions, I received this year, have in total? Oh, I forgot to mention, include all Venmo transactions.", "long_pause_offsets": [ 76 ], "source_sha256": "7c6dca1308e3efcdd4ba82639eb5171795cc859dbffb6130db5b61da9c098d5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the phrase 'include all Venmo transactions', which clarifies that the scope should be broadened to all transactions, not just those previously implied. However, the original instruction already specified 'all Venmo transactions, I received this year', so the addition is redundant but not contradictory. The final goal remains the same." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/addition/addition_01/bb8a20f350262750/speech.wav", "wer": 0.08, "wav_sha256": "3886718cf3f333438c6002bb8fdccdedd0488b9aaa98a8b563686b79468755f6", "seconds": 13.96, "spoken_utterances": [ "How many likes did Venmo transactions, ... I received this year, have in total? ", "Oh, I forgot to mention, include uh, let's see... all Venmo transactions." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many likes did Venmo transactions I received this year have in total?'. After a pause and an interjection, the speaker adds 'include uh let's see, all Venmo transactions.' This addition clarifies and broadens the scope to include all Venmo transactions, which is consistent with the original intent of counting likes for Venmo transactions received this year. The addition does not contradict or remove any prior conditions and is a natural expansion of the initial request. Therefore, the final intent is to execute the combined task.", "heard_transcript": "How many likes did Venmo transactions I received this year have in total? Oh, I forgot to mention, include uh let's see, all Venmo transactions." }, "candidate_id": "wav_3d272b9feaa07cfc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, then adds a detail ('Oh, I forgot to mention...') with a believable hesitation ('uh, let's see') before smoothly incorporating the missing information. The prosody is consistent, and the addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I received last year, have in total? Make that this year, not last year.", "long_pause_offsets": [ 80 ], "source_sha256": "885b71bb96d118ac281c9301e2e7f567b3689ec0568464654612d9a006a26487", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last year' to 'this year', preserving the rest of the query including the app (Venmo), the subject (transactions I received), and the requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/revision/revision_04_time_reference/56342bf9b672ac03/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "8c8ba0423550e7d1578f6187f7d79db3716b83b7c04430e33ee2d09ca71982de", "seconds": 12.88, "spoken_utterances": [ "How many likes did uh... all Venmo transactions, ... I received last year, have in total? ", "Make that um, let me think... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects to 'this year' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to ask about 'this year', which matches the original reference instruction. All other constraints (likes, Venmo transactions, received, total) are preserved.", "heard_transcript": "How many likes did uh all Venmo transactions I received last year have in total? Make that, um, let me think, this year, not last year." }, "candidate_id": "wav_6fdf125ac4e64b7b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then naturally self-corrects the time frame from 'last year' to 'this year' with a plausible 'um' and 'let me think' filler. The correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody. The intonation and rhythm are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "ea23cd321dbf92a213eb3b85255f477639bf7cfc457c91aa2274a6561a0f3617", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/cancellation/cancel_029/d7bea74086de87dc/speech.wav", "wer": 0.05, "wav_sha256": "2a8ca893d52853cff4e0b237a4ff83ae37757e28f3d1996292735a0c9fa2321c", "seconds": 10.64, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?", "Hold on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task and then clearly and unambiguously cancels it with 'Hold on, I'm withdrawing that request.' This matches the expected behavior for a 'cancellation' type, where the final intent is to revoke the task. No execution of the original request is required.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total? Hold on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "e22f4ddda9947eac2465a788726e82d1055b843fa269a9054cbea47546839cd5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/ambiguous_halt/halt_039/b6338bd8f9333a6a/speech.wav", "wer": 0.05, "wav_sha256": "f4ea8f631dba30787f9201d961bf6e99164a9faf370925b320189f07fdd39705", "seconds": 9.64, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, give me a second first,' which is a clear request to pause or hold. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the verdict is pass.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c64bee225adaa125097d3ca322ddf15c64a4adce603751170a58a81792e32c51", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/filler/original/d483edc0f3c8c650/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "9acc0c627aed3a597ea85d13449f8244347c2cae96ef7b131304739b2d1f3bc0", "seconds": 6.12, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction as intended, with no changes to the action, target, or value. The 'filler' type allows for natural pauses or hesitations, but the final goal is clear and matches the original instruction.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How many likes did all CashApp—sorry, I meant Venmo transactions, I sent or received this month, have in total?", "long_pause_offsets": [], "source_sha256": "8e373dd13bf6f48f1bb8d8e624fb9901574dd5bc6e369315962fde7f7493d1b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance. The rest of the instruction, including the action (counting likes), the scope (all transactions sent or received), and the time frame (this month), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/false_start/false_start_02_app/6f86730ee34c0db8/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "6773dc3244dab4b870aa43531ba73a88d7a3b9001156d199191aa956e0bd4203", "seconds": 9.48, "spoken_utterances": [ "How many likes did all CashApp—um... sorry, I meant ... Venmo transactions, I sent or received this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'sorry, I meant'. According to the 'false_start' type criterion, the final corrected value ('Venmo') should be applied, and the original incorrect value ('Cash App') should be discarded. The rest of the instruction, including the scope ('transactions I sent or received this month') and the action ('how many likes... have in total'), remains unchanged. The final audible intent matches the corrected instruction.", "heard_transcript": "How many likes did all Cash App um, sorry, I meant Venmo transactions I sent or received this month have in total?" }, "candidate_id": "wav_34898d611f329976", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a clear, spontaneous false start with 'Cash App' followed by a natural self-correction to 'Venmo transactions'. The speaker's intonation and pacing are consistent with someone recalling the correct term mid-sentence. The correction is immediate and integrated smoothly into the flow, with no awkward pauses or unnatural phrasing. The overall delivery sounds like a genuine, unscripted moment of retrieval." } ] }, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many likes did all Venmo transactions, I this month, have in total? Oh, I forgot to say, include both transactions I sent and received.", "long_pause_offsets": [ 72 ], "source_sha256": "126a05f2e90c691de70d483ee29f8db93acf2ba97c7b844d85e998b258bf38b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'include both transactions I sent and received', which clarifies the scope of transactions to include. The original instruction already specified 'all Venmo transactions' and 'this month', so the final goal is fully specified and consistent with the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/addition/addition_01/0c8eb979721d7b4c/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "3a65828148e9a3ce7b19d3bcfe8dfd1428daa7576489814fae71e008dbbf243f", "seconds": 17.28, "spoken_utterances": [ "How many uh... likes did all um... Venmo transactions, I this month, have in total? ", "Oh, I forgot to say, include both transactions I uh, let me think... sent and received." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is partially garbled but the core intent to count likes on Venmo transactions is clear. The speaker then adds the missing constraint 'include both transactions I sent and received' after a pause and self-correction, which completes the original instruction. The addition clarifies the scope without changing the action or target, so the final goal matches the reference.", "heard_transcript": "How many uh likes did all um Venmo transactions I this month have in total? Oh, I forgot to say, include both transactions I uh let me think, sent and received." }, "candidate_id": "wav_fcd7019524eefa43", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('How many uh likes did all um...') and then adds a clarifying detail ('Oh I forgot to say...') with a believable 'let me think' pause. The addition is smoothly integrated into the same utterance, maintaining a consistent rhythm and intonation. The prosody reflects genuine recall and spontaneous correction, making it sound like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent or received last month, have in total? That should be 'this month', not 'last month'.", "long_pause_offsets": [ 89 ], "source_sha256": "c4a673908b66b3604c0e30d308f824d8dbe4ebfaf78ecff5a00ee6b80614544f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app (Venmo), transaction scope (sent or received), and the requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/revision/revision_04_time_reference/2ad55c56bf20f562/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "7c7e03baf04d1d6077e8d0e06b47d63d6cc4c68e09d21abf0a632d3ad91a46f1", "seconds": 15.84, "spoken_utterances": [ "How many um... likes did all uh... Venmo transactions, I sent or received last month, have in total? ", "That should be 'um, let me think... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'last month' but immediately self-corrects to 'this month' after a pause and repair. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the total likes for Venmo transactions from 'this month', which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "How many um likes did all uh Venmo transactions I sent or received last month have in total? That should be um let me think, this month, not last month." }, "candidate_id": "wav_1694bdc29eeffac2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage revision process. They begin with a hesitant 'um' and 'uh', then correct 'last month' to 'this month' with a plausible 'let me think' filler and a self-correcting 'that should be'. The prosody is continuous and the intonation reflects genuine thought and correction, making the revision sound spontaneous and realistic." } ] }, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "2906abcc9354e74d106ae123bce2a33a93fa4c78222153d8d8cebc7a5d997aec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/cancellation/cancel_029/6fe7a1be4c71d05d/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "622e82a6016d9092ffcb8ca214c98949b4dd685af4acf2e8cd64a7171f3e5e86", "seconds": 14.64, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?", "Wait, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins a request but then clearly states 'Wait, I'm withdrawing that request' as the final interjection, which constitutes a cancellation of the entire task. The initial request is irrelevant once the cancellation is audibly confirmed.", "heard_transcript": "How many uh likes did all, let me think, Venmo transactions I um sent or received this month have in total? Wait, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "cde8ea0d977fd77cd4f046b9d60123b4c95a2df667ce9cf3aa4280dfb2cd63f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/ambiguous_halt/halt_008/dff53c7153202209/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "fbc3b58d0071283730bd910f4942d7b92daf57585ba03db77420ed9052eeb4a9", "seconds": 13.24, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then says 'let me think' and 'wait a second' as final interjections. These are audible requests to hold or pause, with no indication of cancellation or resumption. The final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How many uh likes did all, let me think, Venmo transactions I um sent or received this month have in total? Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "23cf851_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0b997508f1f0ae92b45080ac2ad357a9c5dd182dfc788280e88da8df733857dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/filler/original/f1c904514d4aa08a/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ecfac9d84834fd4905d18c5dc726386063100a79f58a2a281ead41e5b46e3246", "seconds": 10.28, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let me think', 'um') and pauses but does not change the core request. The final audible intent matches the original instruction. The filler type criterion allows ignoring hesitations and backchannels, so the task is executed as intended.", "heard_transcript": "How many uh likes did all let me think Venmo transactions I um sent or received this month have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail inbox—scratch that, make it inbox and outbox before the last calendar month as read.", "long_pause_offsets": [], "source_sha256": "193aed7be3e5cbdc22abba6d7dcec927815c9ddd9b928dd7aca35239bfdc0add", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'inbox and outbox'. The final scope includes both folders, matching the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/false_start/false_start_02_email_folder_scope/fc937eb9cda3e138/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "fe8559e79778a21156aececd60f684bd969610fc9907d773d5383b86c731521c", "seconds": 10.08, "spoken_utterances": [ "Mark everything in my Gmail inbox—hmmm... scratch that, make it ... inbox and outbox before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Mark everything in my Gmail inbox' but immediately self-corrects with 'Scratch that. Make it inbox and outbox before the last calendar month as read.' This is a false_start where the final instruction replaces the initial one. The final instruction matches the original reference in both action ('mark as read') and scope ('inbox and outbox before the last calendar month'). No other constraints differ. Therefore, the final intent is fully aligned with the reference.", "heard_transcript": "Mark everything in my Gmail inbox. Hmm. Scratch that. Make it inbox and outbox before the last calendar month as read." }, "candidate_id": "wav_9228e9abdf8c4fdd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with a thoughtful 'hmm' before self-correcting 'scratch that' and smoothly revising the task to include 'outbox' and 'last calendar month as read'. The correction is immediate and the prosody reflects genuine reconsideration, with a natural pause and intonation shift. The flow is continuous and the repair feels spontaneous." } ] }, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as read. Oh, I forgot to mention, mark everything before the last calendar month as read.", "long_pause_offsets": [ 54 ], "source_sha256": "bc9b6dca7f8e366c5b32bc1f4709ab6b9b2cd644d9cd6b8b76b92bede366b029", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'before the last calendar month' to the initial instruction, which already specified the app (Gmail), locations (inbox and outbox), and action (mark as read). The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/addition/addition_01/d39c2e92da92e1db/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "1fedc40944873724900fd06b5638bef39dd4b49e358622a9f4b5a2c8e5b5fe0f", "seconds": 15.44, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and um... outbox as read. ", "Oh, I forgot to mention, mark everything let's see... before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to mark everything in the inbox and outbox as read. After a pause and self-correction, the speaker added the constraint 'before the last calendar month'. The final instruction combines the original request with the added temporal condition. The final intent matches the original instruction with an additional constraint, satisfying the 'addition' type criterion.", "heard_transcript": "Mark everything in my Gmail inbox and um outbox as read. Oh, I forgot to mention, mark everything, let's see, before the last calendar month as read." }, "candidate_id": "wav_abcb84f6c932262f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly natural, multi-stage thought process. It begins with a clear initial request, includes a believable self-correction ('Oh, I forgot to mention'), and smoothly adds a new temporal condition ('before the last calendar month as read'). The speaker uses natural fillers ('um', 'let's see') and pauses that align with real-time formulation. The prosody is conversational and the transitions between ideas feel spontaneous, not rehearsed." } ] }, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before last week as read. I misspoke about the time filter: it should be before the last calendar month, not before last week.", "long_pause_offsets": [ 71 ], "source_sha256": "27394dd482f0cb0a936090599c6c27b1dc9f2c6de8ed3b6a490d5e4e933222a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time filter from 'last week' to 'last calendar month', while retaining the app (Gmail), locations (inbox and outbox), and action (mark as read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/revision/revision_03_time_filter/90d67470b0cea631/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "39a27a92f75004954d4dd39ddcd001f0b9fd42e970e30abded1e582d17e93f3d", "seconds": 16.04, "spoken_utterances": [ "Mark everything in uh, let's see... my Gmail inbox and um... outbox before last week as read. ", "I misspoke about the time filter: it should be before the last calendar month, not before last week." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'before last week' but immediately corrects themselves with 'I misspoke about the time filter. It should be before the last calendar month, not before last week.' This is a clear revision where the final intent is the corrected version. The original instruction is overridden by the revision, and the final goal matches the reference instruction exactly.", "heard_transcript": "Mark everything in uh let's see my Gmail inbox and um outbox before last week as read. I misspoke about the time filter. It should be before the last calendar month, not before last week." }, "candidate_id": "wav_6fcbb664e07c8fbb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a highly realistic revision. The speaker begins with a natural, slightly hesitant formulation ('mark everything in uh let's see my Gmail inbox and um outbox before last week as read'), including a filler ('uh') and a self-correction ('let's see'). After a brief pause, they clearly state the error ('I misspoke about the time filter') and provide a coherent, grammatically correct revision ('it should be before the last calendar month, not before last week'). The prosody is natural, with appropriate stress and intonation, and the correction flows seamlessly from the initial mistake, mimicking how a real person would spontaneously correct themselves." } ] }, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e68d62ccf751d4a484e56ee00b52b96c4e9ce4fb7f217e324dc0212822cb62ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/cancellation/cancel_009/3c04a28d2108cf6c/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "2a7736d3e8bb9ce00e955182f72c98331756546f1b18a9c66167cdc3a28572ca", "seconds": 12.2, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task and then immediately follows with 'Oh wait, cancel that altogether,' which is a clear and unambiguous cancellation of the entire prior request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "88c51a1334e96e51084b8564158765bba813334cbc8888db33aabe4e8dfbd1f9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/ambiguous_halt/halt_027/250186e6c3fe0a44/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "252a192d98bd0a4f1e94e1cb991feadb253ef4471d5975b3d2f3e9dbb5900529", "seconds": 12.4, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, I need a moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the verdict is pass.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1f57d5a49606b97f0dc379c81ea900135b17ccc8300d8def668b384639c7b4d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/filler/original/372e4ff5758b7235/speech.wav", "wer": 0.0625, "wav_sha256": "109a016d61eb346afa923088f9f987a231d691ef849b75b10695706fd12a6c97", "seconds": 7.92, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core task. The final intent matches the original instruction, with no correction or cancellation. The task is executed as requested.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox in the current calendar month as read—sorry, I meant unread.", "long_pause_offsets": [], "source_sha256": "73b501ba9b6eb62ecd6e6c760377f3de95e6e07506c11c463d7063b33d159d17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'read' but immediately corrects to 'unread' within the same utterance. The final intended action is to mark items as unread, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/false_start/false_start_04_email_state/f6b2d9ec625d6573/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "8e4a97925e83d7e1ae154c7e0cb0e3484b8333d0d1da432f12484a7a5d536699", "seconds": 11.6, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and outbox in the current calendar month as read—uh... sorry, I meant um... unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as read' but immediately self-corrects to 'unread'. According to the false_start type criterion, the final corrected value should be applied. The correction is clear and immediate, and the rest of the instruction remains unchanged. Therefore, the final intent is to mark items as unread, which matches the original instruction after correction.", "heard_transcript": "Mark everything in my um Gmail inbox and outbox in the current calendar month as read. Uh, sorry, I meant um unread." }, "candidate_id": "wav_030f1c48351d46c5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start with 'is read uh sorry I meant um unread'. The self-correction is immediate, the filler 'uh' and 'um' are placed at natural points of hesitation, and the intonation reflects genuine recollection. The prosody is consistent and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as unread. Oh, I forgot to mention, this should apply only to the current calendar month.", "long_pause_offsets": [ 56 ], "source_sha256": "bc66294daf1fdeaf9d74cfaf67ae8b27a9e3a1236a635cbf5dab4e0c879c0d6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time constraint that narrows the scope of the initial instruction. The core action and targets remain unchanged, and the new condition is a restriction, not a contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/addition/addition_01/53b6b7fbefd666eb/speech.wav", "wer": 0.08, "wav_sha256": "4ae3c779ef592120363cc0f981cf45a51f5d63b7da0e1fb4474d128bb0c161a3", "seconds": 13.88, "spoken_utterances": [ "Mark everything in my Gmail inbox and hmmm... outbox as unread. ", "Oh, I forgot to mention, this should apply only to the ... current calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to mark all emails in the inbox and outbox as unread. After a pause, they add a constraint that this should only apply to the current calendar month. Since the addition is a clarifying constraint and not a contradiction, the final intent is the original task with the added restriction. The type is 'addition', and the final goal matches the original instruction with the added condition, so it passes.", "heard_transcript": "Mark everything in my Gmail inbox and outbox as unread. Oh, I forgot to mention, this should apply only to the current calendar month." }, "candidate_id": "wav_90a211a9299633cc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and filler ('hmm') before correcting 'inbox' to 'outbox'. The self-repair is smooth and integrated into the flow. The addition of the calendar month restriction is a believable, contextually relevant follow-up with a natural-sounding 'oh, I forgot to mention'. The prosody is consistent, and the speaker's rhythm reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox in the current calendar month as read. That should be unread, not read.", "long_pause_offsets": [ 84 ], "source_sha256": "d775e9d7c54ea61987d8fc1011166f9107a97b15f2d3a07f93f7df43757470b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'read' to 'unread', preserving all other conditions including the app, locations, and time scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/revision/revision_04_email_state/484b31948c9ce6a0/speech.wav", "wer": 0.0, "wav_sha256": "4c45d1e52947cfd4c99193546abe2c86fe95022bb458b6a821655ed4e7c0c9ee", "seconds": 16.6, "spoken_utterances": [ "Mark everything in my um, let's see... Gmail inbox and ... outbox in the let me think... current calendar month as read. ", "That should be unread, not read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as read' but immediately corrects themselves with 'That should be unread, not read.' This constitutes a revision where the final audible intent is to mark items as unread. The correction is clear and directly modifies the initial incorrect value, satisfying the revision type criterion.", "heard_transcript": "Mark everything in my um, let's see, Gmail inbox and outbox in the current calendar month as read. That should be unread, not read." }, "candidate_id": "wav_dc6b5ea67c8f6c5d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'let me see', 'let me think') that are well-placed and reflect genuine hesitation. The revision ('That should be unread, not read') is a clear, spontaneous correction with a believable change-of-mind intonation and a direct reference back to the initial request. The prosody is continuous and conversational, with a natural rhythm and pacing throughout the entire utterance." } ] }, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "5ccb15ed12c1f9cdba1a3fccd8643e926838a2bc4a5606484222610dfe28a8d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/cancellation/cancel_026/31c7173adc6e6847/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "ca1f135ba0600582b8e325b43c11aaa46a3727e3d475843a8cef7d6e5a0275f2", "seconds": 14.68, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread.", "Oh, wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to mark emails as unread, then immediately says 'forget what I just asked, cancel it,' which is a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread. Oh wait, forget what I just asked, cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "e94fcbc518d3430802af55499806f7a5ad7db3bb07798bbf8b174ec1831324e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/ambiguous_halt/halt_007/8b2f908dac872aa9/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "65e74458d62ff42aa1d0e01c0476439d54aa112e664d63c09d01607ccd83fe54", "seconds": 13.12, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full instruction and then says 'One moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original instruction is fully stated before the pause.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5603b16745a2524a2d9d75ba1fbf752db3932e375344a63733bb049d0b8ad770", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/filler/original/27024453b0168fcf/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "941d827d7f1dd13e6b57f5dad1520141d4a990cca8322653eebdc6be5bfdc9c6", "seconds": 9.64, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'uh') but does not change the core task. The final intent matches the original instruction, with no self-correction or modification. The task is to mark items as unread, which is preserved. The use of 'inbox and outbox' is consistent with the original request.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the current fiscal year—sorry, that should be before the current calendar year as read.", "long_pause_offsets": [], "source_sha256": "36938f00401532d839f65d5dff0e15c819e65bf72ef371087d9fc160de6dcbcd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'fiscal year' but immediately corrects it to 'calendar year'. The rest of the instruction, including the action, scope (inbox and outbox), and target condition (before the year), remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/false_start/false_start_03_time_filter/04d830f3fb508296/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "d15a512ff77ca93f99af45672d208783f56a8f3de434b4142b87020561aa49d4", "seconds": 12.84, "spoken_utterances": [ "Mark everything in my Gmail inbox and outbox hmmm... before the current fiscal year—um... sorry, that should be before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fiscal year' but immediately self-corrects to 'calendar year' and maintains the rest of the instruction. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the current fiscal year. Um, sorry, that should be before the current calendar year as read." }, "candidate_id": "wav_f03934c2008a43b1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker initially says 'fiscal year' and then smoothly corrects to 'calendar year' with a believable 'sorry' and a clear, continuous intonation shift. The pause and filler 'um' are placed naturally, and the overall prosody sounds like a spontaneous, real-time thought process without any robotic or rehearsed qualities." } ] }, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as read. Oh, I forgot to mention, only mark emails from before the current calendar year.", "long_pause_offsets": [ 54 ], "source_sha256": "03b85a336c7b1c65a33fa20b2369d7e7ca4f9c0d752198bfa890891d8a9c68ae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing temporal restriction that narrows the initial broad request. The final goal combines the original scope (inbox and outbox) with the new condition (emails before the current calendar year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/addition/addition_01/0a7d5f154a3f09eb/speech.wav", "wer": 0.1, "wav_sha256": "906fcc4aa294f15b726a930d567f1e18e1c86abea568677bbe391bc7b5104e7d", "seconds": 20.08, "spoken_utterances": [ "Mark everything in um, let me think... my Gmail inbox and uh... outbox as read. ", "Oh, I forgot to mention, only mark emails from um... before the current calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the base instruction to mark everything in the inbox and outbox as read. After a pause and an interjection, they add a new constraint: 'only mark emails from before the current calendar year.' This is a clear addition that narrows the scope of the original task. The final intent combines the original action with the new condition, matching the reference instruction. The addition is spoken after a longer pause, satisfying the 'addition' type criterion.", "heard_transcript": "Mark everything in um, let me think, my Gmail inbox and outbox as read. Oh, I forgot to mention, only mark emails from um, before the current calendar year." }, "candidate_id": "wav_17d4abb5b52f576d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the end of last year as read. I misspoke about the time filter: it should be before the current calendar year, not before the end of last year.", "long_pause_offsets": [ 82 ], "source_sha256": "70ad2060c304488e25d10af46018905500707f560010f4a2746c16704ee2fdfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time filter from 'before the end of last year' to 'before the current calendar year', while retaining the app (Gmail), locations (inbox and outbox), and action (mark as read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/revision/revision_03_time_filter/ddf716836ac42b0b/speech.wav", "wer": 0.05, "wav_sha256": "47d5cde0b5e3b2da9dba6f1075fd72943c7edd14a800c78839e8793801c200e2", "seconds": 21.52, "spoken_utterances": [ "Mark everything in um... my Gmail inbox and outbox ... before the end of last year as read. ", "I misspoke about the time filter: it should be let's see... before the current calendar year, not before the end of last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives an incorrect time filter ('before the end of last year') but immediately self-corrects with a longer pause and revision ('before the current calendar year'). According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to mark items before the current calendar year as read, which matches the original instruction. The correction is clear and unambiguous, and the rest of the task (targeting Gmail inbox and outbox, marking as read) remains unchanged.", "heard_transcript": "Mark everything in um my Gmail inbox and outbox before the end of last year as read. I misspoke about the time filter. It should be, let's see, before the current calendar year, not before the end of last year." }, "candidate_id": "wav_126a73ecbbc9bb12", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker first makes a clear request, then self-corrects with a believable hesitation ('I misspoke about the time filter'), followed by a smooth, self-repairing explanation ('It should be...') that directly references and corrects the earlier value. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "54faded14461597bddbe612c8549f8db3930c201f6e8f0513e46dd158733df03", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/cancellation/cancel_009/989a996d807601a3/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "8cb118a09f9a3bd71b60ba4b4def9a9b7de399399e5821f2306882d19b8526b4", "seconds": 13.96, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly states 'cancel that altogether' as the final interjection, which constitutes a direct cancellation of the entire task. The cancellation is unambiguous and not followed by any resumption or modification of the original request.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "9973d4853a155ef7a2791d617adde81f7aac2b40abbcfc704ff54dfedd1fbb3f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/ambiguous_halt/halt_024/e6dd716031da44ba/speech.wav", "wer": 0.04, "wav_sha256": "64e08172e16314d9fae19427f9e8bb7b730efaa8eaf966ed83ca985c1efa5866", "seconds": 13.08, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler phrase 'um, let's see' and a final interjection 'Give me a moment to think' which clearly indicates a request to pause. The original instruction is not canceled or revised, and the speaker does not execute or correct the task. The final audible intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "277d81d_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8fc2f3ef4a99da50e62fc19a1576e35244aca23beace8eb58604ff4764424715", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/filler/original/f028080affb76b51/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "a19591b65cfd7fbefb67b20a99e297cdd8f3fc0c24885a60a233acbf8a97a4fc", "seconds": 9.32, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see', 'uh') and pauses but does not change the core task. The final intent matches the original instruction exactly, with no self-correction or modification. The filler type allows ignoring hesitations and backchannels, so the task is executed as stated.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_recommendation_order", "proposed_instruction": "Name the artist least recommended—no, I meant most recommended to me on Spotify.", "long_pause_offsets": [], "source_sha256": "6d9f84a485d6197033927e5561f48978c9dfbaa370e2e5bfb36f16e826198b1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects to 'most recommended'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/false_start/false_start_01_recommendation_order/1de6bc11fe5e2d2a/speech.wav", "wer": 0.0625, "wav_sha256": "30924511d230fa159c7e58c6a69fcc24cdff6455d7ae9fcbb5ad3420a0e71bd2", "seconds": 9.72, "spoken_utterances": [ "Name the artist least recommended—um... no, I meant uh... most recommended to me on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately self-corrects to 'most recommended' with a pause and 'no, I meant'. According to the 'false_start' type criterion, the final corrected value should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Name the artist least recommended um, no, I meant uh most recommended to me on Spotify." }, "candidate_id": "wav_0ad733b42165d01b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start and self-correction. The initial 'least recommended' is followed by a clear 'no, I meant' with a believable pause and intonation shift, leading into the corrected 'most recommended to me on Spotify'. The pacing, filler 'um', and the pragmatic flow of the correction feel spontaneous and unscripted, closely mimicking how a person would actually rephrase a thought mid-sentence." } ] }, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Name the artist on Spotify. Oh, I forgot to say, it should be the artist most recommended to me.", "long_pause_offsets": [ 28 ], "source_sha256": "71fc8bdc7aebbbae9c75c188374b9771229dfbbf05d97a79e32663b1e6a52bd9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'most recommended to me' to the initial broad request 'Name the artist on Spotify', refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/addition/addition_01/7689c921b62e771c/speech.wav", "wer": 0.0, "wav_sha256": "9d6860cfe32cd4ec13bf3d356e4a7695a813a2e43d671b7473780455b613fbff", "seconds": 11.76, "spoken_utterances": [ "Name the artist on let me think... Spotify. ", "Oh, I forgot to say, it should be the artist ... most recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Name the artist on Spotify' and then adds the constraint 'most recommended to me' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added information. The final intent is clear and matches the original instruction.", "heard_transcript": "Name the artist on, let me think, Spotify. Oh, I forgot to say it should be the artist most recommended to me." }, "candidate_id": "wav_f8c16c8297cbb068", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') and then clearly adds a crucial detail ('it should be the artist most recommended to me') as a post-hoc clarification. The addition is smoothly integrated into the sentence, with a believable shift in intonation and pacing that suggests spontaneous thought. The prosody is consistent, and the speaker does not sound rehearsed or robotic." } ] }, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_recommendation_order", "proposed_instruction": "Name the artist least recommended to me on Spotify. Wait, the artist should be most recommended to me, not least recommended.", "long_pause_offsets": [ 52 ], "source_sha256": "1c17bfad75e7da5c5801d6eb7e3512f4d3b90fb56808c6336a84aaa83567992d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial instruction by replacing 'least recommended' with 'most recommended', while retaining the app and the scope of 'artist recommended to me'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/revision/revision_01_recommendation_order/94ae1dc2112ab8fd/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ec7c77a891c7a594c86bfc4e144d3b389be162052f6cbf7a60bbd3d4b79bfac6", "seconds": 12.76, "spoken_utterances": [ "Name the artist ... least recommended to me on um... Spotify. ", "Wait, the artist should be ... most recommended to me, not least recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects themselves with 'Wait, the artist should be most recommended to me, not least recommended.' This is a clear revision where the final intent is the corrected version. The original instruction was 'Name the artist most recommended to me on Spotify,' and the final audible intent matches this exactly. Therefore, the revision criterion is satisfied.", "heard_transcript": "Name the artist least recommended to me on um Spotify. Wait, the artist should be most recommended to me, not least recommended." }, "candidate_id": "wav_6405e3580afade86", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural revision pattern: the speaker starts with 'least recommended', then self-corrects to 'most recommended' with a clear, spontaneous 'wait' and a smooth, grammatically correct revision. The prosody is conversational, with a believable hesitation and intonation shift during the correction. The pacing and filler usage ('um') are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "29b7955c72285f66c5452915b1dda3505c9e93274861b609c67c91ef9ed458db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/cancellation/cancel_018/9abd378e83b48fe2/speech.wav", "wer": 0.0, "wav_sha256": "c7104f8fdbb371bce324c5e68eb88994949f25b3e5efb4a3873147b6ba8324e1", "seconds": 11.56, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify.", "Oh, wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to name the artist, but immediately follows with 'Oh wait, I'd rather not do that after all. Please cancel it.' This constitutes a clear and final cancellation of the original request. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "Name the artist most recommended to me on Spotify. Oh wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "df3f7ea240160850c294f650335e65ff5d6a77d884b80293d55c9528d9b98cdf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/ambiguous_halt/halt_028/6a34cec30a019983/speech.wav", "wer": 0.0, "wav_sha256": "aca4de2b1f42d8188f0e8a4cf29f8b3cee3ffb33f386142a35f8640a1f3099a5", "seconds": 10.0, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. No cancellation or execution is requested, and the original task is not revoked.", "heard_transcript": "Name the artist most recommended to me on Spotify. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8ed19a907434b9f76abbe74b8932e01662f36e6a0ac5c4e1c878b35a98c09f04", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/filler/original/d113222461630072/speech.wav", "wer": 0.0, "wav_sha256": "f0dd52b3c0d91ffd5b20e04341c2a84e420ee227e56b50d4698963c43226d848", "seconds": 5.12, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or repair. The speaker delivers the full request without interruption, so the final intent matches the original instruction exactly. As this is a 'filler' type, any initial hesitation or pause within the instruction is ignored, and the final goal is to execute the task as stated.", "heard_transcript": "Name the artist most recommended to me on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Name the artist least recommended to me on Apple Music—sorry, I meant Spotify.", "long_pause_offsets": [], "source_sha256": "043a14bbb5f587d33436dd967e0ef0a9eb0b6af69051a93e1df6bb6d63a97fe2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, direction, and target entity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/false_start/false_start_02_app/d5600b107d75ea33/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8006f136a5cdc8122a49e1672f7c945cffa805fe823cc239090dd1ef22c996d3", "seconds": 11.48, "spoken_utterances": [ "Name the artist um, let's see... least recommended to me on Apple Music—ah... sorry, I meant ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Name the artist, um, let's see, least recommended to me on Apple Music, uh, sorry, I meant Spotify." }, "candidate_id": "wav_3abac10807ec9012", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural, spontaneous flow. The initial hesitation ('um', 'let's see') and the self-correction ('sorry I meant Spotify') are delivered with authentic prosody and pacing. The correction is immediate and integrated smoothly into the ongoing thought process, making it sound like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Name the artist to me on Spotify. Oh, I forgot to mention, it should be the artist least recommended to me.", "long_pause_offsets": [ 34 ], "source_sha256": "0ed74da8d48bba1f2f47ec478dd47da39f8af1962ec2c7988a0184ca984b6e49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'least recommended to me', which refines the initial vague request. The app (Spotify) and action (name the artist) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/addition/addition_01/a4bf53eed120f8e4/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5bc606c07193cbee73b34921bcfee163290fa5000ecf438c9870243096255d33", "seconds": 14.84, "spoken_utterances": [ "Name the let's see... artist to me ... on Spotify. ", "Oh, I forgot to mention, it should be the artist uh, let's see... least recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Name the let's see artist to me on Spotify', which is incomplete and unclear. After a pause, they correct and add the missing constraint 'it should be the artist uh let's see least recommended to me'. This addition completes the original instruction by specifying the correct artist type. The final intent matches the original instruction, and the addition is applied correctly.", "heard_transcript": "Name the let's see artist to me on Spotify. Oh, I forgot to mention, it should be the artist uh let's see least recommended to me." }, "candidate_id": "wav_e1de749aaf860e0b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Name the artist least recommended to me on Apple Music. Let me correct the app: the artist should be least recommended on Spotify, not Apple Music.", "long_pause_offsets": [ 56 ], "source_sha256": "1c358bb0052878859b2a873bc88cc098dd9decf8c40278e0ac9c11d8f50e3b2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the action, the user reference, and the recommendation direction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/revision/revision_02_app/a9cdccc51ccef5e8/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "1847a14b6387eed39d3ec72cf93ebef9445b43b12fb9d74a821a54b0bc80d20f", "seconds": 13.84, "spoken_utterances": [ "Name the ... artist least recommended to me on Apple Music. ", "Let me correct the app: the artist should be least recommended on hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "Name the artist least recommended to me on Apple Music. Let me correct the app. The artist should be least recommended on Spotify, not Apple Music." }, "candidate_id": "wav_5489329460ab7df1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural prosody and a believable self-correction. The pause before 'Let me correct the app' is plausible, and the revision from 'Apple Music' to 'Spotify' is smoothly integrated with a natural change-of-mind intonation and reference back to the earlier statement. The pacing and rhythm are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "6e367a622305cdbfa3b7382a29fd2594fbca0943cbc6a813a9077909bc1154df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/cancellation/cancel_003/b2a04900cbfa552f/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1bc0ec145fc0e47efed92259ba5063dbb010c0277f48e5a9d92ac41bbda69b4a", "seconds": 11.68, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify.", "Wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the original task but immediately corrects themselves with 'Wait, please cancel that,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify. Wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "ead5ac39a51b22d339013d24e69791233ec9a88edc06b6f96e627e1587888271", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/ambiguous_halt/halt_032/4891670b48c7c390/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "ecc04da179438d556114a2e1098d01f4b15225336d7930ab6ce6198c37c31b31", "seconds": 11.64, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Wait, before you do that,' which is a clear request to pause the action. The final interjection is a pause command, not a cancellation or a new instruction. The audio ends with this request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify. Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "71342c7f979cf1e41456b312ff93ad728c4a4d9af10d9b4b2e4310b126e24974", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/filler/original/27968835c2499b93/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "743eb9205c8e29ffcc0fef39a897d20628bd1febf0621efb1e36c1dbea2053f6", "seconds": 7.56, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not change the core request. The final intent remains to name the artist least recommended on Spotify, matching the original instruction. The fillers do not constitute a pause or cancellation, so the task should be executed.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "287e338_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_recommendation_order", "proposed_instruction": "Name the artist least recommended—no, that should be most recommended to me on Spotify.", "long_pause_offsets": [], "source_sha256": "6d368664f10f93ccdac337a3f7630560887d13e37dd414068e63886ae6a59078", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects it to 'most recommended'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/false_start/false_start_01_recommendation_order/18708be9eb9ae1dd/speech.wav", "wer": 0.0, "wav_sha256": "08c26762093e96b9b6547c42088d48807fdfbbe9324342ec0fe00b7fe969b3b3", "seconds": 11.48, "spoken_utterances": [ "Name the artist um, let's see... least recommended—uh... no, that should be most recommended to me on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately self-corrects with 'no, that should be most recommended'. This is a clear false start where the final, corrected value is the intended goal. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Name the artist um let's see least recommended uh no, that should be most recommended to me on Spotify." }, "candidate_id": "wav_be7af67a54344757", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "287e338_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Name the artist on Spotify. I didn't mention this, but it should be the artist most recommended to me.", "long_pause_offsets": [ 28 ], "source_sha256": "d1e815ce0acbacb147df5013e177d7deb660a3cad2031b784b5571565f7e40b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'most recommended to me' to the initial broad request 'Name the artist on Spotify', refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/addition/addition_01/8bf4fe6bb38edf13/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "b114f50bd3111ea5e92d3d440d096ad2827a91914df4532a687f8ea87cf267fa", "seconds": 11.6, "spoken_utterances": [ "Name the artist on um... Spotify. ", "I didn't mention this, but it should be the artist uh, let's see... most recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Name the artist on Spotify' is broad. The speaker then adds the constraint 'most recommended to me' after a pause, refining the request. The final intent combines the original request with the added detail, matching the reference instruction. The addition is clear and not contradicted.", "heard_transcript": "Name the artist on um Spotify. I didn't mention this but it should be the artist uh let's see, most recommended to me." }, "candidate_id": "wav_110bc411b429b529", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "287e338_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_recommendation_order", "proposed_instruction": "Name the artist most listened to me on Spotify. Wait, the artist should be most recommended to me, not most listened.", "long_pause_offsets": [ 48 ], "source_sha256": "a298b33aad55223d21ee04a6c0c54d998ee910806155b73cba76a7cb6a0e29c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'most listened to' to 'most recommended to', preserving the app and the recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/revision/revision_01_recommendation_order/a3eb0c154d4d3493/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "26a72a9ea0012e053391a88d5efd14ae12db99402e638a29ebd679ceaf1b1665", "seconds": 13.76, "spoken_utterances": [ "Name the artist most listened to me on hmmm... Spotify. ", "Wait, the artist should be um... most recommended to me, not most listened." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most listened to me' but immediately corrects themselves with 'Wait, the artist should be um most recommended to me, not most listened.' This is a clear revision where the final intent is to find the most recommended artist, not the most listened-to artist. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Name the artist most listened to me on Spotify. Wait, the artist should be um most recommended to me, not most listened." }, "candidate_id": "wav_2719e7d5503c6c13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly realistic revision pattern. The speaker begins with a plausible initial request ('Name the artist most listened to me on Spotify'), then hesitates with 'Hmm' and a pause, followed by a self-correction ('Wait, the artist should be...') that clearly references and revises the earlier phrasing ('most listened' to 'most recommended to me'). The intonation and pacing of the correction sound spontaneous and natural, with a believable shift in thought. The flow from initial statement to correction is seamless and typical of real human speech." } ] }, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $250 on venmo to Caitlyn—sorry, I meant Catherine.", "long_pause_offsets": [], "source_sha256": "16b44c72224b2f224a55e521a83b3b543f54f36466e2a6200c805f3f898f35db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caitlyn' but immediately corrects to 'Catherine' with 'I meant Catherine'. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/false_start/false_start_03_person_name/de9dafedec603577/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "31ce76f16856ef58eb4277dbda0c349512ec966e178229114231cb4722d8cb05", "seconds": 9.12, "spoken_utterances": [ "Send um... two hundred fifty dollars on venmo to Caitlyn—uh... sorry, I meant Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caitlyn' but immediately self-corrects to 'Catherine' after a pause and apology. According to the 'false_start' type criterion, the final corrected target should be applied, and the original incorrect value discarded. The amount and app remain consistent. The final intent matches the original instruction.", "heard_transcript": "Send um two hundred and fifty dollars on Venmo to Caitlyn. Oh, sorry, I meant Catherine." }, "candidate_id": "wav_567b69d2e129f3a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker initially says 'Caitlyn' and then clearly and smoothly corrects to 'Catherine' with a believable 'sorry, I meant' phrase. The pause before the correction and the slight hesitation in 'um' at the start sound spontaneous. The overall prosody is conversational and the flow from initial error to correction feels authentic." } ] }, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Send $250 on to Catherine. Oh, and one more thing: use venmo to send the money.", "long_pause_offsets": [ 27 ], "source_sha256": "13bb1d60023b950a37108bb6bae33b1e0aa70c5e364e80d52fa12f1c256ead98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the payment method as Venmo, which is consistent with the original instruction. The amount and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/addition/addition_02/f9a38c7fd9ba4fa2/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "5251d0768ea1f738fa05bb5aec06bd8f4deff6ec421debfcc0b34e0ebe1b8f52", "seconds": 11.84, "spoken_utterances": [ "Send uh, let me think... two hundred fifty dollars on to um... Catherine. ", "Oh, and one more thing: use venmo to send the money." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction to send $250 on Venmo to Catherine is completed by the addition of 'use Venmo to send the money' after a pause, which clarifies the method. The final intent combines the initial request with the added detail, matching the original instruction exactly. No contradictions or omissions remain.", "heard_transcript": "Send uh, let me think, two hundred and fifty dollars on to um Catherine. Oh, and one more thing, use Venmo to send the money." }, "candidate_id": "wav_f0d1b01ae0bedb55", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a clear intent, uses a believable 'let me think' filler, and smoothly adds a new detail ('Use Venmo') after the initial request. The prosody is consistent, with natural pauses and intonation that suggest spontaneous thought. The addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_person_name", "proposed_instruction": "Send $250 on venmo to Caleb. Make that Catherine, not Caleb.", "long_pause_offsets": [ 29 ], "source_sha256": "198c0180f149a88e4e9655706d427466ad99ef136eebb85b16daf27f93e1d240", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from Caleb to Catherine, while retaining the amount, app, and action. The original instruction's recipient is replaced as intended." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/revision/revision_03_person_name/03b78c964f79488f/speech.wav", "wer": 0.0, "wav_sha256": "0b4c207cc8148460577fcf44f73a6bf879b09992e4a625f9205897ced87009ff", "seconds": 12.48, "spoken_utterances": [ "Send um... two hundred fifty dollars on venmo to uh, let me think... Caleb. ", "Make that Catherine, not Caleb." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caleb' but immediately corrects to 'Catherine' with 'Make that Catherine, not Caleb.' This is a clear revision where the final target is Catherine. The amount and app remain unchanged. The revision type criterion is satisfied.", "heard_transcript": "Send um two hundred and fifty dollars on Venmo to uh let me think Caleb. Make that Catherine, not Caleb." }, "candidate_id": "wav_8225011549c01bac", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear intent, uses a believable 'let me think' pause to recall the recipient, and then smoothly self-corrects from 'Caleb' to 'Catherine' with a natural intonation shift. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "1f8e60be03eb1b2c63ded40b150aaddd62d693ae0941c0d0b08c39de78cb4435", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/cancellation/cancel_034/4dc062c558a96447/speech.wav", "wer": 0.15, "wav_sha256": "5a266b1ac360b1811462db9f644dad1329cacc996c2278d4165110fdf39a637e", "seconds": 13.12, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task but then clearly states 'Hold on, let's drop the whole task,' which is a direct cancellation of the entire request. The final interjection is a complete revocation, satisfying the cancellation type criterion.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "286591a3e23bdd04366a002169dae7605c9c639046b609ee4425111727ebe6a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/ambiguous_halt/halt_024/dc81311856722a93/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "dd580bfe2dd668ed85b05e081103a3cf956072d28159e3a375af07b7f2ca6b61", "seconds": 12.08, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Give me a moment to think,' which is a clear request to pause. The task is not canceled, and the speaker does not provide a correction or new instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a3683cf762e5c4ea63e9009bacd24c675e591571af89758fd1e774fc12b5ae86", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/filler/original/bc8b982d41857f71/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "11f41f1a093d497a49a6320ceb68c785e3ff68d7076fa39acf57aeb8aafac507", "seconds": 8.24, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'let's see' but ultimately delivers the full original instruction with correct values, target, and action. The hesitations are fillers that do not alter the final goal. The final intent matches the original instruction exactly.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $427 on venmo to Amanda—I meant Anita.", "long_pause_offsets": [], "source_sha256": "cf7d81375eea8017291b5a64245078c5c81401d8a5bc36e408d44ea106f399f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amanda' but immediately corrects to 'Anita' with 'I meant Anita.' The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/false_start/false_start_03_person_name/7478412a2a8a81c0/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "20d737744177da861970365704c3e211384cc499ac45036d45a33565f44be2a5", "seconds": 7.12, "spoken_utterances": [ "Send four hundred twenty-seven dollars on uh... venmo to Amanda—um... I meant Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amanda' but immediately self-corrects to 'Anita' with 'I meant Anita.' This is a clear false start where the final target is corrected. The amount and app remain unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send four twenty seven dollars on uh Venmo to Amanda. Um, I meant Anita." }, "candidate_id": "wav_0e1ee54154b727f4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with a clear self-correction. The initial request is slightly hesitant ('send 427 dollars on uh Venmo to Amanda'), followed by a believable pause and a self-repair ('I meant Anita'). The correction is smoothly integrated into the same utterance with natural intonation and pacing, making it sound like a spontaneous recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Send $427 on venmo to. I left out a detail: send it to Anita.", "long_pause_offsets": [ 23 ], "source_sha256": "fc59e8698b1fc3c6f543da679925c72368dfaa26194ba9164fbce21a3689e412", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing recipient 'Anita' after an incomplete initial utterance. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/addition/addition_03/b57577df505ba05c/speech.wav", "wer": 0.2916666666666667, "wav_sha256": "3b4f7f90b83dd9c7a77f826107d952815876250a83d009c68b4899041575d8bc", "seconds": 13.88, "spoken_utterances": [ "Send let's see... four hundred twenty-seven dollars on um, let me think... venmo to. ", "I left out a detail: send it to ... Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the amount and app, then pauses to correct the recipient by adding 'send it to Anita' after a longer pause. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final intent is clear and matches the reference.", "heard_transcript": "Send, let's see, four twenty seven dollars on um, let me think, Venmo to, I left out a detail, send it to Anita." }, "candidate_id": "wav_cb0b5feeecd53ebc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural-sounding addition with a clear, spontaneous self-correction. The initial request 'send $427 on Venmo to' is followed by a believable 'let me think' pause, then a smooth, pragmatic revision where the speaker explicitly acknowledges the omission ('I left out a detail') and adds the recipient 'Anita'. The prosody is consistent, the repair is integrated into the flow, and the intonation reflects a genuine moment of recall." } ] }, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Send $427 on PayPal to Anita. Wait, send it on Venmo, not PayPal.", "long_pause_offsets": [ 30 ], "source_sha256": "b689f3e69c56c51f42df0952183efbe6776f91449f3a7b7f01c3bea62eb0640a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the amount and recipient. The monetary value is expressed in words but matches the reference amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/revision/revision_02_app/de698c84cf17d8ad/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "fc7b5be2b8650bd395f05397863fb4811446f528f504c83af4bd79998619c8ed", "seconds": 11.88, "spoken_utterances": [ "Send hmmm... four hundred twenty-seven dollars on um... PayPal to Anita. ", "Wait, send it on ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send the money on Venmo, which matches the original reference instruction. All other constraints (amount, recipient) are preserved.", "heard_transcript": "Send four twenty seven dollars on um PayPal to Anita. Wait, send it on Venmo, not PayPal." }, "candidate_id": "wav_7a5b0fe2cd044a2a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and spontaneous revision. The speaker begins with a plausible false start ('Send... um...'), then corrects the payment platform from PayPal to Venmo mid-sentence with a believable 'Wait, send it on Venmo not PayPal.' The hesitation and self-correction are integrated smoothly into the flow, with natural fillers and a clear, pragmatic change-of-mind. The prosody and pacing reflect genuine formulation and correction." } ] }, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "7aff6152b0605fd6d559236cd6f415f83382e29e21b00b7d6d8974fb9fc8b436", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/cancellation/cancel_012/b7c4d303833804b2/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b37efe7029b5c6ec056cb8d4526094fb68993ac4bd137a33a1d3539e36bd1990", "seconds": 10.92, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita.", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests sending $427 to Anita on Venmo, but immediately corrects themselves with 'Wait, actually, I've changed my mind. Don't do that,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly revokes the original instruction, satisfying the cancellation type criterion.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita. Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "87c7164aece390e5396d2115f561490a5f7cd8685ed17e60bd471b036518caf8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/ambiguous_halt/halt_040/e5aa4ae59416ceab/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "984b03957f1c753c7e46d9449174cdf4f1e71cd87bce6a38338a478ff740d6c4", "seconds": 9.64, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction with a final interjection 'Hang on, just a minute' that clearly requests a pause. This matches the 'ambiguous_halt' type criterion, which requires an audible request to hold/wait/pause with no permanent cancellation or permission to resume. The initial request is fully audible and complete, and the final interjection is a clear pause command.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9806321d2f483a1055ce3ab871a72ee4eeb4aea2923bcf5c18b0f0e38e59f4c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/filler/original/62695ae60c34455e/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "887e5efd689190b4fe6e6718b899bbe33bb0ad21ac0ca99aba2dcfd2dd99f0b4", "seconds": 6.04, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'let's see' but ultimately completes the original instruction without changing the action, amount, app, or recipient. The hesitation and backchanneling are typical filler behavior and do not alter the final task goal. The final intent matches the original instruction.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $234 on venmo to Jill—sorry, I meant Jack.", "long_pause_offsets": [], "source_sha256": "4767b2717c08f7c299b2d3a78ff48a1807d813d582a2779f0967a96bc5618a4e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately corrects to 'Jack' with 'I meant Jack'. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/false_start/false_start_03_person_name/ad70a045dfb8d78b/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "cd27e174223ea5ad3fa31706cc9e9a83f95815bd51507184bd8771cad141e7da", "seconds": 10.04, "spoken_utterances": [ "Send two hundred thirty-four dollars on um... venmo to Jill—uh... sorry, I meant um... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately self-corrects to 'Jack' after a pause and 'sorry I meant'. This is a clear false start where the final value replaces the mistaken one. The amount and app remain unchanged. The final intent matches the original instruction.", "heard_transcript": "Send two hundred thirty four dollars on um Venmo to Jill uh sorry I meant um Jack." }, "candidate_id": "wav_91663b0ddb625f8a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'uh' and 'um' fillers, a believable self-correction from 'Jilla' to 'Jack', and maintains a consistent, conversational pace. The repair is integrated smoothly within the utterance, and the prosody reflects genuine hesitation and recollection." } ] }, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Send $234 on venmo to. One more thing: send it to Jack.", "long_pause_offsets": [ 23 ], "source_sha256": "dc527209bdf7b6bf060e502a77109e1f73297a4dfbcd658ad818f690837374c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The first utterance is incomplete, lacking a recipient. The followup adds the missing recipient, Jack, without changing any other parameters. The amount and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/addition/addition_03/c0c7addb4f2d3cee/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "68b3b7cf6d71a8034c63ad86adbd18115e25fd529bbc917981d38cb9f1714de9", "seconds": 11.56, "spoken_utterances": [ "Send let's see... two hundred thirty-four dollars on ... venmo to. ", "One more thing: send it to hmmm... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Send two hundred thirty four dollars on Venmo to' is followed by a pause and the addition 'send it to, hmm, Jack.' The final target 'Jack' is clearly stated after the initial amount and app, satisfying the addition type criterion. The final goal combines the original instruction with the added target, matching the reference.", "heard_transcript": "Send, let's see, two hundred thirty four dollars on Venmo to, one more thing, send it to, hmm, Jack." }, "candidate_id": "wav_13ec3aae73d73cbf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident instruction ('Send let's see two hundred thirty four dollars on Venmo to...'), then pauses thoughtfully ('One more thing') before adding the recipient ('Jack'). The addition is natural, with a slight hesitation and filler ('hmm') that sounds spontaneous. The prosody is consistent, and the flow from initial command to addition feels like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_person_name", "proposed_instruction": "Send $234 on venmo to Jill. Wait, the recipient should be Jack, not Jill.", "long_pause_offsets": [ 28 ], "source_sha256": "33f75de332034e2f319be21d0fc8c4cccd1ad4050e150c2c99dc3b4a65a34f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from Jill to Jack, while retaining the amount and app. The original instruction specified Jack as the recipient, and the final spoken instruction confirms this." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/revision/revision_03_person_name/7934192acbf06618/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "0183ef842caefea151a2710d01de55267da37aa2d250d962280e15bc1e3df7c4", "seconds": 12.76, "spoken_utterances": [ "Send ... two hundred thirty-four dollars on venmo to um... Jill. ", "Wait, the recipient should be um, let's see... Jack, not Jill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately corrects to 'Jack' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final audible intent is to send $234 to Jack, matching the original instruction. All other constraints (amount, app) are preserved.", "heard_transcript": "Send two hundred thirty four dollars on Venmo to um Jill. Wait, the recipient should be um let's see Jack, not Jill." }, "candidate_id": "wav_cf955f0e686c49bc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic revision pattern: the speaker begins with a complete request, then hesitates with 'um' before correcting the recipient from 'Jill' to 'Jack'. The self-correction is natural, with a plausible change-of-mind intonation and a brief pause ('wait') before the revision. The prosody is conversational, with appropriate stress and pacing, and the correction flows seamlessly within the utterance, making it sound like a spontaneous real-time thought process." } ] }, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "e3e7a3cefa0b4af7e9bf3350c8d832956534458abb2ea3673c9e062dd593a13d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/cancellation/cancel_019/115801f389aaad37/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "20460c0903365e2b9c6e399f5a39cba1028868c3acddb38e3e96aeafa5db86ec", "seconds": 11.32, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack.", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hang on, never mind, don't do that.' This constitutes a cancellation of the entire task as per the type-specific criterion. The final control intent is a cancellation, which matches the expected behavior for this type.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on Venmo to Jack. Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "b5d75ae1c989c89c870229f741133d99329a27b9ca657a741ec15888d0c66124", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/ambiguous_halt/halt_003/6958f2549ba16596/speech.wav", "wer": 0.25, "wav_sha256": "410d000060dcb1a46947a19f5663de94be23d119b9dfd014787876cd43401473", "seconds": 9.96, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack.", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then says 'Hang on' as a final interjection, which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the final audible control is a pause.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on venmo to Jack. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "2c544f9_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4732f3943148fac16b1075901de870e11adee0dcaedcbb3a36874b6bcb8c7957", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/filler/original/b94e5a7fa496dde9/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "e55ddd80fb6cf8f519592d36017f792d062aa9c6c05c71832005f2e643306b8c", "seconds": 7.08, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'um, let me think' but does not change the core task. The final audible instruction matches the original request in action, amount, app, and recipient. The hesitation is a common backchannel and does not alter the goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on Venmo to Jack." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 10 minutes—no, that should be 5 minutes.", "long_pause_offsets": [], "source_sha256": "2f11905e0678ff5be05a4843d37db6992ecad33719cdc3dc310289d2f84c42dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces ten minutes with five minutes; the action, alarm type, and day scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/false_start/false_start_03_duration/3864642d70a3af4c/speech.wav", "wer": 0.0, "wav_sha256": "1e666351a11a981c5e4bf43b403eadc87d267776bd3868b09a3174db66311d4f", "seconds": 9.48, "spoken_utterances": [ "Set my um... weekday wake up alarm snooze to ten minutes—uh... no, that should be ... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'ten minutes' but immediately self-corrects to 'five minutes'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Set my um weekday wake up alarm snooze to ten minutes uh no, that should be five minutes." }, "candidate_id": "wav_1818603b9080e8e7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with a clear request, then after a brief pause and filler ('uh'), corrects the value from 'ten minutes' to 'five minutes' with a believable change-of-mind intonation. The prosody is consistent, and the correction is integrated smoothly into the utterance without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Set my weekday wake up alarm snooze to. I forgot to add, set the snooze to 5 minutes.", "long_pause_offsets": [ 40 ], "source_sha256": "61cbf40105f050bf6026f24b4eeddf0bb0d6cd7601434df68d991e4201f23876", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing snooze duration, preserving the original alarm type and schedule. The initial incomplete utterance is completed by the second." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/addition/addition_02/dbad033bac8f2260/speech.wav", "wer": 0.1, "wav_sha256": "1b9384f5903b45c6044175fa4b042eef5bb24461d1ff52677a2f84b92e2ab11c", "seconds": 11.28, "spoken_utterances": [ "Set my uh... weekday wake up alarm snooze to. ", "I forgot to add, set the snooze to hmmm... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to set the snooze, then adds the specific value 'five minutes' after a pause and self-correction. This matches the addition type criterion: the original instruction is combined with the added information to form the final goal, which aligns with the reference instruction.", "heard_transcript": "Set my uh weekday wake up alarm snooze to. I forgot to add, set the snooze to five minutes." }, "candidate_id": "wav_fd11f3e0a4f2b61d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_day_category", "proposed_instruction": "Set my weekend wake up alarm snooze to 5 minutes. Sorry, that should be my weekday alarm, not my weekend alarm.", "long_pause_offsets": [ 50 ], "source_sha256": "ec5b778950d456218c3ce4be3c426801505d9087f45a1264f8fe9c2f2133f878", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm category from 'weekend' to 'weekday', while retaining the snooze duration of five minutes. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/revision/revision_01_day_category/06e6a78df7e89029/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "fb720c7d1d4f64fc290b1be167ffb447e19df3007ba75f5091b275ac13854b29", "seconds": 14.72, "spoken_utterances": [ "Set my uh... weekend wake up alarm snooze to um... five minutes. ", "Sorry, that should be my weekday alarm, not my uh, let's see... weekend alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekend alarm' but immediately corrects it to 'weekday alarm' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original instruction exactly.", "heard_transcript": "Set my uh weekend wake up alarm snooze to um five minutes. Sorry, that should be my weekday alarm, not my uh let's see, weekend alarm." }, "candidate_id": "wav_e1ab1ae2bc95bcd8", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then self-corrects from 'weekend' to 'weekday' with a believable hesitation ('uh', 'let's see') and a smooth, pragmatic repair. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a forced edit. The pacing and intonation suggest a real person formulating and adjusting a command." } ] }, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_022", "hide_variant_original": true, "source_sha256": "aff8d50bd84b59495104b565c3c4fd4cb9bd364afbde19dbe14cd67451f23a75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/cancellation/cancel_022/a7dd8bce01fc961d/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "f930058348e44c4d1422f61376a817810998ecc3f98437bb84d6c8f786e83692", "seconds": 11.2, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes.", "Wait, actually, never mind. Cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'wait, actually never mind, cancel the task.' This constitutes a final, unambiguous cancellation of the entire task. The initial partial execution is irrelevant once the cancellation is stated.", "heard_transcript": "set my uh weekday wake up alarm snooze to um five minutes. wait, actually never mind, cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_035", "hide_variant_original": true, "source_sha256": "d17c0a053b4e202632fde18999a4bfcd9dc2611d7edbd773843b127f9c119e50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/ambiguous_halt/halt_035/4d33c046e15ae900/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "aa51d19d6ea710bc8d54ee756122a550c7880fcdba0e5a7747d5ab1ee8928e12", "seconds": 10.36, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes.", "Hold on before you do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to set the alarm snooze to five minutes but then immediately says 'Hold on, before you do that,' which is a clear request to pause or wait. This constitutes an ambiguous_halt as the final interjection is a pause command with no cancellation or resumption permission. The original instruction is not executed, but the intent to pause is audible and final.", "heard_transcript": "Set my uh weekday wake up alarm snooze to um five minutes. Hold on, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0c59011e4a92a30237ea037656a2d741784844083a2b4ec7ac3bc86ce5d471de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/filler/original/34df81ee479e04d0/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "2e2bf97aa23ca3b2b0ee0b1d87e7f82ba3fcd2184464d2294a30693042b28bbf", "seconds": 6.24, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core instruction. The final audible intent matches the original instruction with no correction or modification. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Set my uh weekday wake up alarm snooze to um five minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_day_category", "proposed_instruction": "Set my weekday—I meant weekend wake up alarm snooze to 15 minutes.", "long_pause_offsets": [], "source_sha256": "77c198e84411988185360fbf09303db5c5965bdd61986447271d854a0c9c9fe9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'weekday' to 'weekend' within the same utterance, preserving the action, snooze duration, and alarm type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/false_start/false_start_01_day_category/39f121d3bb6eb98e/speech.wav", "wer": 0.0, "wav_sha256": "39c73e679b548438325f442c91de61bf2a8c1c18516372ac232a8b86b9b9c099", "seconds": 7.76, "spoken_utterances": [ "Set my ... weekday—uh... I meant weekend wake up alarm snooze to fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday' but immediately self-corrects to 'weekend'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Set my weekday uh, I meant weekend wake up alarm snooze to fifteen minutes." }, "candidate_id": "wav_afc4557d208d0734", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start. The speaker begins with 'weekday', hesitates with 'uh', then self-corrects with 'I meant weekend' before continuing the full request. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and correction. The pacing and intonation are consistent with someone thinking aloud while formulating a command." } ] }, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Set my weekend wake up alarm snooze to. I left out a detail: set the snooze to 15 minutes.", "long_pause_offsets": [ 40 ], "source_sha256": "93c8d604c917552f0daa665091c2458ad9f0f684ba63155dae330e016c8e08ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing snooze duration value, which is consistent with the original instruction's context and target. The initial incomplete utterance is completed by the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/addition/addition_02/6945501082dadda6/speech.wav", "wer": 0.2, "wav_sha256": "0a6dbc09b2b91749b58c91729eea6dffa979dab3c05a36eba3f3ccf443703d1a", "seconds": 11.68, "spoken_utterances": [ "Set my um... weekend wake up alarm snooze to. ", "I left out a detail: set the snooze to ... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Set my um weekend wake up alarm snooze to' and then corrects by adding 'Set the snooze to 15 minutes.' This is a clear addition that completes the original instruction. The final audible intent matches the original instruction exactly, with the addition of the specific snooze duration. The 'um' and the self-correction are typical of natural speech and do not alter the final goal.", "heard_transcript": "Set my um weekend wake up alarm snooze to. I left out a detail. Set the snooze to 15 minutes." }, "candidate_id": "wav_dcaa739e972a865b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant 'um' and a clear false start ('Set my um...'), followed by a natural self-correction ('I left out a detail') and a smooth, additive continuation ('Set the snooze to 15 minutes'). The prosody is conversational, with a believable pause and filler before the correction, and the final instruction is delivered with consistent intonation and rhythm, making the entire sequence sound like a spontaneous, real-time thought process." } ] }, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_day_category", "proposed_instruction": "Set my weekday wake up alarm snooze to 15 minutes. Wait, set the alarm for the weekend, not the weekday.", "long_pause_offsets": [ 51 ], "source_sha256": "b398f9bf2d9f194a156ffa54a093663df847031bf96b16a2000da3779013077a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm category from 'weekday' to 'weekend' while retaining the snooze duration of fifteen minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/revision/revision_01_day_category/684218ce6d6b5837/speech.wav", "wer": 0.12, "wav_sha256": "543ffb825db095cf39bd2bc3e96d438ed21762b50e8b1e6173edd35a4450d719", "seconds": 12.2, "spoken_utterances": [ "Set my um, let me think... weekday wake up alarm snooze to fifteen minutes. ", "Wait, set the alarm for the hmmm... weekend, not the weekday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the target as 'weekday' but immediately corrects it to 'weekend' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The value 'fifteen minutes' remains consistent. The final intent matches the original instruction, so the task passes.", "heard_transcript": "Set my um, let me think, weekday wake up alarm snooze to fifteen minutes. Wait, set the alarm for the weekend, not the weekday." }, "candidate_id": "wav_fc538e738113fec0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a plausible 'let me think' filler, formulates a clear initial request, then pauses and self-corrects with a believable change-of-mind intonation ('Wait, set the alarm for the weekend, not the weekday'). The prosody is consistent, the correction is smoothly integrated, and the pacing reflects genuine thought processing without robotic or disjointed transitions." } ] }, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "28cf7b256b8c7469c1e9b7734b92d8d651fa53ab4db1ae66afa930e7cb8c8a56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/cancellation/cancel_008/c857733e7406ce14/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "955ad8203517d22e5be2cf9c4cab770c2d3833a638edf5ebb7c17ed04cdfde88", "seconds": 11.44, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes.", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first attempts to set the alarm snooze to 15 minutes, then immediately follows with 'Hang on, I want to cancel that task.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, not to execute the original instruction. The cancellation is explicit and complete, satisfying the 'cancellation' type criterion.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes. Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "2a0265d6d7706eae02849e2e3def86d58a0c9ea07dad95dd8d64fa3e2f917d67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/ambiguous_halt/halt_019/61b9f12208bb473a/speech.wav", "wer": 0.1, "wav_sha256": "f39914d94686fd6c45f4234af5b6db6ae62b60e603815241b1f22f40114e177d", "seconds": 10.52, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The user begins the task but then explicitly requests a pause with 'Can you give me a moment?'. This is a clear, final interjection to hold the process, fitting the 'ambiguous_halt' type. The initial request is not canceled, but execution is paused. The final control intent is to pause, which matches the type's criterion.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6c205f559ac9e4dc752942274c3a54fc8fbfb42e9bc1318fb44fe0e10b65b75b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/filler/original/f1cad63ed0929daf/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "cdbded738128870043ebbff624a25f149b0254fa60380ec540037b91da3e4ab0", "seconds": 6.76, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') and pauses but does not change the core request. The final audible instruction matches the original task intent, with 'fifteen minutes' confirming the value. The filler type allows ignoring hesitations and backchannels, so the task is executed as intended.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_alarm_description", "proposed_instruction": "Set my weekday morning alarm—sorry, I meant wake up alarm snooze to 20 minutes.", "long_pause_offsets": [], "source_sha256": "4a8548fdfec8f9f5d2a0bdc8ed74cc32ceeb27c9fe3eb70f0b73c86c42340052", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately corrects to 'wake up alarm'. The snooze duration is specified as 'twenty minutes' in both the original and spoken instruction. The final goal matches the original instruction in action, target, and value." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/false_start/false_start_02_alarm_description/501bb83533b2b365/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "8f3c19f4d9b186d52978d7d9f13c9dd466a61a617c0e871d0ab1a1f23d8faed6", "seconds": 8.64, "spoken_utterances": [ "Set my weekday morning alarm—uh... sorry, I meant ah... wake up alarm snooze to twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately self-corrects to 'wake up alarm'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistaken value should be discarded. The final instruction matches the original reference in intent and value, with only a minor paraphrase ('weekday morning alarm' → 'wake up alarm') that does not change the core task.", "heard_transcript": "Set my weekday morning alarm uh, sorry, I meant, uh, wake up alarm snooze to twenty minutes." }, "candidate_id": "wav_fd18c03b15071a7e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'weekday morning alarm', hesitates with 'uh', then clearly self-corrects to 'wake up alarm' with a smooth, conversational 'sorry, I meant'. The correction is integrated into the sentence without awkwardness, and the final instruction is delivered with consistent prosody and a believable, slightly rushed but coherent flow." } ] }, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Set my wake up alarm snooze to 20 minutes. Oh, right, this is just for my weekday wake-up alarm.", "long_pause_offsets": [ 43 ], "source_sha256": "2ec03544c4397844808e2f3ac9cb979d2ffed3956d16ef550b597c1a5165891d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup restricts the initial broad alarm to only the weekday wake-up alarm, preserving the snooze duration of twenty minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/addition/addition_01/624a8c5eacbdd0d8/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "95c04c8ac5ee254554d9b2a1bb71c555228a7522db4762f865d1d1e7fd67e279", "seconds": 10.8, "spoken_utterances": [ "Set my wake up alarm snooze to um... twenty minutes. ", "Oh, right, this is just for my uh... weekday wake-up alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to set the wake up alarm snooze to 20 minutes. The speaker then adds a clarifying constraint, specifying that this should apply only to the 'weekday wake up alarm'. This addition narrows the scope of the original instruction without contradicting it. The final intent is a combination of the original request and the added constraint, matching the expected behavior for an 'addition' type. The value '20 minutes' is correctly heard and preserved.", "heard_transcript": "Set my wake up alarm snooze to um twenty minutes. Oh right, this is just for my uh weekday wake up alarm." }, "candidate_id": "wav_9c3e7b8b3e2b8690", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um') and a plausible self-correction ('Oh right, this is just for my...'). The addition of 'weekday wake-up alarm' is smoothly integrated as a clarifying detail after the initial command, maintaining a consistent rhythm and intonation. The prosody is fluid, and the pause before the addition feels like genuine recollection rather than a scripted edit." } ] }, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 15 minutes. Make that 20 minutes, not 15 minutes.", "long_pause_offsets": [ 51 ], "source_sha256": "94a76e4f16debf14760c2011f3872a3b447f15590a0b3655ab23720c7f7f3f86", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the snooze duration from fifteen to twenty minutes, while retaining the alarm type (weekday wake up) and action (set snooze)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/revision/revision_03_duration/04779ae2dfe4e0b8/speech.wav", "wer": 0.0, "wav_sha256": "f5b015fece016ddaac73636fde1794b79c59e8568305016c2d00821d3e4cd184", "seconds": 11.0, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to ... fifteen minutes. ", "Make that twenty minutes, not fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen minutes' but immediately corrects with 'Make that twenty minutes, not fifteen minutes.' This constitutes a revision where the final audible intent is to set the snooze to twenty minutes, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Set my weekday wake up alarm snooze to fifteen minutes. Make that twenty minutes, not fifteen minutes." }, "candidate_id": "wav_577c64505e7102d6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural revision flow. The speaker first makes a clear request, then immediately corrects the value from '15 minutes' to '20 minutes' with a smooth, self-correcting intonation. The phrasing 'Make that 20 minutes, not 15 minutes' is a highly plausible, spontaneous way to revise a number, and the delivery is continuous and fluent without awkward pauses or fillers." } ] }, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "64390c0bfa798cc329b02e77ec85c7ed9254f7a91b4901e8f476f749bc730944", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/cancellation/cancel_009/acc675c296c30ca3/speech.wav", "wer": 0.125, "wav_sha256": "77dcc9f71a6d59c6fc674a1810c898fb20dba2f1da16e7c6944731feee990b78", "seconds": 10.28, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially attempts to set the snooze duration but immediately corrects themselves with 'Oh wait, cancel that altogether,' which is a clear and final cancellation of the original task. The cancellation is not ambiguous and applies to the entire task. The audio ends with this cancellation, satisfying the cancellation type criterion.", "heard_transcript": "Set my weekday wake up alarm snooze to um twenty minutes. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "23735584050652d0dae4a670af930f7dff77d48bf08dd55c3e86ff1c68fb7fa3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/ambiguous_halt/halt_029/6be435cb28c1afb5/speech.wav", "wer": 0.0, "wav_sha256": "c8293f580fd8df1466b5de87961725c057996a9ba76f0bd5546b97af1ab2f2d0", "seconds": 9.48, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, actually...', which is a clear interjection indicating the speaker is halting the current task to revise or reconsider. This matches the 'ambiguous_halt' type, where the final intent is to pause, not execute or cancel. The initial instruction is not completed, and no new instruction is given, so the task is paused.", "heard_transcript": "Set my weekday wake up alarm snooze to um, twenty minutes. Wait, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "31dc501_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ebe9674d76f0746fa8bdeab37f03266167f758d078a6bb5df41578d507a38c75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/filler/original/0f6aab46a191d28a/speech.wav", "wer": 0.0, "wav_sha256": "1f4780932a139801399404e8cb335c4008f13157b925bbc1da01ddafb58d54a8", "seconds": 6.04, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'um' but immediately corrects and completes the original instruction with the correct value 'twenty minutes'. The final audible intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Set my weekday wake up alarm snooze to um, twenty minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_note_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Gmail—sorry, I should have said Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "c87283ea1cbc335a5cdab3f4e0abd5657a64ea69e8e1908706b488153151751f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately corrects to 'Simple Note'. The final intended source of information is Simple Note, which matches the original instruction. All other conditions, including the action, destination app, and context, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/false_start/false_start_02_note_app/1abf125d08b6e0d8/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "f08b5e917a6315ac4ee26058db113e9964b0f374252e47740716e7a2b5e41486", "seconds": 22.12, "spoken_utterances": [ "I went on a few trips each with some of my friends. My Gmail—ah... sorry, I should have said Simple Note has information on who owes whom what um, let's see... from each trip. I have already created Splitwise groups for the trips. uh... Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately self-corrects to 'Simple Note'. The final instruction is to 'record the expenses accordingly in the respective groups', which matches the original instruction. The correction is immediate and does not alter any other conditions. Therefore, the false_start type is satisfied with a pass.", "heard_transcript": "I went on a few chips each with some of my friends. My Gmail uh sorry I should have said Simple Note has information on who owes whom what um let's see. From each trip, I have already created Splitwise groups for the trips. Uh record the expenses accordingly in the respective groups." }, "candidate_id": "wav_72cc4d62bb0d5145", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and credible false start. The speaker begins with 'Gmail', then self-corrects to 'Simple Note' with a clear, spontaneous 'sorry I should've said' phrase. The correction is integrated smoothly into the flow, and the 'let's see' filler after the correction sounds like genuine recollection. The prosody is consistent, and the pacing feels like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. One more detail: it's about who owes whom what.", "long_pause_offsets": [ 215 ], "source_sha256": "0637b2f07a99fa002c84dd63568ca1d17867b1d332fb692bce3cf301f531e2dc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail about the content of the Simple Note, specifying that it contains 'who owes whom what'. This clarifies the information source for the expense recording task, without changing the action, app, or groups. The original instruction already specified recording expenses in the groups, so the addition completes the missing condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_03/d6d2961b35296b08/speech.wav", "wer": 0.0196078431372549, "wav_sha256": "b4c8ecc4926a7e63f483b3774ddda0777e7138930600b7bdc75851f07d871f16", "seconds": 23.64, "spoken_utterances": [ "I went on ... a few trips each with some of my friends. My Simple Note has information on um, let me think... from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "One more detail: it's about ... who owes whom what." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and completed. The speaker adds the detail 'who owes whom what' after a pause, which is an addition to the original request. The final intent combines the original task with the added detail, matching the expected behavior for an 'addition' type. No contradictions or cancellations are present.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on um, let me think. From each trip, I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. One more detail, it's about who owes whom what." }, "candidate_id": "wav_8a5ed8fa03eb87be", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear statement, then uses a believable 'let me think' filler pause before continuing. The addition of 'one more detail' followed by the clarification about 'who owes whom what' is a pragmatic and plausible revision. The intonation and pacing are consistent with someone recalling information in real time, and the flow feels unscripted and conversational." } ] }, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my colleagues. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be 'friends', not 'colleagues'.", "long_pause_offsets": [ 237 ], "source_sha256": "50a91212fc3bc9112460ef9f44053ed816dcf1030e8fbbf30ffbc52637c07851", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up revision corrects the companion group from 'colleagues' to 'friends', while retaining all other requirements: recording expenses in the respective Splitwise groups based on information in Simple Note. The final goal is consistent with the original instruction except for the corrected group identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/revision/revision_01_companion_group/588a8a45e06c364f/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "84a5ebb5d5b26b590f1e7b2232838b6fc91cdfc04afa15921810aab235740e9b", "seconds": 21.44, "spoken_utterances": [ "I went on um... a few trips each with some of my colleagues. My Simple Note has information on ... who owes whom what hmmm... from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the companion group should be 'friends', not 'colleagues'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'colleagues' but immediately corrects it to 'friends' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to record expenses in the groups, with the companion group being 'friends'. The original instruction is revised to match the final audible intent, so the task is executed as revised.", "heard_transcript": "I went on um a few trips each with some of my colleagues. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be friends, not colleagues." }, "candidate_id": "wav_eb59a1c59f64ab91", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a plausible hesitation ('um'), uses a realistic self-correction ('Wait, the companion groups should be friends, not colleagues') that sounds like a genuine moment of recollection, and maintains a consistent, conversational pace. The correction is integrated smoothly into the utterance, and the overall prosody is fluid and human." } ] }, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "8d73faf236c1028b8ed15ca58912eae76ec970ae5b001a6e0236c0599d86e4b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/cancellation/cancel_026/00d77ece550875a7/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "b6e68cde3cf85258ad7a368432bbb505281888447d9d99054784fe142f829197", "seconds": 22.4, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Wait, forget what I just asked. Cancel it.' This constitutes a final, unambiguous directive to cancel the entire task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "fc0f21bfb3c1fbdcdc0c78d76d0d7af8c85b8595f84ca1a3e53eba2ef1a6aa9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/ambiguous_halt/halt_026/2afd7bc1ab14ba51/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "f8fd75918a345cee57e061492c43bcdcee1a948e30180a84bb7616fb159def72", "seconds": 22.0, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d1e78497926c98f17b315de07809b6261cb7f892eb6c156db19835f1d442d7f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/filler/original/d3d68c3c438dfd27/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "7658266d1707dbd51d91b556267fd35d398eb6dee13b8f985b17765ec0ac3180", "seconds": 17.64, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('Um', 'Uh', 'let's see') and pauses but does not change the core task. The final instruction is to 'Record the expenses accordingly in the respective groups,' which matches the original intent. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my friends—wait, I meant coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "30adb8dab03143c9b3b48123007310688081db65c51144bd73c6c4179482a7f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'friends' to 'coworkers' within the same utterance. The final intended group is 'coworkers', matching the original instruction. All other conditions—recording in Splitwise groups, using Simple Note data, and the context of multiple trips—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/false_start/false_start_01_companion_group/75979b99ef20f2a7/speech.wav", "wer": 0.125, "wav_sha256": "131ce9797fdc361d3fe2c89e41b64a9e6064de5f300c484928e0b2143c95059f", "seconds": 21.04, "spoken_utterances": [ "I went on a few trips each with some of my um... friends—uh... wait, I meant coworkers. My ... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'coworkers'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction remains unchanged and fully matches the original. The correction is immediate and does not alter any other conditions, so the task is valid and complete.", "heard_transcript": "I went on a few trips each with some of my um, friends uh, wait, I meant coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_2a54a61ac4d2b971", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'friends', hesitates with 'uh', then self-corrects with 'wait, I meant coworkers'—a very credible and common pattern of self-correction. The intonation and pacing are consistent with someone recalling and adjusting their words in real time. The pause and filler are appropriately placed and not overly long, contributing to a realistic delivery." } ] }, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I went on a few trips each with. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, it's only some of my coworkers.", "long_pause_offsets": [ 215 ], "source_sha256": "76294d77d9cd5aeb96a2a07b9966450c9e239984c123afd38d26962f567b4945", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the scope to 'some of my coworkers', which is a narrowing of the original instruction's broader implication of all coworkers. The core action—recording expenses in Splitwise groups based on Simple Note data—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/addition/addition_01/48e1055e85e9b01b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "8c2972417ffabe0a0768d1f4d5cd43b3bb1cef39953e285a04f1d56a3105e8dc", "seconds": 23.48, "spoken_utterances": [ "I went on a few trips each with. My let's see... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Oh, I forgot to mention, it's only hmmm... some of my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition 'it's only some of my coworkers' narrows the scope of the initial request. The final intent is to record expenses in the respective groups, but limited to 'some of my coworkers'. This matches the addition type criterion, where the final goal combines the original instruction with the added constraint.", "heard_transcript": "I went on a few trips each with my let's see. Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, it's only some of my coworkers." }, "candidate_id": "wav_0cfba5590bc4f85e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see') and a filler ('um') before continuing. The addition ('Oh, I forgot to mention...') is delivered with a believable, slightly apologetic tone and a natural pause before the new information. The phrasing 'it's only some of my coworkers' flows smoothly as a remembered detail, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the group should be coworkers, not friends.", "long_pause_offsets": [ 234 ], "source_sha256": "87805c210882c5dffb9d5eb4be4c4bbe855a91cca1c338806fb91774a3d8aa44", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group membership from 'friends' to 'coworkers', preserving the action, app, and context of recording expenses in existing groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/revision/revision_01_companion_group/877a13271cf322b8/speech.wav", "wer": 0.09615384615384616, "wav_sha256": "4c79ac5f7c195e19efc5aeee5901c36a7bb9afbab0ce1f41951613647905a268", "seconds": 23.52, "spoken_utterances": [ "I went on a few trips each with some of my ... friends. My Simple Note has information on who owes whom what from each trip. I have already created hmmm... Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the group should be um... coworkers, not friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially refers to 'friends' but immediately corrects themselves to 'coworkers' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to record expenses in the groups, but the group context is revised from 'friends' to 'coworkers'. All other conditions remain unchanged, so the final task is valid and matches the revision type requirement.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the groups should be um coworkers not friends." }, "candidate_id": "wav_4e5e5b94d518c369", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and pragmatic flow. The speaker begins with a plausible initial statement ('friends'), then after a realistic pause and self-correction ('wait, the group should be...'), corrects the group type to 'coworkers'. The correction is integrated smoothly into the ongoing utterance with a believable change-of-mind intonation and reference back to the earlier request. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "deb711d8345e5deeb65d051c035ae7089f31e1e172cd6e76777a3b19c567b457", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/cancellation/cancel_023/4875dfcccab64b38/speech.wav", "wer": 0.1, "wav_sha256": "721bb33ba1c1320410b7f9af0b1ffb70643e59e99e08d6c113694bded85d2424", "seconds": 24.2, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker explicitly states 'Hang on, never mind about that. Don't carry out the request' after completing the full description of the task. This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "I went on a few trips each with some of my uh, let me think, coworkers. My Simple Note has information on who owes whom what from each trip. I have already created, let me think, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "1349931c9f705a4da641e42abca056bd8245f7666d9903a9ef520fa9c635333c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/ambiguous_halt/halt_033/7df83e870bcdc91b/speech.wav", "wer": 0.11320754716981132, "wav_sha256": "70a9b6c360f6019387005be24b3c435451b6d12d974dbe72d8a58b6aa5fa382c", "seconds": 22.44, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something...', which is a clear request to pause or hold the current task. The speaker does not cancel the task or provide a new instruction. The initial request is fully stated and understood, but the final interjection indicates a need to pause before proceeding. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "I went on a few trips, each with some of my uh, let me think, coworkers. My Simple Note has information on who owes whom what from each trip. I have already created, let me think, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Hold on, there's something..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "feaa852ddcab90acb767768909aedb4b5e5c7aeef0040d3a966a6a6edc73318b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/filler/original/c059555148844a80/speech.wav", "wer": 0.12244897959183673, "wav_sha256": "0ab23a3c8db5514ab20b61fd130c299f1209208134b21eb19190e71ec6df2f92", "seconds": 18.52, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes two filler phrases ('uh let me think') but does not change the core task. The final instruction is to record expenses in the respective groups, which matches the original instruction. The filler phrases do not alter the goal, so the task should be executed.", "heard_transcript": "I went on a few trips each with some of my uh let me think coworkers. My simple note has information on who owes whom what from each trip. I have already created let me think splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my roommates—I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "0502c2a9b1c067c1b7dae2eb226750ccc2b730d1be28f02efee73f3e37c12e37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'roommates' to 'friends' within the same utterance. The final intended group is 'friends', but the core task—recording expenses in the appropriate Splitwise groups based on Simple Note data—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/false_start/false_start_01_companion_group/5482c1f608a0793d/speech.wav", "wer": 0.0425531914893617, "wav_sha256": "a514c2f9206053e4a93a3ad9927c4ff834c5091a75488206b0b5b5d498f4b61f", "seconds": 20.32, "spoken_utterances": [ "I went on a few trips each with some of my roommates—um... I meant friends. My hmmm... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'roommates' but immediately self-corrects to 'friends'. The final intent is to record expenses in Splitwise groups as per the original instruction. The correction is a false start, and the final goal matches the original instruction.", "heard_transcript": "I went on a few trips each with some of my roommates, some. I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_f1af66694fe73c27", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with 'roommates' but immediately self-corrects to 'friends' with a believable 'I meant' repair, followed by a realistic 'hmm' pause that suggests genuine thought. The prosody is varied and conversational, with appropriate intonation and pacing. The correction is integrated smoothly into the flow, and the overall rhythm feels like a person recalling and refining a task in real time." } ] }, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. One more thing: the Splitwise groups for the trips are already set up.", "long_pause_offsets": [ 179 ], "source_sha256": "ef222cea2cd05945d77906bfdc9c5e27a349b8a76c5a3bec4ae0fe6c378f3d30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds confirmation that the Splitwise groups are already created, which supports but does not alter the core task of recording expenses in those groups. The original instruction already assumed the groups existed, so this addition is consistent and non-contradictory." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/addition/addition_02/2e55e7ee297027af/speech.wav", "wer": 0.02127659574468085, "wav_sha256": "2fd2a358eaaf8c7fbcfd41c13790423c678d31db09d80f52c30fc10c4aba2d4d", "seconds": 19.8, "spoken_utterances": [ "I went on a few trips each with some of my um... friends. My ... Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. ", "One more thing: the ... Splitwise groups for the trips are already set up." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to record expenses in respective groups. The speaker adds the information that the Splitwise groups are already set up, which is an additional constraint that narrows the scope of the task. The final intent is to execute the original task with this added detail, matching the 'addition' type criterion.", "heard_transcript": "I went on a few trips each with some of my um friends. My Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. One more thing, the Splitwise groups for the trips are already set up." }, "candidate_id": "wav_d16c74c6c9c2bfeb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker's delivery is highly natural, with a clear, conversational rhythm and plausible fillers ('um'). The addition of 'One more thing' followed by a revision of the group setup information is a believable pragmatic flow. The intonation and pacing suggest spontaneous thought, and the correction ('the splitwise groups... are already set up') is smoothly integrated without awkwardness. The voice is consistent and the speech is free of distortions or unnatural pauses." } ] }, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be 'friends', not 'roommates'.", "long_pause_offsets": [ 236 ], "source_sha256": "e6a0913279ff58b17f903dbc6473cd89bc4a1531d02bdbcc169aed8aea101beb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the companion group from 'roommates' to 'friends', while retaining all other requirements including the action, app, and context of recording expenses based on Simple Note data." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/revision/revision_01_companion_group/92b4f61d7e408cd1/speech.wav", "wer": 0.0, "wav_sha256": "6d8e053a5acedbb8cd41c6d958ce7961eba6b19a1f1bb3115674207dd1387b2b", "seconds": 23.36, "spoken_utterances": [ "I went on a few trips each with some of my um... roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the companion group should be '... friends', not 'roommates'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'roommates' but immediately corrects to 'friends' after a pause, which is a revision. The final intent is to execute the task with the corrected target group name. The rest of the instruction remains unchanged and is executable.", "heard_transcript": "I went on a few trips each with some of my um, roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion groups should be, friends, not roommates." }, "candidate_id": "wav_179069e10bae4bd0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous initial request with a believable self-correction. The initial mention of 'roommates' is followed by a pause and a clear, spontaneous realization that it should be 'friends', with a smooth back-reference to the earlier term. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The pacing and intonation are very human." } ] }, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "20ec33131e13767765713ee5e8c6309ba44e60759878c9d31b26c6e0d7588744", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/cancellation/cancel_039/7206ac20852a46d0/speech.wav", "wer": 0.057692307692307696, "wav_sha256": "565a08802af3d14e3d7628354fd140c526916d463915ac8c5ac05d4330bfd503", "seconds": 22.6, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Wait, don't go ahead with that. Cancel it.' This constitutes a final, unambiguous directive to cancel the previously stated task. The initial instruction is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "I went on a few trips each with some of my friends. My uh Simple Note has information on who owes whom what from each trip. I have already created um Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "ad7c8fd373a803906e84270cae123dbd10feb7f352ab0202ab0be4fc2e3c73d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/ambiguous_halt/halt_036/8f127384746ad5da/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "9dbd1647ebf6f794f1871a1d72743adedf513441cbbf8f0b4c61c55fc646efb7", "seconds": 22.64, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly says 'Wait a second, before you go ahead,' which is a direct request to pause. The final interjection is a pause command, not a cancellation or a new instruction. The type is 'ambiguous_halt', and the audio ends with a valid pause request, satisfying the criterion.", "heard_transcript": "I went on a few trips each with some of my friends. My uh, Simple Note has information on who owes whom what from each trip. I have already created um, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "32616b5_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "29051ac7a1667bf0d4f93a734601321538306c0daf4a8b38af096c94715d72d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/filler/original/e2dd72751961c680/speech.wav", "wer": 0.06818181818181818, "wav_sha256": "7bc4ceca42f4492ec8a9f467478b47b0bfa324390e377639e130b2121a135c5b", "seconds": 18.04, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') and pauses but does not change the core request. The final instruction remains to record expenses in the respective groups, matching the original intent. No correction, addition, or cancellation is made. The filler words do not alter the task goal.", "heard_transcript": "I went on a few trips each with some of my friends. My uh, Simple Note has information on who owes whom what from each trip. I have already created um, Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_subscription_type", "proposed_instruction": "How much have I paid in basic membership—no, I meant prime membership since I made the amazon account?", "long_pause_offsets": [], "source_sha256": "21099084550af770e7b165c8b7a01ca40feb17e5cf4c6ed335585f6a239a5a1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'basic membership' to 'prime membership' within the same utterance, preserving the rest of the query including the app (amazon), the financial scope (how much paid), and the time frame (since I made the account)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/false_start/false_start_01_subscription_type/a480a5557ece9215/speech.wav", "wer": 0.0, "wav_sha256": "315d8f51ec6a4a9dfdc9f6e910d4b6c122b6149b553aac9371a018339c3768c3", "seconds": 8.92, "spoken_utterances": [ "How much have I paid in basic membership—um... no, I meant ... prime membership since I made the amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately self-corrects to 'prime membership'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in basic membership, um, no, I meant prime membership since I made the Amazon account." }, "candidate_id": "wav_efe94806a1d02055", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects 'basic membership' to 'Prime membership' with a natural 'no, I meant' transition. The correction is immediate and integrated smoothly into the sentence, maintaining a consistent pace and intonation. The overall delivery is fluent and sounds like a spontaneous, real-time thought process." } ] }, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in since I made the amazon account? Oh, I forgot to mention, it's specifically for prime membership.", "long_pause_offsets": [ 57 ], "source_sha256": "8ea88290d431d37fd0c533911c85a3eb9822f3babac38a0b81286834514aeda9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specificity about 'prime membership', which clarifies the initial broad 'paid in' request. The app (Amazon), the financial context (paid), and the time frame (since I made the account) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/addition/addition_01/23fcc1ea2974e032/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5b7b1c1eb49491538c563fe6a42e525bf89f40f7d34022caac75e4bbf3f1b1ca", "seconds": 12.84, "spoken_utterances": [ "How much have I paid in um... since I made the amazon account? ", "Oh, I forgot to mention, it's specifically for um, let me think... prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('How much have I paid in um since I made the Amazon account?'), but the speaker immediately adds the missing constraint 'Prime membership' after a pause and self-correction ('let me think'). This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is fully specified and matches the reference.", "heard_transcript": "How much have I paid in um since I made the Amazon account? Oh, I forgot to mention, it's specifically for um, let me think, Prime membership." }, "candidate_id": "wav_25cd55da92c5b0ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much have I paid in...') and then adds a clarifying detail ('Oh, I forgot to mention...') about Prime membership. The addition is smoothly integrated with a realistic 'let me think' filler, and the prosody reflects genuine recall rather than a rehearsed or robotic delivery. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the ebay account? Let me correct the app: look for the payment history in Amazon, not eBay.", "long_pause_offsets": [ 72 ], "source_sha256": "18122217fdd76117f8eaef0dc2ac1d2af63cbede9311c9e89619c63a14c3893f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the original intent of querying the payment history for Prime membership since account creation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/revision/revision_02_app/6c465eb7dd1bd3b6/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "35e0cbc85a9f2b0936bcf048321a42343db4d01e0df43631fd7cdcaa43ff820b", "seconds": 13.24, "spoken_utterances": [ "How much have I paid in uh... prime membership since I made the ebay account? ", "Let me correct the app: look for the payment history in um... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions eBay but immediately corrects the app to Amazon, which is the original target. The revision is clear and final, and the core request (total Prime payment since account creation) remains unchanged. The correction applies to the target app, not the action or value, so the final intent matches the original instruction.", "heard_transcript": "How much have I paid in uh prime membership since I made the eBay account? Let me correct the app, look for the payment history in um Amazon, not eBay." }, "candidate_id": "wav_f00b36799954f820", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much have I paid in uh...') with a plausible filler and a clear, self-corrected revision ('let me correct the app, look for the payment history in um... Amazon not eBay'). The correction is smoothly integrated into the same utterance, with a believable shift in focus and a clear reference back to the initial error. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "15854cd31cbd9a5fcc9b2dc7065d7e98a5e0627db9429bfb8da84b07195ef966", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/cancellation/cancel_026/ee5144c702d82247/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "83568747fd9f8c67f9d1a0ee1e4abd113b0f20c4f3894754d854e14862489d77", "seconds": 12.48, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?", "Wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question, then pauses and says 'forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How much have I paid in, um, prime membership, uh, let me think, since I made the amazon account? Wait, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "5f0dd743ea0fa4f70ed490c33d30cfc56070af47feea4c774b66a5da6df94843", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/ambiguous_halt/halt_002/349e836d524f47a5/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "a9ac1c53ace03c608486695ac2e8ade2883d971606a65782140a4dec690fb3ae", "seconds": 10.48, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, but the speaker explicitly halts further action. No execution occurs, and the final intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "How much have I paid in prime membership, uh, let me think, since I made the Amazon account? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "365e0a3_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4040c038dcc8e4e5716efa96414ad43d2098dc3567e09400fdaa102c2d4c86a0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/filler/original/41700c140bbf3565/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "71500f75ecccf452c46dd6d408ea7caf54574c735549feb9c28220c441e69aca", "seconds": 7.56, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('mm', 'uh', 'let me think') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "How much have I paid in mm prime membership uh let me think since I made the amazon account?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "365e0a3_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in premium membership since I made the Apple Music—I should say spotify account?", "long_pause_offsets": [], "source_sha256": "a7dd875b386b4e68ede8ba05281829d52f5857655d3a4d7fec0597249532e8bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the app, the financial aspect, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/false_start/false_start_02_app/8a4e81b86449af47/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "b079764d2f1f4d39d2c75519037fc98f61cd8db64dd220a8c493d17209bd4fd0", "seconds": 12.24, "spoken_utterances": [ "How much have I paid in premium membership ... since I made the um... Apple Music—uh... I should say spotify account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify account'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in premium membership since I made the um, Apple Music uh, I should say Spotify account?" }, "candidate_id": "wav_78a52851b8d6b4b6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start with a clear self-correction from 'Apple Music' to 'Spotify'. The speaker's intonation and pacing are consistent with spontaneous speech, including a realistic pause and filler ('um') before the correction. The correction is immediate and smoothly integrated into the sentence, maintaining a natural flow without any robotic or rehearsed quality." } ] }, "is_default": true } }, { "task_id": "365e0a3_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in since I made the spotify account? Oh, I forgot to mention, it's specifically for premium membership.", "long_pause_offsets": [ 58 ], "source_sha256": "df6932de8c3cb473d8865143050e138c5d052efabcfe89c873756a66a1dddc8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'premium membership' to the initial broad question about payments, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/addition/addition_01/746e5f000371231b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "4868187d2ef6d150d10f565ce7d760aab14cabcac93694843b35e9dbc87ed990", "seconds": 12.44, "spoken_utterances": [ "How much have I let me think... paid in ah... since I made the spotify account? ", "Oh, I forgot to mention, it's specifically for premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about payments since making the Spotify account, then adds the specific constraint 'for premium membership' after a longer pause. This matches the 'addition' type criterion, where the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "How much have I let me think paid in uh since I made the Spotify account. Oh, I forgot to mention it specifically for premium membership." }, "candidate_id": "wav_320066d5154d1309", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "365e0a3_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in premium membership since I made the apple music account? Wait, the question is about Spotify, not Apple Music.", "long_pause_offsets": [ 81 ], "source_sha256": "b126f8c9e25c3831fca3df47779c98f33f555f5f0dad4c383add0e3e66852a63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the rest of the query including the action, financial scope, and time reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/revision/revision_02_app/5c9cdb782df5b92f/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "18ebf693ee44ee3daf38aedc2bb3010113b6f4e9a7fd2eb06436a7fc69ef88d7", "seconds": 16.76, "spoken_utterances": [ "How much have I hmmm... paid in premium membership uh... since I made the apple music account? ", "Wait, the question is about hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects themselves to Spotify after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to inquire about Spotify, which matches the original reference instruction. All other constraints are preserved, and the correction is clear and unambiguous.", "heard_transcript": "How much have I hmm paid in premium membership uh since I made the Apple Music account. Wait, the question is about hmm Spotify, not Apple Music." }, "candidate_id": "wav_268522adc155e6f6", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "365e0a3_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the apple—sorry, make that amazon account?", "long_pause_offsets": [], "source_sha256": "c925f8848a55bd9b0a3a31b78b151e0131bc04776cf3738c4f18e645127b2342", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'apple' to 'amazon' within the same utterance, preserving the rest of the query including the subject (prime membership), the action (how much paid), and the time reference (since I made the account)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_3/audio_natural/false_start/false_start_02_app/992820b8249d4389/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c5ad601c77a750791c9ec661fa38d629dc4e4e928fa45ee2fc3c730b6f3bdb86", "seconds": 10.96, "spoken_utterances": [ "How much have I paid in ah... prime membership since I made the uh... apple—um... sorry, make that amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'apple' but immediately self-corrects to 'Amazon' with 'sorry, make that Amazon account'. This is a clear false start where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to calculate the total paid for Prime membership since the Amazon account was made.", "heard_transcript": "How much have I paid in uh prime membership since I made the uh apple um sorry, make that Amazon account?" }, "candidate_id": "wav_eb841d51a4a2f0f6", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_visibility", "proposed_instruction": "Send $91 publicly—sorry, I meant privately on Venmo to the person with this phone number 2192158297.", "long_pause_offsets": [], "source_sha256": "97385442d3b17b255f4b393f4f78567be21d3934396f9ba7795285143a8870ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects to 'privately' within the same utterance. The amount and recipient phone number are correctly stated and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/false_start/false_start_02_visibility/48d7c0df5c099b2e/speech.wav", "wer": 0.0, "wav_sha256": "d4276a256f8006b8af3ef5fb2a23b92e447913bfefb0d70a5c694cef04123769", "seconds": 11.48, "spoken_utterances": [ "Send ninety-one dollars publicly—um... sorry, I meant ... privately on Venmo to the person with this phone number two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately self-corrects to 'privately'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is overridden. The amount ($91) and recipient (phone number 2192158297) remain consistent. The correction is immediate and unambiguous, so the final intent matches the original instruction except for the privacy setting, which is corrected as intended.", "heard_transcript": "Send ninety one dollars publicly, um, sorry, I meant privately on Venmo to the person with this phone number two one nine, two one five, eight two nine seven." }, "candidate_id": "wav_774d05c8e9e651aa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a clear initial request, then self-corrects with a believable 'sorry, I meant' phrase, smoothly transitioning to the correct value. The prosody is consistent, with a natural pause and intonation shift during the correction. The delivery is fluent and sounds like a genuine moment of recall, with no awkwardness or hesitation beyond what's expected in real speech." } ] }, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Send $91 on Venmo to the person with this phone number 2192158297. Oh, and one more thing: send it privately.", "long_pause_offsets": [ 67 ], "source_sha256": "ed2432278eaebd7d4341459fd8dc419ff61fbd1709f30ab65fd6234ec7977e3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing privacy constraint; all other values including amount, app, recipient identifier, and phone number are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/addition/addition_01/f5177d0b5e282ee5/speech.wav", "wer": 0.0, "wav_sha256": "460691986589e5db677fbc58af2dfd2362a82ef552c4648334c877cf6db181db", "seconds": 15.36, "spoken_utterances": [ "Send ... ninety-one dollars on ... Venmo to the person with this phone number let me think... two one nine, two one five, eight two nine seven. ", "Oh, and one more thing: send it privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send $91 on Venmo to the person with the given phone number. After a brief pause and hesitation, they add the constraint 'send it privately'. This is a clear addition to the original instruction, and the final goal combines both the original and the added condition. The addition is not a correction or cancellation, but an extra constraint. Therefore, the final intent matches the expected behavior for an 'addition' type task.", "heard_transcript": "Send ninety one dollars on Venmo to the person with this phone number, let me think, two one nine, two one five, eight two nine seven. Oh, and one more thing, send it privately." }, "candidate_id": "wav_9d8db92f0b575325", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then uses a believable 'let me think' filler pause before providing the phone number. The delivery includes a smooth, pragmatic addition ('Oh, and one more thing') followed by a coherent correction of the initial instruction. The prosody is consistent, and the flow reflects a realistic thought process with a minor correction and addition." } ] }, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Send $91 privately on CashApp to the person with this phone number 2192158297. Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 79 ], "source_sha256": "58b3788d8b9496cfb7a5d7761f812a6fb1c8eaad03c4d1ae20da2ead9a493e7b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the amount, privacy setting, and recipient phone number. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/revision/revision_03_app/4b0140febf88bc60/speech.wav", "wer": 0.38461538461538464, "wav_sha256": "d443d30616135b0cad0175d7f796109b13bab86a12197a799b4d6824f85410d1", "seconds": 19.92, "spoken_utterances": [ "Send ninety-one dollars privately on um, let me think... CashApp to the person with this phone number hmmm... two one nine, two one five, eight two nine seven. ", "Let me correct the app: use um... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, privacy setting, app, and recipient identifier.", "heard_transcript": "Send ninety one dollars privately on um, let me think, Cash App to the person with this phone number, two one nine, two one five, eight two nine seven. Let me correct the app, use um, Venmo not Cash App." }, "candidate_id": "wav_1a3d61719242c274", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear request, then uses a believable 'let me think' filler before correcting the app from 'cash app' to 'Venmo'. The correction is integrated smoothly within the same utterance, with a natural self-repair and no awkward pauses. The prosody and pacing mimic genuine hesitation and recollection, making it sound like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "27ba739a886c624d977ee97001c5ac5ca3c03b631567481c574e3173340af727", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/cancellation/cancel_018/d5e90d1e78ae2aa3/speech.wav", "wer": 0.1, "wav_sha256": "733022540b37730e211a015fbd56bd08d1ca8263ebc305a7b5a01114c868ef9e", "seconds": 17.72, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven.", "Wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the full transaction request, then clearly revokes it with 'Wait, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the task. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send ninety-one dollars privately on Venmo to the person with this phone number two one nine two one five eight two nine seven. Wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "24c090675eb12bab9c6b579ee27de61d2733b3447a40c08a2dd3be32983123c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/ambiguous_halt/halt_008/b1053211861bcb96/speech.wav", "wer": 0.125, "wav_sha256": "72a5dff15ebf9ec6ab552fa3a3318bdcbe2e6581daf3c65d457e463d42944f7f", "seconds": 15.48, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction with a final interjection 'Wait a second,' which is an audible request to pause or hold. There is no cancellation or permission to resume, and the original task is not executed. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Send ninety one dollars privately on uh Venmo uh let me think to the person with this phone number two one nine two one five eight two nine seven. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e7b2bce7c6e6fb8993ea1813175c9d5fb47220a3508e37e3a165b9b40f52cd57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/filler/original/e32fa0e452d7e06a/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "9b5974af3bd6aba289dd072ef5d7dd33c404363f1e8d10fa96924948af5b7f5e", "seconds": 12.44, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'uh, let me think' but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification to the amount, app, privacy setting, recipient, or phone number. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Send $91 privately on Venmo, uh, let me think, to the person with this phone number 2192158297." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_app", "proposed_instruction": "Send $100 publicly on CashApp—I should say Venmo to the person with this phone number 5734599766.", "long_pause_offsets": [], "source_sha256": "092246b3e81ce132ca067ae32fa798171956cd076b6ead14720ef4d89c6ef720", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'CashApp' but immediately corrects to 'Venmo'. The amount and recipient phone number are correctly stated in the final version." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/false_start/false_start_03_app/a4157bb11340ca83/speech.wav", "wer": 0.09375, "wav_sha256": "9441f7a265326c5e1f97617b1794c07ecb206c75997fef8ba9e61da09186feaf", "seconds": 13.08, "spoken_utterances": [ "Send one hundred dollars publicly on CashApp—hmmm... I should say Venmo uh, let's see... to the person with this phone number five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. The final intent is clear and matches the original instruction with the correct app. The value and phone number are correctly stated. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Send one hundred dollars publicly on Cash App. Perm. I should say Venmo. Uh, let's see. To the person with this phone number five seven three four five nine nine seven six six." }, "candidate_id": "wav_c28ada8d792cc421", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident request, then self-corrects 'Cash App' to 'Venmo' with a natural hesitation ('uh') and a brief pause ('let's see'), followed by a smooth continuation. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The number is spoken clearly and without hesitation." } ] }, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Send $100 on Venmo to the person with this phone number 5734599766. Oh, and one more thing: make the payment public.", "long_pause_offsets": [ 68 ], "source_sha256": "e4d2c154aaaae3a37c8f3fa82aaa7e021fb4e226863c9ce9b0151258e9a6ae74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'public' constraint, which was not in the initial utterance but is present in the original instruction. The amount, recipient, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/addition/addition_01/215bc7c0552a9356/speech.wav", "wer": 0.0, "wav_sha256": "2a4e3c650aa400bcc301437aeca0ffa102bae215a4b2635e503bac8d56faa9d6", "seconds": 15.76, "spoken_utterances": [ "Send ... one hundred dollars on Venmo ... to the person with this phone number five seven three, four five nine, nine seven six six. ", "Oh, and one more thing: make the payment public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to send $100 on Venmo to a person with a given phone number. After a pause and an 'Oh, and one more thing' interjection, the speaker adds the constraint 'make the payment public'. This is a clear addition to the initial request, and the final intent combines both the original and the added condition. The addition is not a correction or cancellation, and the final goal matches the reference instruction with the added public visibility requirement. Therefore, the task is executed with the full intent.", "heard_transcript": "Send $100 on Venmo to the person with this phone number 5734599766. Oh, and one more thing, make the payment public." }, "candidate_id": "wav_5a795051f67463fa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the request with a natural, conversational rhythm and clear prosody. The initial command is fluent, followed by a plausible 'Oh, and one more thing' addition that smoothly introduces the extra instruction. The intonation and pacing suggest spontaneous thought, and the number is spoken clearly without hesitation or distortion." } ] }, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_visibility", "proposed_instruction": "Send $100 privately on Venmo to the person with this phone number 5734599766. Wait, the payment should be publicly, not privately.", "long_pause_offsets": [ 78 ], "source_sha256": "e9e57ce248e4bb835912fda363bf28fda62dd5845ba86fc72bd6a3d3c0a3d402", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the privacy setting from 'privately' to 'publicly', while retaining the amount, app, recipient phone number, and action. The final goal matches the original instruction except for the corrected visibility." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/revision/revision_02_visibility/f2a1b0ae323bf26d/speech.wav", "wer": 0.0, "wav_sha256": "3c624d1072346ab82aba7bb4172258d8d319a48f730b86d366db1c56458ae351", "seconds": 17.44, "spoken_utterances": [ "Send one hundred dollars privately on Venmo ... to the person with this phone number uh... five seven three, four five nine, nine seven six six. ", "Wait, the payment should be um... publicly, not privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested a private payment but immediately corrected it to 'publicly, not privately' after a pause. This is a clear revision of the constraint. The final audible intent matches the original instruction with the corrected privacy setting, so the revision criterion is satisfied.", "heard_transcript": "Send one hundred dollars privately on Venmo to the person with this phone number, uh, five seven three, four five nine, nine seven six six. Wait, the payment should be um, publicly, not privately." }, "candidate_id": "wav_459b01f69a498b6f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear, fluent request, then pauses and revises the privacy setting. The 'Wait' is a natural, pragmatic cue for a change of mind. The correction is delivered with a slight shift in tone and a brief pause, suggesting a moment of recollection. The flow is coherent, though the initial delivery is slightly more rehearsed than audio_01." } ] }, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "653efc5e346080a8f70acd75899ae9a7e2f74b060ea2ae575a09e4e5efef558a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/cancellation/cancel_040/87e5cb07388ed68a/speech.wav", "wer": 0.0, "wav_sha256": "52419912032b766c98fd474b9c9f8937bb76e6877787f28fd34d03eee5e8239c", "seconds": 17.64, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six.", "Wait, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly states 'Wait, I don't want to proceed with that task. Please cancel it.' at the end, which is a direct and unambiguous cancellation of the previously stated task. The cancellation is not conditional or followed by a new instruction, and the original task is fully described before the cancellation is issued. Therefore, the final control intent is to cancel, matching the 'cancellation' type criterion.", "heard_transcript": "Send one hundred dollars publicly on Venmo to the person with this phone number five seven three four five nine nine seven six six. Wait, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "0b07d526b1b44c5ad858e42e4227eb1c8dd53a596dd944d171ee484a41a09c0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/ambiguous_halt/halt_017/8949775e5cbd391f/speech.wav", "wer": 0.0, "wav_sha256": "c0698853f61f63a113c51e56d40f6665cc7bbfe787b7e905ac308719401e4c0f", "seconds": 15.64, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and correct, and no correction or cancellation is made. The final control request is a pause, satisfying the type-specific criterion.", "heard_transcript": "Send one hundred dollars publicly on Venmo to the person with this phone number five seven three four five nine nine seven six six. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c62ae424670dd2061a564468ff9975682d30d7ea820e886fc155f7cc7f87e7d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/filler/original/68c616eec1db8e20/speech.wav", "wer": 0.0, "wav_sha256": "5fa07ff1ca42a00a60e2a8be50e1f8ec2f75ff80c5c50cf6a6cb7617997a1a49", "seconds": 12.4, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio matches the original instruction exactly, with no hesitation, correction, or change in intent. The 'filler' type allows for natural speech including pauses or fillers, but the final task goal is clear and fully specified. No issues are present.", "heard_transcript": "Send $100 publicly on Venmo to the person with this phone number 5734599766." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $75—no, that should be $68 privately on Venmo to the person with this phone number 2114992604.", "long_pause_offsets": [], "source_sha256": "2c738f80e687daa1b6782b2940b8138edc03eb6489e78caa1097b6acf824a922", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces seventy-five with sixty-eight; the recipient, app, privacy setting, and phone number remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/false_start/false_start_01_amount/fb50f108297a2747/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2620d2a6c7ce548d10a610a9f40f3d23e434d003ea060c5fb1f171689b37806c", "seconds": 14.92, "spoken_utterances": [ "Send seventy-five dollars—um... no, that should be hmmm... sixty-eight dollars privately on Venmo to the person with this phone number ... two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'seventy five dollars' but immediately self-corrects to 'sixty eight dollars'. According to the 'false_start' type criterion, the final intent is the corrected value, and the original incorrect value is discarded. All other conditions (private, Venmo, recipient by phone number) are preserved. The final amount and recipient match the original instruction.", "heard_transcript": "Send seventy five dollars dollars um no that should be sixty eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four" }, "candidate_id": "wav_9e4f94cd33bb3b9f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent false-start repair. The speaker begins with 'send seventy-five dollars' and immediately self-corrects with 'no, that should be sixty-eight dollars' in a smooth, conversational tone. The correction is integrated into the same utterance without awkward pauses or stutters, and the prosody reflects genuine hesitation and correction. The delivery is continuous and credible, with no signs of artificiality or forced pacing." } ] }, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Send $68 on Venmo to the person with this phone number 2114992604. Oh, I forgot to mention, send it privately.", "long_pause_offsets": [ 67 ], "source_sha256": "ad75f794538fea9dc016f1f497085a337be41b8474955461ac5c0c0a620fe48e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing privacy requirement; all other values including amount, app, recipient identifier, and phone number are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/addition/addition_01/c873b5515dd0715c/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f3a6708189232317277803727a488e0a2a65cb96f10e775c1cab1d9cf6716616", "seconds": 18.2, "spoken_utterances": [ "Send ... sixty-eight dollars on Venmo uh... to the person with this phone number um... two one one, four nine nine, two six zero four. ", "Oh, I forgot to mention, send it privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction to send $68 on Venmo to the specified phone number is followed by an addition that explicitly requests the transaction be sent privately. This addition modifies the original instruction by adding a constraint, and the final audible intent is to execute the original task with the added privacy condition. The addition is clear and does not contradict the original request, satisfying the 'addition' type criterion.", "heard_transcript": "Send sixty-eight dollars on Venmo uh to the person with this phone number um two one one four nine nine two six zero four. Oh, I forgot to mention, send it privately." }, "candidate_id": "wav_888e21af667fd9ac", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('uh', 'um') and a believable self-correction ('Oh, I forgot to mention...'). The addition of 'send it privately' is smoothly integrated with a change in intonation and pacing, sounding like a spontaneous afterthought rather than a rehearsed or robotic insertion. The prosody is consistent, and the number is pronounced clearly without distortion." } ] }, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_amount", "proposed_instruction": "Send $75 privately on Venmo to the person with this phone number 2114992604. Sorry, I meant $68, not $75.", "long_pause_offsets": [ 77 ], "source_sha256": "ec75aae00517c04772ad0ecf5564243022a4d38a49195b711c8c3db13afcfecc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from seventy-five to sixty-eight, while retaining the app, privacy setting, recipient phone number, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/revision/revision_01_amount/e34b58b0c87330ee/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "4df0a41eefcbb2cb7298064ab72587730b21bb4cbfd14dbf04351f144d3bbfe5", "seconds": 23.84, "spoken_utterances": [ "Send seventy-five dollars privately on ah... Venmo let me think... to the person with this phone number uh... two one one, four nine nine, two six zero four. ", "Sorry, I meant sixty-eight dollars, not seventy-five dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states $75 but immediately corrects to $68 after a pause, which is a clear revision. The final intent matches the original instruction exactly, including the app (Venmo), amount ($68), privacy setting (privately), and recipient (by phone number). The revision type criterion is satisfied.", "heard_transcript": "Send seventy five dollars privately on uh Venmo, let me think, to the person with this phone number uh two one one four nine nine two six zero four. Sorry, I meant sixty eight dollars not seventy five dollars." }, "candidate_id": "wav_d056b069d5032577", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, inserts a believable 'let me think' filler, and then self-corrects the amount from $75 to $68 with a smooth, spontaneous-sounding 'sorry, I meant...'. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_033", "hide_variant_original": true, "source_sha256": "b763cd235f66dad164df022079ff1b00f54fdc9d3a30cb619b967ea2fa55216c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/cancellation/cancel_033/47b053ffcbf4157e/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "77b2a9c833c5b88ac97365dc3db418d2f2837092309a200223b4460309a23ab5", "seconds": 15.96, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four.", "Hang on, that request is no longer needed. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker stating 'Cancel it' after acknowledging the initial request, which constitutes a direct cancellation of the task. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four. Hang on, that request is no longer needed. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "e5e76b1e922f95a1eb5721fc6043c5047311d876b00874e38cf7050ef859ef3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/ambiguous_halt/halt_037/f8c5536a5863d41f/speech.wav", "wer": 0.1, "wav_sha256": "cddd75c1bddd19053e14783808b302c7bd73e11724c5f3a8a6a5e7846c7fd7c1", "seconds": 14.16, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the final interjection 'Pause for a moment,' which is a clear request to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and complete, but the final spoken control is a pause request.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "37a8675_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ec0741a1d9cc5a16338098ee7e9ee75a3f80f407f1bf4e8e089056b4012c34bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/filler/original/d930e7c8fefaa0d2/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "67cbda7c3f9f78d3fa856e8f9a07ed6d48e3a38b9a4eb81fe5ee097178b9b6a9", "seconds": 10.64, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'Send uh' but then clearly states the full correct instruction. The hesitation and filler sound are ignored as per the 'filler' type criterion, which allows for such interruptions as long as the final task meaning matches the original instruction. The amount, action, privacy setting, app, and recipient identifier are all correctly stated and match the reference.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_invitation_app", "proposed_instruction": "I got some Venmo—I should say Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fb03a6be220edbafccc9ec7a502d0ee06dfe18d92e984dc33ebe9f42486b77f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Venmo' to 'Splitwise' within the same utterance, preserving the rest of the instruction including the app, message type, condition, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_01_invitation_app/73dcea0d09bf08b0/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "90d288739b2800c1d40be92f9f5b026993bfe9d37faea6211b08d30edae78903", "seconds": 16.28, "spoken_utterances": [ "I got some Venmo—uh... I should say Splitwise group invitations ... over phone text messages yesterday. ... If their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise'. This is a false start, and the final, corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "I got some Venmo uh, I should say Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_5dc224bfb410b532", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers the request with a natural, spontaneous flow. The initial false start 'Venmo uh' is immediately self-corrected to 'Splitwise', which sounds like a genuine moment of recall. The rest of the sentence is spoken at a steady, conversational pace with no awkward pauses or stutters. The prosody is consistent and the correction feels integrated, not forced." } ] }, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday.. Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [ 76 ], "source_sha256": "c66e60a6f1eea8e62f925f3af2e30c500637eb7302fd6fd8f64c940b9c5158e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the conditional logic that was missing in the initial utterance. The final goal combines the initial context (Splitwise group invitations via text) with the added condition, preserving the original intent of handling messages based on contact book presence." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/addition/addition_01/f22096089df02c59/speech.wav", "wer": 0.029411764705882353, "wav_sha256": "84033607673a97a5d23f9ff39490de74e51ea42e516c78dbed4e620f06955029", "seconds": 16.4, "spoken_utterances": [ "I got some uh... Splitwise group invitations over phone text messages let's see... yesterday.. ", "Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task and then adds a condition ('if their number is in my phone contact book, accept it, otherwise delete those messages') after a pause. This is a clear addition that refines the original instruction. The final intent is to accept invitations from contacts and delete those from non-contacts, which is a valid execution of the combined request.", "heard_transcript": "I got some uh Splitwise group invitations over phone text messages. Let's see. Yesterday. Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_3ecaca228b6b52fc", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_invitation_app", "proposed_instruction": "I got some Venmo group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Sorry, those are Splitwise group invitations, not Venmo invitations.", "long_pause_offsets": [ 160 ], "source_sha256": "ed6fc912c7e94b9444028979860edbfe5634ed40b3a3756a48c42bfd88ca164b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Venmo to Splitwise, which matches the original instruction. The rest of the logic, including the condition based on the phone contact book and the actions (accept or delete), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_03_invitation_app/505bc0fd7d73c072/speech.wav", "wer": 0.1282051282051282, "wav_sha256": "d3aea5b621b643850599c70f428cca646fd296abd1f32393959a8c52b82365f9", "seconds": 23.76, "spoken_utterances": [ "I got some ... Venmo group invitations over phone text messages um, let's see... yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "Sorry, those are Splitwise group invitations, not uh... Venmo invitations." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the app as 'Venmo' but immediately corrects it to 'Splitwise' after a pause. According to the 'revision' type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to handle Splitwise invitations with the same conditional logic (accept if in contacts, delete otherwise). All other conditions from the original instruction are preserved. The correction is clear and unambiguous, satisfying the revision type requirement.", "heard_transcript": "I got some Venmo group invitations over phone text messages, um, let's see, yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Sorry, those are Splitwise group invitations, not uh, Venmo invitations." }, "candidate_id": "wav_bdc355a470268b7a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'I got some Venmo group invitations...' and uses a genuine 'um, let's see' pause before continuing. The self-correction from 'Venmo' to 'Splitwise' is a plausible false start with immediate, smooth repair. The intonation and rhythm are consistent with someone recalling and adjusting a task on the fly, and the final sentence flows naturally without awkwardness." } ] }, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "85cad0d45022cdb8c9817a72a0b75dfe7ad720f64b858e0b28219c656d83a8c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/cancellation/cancel_031/c0d8e9c12dc52586/speech.wav", "wer": 0.18604651162790697, "wav_sha256": "32821c9e481dec05fadf4ed89ca916a1502fe315399f52355cb402cef49b3920", "seconds": 21.56, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages.", "Hang on, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Hang on, there's no need to do that anymore. Cancel it.' This constitutes a final, unambiguous cancellation of the entire task. The initial instruction is irrelevant once the cancellation is stated.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages. Hang on, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "c3c88df45260532752ddd986bb7b9ef12e874393d89d7a562d855faa6c32365b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/ambiguous_halt/halt_008/01b4d6df3529685f/speech.wav", "wer": 0.22857142857142856, "wav_sha256": "e1321f170707a7e779c06fcbb96910537096361dcc98d5f7156cfc8bf58fb451", "seconds": 19.24, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait a second,' which is a clear request to pause or hold. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled or revised, and the speaker does not provide a new instruction or cancel the task. Therefore, the final control intent is to pause.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f553ce460479f681d333d37bba3c492dd97070752ac9a470f90cb144fe2ef0fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/filler/original/a1a9aba583fc51b9/speech.wav", "wer": 0.25, "wav_sha256": "574d618966f8a412eefed6ba3fc86d981f7885d002bd4dedd3a89020c61d4b20", "seconds": 15.96, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') and a pause but does not change the core task. The final instruction matches the original request. The hesitation and self-repair do not alter the goal, so the 'filler' type criterion is satisfied.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages yesterday—no, I mean the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "78aaa23492308a9c3b9d195b1c83b445c278763eb3d57f956c4c3da2dc5aa7e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'yesterday' to 'the day before yesterday' within the same utterance. The final time reference matches the original instruction, and all other conditions (app, message type, recipient condition, and action) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_03_time_reference/3d7fa312b9b152c4/speech.wav", "wer": 0.0, "wav_sha256": "938a02148faaf63418bca7d2701b4f4d3ae8cc570cdc1e79522812587c694d71", "seconds": 15.08, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages yesterday—um... no, I mean ... the day before yesterday. If their number is in my phone contact book, uh... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yesterday' but immediately self-corrects to 'the day before yesterday'. This is a false start, and the final, corrected time reference matches the original instruction. All other conditions (Splitwise group invitations, phone voice messages, accept if number is in contact book, delete otherwise) are preserved. The final intent is clear and executable.", "heard_transcript": "I got some splitwise group invitations over phone voice messages yesterday. Um, no, I mean the day before yesterday. If their number is in my phone contact book, uh, accept it, otherwise delete those messages." }, "candidate_id": "wav_a04167e0574414b1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker initially says 'yesterday', then self-corrects to 'the day before yesterday' with a clear, spontaneous 'no, I mean' transition. The prosody is consistent, and the correction feels like a genuine moment of recall rather than a rehearsed edit. The 'uh' pause before 'accept it' is also well-placed and natural." } ] }, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I left out a detail: these were the invitations from the day before yesterday.", "long_pause_offsets": [ 155 ], "source_sha256": "79ee9763ad74725b659b329b0b30f6d9c1524e658bff73cd3a41210f60fa7163", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time detail 'from the day before yesterday' to the initial instruction, which was otherwise complete. The action, condition, and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/addition/addition_02/a6429b7cc1b245d6/speech.wav", "wer": 0.06976744186046512, "wav_sha256": "757b53b728f6553c2cb2d0ba4be0dd0a1fe7df5b67b12772f23c59d2bfe4277b", "seconds": 25.0, "spoken_utterances": [ "I got some let's see... Splitwise group invitations hmmm... over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "I left out a detail: these were the invitations from the uh... day before yesterday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the core instruction to accept or delete Splitwise invitations based on contact book presence. After a pause and self-correction ('I left out a detail'), they add the time constraint 'from the day before yesterday'. This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the new temporal condition, matching the expected behavior for an 'addition' type. The control intent is to execute the combined task.", "heard_transcript": "I got some let's see, Splitwise group invitations, hmm, over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I left out a detail, these were the invitations from the day before yesterday." }, "candidate_id": "wav_2cdac0284a47fef4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear initial request, uses a plausible 'hmm' pause to transition, and then adds a detail ('these were the invitations from the day before yesterday') in a way that sounds like a spontaneous memory. The prosody is varied and the pacing is conversational, with a natural self-correction from 'accept it' to 'accept it otherwise delete those messages' that fits the context. The addition feels integrated, not tacked on." } ] }, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. I misspoke about the time: it should be the day before yesterday, not yesterday.", "long_pause_offsets": [ 165 ], "source_sha256": "878201181fdaf3b7ffdc53cb20a8e790f600bfe29807570788f2dc3dc64c02d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'yesterday' to 'the day before yesterday', which matches the original instruction. All other conditions and actions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_03_time_reference/c31f4c711f45f71c/speech.wav", "wer": 0.0, "wav_sha256": "49b9d483c03d35495c70adc5a56c6ef9a9ead9ed796bf8f61a6fd4e71090ff2a", "seconds": 24.6, "spoken_utterances": [ "I got some Splitwise group invitations uh, let's see... over phone voice messages yesterday. uh, let me think... If their number is in my phone contact book, accept it, otherwise delete those messages. ", "I misspoke about the time: it should be the day before yesterday, not yesterday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yesterday' but immediately corrects it to 'the day before yesterday' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original instruction exactly, so the task is executed as revised.", "heard_transcript": "I got some Splitwise group invitations uh let's see over phone voice messages yesterday. uh let me think. if their number is in my phone contact book, accept it, otherwise delete those messages. I misspoke about the time, it should be the day before yesterday, not yesterday." }, "candidate_id": "wav_4252315a2428d0c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and continuous flow. The speaker uses a realistic 'uh' and 'let me think' filler, followed by a self-correction of the time ('day before yesterday, not yesterday') that is integrated smoothly into the sentence. The prosody is conversational, with appropriate pauses and intonation, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "883115327695a2fd6a85868507f8118d6f6366764cfb21008fe593f8fc00fbb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/cancellation/cancel_021/5fb8c9a529781de3/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "8eff0b30305c08b01c07670924301ce4f9d11daf1201189384a038aa3b0cd4fc", "seconds": 22.12, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages.", "Oh, wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Oh wait, never mind. I don't need that done anymore,' which is a clear and final cancellation of the entire task. The initial instruction is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages. Oh wait, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "8620cabf17d919cde005b90b5e08235f3e743269aee28d801cd1829e38b2ba88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/ambiguous_halt/halt_019/bde1aa837929286d/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "9e09a1373b3e9a1190a57d1e2327775d761ec41eba7356fc517086a4c6bd6405", "seconds": 20.56, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Can you give me a moment?' is a clear request to pause or wait, with no indication of cancellation or a new instruction. The user's hesitation and backchanneling ('um', 'let's see') are internal processing cues, not a change in the task goal. The final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "df484d4b8146c508ee72c454695ad83a0276cb56e298bcda2aaec196c1240e35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/filler/original/1b106929953323d1/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "bb35edc055048afa1e520a5502814cf2d3120cde7a9b973f767ddf3ced66b119", "seconds": 16.56, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final instruction remains consistent with the original, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages last month—no, that should be this week. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fc7d89e007d71b7523c6d232c3e6fa150cd2d10fd94ce47358c99d413051311f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this week' within the same utterance. The rest of the instruction, including the condition and action for handling invitations, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_03_time_reference/87ea9ac6b3f5c3e8/speech.wav", "wer": 0.08571428571428572, "wav_sha256": "8bdc5984f1ef2cb986aace2b8b7fa48e5453376f8f0eb4934b19b9679dc392e3", "seconds": 14.24, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages ... last month—um... no, that should be this week. If their number is in my phone contact book, uh... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this week'. The final instruction matches the original reference exactly in action and conditions. The false start is corrected, so the final intent is clear and executable.", "heard_transcript": "I got some Splitwise group invitations over phone text messages last month. Um, no, that should be this week. If their number is in my phone contact book, uh, accept it, otherwise delete those messages." }, "candidate_id": "wav_e3e6c0b3411d9357", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural revision pattern: the speaker starts with 'last month', immediately self-corrects to 'no, that should be this week' with a believable hesitation and intonation shift, then continues the sentence smoothly. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause before the correction is also well-placed, suggesting genuine recall." } ] }, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "I got some Splitwise group invitations over phone text messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I didn't mention this, but it's specifically for invitations received this week.", "long_pause_offsets": [ 154 ], "source_sha256": "8fa21282734f1a5a08503ebfeda12e7a2520f8e8b44fdb8b8594b425383a39ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this week') that narrows the scope of the initial instruction. The core logic (accept if in contacts, delete otherwise) and the context (Splitwise group invitations via text) remain unchanged. The addition is a clarifying constraint, not a contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/addition/addition_02/9dd120660a2ff927/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "0fe93dc44bf07a34769e611bf4b4a6d5af5b37444a4d238e1c3dd8139219b655", "seconds": 21.56, "spoken_utterances": [ "I got some ah... Splitwise group invitations over phone text messages. uh... If their number is in my phone contact book, ... accept it, otherwise delete those messages. ", "I didn't mention this, but it's specifically for invitations received this week." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition of 'this week' is a narrowing constraint that is clearly stated after a pause. The final intent combines the original task with the new temporal condition, matching the expected behavior for an 'addition' type. The addition is not a correction or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "I got some uh Splitwise group invitations over phone text messages. Uh if their number is in my phone contact book, accept it, otherwise delete those messages. I didn't mention this, but it's specifically for invitations received this week." }, "candidate_id": "wav_6d20a204a7ec8343", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural prosody and a single, plausible addition at the end. The phrase 'I didn't mention this, but it's specifically for invitations received this week' is smoothly integrated as a post-hoc clarification, with appropriate intonation and pacing. The delivery is fluent and sounds like a spontaneous thought being added, fitting the 'addition' augmentation type well." } ] }, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_invitation_app", "proposed_instruction": "I got some Venmo group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Wait, the app should be Splitwise, not Venmo.", "long_pause_offsets": [ 160 ], "source_sha256": "d537218ec93f940153333396f85740afe534978f92108774e6573f57341e5204", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the app from Venmo to Splitwise, which matches the original instruction. All other conditions, including the action logic based on contact book presence, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_01_invitation_app/c5d85d866d716a1d/speech.wav", "wer": 0.09302325581395349, "wav_sha256": "74ed1087eeaa2daf4bd2e7c793d69cc22d3347ccd5ebd24f0aee848759e40691", "seconds": 22.96, "spoken_utterances": [ "I got some Venmo group invitations over phone text messages this week. um... If their number is in my phone contact book, let me think... accept it, otherwise delete those messages. ", "Wait, the app should be um, let me think... Splitwise, not Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately corrects to 'Splitwise' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "I got some Venmo group invitations over phone text messages this week. Um, if their number is in my phone contact book, let me think, accept it, otherwise delete those messages. Wait, the app should be, um, let me think, Splitwise, not Venmo." }, "candidate_id": "wav_1574fffe6d36df75", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and complex revision pattern. It begins with a plausible false start ('Venmo group invitations'), self-corrects to 'Splitwise', then pauses with 'let me think' before stating the condition. It later revises the app name again with 'wait, the app should be... let me think', showing a believable moment of recollection. The prosody includes natural hesitations and a genuine-sounding correction process, making it the most realistic." } ] }, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "416f857b13a06192c858b1aed8b0eee0f27f0a377b3017cf6301e85a2aa84a0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/cancellation/cancel_034/a4742a150ae71e33/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "3b3c397393e5d0a95afc3e854324307d57d193024ad1882979633cda5a3e2844", "seconds": 19.6, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then concludes with the explicit cancellation phrase 'Hold on, let's drop the whole task.' This clearly revokes the task, satisfying the cancellation type criterion. The final intent is to cancel, not to execute.", "heard_transcript": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "26568bfc60f4af301900fd4c0a68de5ac1b3c71e0000921aa69015d4c6429da0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/ambiguous_halt/halt_007/1433befc45f1effa/speech.wav", "wer": 0.12121212121212122, "wav_sha256": "9d93e961081b4587ed8b04490efc2a9e5c638dad034c6c51391c70f8f4357eaf", "seconds": 18.08, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment,' which is a clear request to pause or wait. The user does not cancel the task or provide a final instruction. The initial part of the audio is a filler ('let me think') followed by the original instruction, but the final control request is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "I got some Splitwise group invitations over phone text messages. Uh, let me think. This week, if their number is in my phone contact book, accept it. Otherwise delete those messages. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3aa1a22_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c60a71174929a04661fdbf593f9bbbaa348db5e2efa9f04953dc8818ab669d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/filler/original/3e9ed68dca2b9326/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "13aed50752dfdd73db4ce55ace2a2e6a03cf04fb6224a06dc244aedfbfe4c893", "seconds": 14.96, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'let me think' but continues with the original instruction without changing the goal. The final intent matches the original instruction, and the filler does not alter the task meaning. Therefore, the task should be executed as intended.", "heard_transcript": "I got some Splitwise group invitations over phone text messages. Uh, let me think. This week, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify playlists that I have saved—sorry, I meant liked.", "long_pause_offsets": [], "source_sha256": "c1044f38a8b3d57e845dca1b5f72458956b58f485fc576c735458d451a436737", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended target is 'liked', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/false_start/false_start_03_interaction_type/6f475b63ef393876/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "9c4b702e4ff04d3e62ae121bccf66681e8a874508aed022496fa0216f82566c7", "seconds": 10.4, "spoken_utterances": [ "Download all the songs from my Spotify playlists uh... that I have saved—um... sorry, I meant ... liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value 'liked' should be applied, and the original incorrect value 'saved' should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Download all the songs from my Spotify playlist, uh, that I have saved, um, sorry, I meant liked." }, "candidate_id": "wav_51274bbd52730bc7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'playlists uh' and then smoothly transitions to 'saved um' before correcting to 'liked'. The false start and repair are integrated within a single, continuous utterance with natural hesitation and intonation. The pacing and prosody are consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Download from my Spotify playlists that I have liked. I left out a detail: download all the songs.", "long_pause_offsets": [ 54 ], "source_sha256": "e581854fe0b7d7cac6a7673661d074b7c10e104becfd8c6435a1732cf5d05a66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'all the songs' requirement, completing the initial instruction which lacked the scope of the download. The app, source, and condition (liked playlists) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/addition/addition_02/d234a79b9ebc7e9c/speech.wav", "wer": 0.2, "wav_sha256": "30fb8f3bd43f93250dc8cad6f8241731187fd2516cbacf7c299dae8ed5b45e2d", "seconds": 11.44, "spoken_utterances": [ "Download from my Spotify playlists uh... that I have liked. ", "I left out a detail: download ah... all the songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Download from my Spotify playlist uh that I have liked.' They then add, 'I left out a detail, download uh all the songs.' This is a clear addition that expands the initial request to include 'all the songs.' The final intent combines the original instruction with the added detail, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Download from my Spotify playlist uh that I have liked. I left out a detail, download uh all the songs." }, "candidate_id": "wav_74b2f8b46248cc1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then adds a detail ('download all the songs') after a brief, plausible self-correction ('I left out a detail'). The prosody is consistent, and the addition flows smoothly without awkward phrasing or unnatural pauses. The speaker sounds like they are genuinely remembering a missing part of the task." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify playlists that I have saved. That should be liked, not saved.", "long_pause_offsets": [ 68 ], "source_sha256": "f3ccf592afe11ea3bf7222b04f25521670c793319d995c2c5163c848798a52cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the condition from 'saved' to 'liked', preserving the app, action, and scope of 'all songs from playlists'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/revision/revision_03_interaction_type/bc4d104cbf957d57/speech.wav", "wer": 0.2, "wav_sha256": "6f1d9a454e89c291fc59925f1dfab4cd207941e264870be9c845307d98f2703e", "seconds": 12.84, "spoken_utterances": [ "Download all the songs from my hmmm... Spotify playlists um... that I have saved. ", "That should be liked, not saved." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is the revision type. The final intent is to download songs from playlists that are liked, matching the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Download all the songs from my um Spotify playlists um that I have saved. That should be liked, not saved." }, "candidate_id": "wav_7b25c8dd02a98d66", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic revision pattern. The speaker begins with a plausible false start ('my um Spotify playlists um'), uses filler words and pauses that sound spontaneous, and then clearly self-corrects 'saved' to 'liked' with a believable change-of-mind intonation. The correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "1c59f5f7113140e39560a28b06eb5840b5afb9ce7dae44f624e2a551989aae27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/cancellation/cancel_003/22e0cbdefa0e5d5c/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "9b7d875269666bd766a352544482dc18275df171c54ae877bcb7ecd81bae152f", "seconds": 11.04, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked.", "Hold on, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting a task, then immediately follows with 'Hold on, please cancel that,' which is a clear and unambiguous cancellation of the previously stated request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked. Hold on, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "d4787ebec7f894d8df532d0da5378c85582ad14a68d39caee006a4856c6d215c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/ambiguous_halt/halt_040/92a446d6666a4439/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "adcfbbfae220b578a16142d6064c20027f090ebee8d6bb86530981f5afd46949", "seconds": 10.4, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly requests a pause with 'Hang on, just a minute.' This is a final interjection that indicates a temporary halt, not a cancellation or a change in the task goal. The original instruction is not revoked, but the speaker explicitly asks to wait, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "266b76f2aff1cbc6825c601ec436507c7a1902e0994aebf4ec064cca5ac516ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/filler/original/9c6489e39e4a14e6/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "871a4945b23c94039348e34607788e26a1776b3613bbe256b88154f2704f9788", "seconds": 6.32, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent is identical to the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify song library that I have saved—I meant liked.", "long_pause_offsets": [], "source_sha256": "5d2365743b3884153e0c580d455fa1cd94c96ad96164c0c156e3485479a6b79b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended action is to download songs that have been liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/false_start/false_start_03_interaction_type/cef89a6eab7d6767/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8b164b2a94500d0561590509ee92726a82ffd4d3d0a6994ec625ba7d47e64c2b", "seconds": 7.12, "spoken_utterances": [ "Download all the songs from my um... Spotify song library that I have saved—uh... I meant liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects with 'I meant liked'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is discarded. The final audible task matches the original instruction exactly after the correction.", "heard_transcript": "Download all the songs from my um Spotify song library that I have saved uh I meant liked" }, "candidate_id": "wav_081be59c9cea1e6a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with 'saved' and immediately self-corrects to 'liked' with a clear, audible 'uh' and a slight pitch drop, mimicking genuine hesitation. The prosody is fluid, with a natural rhythm and appropriate pauses. The correction feels integrated into the thought process rather than tacked on, and the overall delivery is continuous and believable." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Download all the songs from my Spotify song library. Oh, I forgot to mention, only download the songs that I have liked.", "long_pause_offsets": [ 53 ], "source_sha256": "467c4c5567706a8e2c3ed5728669a9930ad37e694e33ccac441adac209e0553b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'that I have liked', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/addition/addition_01/123d2501c0820b00/speech.wav", "wer": 0.04, "wav_sha256": "0269c92c3ea85aa0dcdc10b0f0a6a30ced5ae3b5930645572eab2f53f0d8ae24", "seconds": 13.76, "spoken_utterances": [ "Download all the songs from my uh... Spotify song library. ", "Oh, I forgot to mention, only download the songs let's see... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to download all songs from the Spotify library. After a pause and a self-correction, the speaker added the constraint 'only download the songs that I have liked'. This is a clear addition that narrows the scope of the original request. The final intent is to execute the download with the added constraint, which matches the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Download all the songs from my uh Spotify song library. Oh, I forgot to mention, only download the songs, let's see, that I have liked." }, "candidate_id": "wav_5d31de81afa6c98c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, uses a plausible filler ('uh') and a brief pause, then smoothly adds a restriction ('only download the songs... that I have liked') after a moment of hesitation ('let's see'). The self-correction from 'Spotify song library' to 'that I have liked' is integrated naturally within the flow, with a believable change of mind and intonation. The pacing and prosody are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify song library that I have saved. I misspoke about the interaction: it should be 'liked', not 'saved'.", "long_pause_offsets": [ 71 ], "source_sha256": "4ac4d1b6a69f546497eee112d43487073debb76faceec3053b5c34a4580b2f8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'saved' to 'liked', while retaining the app, scope ('all songs from my library'), and action ('download')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/revision/revision_03_interaction_type/c5145394616ed46f/speech.wav", "wer": 0.12, "wav_sha256": "c082914b72b255cb5325605afda711c649ab54468b072163c956c11aacc87c5c", "seconds": 13.88, "spoken_utterances": [ "Download all the songs from my Spotify song library ... that I have saved. ", "I misspoke about the interaction: it should be 'ah... liked', not 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is a revision. The final audible intent is to download songs that are 'liked', matching the original instruction. The revision is clear and the correction is applied to the earlier instruction.", "heard_transcript": "Download all the songs from my Spotify song library that I have saved. I misspoke about the interaction, it should be uh liked, not saved." }, "candidate_id": "wav_a092ea94f1db1db7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a natural, spontaneous-sounding revision. The speaker begins with a clear request, then hesitates with 'I misspoke about the interaction,' followed by a smooth correction from 'saved' to 'liked.' The intonation and pacing are consistent with a real-time self-correction, and the pause before the correction feels like genuine recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "91fff0a3b381b5d4bd16f9b0eb7fa078c4d46fc1a02a4667a2d5e33e6d200f4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/cancellation/cancel_024/aa006c54858aa1fc/speech.wav", "wer": 0.0, "wav_sha256": "f6b4228d5e1da5196a522a1e42dfe0d1f2499dcd2aaf2f7b7f7721fa7b8eab29", "seconds": 13.16, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked.", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to download liked songs from Spotify. After a brief pause and the interjection 'Hang on', they explicitly state 'Oh, never mind. Cancel the whole thing.' This constitutes a clear cancellation of the entire task as per the 'cancellation' type criterion. The final audible intent is to cancel, not to execute or pause.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked. Hang on. Oh, never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "2eac5f3457f53b29a3e00331640029f1d7ad5de2748dbce9036899b1f5f52e38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/ambiguous_halt/halt_034/ebf5b95e520c1bf4/speech.wav", "wer": 0.0, "wav_sha256": "7e2d11aba5ec546419d763c21605dd1b923e1a6435a569ddf9bfb70b639c27e9", "seconds": 12.48, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation. No further instruction is given to resume or cancel, so the agent should pause.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2a1b314753a067b7c08a5cd5bbccec865190b76b3bbc56c253ba6989e45d1977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/filler/original/38a3d3c0ad67c3ca/speech.wav", "wer": 0.0, "wav_sha256": "c31d79abe73e5644d5cc989331d1850244c8bb54df06b541ec87378164c8afc7", "seconds": 7.76, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchannel, or pause that would indicate a change in intent. The instruction is delivered as a single, complete utterance. Since the type is 'filler', any initial hesitation or pause within the instruction is ignored, and the final task meaning must match the original instruction. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify playlist collection—or rather, album library that I have liked.", "long_pause_offsets": [], "source_sha256": "e8dadc757b101f469c06e5855f4f467c549ff0c79bfde660ecaa6753f0ffc36e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately corrects to 'album library'. The final target is the album library, and all other conditions (download all songs, Spotify, liked status) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/false_start/false_start_02_source_collection/81d1b432034fb8ec/speech.wav", "wer": 0.1, "wav_sha256": "fda8db22f8be95ec68c6ce6ce73697fb24aedee1d23d3b55c7e7cf08558067e1", "seconds": 10.04, "spoken_utterances": [ "Download all the songs from my let's see... Spotify playlist collection—um... or rather, album library that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately self-corrects to 'album library'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction, with 'album library' replacing 'playlist collection'. All other constraints are preserved.", "heard_transcript": "Download all the songs from my let's see, Spotify playlist collection or rather album library that I have liked." }, "candidate_id": "wav_e021215093068811", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. It begins with a filler ('let's see'), includes a clear self-correction from 'playlist collection' to 'album library', and maintains a consistent, conversational pace. The intonation and rhythm are highly realistic, with the correction sounding spontaneous and integrated into the sentence rather than tacked on." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Download all the songs from my Spotify that I have liked. I left out a detail: download from my album library.", "long_pause_offsets": [ 58 ], "source_sha256": "61c81f60bff15e57c4f717f9e346625ace64bade92f8b1f22aab036c20242300", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source 'album library' to the initial request, combining it with the existing condition 'that I have liked'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/addition/addition_02/5877bc1057888398/speech.wav", "wer": 0.08, "wav_sha256": "368eff07136b874c57d95ab0d6f8e41994f695033612bfb2610a605bc91a69fa", "seconds": 13.48, "spoken_utterances": [ "Download all the songs from my um, let's see... Spotify that I have liked. ", "I left out a detail: download from my um... album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Download all the songs from my Spotify that I have liked.' The speaker then adds a clarifying detail, 'Download from my um album library,' which narrows the scope of the original request. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent is to download liked songs specifically from the 'album library' on Spotify, which matches the reference instruction. The speaker's hesitation ('um') and self-correction are typical of natural speech and do not alter the final intent.", "heard_transcript": "Download all the songs from my um let's see Spotify that I have liked. I left out a detail. Download from my um album library." }, "candidate_id": "wav_622bbf4d890600a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, uses a plausible filler ('um') and a self-correction ('I left out a detail') that leads to a smooth, additive revision of the source ('download from my album library'). The prosody is consistent, and the pause before the correction feels like genuine recollection. The delivery is continuous and sounds like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify album library that I have saved. I misspoke about the interaction: it should be liked, not saved.", "long_pause_offsets": [ 72 ], "source_sha256": "7b0182507bcf61c921ad4bd38e320e1c382db0b373839086a819708285dd44e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'saved' to 'liked', while retaining the app, scope (album library), and action (download all songs)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/revision/revision_03_interaction_type/86e52cf0d505e460/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "43d57966d0d819c0c7a3c570af92798de24cb17289957370f7636d913914ae31", "seconds": 15.64, "spoken_utterances": [ "Download all the songs from my uh... Spotify album library um... that I have saved. ", "I misspoke about the interaction: it should be uh... liked, not saved." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is a clear revision. The final intent is to download songs that are 'liked', matching the original instruction. The correction is applied to the earlier instruction, and all other conditions remain unchanged.", "heard_transcript": "Download all the songs from my uh Spotify album library um that I have saved. I misspoke about the interaction. It should be uh liked, not saved." }, "candidate_id": "wav_b8329c9b5a04e3fe", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a clear request, then self-corrects 'saved' to 'liked' with a believable, slightly apologetic tone ('I misspoke about the interaction'). The correction is integrated smoothly into the utterance, and the prosody remains consistent and conversational throughout. The pacing and intonation suggest genuine spontaneous thought, making it the most realistic." } ] }, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "96ceb9f9a8f3edfeeeb5c9a4a0143a657025081c5c6b45c026e18f92d61082ce", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/cancellation/cancel_027/a91aaf28251cca85/speech.wav", "wer": 0.05, "wav_sha256": "8ff642de8a255d40b1a0799650770ee2252c00f49aaf05f13fd6f5a81f601c71", "seconds": 13.08, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked.", "Wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to download songs, then immediately follows with 'Wait, please disregard my last request,' which is a clear and unambiguous cancellation of the prior instruction. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked. Wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "8b7367b798ed6b08162d6d7eaf280be8d21faed4f0fb63e4036c2c679419c730", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/ambiguous_halt/halt_007/c37ee94b5bb2bf56/speech.wav", "wer": 0.0625, "wav_sha256": "423907938b792f290313dbb3fd77d95f6b613ac8bb0902ebcdb6ed174b5def53", "seconds": 11.12, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request and then clearly says 'One moment,' which is an audible request to hold or wait. This constitutes a final interjection that aligns with the 'ambiguous_halt' type, indicating a temporary pause rather than cancellation or execution. The original instruction is not revoked, and the speaker does not provide a new command, so the intent is to pause.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3ab5b8b_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "67103aabcc5101a98169d78af478c3952728bbcaed609d24d67f90c97f4fa22c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/filler/original/7d267fe5c412c938/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "22594376761915329a92c2864258d66d0578d74e887e8f0c8a8b3ac55c688541", "seconds": 8.0, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler word ('um') but the final, complete instruction matches the original request. The user does not correct or modify the task, and the intent to download liked songs from the Spotify library is clear and unchanged.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts—sorry, that should be payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "dffbb5b2d16efc381689ed8fbadc495ccef4a12a70bca88e87e7037977261aeb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'bank accounts' to 'payment cards' within the same utterance. The final target is 'expired payment cards', and the scope 'all my app accounts that have payment cards' remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/false_start/false_start_02_payment_method_type/8e953785046bb222/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "b8c3fe8b37ae8a5a273f9071ef141e25a674ac942fda86032cd2adec9d6a37e7", "seconds": 9.24, "spoken_utterances": [ "Remove expired bank accounts—um... sorry, that should be payment cards uh... from all my app accounts ... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bank accounts' but immediately self-corrects to 'payment cards' with 'sorry, that should be payment cards'. This is a false start where the corrected value replaces the mistaken one. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Remove expired bank accounts. Um, sorry, that should be payment cards. Uh, from all my app accounts that have payment cards." }, "candidate_id": "wav_1d3a16620ba598ae", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then naturally self-corrects from 'bank accounts' to 'payment cards' with a plausible 'um' and 'sorry' filler. The correction is immediate and integrated smoothly into the sentence, followed by a coherent continuation. The prosody is natural, with appropriate pauses and intonation that suggest spontaneous thought, making the false start sound authentic and well-executed." } ] }, "is_default": true } }, { "task_id": "3f3c139_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_account_scope", "proposed_instruction": "Remove expired payment cards from all my social media accounts that have payment cards. Let me correct the scope: remove expired payment cards from all my app accounts, not social media accounts.", "long_pause_offsets": [ 88 ], "source_sha256": "45b69e6dd62daea75fec68ae4e44a02147cb3aab525793de7d532589c81fb4f7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the scope from 'social media accounts' to 'app accounts', which matches the original instruction. The action, condition, and target type are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/revision/revision_03_account_scope/ac362c3f2b0651f5/speech.wav", "wer": 0.027777777777777776, "wav_sha256": "9647e62218992fe1ee9db7fa8664b963e7bd2cfd3684fef0d93c3bca55abe1f5", "seconds": 17.12, "spoken_utterances": [ "Remove expired payment cards uh... from all my social media accounts ... that have payment cards. ", "Let me correct the scope: remove expired payment cards from let me think... all my app accounts, not social media accounts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests removal from 'social media accounts' but then corrects the scope to 'all my app accounts, not social media accounts' after a pause. This constitutes a revision where the final intent is the corrected version. The final instruction matches the original reference instruction exactly.", "heard_transcript": "remove expired payment cards uh from all my social media accounts that have payment cards. let me correct the scope, remove expired payment cards from, let me think, all my app accounts, not social media accounts." }, "candidate_id": "wav_b7f21f154b412e44", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "3f3c139_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "24c28bd694c8235b1a533bc4eed6544fba627bd74447c5cf7094050f1db58c3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/cancellation/cancel_021/cc85a22564eb8607/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "c4e58b76283bb10f7f925cc4f77020274341768ca83a278720a826032892b9f2", "seconds": 13.72, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards.", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and self-correction ('Hold on, never mind'), explicitly states 'I don't need that done anymore,' which is a clear cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards. Hold on, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "f10ab6f38c37e3865c92356fbaabc67e07c7826b88b60b36c61ee7091723706f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/ambiguous_halt/halt_023/136a329e8d2d7186/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "8dce352312d4484d6f46f379c16cfbb5c7209a84fc5341b43bb68cd1b42c9ce3", "seconds": 12.72, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Hold on, I need a second,' which is a clear and audible instruction to pause. This matches the 'ambiguous_halt' type criterion, where the final interjection is a request to wait or pause without canceling the task. The initial request is not executed, but the final intent is to pause, not to cancel or execute. Therefore, the control intent is correctly identified as 'pause' and the verdict is 'pass'.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "85d1c62aa2b95d1218cb3ff18839d29359ff10cb1671a085ac8754981723be25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/filler/original/26809c68e07d5393/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "7ea91e7cba661da669fe690733f91a77e9530b551a706242b389aa5713b101ee", "seconds": 8.4, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitations ('uh', 'um') and a self-pause ('let's see'), but the final, complete instruction matches the original reference. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_account_scope", "proposed_instruction": "Remove expired payment cards from all my email accounts—no, that should be all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "93ef8d961e127a1cc87de19cacc7bbdad05bbd3d929499cb749ba57e7b163b9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email accounts' to 'app accounts' within the same utterance, preserving the action, condition, and scope of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/false_start/false_start_03_account_scope/212d8872b82ed50c/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5a8f7b77e7700757fc2ac46dabaa657c190f266ad29aeb99c8f3a600738b2d79", "seconds": 12.12, "spoken_utterances": [ "Remove expired payment cards uh, let me think... from all my email accounts—hmmm... no, that should be all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email accounts' but immediately self-corrects to 'app accounts' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Remove expired payment cards, uh, let me think, from all my email accounts. Hmm. No, that should be all my app accounts that have payment cards." }, "candidate_id": "wav_5ca7a71a96245eb0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage thought process. The initial request is made, followed by a believable 'let me think' filler and a pause, then a self-correction from 'email accounts' to 'app accounts'. The correction is smoothly integrated into the ongoing utterance with a natural prosodic shift, and the final sentence is delivered with a confident, continuous flow. The pacing and intonation suggest genuine spontaneous formulation." } ] }, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Remove payment cards from all my app accounts that have payment cards. Oh, I forgot to mention, only remove expired payment cards.", "long_pause_offsets": [ 71 ], "source_sha256": "436e5045f37c277c933fb38e0f70d4ff1f31d51ceec6473b15a94cf85b3ac0b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the removal to only expired cards, which is consistent with the original instruction's scope and does not contradict it. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/addition/addition_01/856a19192007209f/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "5f66fbae62d3c5f07d54c8f21af08eccaf5c200cbec644afa5349e478fd9c086", "seconds": 13.04, "spoken_utterances": [ "Remove payment cards uh, let's see... from all my app accounts uh... that have payment cards. ", "Oh, I forgot to mention, only remove expired payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a broad instruction to remove payment cards from all accounts that have them. After a pause, they add the constraint 'only remove expired payment cards'. Since this is an addition that narrows the scope of the original instruction, the final intent is to execute the original task with the added constraint. The final audible intent matches the reference instruction, which is to remove expired payment cards from all app accounts that have payment cards.", "heard_transcript": "Remove payment cards, uh, let's see, from all my app accounts, uh, that have payment cards. Oh, I forgot to mention, only remove expired payment cards." }, "candidate_id": "wav_6857e653823d63f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural hesitation ('uh', 'let's see') and a plausible, smoothly integrated revision. The correction ('only remove expired payment cards') is clearly a follow-up thought, not a mechanical insertion, and the prosody reflects a genuine change of mind. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_account_scope", "proposed_instruction": "Remove expired payment cards from only my primary app accounts that have payment cards. Let me correct the scope: remove expired payment cards from all my app accounts, not just the primary ones.", "long_pause_offsets": [ 88 ], "source_sha256": "1232aa2056c23651ca82e3369f2df5c7d1e46b405dbcd45593d96555b75f44e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the scope from 'only my primary app accounts' to 'all my app accounts', which matches the original instruction. The action, target entity, and condition remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/revision/revision_03_account_scope/b113100ba1c7f878/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "ef4616f2e15ae2c92aa272d3c61a608a536a224773be5222cc160d3c82730509", "seconds": 17.84, "spoken_utterances": [ "Remove uh... expired payment cards um... from only my primary app accounts ... that have payment cards. ", "Let me correct the scope: remove expired payment cards from all my app accounts, not just the primary ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially restricts the action to 'primary app accounts' but then explicitly revises the scope to 'all my app accounts, not just the primary ones' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "remove expired payment cards from only my primary app accounts that have payment cards. let me correct the scope, remove expired payment cards from all my app accounts, not just the primary ones." }, "candidate_id": "wav_4e33d1ce65bd913c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural hesitation ('uh', 'um') and a smooth, self-corrected revision. The correction ('not just the primary ones') is integrated into the same utterance with a believable shift in intonation and pacing, showing a realistic thought process. The prosody is consistent, and the flow from initial statement to revision feels spontaneous and complete." } ] }, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "e6e327cbc3b4e16305e1069f4eabed670e19c8ce6665aa3aca5de766e77e69d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/cancellation/cancel_006/b080d592138544c2/speech.wav", "wer": 0.0, "wav_sha256": "0da80eb6abf980d1936219a1dbfbb32a5ea8b8df07f495cbcaf18562974335ca", "seconds": 13.72, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards.", "Hold on, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hold on, please cancel the entire task,' which is a direct and unambiguous cancellation of the previously stated task. The initial part of the audio is a false start or filler, but the final interjection is a complete and explicit cancellation command. According to the rubric, a cancellation must clearly revoke the task, and this does so without requiring execution of the original request.", "heard_transcript": "remove, let me think, expired payment cards from all my app accounts that have payment cards. Hold on, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "f6e8a00d89a830ce9b7d4d19b5e63b560b6ab7b22c0d945c0f583ab00b2c1e15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/ambiguous_halt/halt_026/d523abed0c3db472/speech.wav", "wer": 0.0, "wav_sha256": "c31103b34839c6d1611e0469e477d36c7cfb460b0e649dba13b0dd87cbc01f63", "seconds": 13.4, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The initial request is not canceled or revised, and the speaker does not execute or complete the task. The final control intent is to pause, which aligns with the type-specific criterion.", "heard_transcript": "Remove, let me think, expired payment cards from all my app accounts that have payment cards. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "eccaf1e4ef9bb43e06c5d95cc6779d5e62d6d09fb52f7eda4d00eeadb67756b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/filler/original/f1dde9ca7ecc370e/speech.wav", "wer": 0.0, "wav_sha256": "e2c386b3f4c15664fe8f7c26edb12f58669ffea17b32c16013f8c417a8f38395", "seconds": 8.68, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial hesitation and filler phrase 'let me think' are internal processing cues and do not alter the final task goal. The speaker completes the original instruction without changing the action, target, or constraints. The final intent matches the reference instruction, so the task should be executed.", "heard_transcript": "Remove, let me think, expired payment cards from all my app accounts that have payment cards." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "3f3c139_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_card_status", "proposed_instruction": "Remove active—no, I meant expired payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "6222063573d3ed9a91b76e57ec93133994b1fc05b232974c053e77298034d9f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'active' to 'expired' within the same utterance, preserving the target action, scope of accounts, and the condition of having payment cards." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/false_start/false_start_01_card_status/3d045b0e2888b7cd/speech.wav", "wer": 0.15, "wav_sha256": "d455c38f6d7b3fb1af6ac231fbd7557619cc8cbf8d0f221d40335d5022adfe5a", "seconds": 9.8, "spoken_utterances": [ "Remove active—hmmm... no, I meant expired payment cards uh... from all my app accounts um... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'active' but immediately self-corrects to 'expired' with 'no, I meant expired payment cards'. This is a clear false start where the final corrected value is 'expired'. The rest of the instruction matches the original. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Remove active, no, I meant expired payment cards, uh, from all my app accounts, um, that have payment cards." }, "candidate_id": "wav_0832c444ec17fe3f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a false start ('active'), immediately self-corrects to 'expired payment cards' with a natural-sounding hesitation and intonation shift. The correction is followed by a smooth continuation of the full request. The pacing, filler use ('uh', 'um'), and prosody are highly consistent with spontaneous speech, and the correction feels genuinely integrated into the flow rather than tacked on." } ] }, "is_default": true } }, { "task_id": "3f3c139_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Remove payment cards from all my app accounts that have payment cards. Oh, I forgot to mention, only remove expired payment cards.", "long_pause_offsets": [ 71 ], "source_sha256": "436e5045f37c277c933fb38e0f70d4ff1f31d51ceec6473b15a94cf85b3ac0b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the removal to only expired cards, which is consistent with the original instruction's scope and does not contradict it. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/addition/addition_01/ba14515ecc19b323/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "64cfca7a9078bd6c083e147f876bb3e053d929b4bce0d4c806e0324407790c8a", "seconds": 13.92, "spoken_utterances": [ "Remove payment cards um, let me think... from all my app accounts um... that have payment cards. ", "Oh, I forgot to mention, only remove hmmm... expired payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Remove payment cards from all my app accounts that have payment cards.' After a pause and self-correction, the speaker adds the constraint 'only remove expired payment cards.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Remove payment cards, um, let me think, from all my app accounts, um, that have payment cards. Oh, I forgot to mention, only remove expired payment cards." }, "candidate_id": "wav_1a09e59fbf0084da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with plausible hesitation and self-correction. The initial request is clear, followed by a filler 'um' and a pause, then a revision ('Oh, I forgot to mention...') that smoothly adds a restriction. The prosody is consistent, and the flow reflects a realistic thought process. The correction is integrated naturally into the utterance, making it sound like a spontaneous clarification rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the least—no, make that most in my Spotify song library?", "long_pause_offsets": [], "source_sha256": "ee135a49efa2d2af8384a0568b8cd680797a25a353a86ed32442261b8cd42733", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately corrects it to 'liked the most'. The final intended action is to determine the genre with the most liked songs, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/false_start/false_start_02_aggregation_order/eea4bfdb27020762/speech.wav", "wer": 0.1, "wav_sha256": "29b2701573f21c6a434c77e2d92650aa767f8fa78eed03cf526c505c6e39b84d", "seconds": 11.16, "spoken_utterances": [ "Songs of which ... genre have I um... liked the least—uh... no, make that most in my Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately self-corrects with 'no make that most'. According to the 'false_start' type criterion, the final corrected value ('most') should be applied, and the original incorrect value ('least') should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Songs of which genre have I um liked the least uh no make that most in my Spotify song library?" }, "candidate_id": "wav_252fc6e11d110f38", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody with a believable self-correction. The speaker starts with 'liked the least' and immediately corrects to 'most' with a clear change in intonation and a smooth transition. The filler 'um' is placed naturally, and the overall rhythm mimics genuine spontaneous speech. The correction is integrated into the same utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Songs have I liked the most in my Spotify song library? Oh, I forgot to mention, I want to know by genre.", "long_pause_offsets": [ 56 ], "source_sha256": "592c80e4051b0360d13a4b554be55e9b5521a4dda49945fa27de3b63bfb8c97c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'by genre' to the initial question, clarifying the desired grouping for the comparison. The core action (identifying the most-liked genre) and context (Spotify song library) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/addition/addition_01/9ef43d142aee5130/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "cc579b25df0bbbaf79fd659fe831cddf978de516522e342bf1acd4dcb455ac2a", "seconds": 14.84, "spoken_utterances": [ "Songs have I uh, let's see... liked the most in my Spotify song library? ", "Oh, I forgot to mention, I want to know by uh... genre." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about songs liked the most, then adds the constraint 'by genre' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "Songs have I uh let's see liked the most in my Spotify song library. Oh, I forgot to mention, I want to know by uh genre." }, "candidate_id": "wav_5e3d7dc24aef560d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see') and a self-correction ('songs have I...'), which is followed by a plausible addition of a missing detail ('Oh, I forgot to mention...') and a smooth continuation. The prosody is conversational, with appropriate pauses and filler words that suggest spontaneous thought. The addition feels integrated into the flow rather than tacked on." } ] }, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_grouping_attribute", "proposed_instruction": "Songs of which artist have I liked the most in my Spotify song library? Wait, I meant genre, not artist.", "long_pause_offsets": [ 72 ], "source_sha256": "173ede0c3ef6b04fcbd8d3511b5e48cac5700e9d575a3871b05cc29fe6c85441", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial request from 'artist' to 'genre', preserving the rest of the structure and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/revision/revision_01_grouping_attribute/17eb6bfa17e3eb89/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "b39a8190aebe3060c40e43199c081abb90a40839a542a7075454d49aa4054afd", "seconds": 13.28, "spoken_utterances": [ "Songs of which ah... artist have I ... liked the most in my uh... Spotify song library? ", "Wait, I meant genre, not artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'artist' but immediately corrects themselves to 'genre' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the genre, not the artist, which matches the original reference instruction.", "heard_transcript": "Songs of which uh artist have I liked the most in my uh Spotify song library? Wait, I meant genre not artist." }, "candidate_id": "wav_ed9c9bdc9ae736bd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('artist'), self-corrects to 'genre' with a natural-sounding hesitation and intonation shift, and continues the sentence smoothly. The correction is integrated into the flow without awkwardness, and the overall prosody sounds like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "f691939bf8b9414b16a088270febc9360ca78e9d80dc533d5110f8b570c27b77", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/cancellation/cancel_011/818b5210ff7f1572/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "359474e8d86e5cee8b4b28debb5775926c2bea21a2bd237b788fee2355bbd773", "seconds": 14.12, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?", "Hold on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question about liked genres in their Spotify library. After a brief pause and the phrase 'Hold on, I've changed my mind,' they explicitly state 'Cancel that.' This constitutes a clear cancellation of the initial task. The final intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library? Hold on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "6716a76d5ecd5486635fca476c14b1432a95c49b523b4b1c83b15951019f2926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/ambiguous_halt/halt_010/410943e799ea9864/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "9af26c97313c213ebfd694f8aa36f3a650a5a5f33a477b1fe69f6174eb4268e9", "seconds": 12.24, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled, but the speaker indicates a need to pause before proceeding.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library? Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "063f534b7c66a106fca4dcb9a81aa6c6876f9bcaabffdda36c2a7499d1b0654f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/filler/original/e4db8d33b59b7174/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "405d9217d6680f16590c2ecbcd1d0aed006518ed3c3e09d9e1b3a487b29768af", "seconds": 8.88, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Songs of which um genre...' and then continues the original instruction without any correction or change in intent. The 'um' is a filler word and does not alter the task goal. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the least in my Spotify playlist—or rather, album library?", "long_pause_offsets": [], "source_sha256": "dec6df954f7d20491961585e15ec33d42d68dd31f4c33a3d27c4d7759c39d7f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects it to 'album library' using 'or rather', preserving the rest of the query including the app (Spotify), the action (finding the genre liked the least), and the scope (library)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/false_start/false_start_04_source_collection/e0a419e938614e46/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "9363de1402175fb4428745a5ca44e9ef246bc810c75b182307ebe91cf3566172", "seconds": 10.44, "spoken_utterances": [ "Songs of which genre have I liked the least in my hmmm... Spotify playlist—uh... or rather, album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects to 'album library' with 'Or rather'. This is a clear false start where the final, corrected target is 'album library'. The rest of the instruction matches the original. The final intent is to query the genre with the least liked songs in the album library, which aligns with the original instruction.", "heard_transcript": "Songs of which genre have I liked the least in my Spotify playlist? Or rather, album library?" }, "candidate_id": "wav_3d32ea07b30d49a9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear request, then self-corrects from 'playlist' to 'album library' with a believable 'uh' and 'or rather' transition. The prosody is fluid, with a natural hesitation and a smooth, pragmatic correction that sounds like a genuine moment of recall. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Songs have I liked the least in my Spotify album library? Oh, I forgot to mention, I want to know which genre I've liked the least.", "long_pause_offsets": [ 58 ], "source_sha256": "fb090e7a0f0c6fcdf40339157d7d2490fef63f99e59d02d24ec4031c082162b8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'genre' specification to the initial question, clarifying that the user wants to know the genre associated with the least-liked songs. The core intent—identifying the least-liked genre—is preserved and completed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/addition/addition_01/7988c545725234a1/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "f7c0cbb83573643481426487f2c575aa9ce93ffe5b1a2bdc3545ea8b7320fba4", "seconds": 13.28, "spoken_utterances": [ "Songs have I liked the least in my ... Spotify ... album library? ", "Oh, I forgot to mention, I want to know which genre I've liked the least." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Songs of which genre have I liked the least in my Spotify album library?'. This is followed by a longer pause and an addition: 'Oh, I forgot to mention, I want to know which genre I've liked the least.' The addition clarifies and refines the original request, making it more direct and removing the slightly awkward phrasing 'Songs of which genre'. The final intent is clearly to find the genre liked the least, which matches the original instruction's goal. The addition does not introduce any conflicting or contradictory information. Therefore, the final intent is consistent with the original instruction and the addition serves to clarify it, satisfying the 'addition' type criterion.", "heard_transcript": "Songs of which genre have I liked the least in my Spotify album library? Oh, I forgot to mention, I want to know which genre I've liked the least." }, "candidate_id": "wav_df74918a7ad3012d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('Songs have I liked the least in my Spotify album library'), then naturally adds a clarification ('Oh, I forgot to mention...') with a believable filler and a smooth, additive correction. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition feels like a genuine memory or oversight, not a rehearsed or robotic insertion." } ] }, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Songs of which genre have I liked the least in my Apple Music album library? Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "25f261a0b3c00e607dc0739809ef218886aff9f5fc38a718417f4b0816b107ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, preserving the rest of the query including the action, scope, and condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/revision/revision_03_app/3f93ccc20de5a9e1/speech.wav", "wer": 0.04, "wav_sha256": "df1f632b85360a9481fba118e6d0482b7c80ad39c17f1f60eb3e87c3393926c7", "seconds": 13.84, "spoken_utterances": [ "Songs of ... which genre have I liked the least in my hmmm... Apple Music album library? ", "Wait, the app should be Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Songs of which genre have I liked the least in my Apple Music album library? Wait, the app should be Spotify, not Apple Music." }, "candidate_id": "wav_b5be0fc82970dcec", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then hesitates with a realistic 'um' before correcting the app name from 'Apple Music' to 'Spotify'. The correction is delivered with a believable shift in intonation and pacing, sounding like a spontaneous realization rather than a rehearsed edit. The flow is continuous and the self-correction feels integrated into the thought process." } ] }, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "a87512bd6415c70502f13164a0f1d7fe3b662158793bb70cf87c04e98a23e753", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/cancellation/cancel_024/a2974420db46987e/speech.wav", "wer": 0.0, "wav_sha256": "d5ef4e744836defcdbccfb93452f0beb8fe0bbeefaeea293aef72bd7768181ca", "seconds": 13.08, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?", "Hold on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly states 'Hold on, oh never mind, cancel the whole thing' as the final interjection, which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Songs of uh which genre have I liked the least in my um Spotify album library? Hold on, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "12db936dd09889e079c17b6f0e2dd9399e82a79ebc135cad888e2dc5b2dd92b5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/ambiguous_halt/halt_005/c92a00fb1cc0af83/speech.wav", "wer": 0.0, "wav_sha256": "38682a5f3a95bde19e22ca196602b70d41298d21f67e7b8c476121af710b9ec2", "seconds": 10.92, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'One second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, but the speaker explicitly halts the process.", "heard_transcript": "Songs of uh, which genre have I liked the least in my um, Spotify album library? One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e98bfb3414545339d56d8a884cab320298d43ca823c30fff33f1f278d9f9fa1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/filler/original/66baf2865b78e133/speech.wav", "wer": 0.0, "wav_sha256": "50843cd29ddaa805aaa094ec6f3a83b048fc1cba324afdf9c985cd6ba4de8694", "seconds": 7.76, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Songs of uh' and 'in my um', which are fillers. The core instruction is clearly stated as 'which genre have I liked the least in my Spotify album library?'. The final intent matches the original instruction, and the fillers do not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Songs of uh, which genre have I liked the least in my um, Spotify album library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify albums—or rather, playlists?", "long_pause_offsets": [], "source_sha256": "8dde9977084aab9278d963a249597c6b47e3039550c237d15e7f7e3f63b76131", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'albums' to 'playlists' within the same utterance, preserving the rest of the query including the app (Spotify), the action (finding liked songs), and the scope (genre)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/false_start/false_start_04_source_collection/2b35a51415b5b8b1/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "362f127bff3c3d780e4cecd075b176732a68c5a19bf47c705527be338895e999", "seconds": 8.24, "spoken_utterances": [ "Songs of which genre have I uh... liked the most in my Spotify ... albums—ah... or rather, playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with 'uh rather playlists'. This is a false start where the final, corrected target is 'playlists'. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final intent is to query liked genres in Spotify playlists.", "heard_transcript": "Songs of which genre have I uh liked the most in my Spotify albums, uh rather playlists." }, "candidate_id": "wav_f02fb929d916e6a6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding self-correction from 'albums' to 'playlists' with a clear, spontaneous 'uh' and 'rather' as a connector. The prosody is consistent, the pace is conversational, and the repair feels like a genuine moment of recall rather than a rehearsed edit. The intonation and rhythm are the most fluid and human-sounding among all clips." } ] }, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Songs of which genre have I in my Spotify playlists? I left out a detail: find the genre I've liked the most.", "long_pause_offsets": [ 53 ], "source_sha256": "ddc310e3965182fc9b1531007b64bdd90b0f0cf1bf1a1f0c255c97d8d28ca9f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'liked the most' to the initial incomplete question, completing the intended goal. The initial utterance is grammatically incorrect but the final goal is unambiguously reconstructed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/addition/addition_02/521147776fa6a704/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "8d6d53513851f1ebc4e756a48b88daa0ad8cf64b4d81c9c7e30b3a1339e6f530", "seconds": 13.6, "spoken_utterances": [ "Songs of which genre have I in my um, let's see... Spotify playlists? ", "I left out a detail: find the genre I've uh... liked the most." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Songs of which genre have I and my, um, let's see, Spotify playlists?' which is incomplete and unclear. The speaker then adds 'I left out a detail. Find the genre I've uh liked the most.' This addition clarifies and completes the original request. The final intent is to identify the genre of songs the user has liked the most in their Spotify playlists, which matches the original instruction. The addition corrects and completes the initial incomplete thought, satisfying the addition type criterion.", "heard_transcript": "Songs of which genre have I and my, um, let's see, Spotify playlists? I left out a detail. Find the genre I've uh liked the most." }, "candidate_id": "wav_b1b9c7d151ed2439", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('Songs of which genre have I and my...') that transitions into a plausible self-correction ('let's see... I left out a detail'). The addition of 'find the genre I've uh like the most' flows naturally from the prior hesitation, maintaining a consistent rhythm and intonation. The pauses and fillers are well-placed and sound spontaneous, contributing to a realistic, conversational delivery." } ] }, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Songs of which genre have I liked the most in my Apple Music playlists? Let me correct the app: look at Spotify playlists, not Apple Music.", "long_pause_offsets": [ 72 ], "source_sha256": "a1463cce8df18ed1396bba23080b9694141fcedd3b32d982e8514df576ded8d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, preserving the original intent of identifying the genre with the most liked songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/revision/revision_03_app/60970cc44801c9ab/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "de14826d0365f7f6694018df455a1c27dce154e07e9f2ff7c05cda1f0750bba7", "seconds": 18.72, "spoken_utterances": [ "Songs of which uh, let me think... genre have I ... liked the most in my Apple Music playlists? ", "Let me correct the app: look at hmmm... Spotify playlists, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music playlists' but immediately corrects to 'Spotify playlists, not Apple Music' after a pause. This is a revision where the final intent matches the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Songs of which uh let me think genre have I liked the most in my Apple Music playlists? Let me correct the app, look at um Spotify playlists, not Apple Music." }, "candidate_id": "wav_e3a522ca5a7b69c7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker begins with a hesitant, self-correcting thought process ('uh, let me think'), then makes a clear, spontaneous correction from 'Apple Music' to 'Spotify' with a natural pause and intonation shift. The prosody is continuous and human-like, with believable fillers and a smooth transition into the corrected request. The revision is integrated into the same utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "a4af815b0b514c68e8b947dc2c2e5536cfe7850755a45462db5f4a2aa35f3ef6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/cancellation/cancel_010/3215b20af115f662/speech.wav", "wer": 0.08, "wav_sha256": "2b5210ec2bdbdcf3fe0150ded4a92b9427f19c85363959788abad6ad10b07915", "seconds": 12.6, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?", "Hold on, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question about liked genres in Spotify playlists, then immediately follows with 'Hold on, cancel the request I just made.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and not followed by any resumption or new instruction.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists? Hold on, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "d18d76ccbb83be32c8bcf18e71ac4506f8f89794f6f95293d2f58daee5728643", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/ambiguous_halt/halt_040/6b2aa2c1b29e7320/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "a98c24c601e2b344c3d18d3c9976d67bf4dcf6d5936c9583ba1821b794670e84", "seconds": 11.8, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Hang on, just a minute,' which is a clear request to pause. The final interjection is a pause directive, not a cancellation or a new instruction. The original task remains unchanged and unexecuted, which is consistent with the 'ambiguous_halt' type. No issues are present.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists? Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "425a494_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f24606e86b966cb7a88f8a62a3bfb507a2cd844b91b51f3543d7e9c14594718c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/filler/original/5c5f8b621d4f512b/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "bc8437deb7fa4c526b17e7fdd98660ba81034b56d9738f1a792956bf0b5f2d2e", "seconds": 8.0, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') and pauses but does not change the core request. The final intent remains the same as the original instruction, and the fillers do not constitute a directive to pause or cancel. Therefore, the task should be executed.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Tell me the card name I used for my last PayPal—I meant amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "ac77fa7d408cd8333dcbd6ae876788b24a3aa48831bcebf6dd17ecb7d6ab5769", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects it to 'Amazon Prime membership payment' within the same utterance. The final intended service is Amazon Prime, not PayPal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/false_start/false_start_02_app/c157a9f524b3070b/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "4d7b7d619d77b432a41f2f5fa83e484deaddcd27549cb2d9584dc0a24b637821", "seconds": 11.08, "spoken_utterances": [ "Tell me the let me think... card name I used for my um... last PayPal—ah... I meant amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Amazon Prime membership payment'. According to the 'false_start' type criterion, the final corrected intent should be executed, and the initial mistaken value should be discarded. The final audible task matches the original instruction exactly.", "heard_transcript": "Tell me the, let me think, card name I used for my um, last PayPal, I meant Amazon Prime membership payment." }, "candidate_id": "wav_0e270edcd4ada749", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start. The speaker begins with 'tell me the' and then inserts 'let me think' as a believable hesitation filler. The initial request is 'card name I used for my um last PayPal'—a plausible but incorrect memory. The self-correction 'I meant Amazon Prime membership payment' is smoothly integrated with a natural-sounding 'I meant' phrase, and the correction flows directly from the prior error. The prosody is consistent, and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Tell me the card name I used for my amazon prime membership payment? Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment.", "long_pause_offsets": [ 69 ], "source_sha256": "8dbf5c958484556af645b6be64552894bcd92739f0764c99846d8d39184b375e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the constraint 'most recent', which refines the original request without contradicting it. The core action (retrieve card name) and context (Amazon Prime membership payment) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/addition/addition_01/6fd1c19f5c6df50e/speech.wav", "wer": 0.08108108108108109, "wav_sha256": "f04d451a594eb9605049f6a86219402757233ecc1aff37007f9f84688dbdb879", "seconds": 16.16, "spoken_utterances": [ "Tell me the um... card name I used for my uh, let me think... amazon prime membership payment? ", "Oh, I forgot to mention, I want the card name for my hmmm... most recent Amazon Prime membership payment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete and hesitant, but the speaker adds the clarifying phrase 'most recent' after a pause and self-correction. This addition refines the original instruction to specify recency, which is a narrowing of the scope. The final audible intent matches the addition type criterion: the original request is combined with the added information to form the complete goal.", "heard_transcript": "Tell me the um card name I used for my uh let me think. Amazon Prime membership payment. Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment." }, "candidate_id": "wav_9c95f7dc76e5da41", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with a believable 'let me think' filler, followed by a self-repair ('Oh, I forgot to mention') and a smooth continuation that adds the missing detail ('most recent Amazon Prime membership payment'). The prosody includes hesitations and intonation shifts consistent with spontaneous recall, and the addition feels integrated rather than tacked on." } ] }, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last amazon basic membership payment? I misspoke about the subscription type: it should be prime membership, not basic membership.", "long_pause_offsets": [ 74 ], "source_sha256": "473a7e3cd2ab0da55d2f75f2bff3e4e583997a10b2a46e0ad4d26bbdc5e1f652", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the subscription type from 'basic' to 'prime', which is the only change from the original instruction. The action, entity (Amazon), and the request for the card name remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/revision/revision_03_subscription_type/1d6392454551cb12/speech.wav", "wer": 0.03125, "wav_sha256": "1e7c99b89f2a164a501f99a92c35951c4d17f07c656c4e57fef0a892534f5858", "seconds": 14.64, "spoken_utterances": [ "Tell me the card name I used for my last um... amazon basic membership payment? ", "I misspoke about the subscription type: it should be um, let's see... prime membership, not basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon basic membership payment' but then corrects themselves to 'Prime membership, not basic membership' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to retrieve the card name for the 'Prime membership' payment, which matches the original reference instruction.", "heard_transcript": "Tell me the card name I used for my last um Amazon basic membership payment. I misspoke about the subscription type, it should be um let's see, Prime membership, not basic membership." }, "candidate_id": "wav_3754d48da7dd241c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, then hesitates with 'I misspoke' before self-correcting 'subscription type' from 'basic' to 'Prime membership'. The correction is integrated smoothly into the same utterance with a natural 'let's see' filler, and the prosody reflects genuine recollection and correction without robotic pacing. The flow is continuous and the intonation matches a real-time thought process." } ] }, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "03f0f4d2b59a3998f05ec5c9e8a73cc42a2da08aadbac1c3c53da55d2e8c2766", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/cancellation/cancel_021/f36da7be43c4297e/speech.wav", "wer": 0.04, "wav_sha256": "71832664bb0685b88469d225ed7cc0aca353e343660cef9ccd6897603ab20b60", "seconds": 11.76, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly revokes it with 'Hold on. Never mind. I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is a complete cancellation, not a pause or a request to wait.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment. Hold on. Never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "39be7443c28020e4713ceb3e0d435d8f6002d88271d6b4cae1d232f19448ad25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/ambiguous_halt/halt_008/8ec8950fa84ccce0/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "bc5cca8dd8ab055df30102bdbc5f898369176bf8111c53b8f992b208706b10bf", "seconds": 10.08, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then clearly interjects 'Wait a second' as a final interjection, indicating a temporary halt. There is no cancellation or execution of the task. The final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment? Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c3636fc0150f66b5ad086cc2631a517afe375faaec4fefc911a7c27bfc062f0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/filler/original/e4d2fd528b467ec4/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "8317943fd3fae5bf26bba8620e99bc7ba17d616e9a04c5baaed17905a73c2c8f", "seconds": 6.96, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the filler type allows for such hesitations without altering the task goal.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last spotify free trial—sorry, make that premium membership payment?", "long_pause_offsets": [], "source_sha256": "22891cbdcf3638077d86a39b36ad88b5151179e8328e4e136032c79383da7a57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'free trial' but immediately corrects it to 'premium membership payment'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/false_start/false_start_03_subscription_type/d9efbd39452d3c22/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b746251565e14c17f0089f8f497b3a4ce45bb777aa32025e63d48788f3be4d2e", "seconds": 12.68, "spoken_utterances": [ "Tell me the um... card name I used for uh... my last spotify free trial—hmmm... sorry, make that premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify free trial' but immediately self-corrects to 'premium membership payment'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Tell me the um card name I used for uh my last Spotify free trial. Hmm. Sorry, make that premium membership payment." }, "candidate_id": "wav_4c6ea933f0a1483e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'Spotify free trial', hesitates with 'hmm', then clearly self-corrects to 'premium membership payment'. The correction is integrated smoothly into the sentence with a natural 'sorry' and a clear rephrasing. The prosody, including the pause and filler, sounds spontaneous and not rehearsed." } ] }, "is_default": true } }, { "task_id": "476b213_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Tell me the card name I used for my last apple music premium membership payment? Make that 'spotify', not 'apple music'.", "long_pause_offsets": [ 81 ], "source_sha256": "b9ba8cd5e76a8eb0b2d3fbedd68115ce73ac1a3c81d10ac9d0978aa60fa91cfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from 'apple music' to 'spotify', preserving the action, the context of the last premium membership payment, and the requested information about the card name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/revision/revision_02_app/dfd38adc164657df/speech.wav", "wer": 0.0, "wav_sha256": "9ad7e00f02a4e2e54fef30a73182e751a07b3292ab50ce8e4478529b39cecada", "seconds": 13.68, "spoken_utterances": [ "Tell me the um... card name I used for ... my last apple music premium membership payment? ", "Make that '... spotify', not 'apple music'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' with 'Make that Spotify, not Apple Music.' This is a clear revision where the final intent is to query the card name for the Spotify payment. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Tell me the um card name I used for my last Apple Music premium membership payment. Make that Spotify, not Apple Music." }, "candidate_id": "wav_dbd033f958830451", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "476b213_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "c174ff964e267efef1e42d6b6d1365a6c12b0b216371f4da1ca5cdccb080a0bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/cancellation/cancel_034/7c9d8c3b03c8396d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b8dd34017de0ebe9143f4c7bc817efd97f8706a8785d9a97390afd312f994a77", "seconds": 10.32, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?", "Wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking for information but then clearly revokes the entire request with 'Wait, let's drop the whole task.' This constitutes a direct cancellation of the original task. The final intent is to cancel, not to execute or pause, satisfying the cancellation type criterion.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify Premium membership payment. Wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "2822eb6b12c92427a8244b8f778e96738a01cb463b8d75d280391c73cac136af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/ambiguous_halt/halt_038/016818575c8c9e82/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "0d9585f68ae420943822e3b233def649ce5a90877290358f1595447d0eb12d79", "seconds": 10.16, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original request and then makes a clear, final interjection to pause the process. The final audible intent is to hold the task, not to execute it or cancel it. This matches the 'ambiguous_halt' type criterion, which requires a final request to wait or pause without cancellation.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify premium membership payment. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f5f6ed3406ee6fbaec9b8e11f1a4cf0e173db9cf5578fd5dec5fb953db852160", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/filler/original/a49983dbe84f3aac/speech.wav", "wer": 0.0625, "wav_sha256": "25d26aa26010a40290bdb6b5142ab8070e8e56db915ffed4d230a684908686ea", "seconds": 5.68, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify premium membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last—no, I mean last amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "bdbac8ea9465e87ba7e8066eeb84aba0873f5d4aeb1381258b3581e85df9a366", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'second-to-last' but immediately corrects to 'last' within the same utterance. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/false_start/false_start_01_payment_recency/6231e3b24b188604/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "fa357ac95fe4751730590ae06278344af0dcf53365a1a84394e7ca3089a9abaa", "seconds": 12.72, "spoken_utterances": [ "Tell me the ah... card name I used for my second-to-last—um... no, I mean ... last amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'second to last' but immediately self-corrects to 'last' after a pause and a 'no, I mean' repair. According to the 'false_start' type criterion, the final corrected value ('last') should be applied, and the initial incorrect value ('second to last') should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Tell me the uh card name I used for my second to last um, no, I mean last Amazon Prime membership payment?" }, "candidate_id": "wav_e4fa0af1b08a264c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible false start. The initial 'uh' and 'second to last' are followed by a clear self-correction with 'no, I mean last' that sounds like a genuine moment of recollection. The prosody is consistent, the pace is conversational, and the correction flows smoothly within the same utterance without any robotic or disjointed feel." } ] }, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Tell me the card name I used for my last payment? I should also mention, this is specifically for my Amazon Prime membership.", "long_pause_offsets": [ 50 ], "source_sha256": "787e4ed721a9520aa1ccc61bf38cbd0ad6ac36ad1439a049dc47a1ec8c975e95", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the specific context that the payment was for an Amazon Prime membership, which refines the initial broad request about the 'last payment'. The final goal combines both the action and the specific scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/addition/addition_02/28fbd888439765a9/speech.wav", "wer": 0.08, "wav_sha256": "2dd10da7c7c697caa35945b039cf21a94e8ca98ab72036a8de3be8f1ceafc10d", "seconds": 13.48, "spoken_utterances": [ "Tell me the card name I used for my um... last payment? ", "I should also mention, this is specifically for my ah... Amazon Prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'Tell me the card name I used for my last payment.' This is followed by an addition specifying 'this is specifically for my Amazon Prime membership.' The addition clarifies the scope of the payment, narrowing the initial broad request. The final intent combines the original request with the added constraint, matching the reference instruction. The type is 'addition', and the final goal is fully aligned with the reference.", "heard_transcript": "Tell me the card name I used for my um last payment. I should also mention this is specifically for my uh Amazon Prime membership." }, "candidate_id": "wav_c2402210116ad60d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a slightly broken flow ('for my um...'), which is followed by a clear, additive clarification ('I should also mention...') that is smoothly integrated into the sentence. The prosody is conversational, with a believable pause and filler before the second clause, and the intonation suggests genuine recollection rather than a rehearsed or robotic delivery. The addition feels like a spontaneous memory." } ] }, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last amazon prime membership payment? Make that the card name for my last amazon prime membership payment, not second-to-last.", "long_pause_offsets": [ 84 ], "source_sha256": "de325511ac7b3cf345e65e2b056ac30a2a178292dbdd489fe1d85dd0698ceb9e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'second-to-last' to 'last', preserving the app (Amazon Prime), the requested information (card name), and the action (tell me)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/revision/revision_01_payment_recency/98bb03af87382838/speech.wav", "wer": 0.0, "wav_sha256": "fa05868dc2da530e0fa0f021c67ea66d1424e193593f48d57d234b872bd2c735", "seconds": 17.2, "spoken_utterances": [ "Tell me the card name I used for my uh... second-to-last amazon prime membership payment? ", "Make that the card name for my um... last amazon prime membership payment, not second-to-last." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'second to last' but immediately corrects themselves to 'last' after a pause, which is the final and intended instruction. The revision is clear and unambiguous, and the final goal matches the original reference instruction. The correction is applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "Tell me the card name I used for my uh second to last Amazon Prime membership payment. Make that the card name for my um last Amazon Prime membership payment, not second to last." }, "candidate_id": "wav_775b917fbc112cde", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding 'uh' and 'um', then makes a clear revision from 'second to last' to 'last' with a smooth self-correction. The prosody is consistent, the pace is conversational, and the revision feels like a genuine moment of recollection rather than a rehearsed correction. The intonation and rhythm are very natural." } ] }, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "d880d12166a9bec4d41321b248ac542fa13ee6d49d218ae4bab04e00f4f114b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/cancellation/cancel_019/339b11d9e0d86095/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "5cd804af27641cfc00d740e8cb6d2cc3c7ce6dbfb2a44fc60168c0eecc906008", "seconds": 11.56, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?", "Wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the original request but immediately self-cancels with 'Wait, never mind. Don't do that.' This constitutes a clear cancellation of the entire task as per the 'cancellation' type criterion. The final interjection explicitly revokes the initial request, and no other task is initiated.", "heard_transcript": "Tell me the card name I used for my last Amazon Prime membership payment. Wait, never mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "e8beef3d69b37c716782d9144b6acc9027aad08850a2bf39b17714befa6f1aa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/ambiguous_halt/halt_018/345326ac443beb7d/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "a38db98f4d731f33b4d453f2f83b7d2109ee427766811344f9c2e00c9559016c", "seconds": 10.76, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which is a clear instruction to hold the current task. The initial request to 'Tell me the card name...' is not canceled, but the final interjection is a valid pause command. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Tell me the card name I used for my uh last Amazon Prime membership payment. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "476b213_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7c70ae27e55d3190cff1dabf3b5d36ab9f303e829cc4e3e3da1e1ccbc54e3f4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/filler/original/779d658680d67737/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "1d661f6e63eefce5ac18e85b9bf21503906c48a3c101a591ac835e2bf44d2713", "seconds": 6.68, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent is to retrieve the card name for the last Amazon Prime payment, which matches the original instruction. The filler does not alter the task goal.", "heard_transcript": "Tell me the card name I used for my uh last Amazon Prime membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_email_state", "proposed_instruction": "Delete all my sent—sorry, I meant archived gmail threads that are from before this calendar month.", "long_pause_offsets": [], "source_sha256": "e31b227ff52caea589865d108b091b6b5261b4acae352b6242bd59395560d4fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'archived' within the same utterance, preserving the app (Gmail), action (delete), scope (all), and time condition (before this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/false_start/false_start_01_email_state/15ca6031f0d944ff/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "e021aef9b822134ac2917c7211091d2e9de3d21d9e54ff4dc57cc7da2d2be58a", "seconds": 6.8, "spoken_utterances": [ "Delete all my sent—um... sorry, I meant archived gmail threads that are ... from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'archive'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final task matches the original instruction's intent, with only the action changed from 'delete' to 'archive'. The target ('Gmail threads') and condition ('from before this calendar month') are preserved. Therefore, the final intent is correctly executed.", "heard_transcript": "Delete all my sent um, sorry, I meant archive Gmail threads that are from before this calendar month." }, "candidate_id": "wav_cbdd49596e8897e0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('delete all my sent them'), immediately self-corrects with a natural-sounding 'sorry, I meant', and smoothly continues with the full, coherent request. The prosody is consistent, the repair is integrated into the flow without awkwardness, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Delete all my archived gmail threads that are. Oh, and one more thing: only delete threads from before this calendar month.", "long_pause_offsets": [ 47 ], "source_sha256": "d03cf5e52808628049c46a7da406fd7720b4b8e2e85db35354dae559896c489a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time condition 'from before this calendar month', which refines the incomplete initial request. The action, target app, and archival status remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/addition/addition_03/d4f24eaecffdc547/speech.wav", "wer": 0.2692307692307692, "wav_sha256": "e59dcb7aadd53efd76de617f93fb90a29c7a2c42f2972bda98ed15cc113c0f0b", "seconds": 13.48, "spoken_utterances": [ "Delete all my um... archived gmail threads that are. ", "Oh, and one more thing: only delete threads from before um, let me think... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Delete all my archived Gmail threads that are from' and then adds 'only delete threads from before this calendar month' after a pause and self-interruption. This matches the 'addition' type, where the final goal combines the initial request with the added constraint. The final intent is clear and matches the original instruction.", "heard_transcript": "Delete all my um archived Gmail threads that are from oh and one more thing only delete threads from before um let me think this calendar month." }, "candidate_id": "wav_55fde216cc129930", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_email_state", "proposed_instruction": "Delete all my sent gmail threads that are from before this calendar month. Wait, delete all my archived gmail threads that are from before this calendar month.", "long_pause_offsets": [ 75 ], "source_sha256": "a2d8a3288499d6dc874efe9479a1a6f7d4b1eecdb6cbaa9b9b26d478d2600bb3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the thread category from 'sent' to 'archived', while preserving the app (Gmail), action (delete), and time condition ('before this calendar month')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/revision/revision_01_email_state/4461cba76fbad9eb/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "3ac4b8cd43ca5f6c918852f697abfa964e7b4a338bcc62cce12ad2b87bb03ba6", "seconds": 13.84, "spoken_utterances": [ "Delete all my hmmm... sent gmail threads that are let's see... from before this calendar month. ", "Wait, delete all my archived gmail threads that are from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent Gmail threads' but immediately corrects to 'archived Gmail threads' after a pause and self-correction ('Wait, delete all my archived...'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Delete all my hmm, sent Gmail threads that are let's see, from before this calendar month. Wait, delete all my archived Gmail threads that are from before this calendar month." }, "candidate_id": "wav_05c5760c0213d6ea", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. It begins with a false start ('Sent'), self-corrects to 'Archive', then pauses and formulates a new request ('Wait, delete all my archived Gmail threads...'). The correction is integrated into a single, continuous utterance with a believable change-of-mind intonation and a clear reference back to the initial idea. The 'Hmm' and 'let me see' are placed naturally, and the overall prosody is fluid and human." } ] }, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "de5c2a9aa4fe56f1fe2ff72a9eff7c554e8b1c95b331c3bcac50ee32e0bd2584", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/cancellation/cancel_024/e53028960ccecee7/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "f421dfb5f819fc8e977a914a00ea9accaaac3c19379cff7036f2b6cf3c59ff26", "seconds": 12.48, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month.", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'hang on, oh never mind, cancel the whole thing.' This constitutes a final, unambiguous cancellation of the task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "delete all my um archived gmail threads that are um let's see from before this calendar month. hang on, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "178dc546f052a3cf68b23379301cbdcb176a64496410418be76d6365e91f4463", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/ambiguous_halt/halt_002/4a9163fefdd63e86/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "132a5a8c523788dd2906c346768a9a5301a88f5318b78e9334869d46ab03585e", "seconds": 10.52, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month.", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'hold on,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there a new instruction to resume. The final interjection is a pause command, consistent with the 'ambiguous_halt' type. The initial request is not executed, but the final intent is to halt, not to cancel or execute.", "heard_transcript": "delete all my um archived gmail threads that are um let's see from before this calendar month. hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7657680ec4f5dc6c028a853ba02a352fd074a2b161dafe4010dca5bab3467c2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/filler/original/1c0f6167e45c1fbc/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "82b60245ddc8d0638ac51052edb788de8c8a825748dff690047b454e43d5dddf", "seconds": 7.48, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final intent matches the original instruction, with no correction or cancellation. The filler words do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Delete all my um archived Gmail threads that are um let's see from before this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from last calendar month—no, that should be this calendar month.", "long_pause_offsets": [], "source_sha256": "bc5ed21658f05efbaa646e8d8482627c35fb00eb8f3d4b3b46b2765ffb42df73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'this calendar month'. The final intended time filter is 'this calendar month', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/false_start/false_start_03_time_filter/e3c62b915b8a4118/speech.wav", "wer": 0.1, "wav_sha256": "7674ab4805c17e30e192c6cafa80e6c4fbbda1c482243054430df4fd53403381", "seconds": 11.96, "spoken_utterances": [ "Delete all my archived ... gmail threads that are from last calendar month—hmmm... no, that should be ... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately self-corrects to 'this calendar month'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final audible instruction matches the original reference instruction, so the task is executed as intended.", "heard_transcript": "Delete all my archived Gmail threads that are from last calendar month. No, that should be this calendar month." }, "candidate_id": "wav_068e6f4b497d5e32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'last calendar month', then immediately self-corrects to 'this calendar month' with a clear, spontaneous-sounding 'No, that should be...' followed by a smooth rephrasing. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels like genuine real-time thought processing." } ] }, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Delete all my archived threads that are from this calendar month. Oh, right, this is specifically for my Gmail threads.", "long_pause_offsets": [ 66 ], "source_sha256": "9d9b5b5041604f01acbdc91aedacc3b10b6f85371a43199bcd2866f38d7aef70", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the initial 'threads' refer specifically to Gmail threads, which matches the original instruction's app context. The action, scope ('all'), archive status, and date condition remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/addition/addition_02/6b842c656d800592/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "85f601a76e2966300accd36a28bb8053d69e5340cbdcf5261850ed713b6b1931", "seconds": 14.44, "spoken_utterances": [ "Delete all my uh... archived threads that are from this calendar month. ", "Oh, right, this is specifically for my let me think... Gmail threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'delete all my archived threads that are from this calendar month'. After a pause and self-correction, the speaker adds 'this is specifically for my gmail threads'. This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "delete all my uh archived threads that are from this calendar month. oh right, this is specifically for my, let me think, gmail threads." }, "candidate_id": "wav_7c44980bfb40326f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural, slightly hesitant 'uh' and a believable self-correction ('Oh right, this is specifically for my... let me think... Gmail threads'). The 'let me think' filler is placed at a plausible moment of cognitive processing, and the correction flows smoothly without awkwardness. The prosody is consistent, and the overall delivery sounds like a spontaneous, real-time formulation of a task." } ] }, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Delete all my archived outlook threads that are from this calendar month. Let me correct the app: delete all my archived threads from Outlook, not Gmail.", "long_pause_offsets": [ 74 ], "source_sha256": "652c65c9224939bdd8a15023fcceed97c282063d8e8b65c4fd5103d50f2cbb2d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to Outlook, while retaining the action, archive status, and calendar month condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/revision/revision_02_app/36adec2b85f8c725/speech.wav", "wer": 0.09375, "wav_sha256": "9e4dad255edd69147cb870fa878a4964a82e5a52b4a24c003f36bd6a745f74c4", "seconds": 18.32, "spoken_utterances": [ "Delete all my archived uh... outlook threads that are from um, let me think... this calendar month. ", "Let me correct the app: delete all my archived threads from Outlook, not uh... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'gmail' but immediately corrects to 'Outlook' after a pause, which is a revision. The final instruction matches the original intent of deleting archived threads from the current calendar month, but now correctly specifies the app as Outlook instead of Gmail. The revision is clear and final, so the task should be executed with the corrected target.", "heard_transcript": "Delete all my archived uh Outlook threads that are from um, let me think, this calendar month. Let me correct the app. Delete all my archived threads from Outlook, not uh Gmail." }, "candidate_id": "wav_351031ffe7ee36b9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a clear request but then introduces a significant error ('Outlook threads') and a long, unnatural pause ('let me think'). The correction ('Let me correct the app') is awkward and sounds like a meta-comment rather than a spontaneous self-repair. The final correction ('delete all my archived threads from Outlook, not Gmail') is delivered with a flat, robotic tone and lacks the natural prosody of the others. The overall delivery feels less like a spontaneous thought process and more like a rehearsed or synthesized correction." } ] }, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_020", "hide_variant_original": true, "source_sha256": "792dc5b968d8d0c63d605efdea97ed26068f4b047eb104e343433c51c45ede46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/cancellation/cancel_020/e82f0887963cd3d6/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "82e6c1421fbcb02624c4e55c60a293c171ce5eae3207e601088a2fef2add6517", "seconds": 12.0, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month.", "Hang on, never mind, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete request, then immediately follows with 'Hang on. Never mind. Cancel that request.', which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the prior instruction, satisfying the cancellation type criterion.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month. Hang on. Never mind. Cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "bd955588bcbc1a1dd834c3b6b98c4b3609f5adea2920485d12473856b6f9f9e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/ambiguous_halt/halt_028/c8af004c540dd40e/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "bb4732baf97ded1525201c6bb1406188357ad83e38efcadfa73d2a9b8c5194a4", "seconds": 11.32, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption. The initial instruction is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8fa56f255e99e4ce765474f01043e52628620a6ee1afc25331416525f58b2f0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/filler/original/2ab36f6c76502965/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "e3c696d968644974c528ae53fb9c2635405637b57b1a3891e7f6d300a3a5ac6e", "seconds": 6.68, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, with no correction or cancellation. The task is executed as intended.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from this or the last two calendar months—no, that should be this or the last calendar month.", "long_pause_offsets": [], "source_sha256": "11ce37bfc6c97f85a0f7b7cbff34b20dea774ea849e3f5e4f3959524c1465941", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'two calendar months' but immediately corrects it to 'one calendar month'. The final value is the corrected one, and all other conditions (archived status, app, action, and time scope) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/false_start/false_start_03_time_filter/07fe40159b04fe6e/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "1c387209045de8543e72370edf7aee41db1feb74c5d7574133472f3a7a79ff81", "seconds": 10.8, "spoken_utterances": [ "Delete all my archived um... gmail threads that are from ... this or the last two calendar months—uh... no, that should be this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last two calendar months' but immediately self-corrects to 'last calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Delete all my archived um Gmail threads that are from this or the last two calendar months uh no, that should be this or the last calendar month." }, "candidate_id": "wav_bd6c30ab72138b5f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then self-corrects 'two calendar months' to 'calendar month' with a natural hesitation ('uh') and a smooth, immediate repair. The prosody is consistent, and the correction feels spontaneous rather than forced. The pacing and intonation suggest genuine recall and correction." } ] }, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Delete all my gmail threads that are from this or the last calendar month. Oh, I forgot to mention, only delete archived threads.", "long_pause_offsets": [ 75 ], "source_sha256": "640163df137e50cc2cc63ebd65ce588e727dd1f4d11131b522e7d5c0488c0e61", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the threads must be archived, which refines the initial broad request. The time range and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/addition/addition_01/8446c6ee654fa1b0/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "a5604eec03d462cf14a230803e3530447685fea7094018e7a51cca47b33bdc7e", "seconds": 14.04, "spoken_utterances": [ "Delete all my uh, let me think... gmail threads that are from ... this or the last calendar month. ", "Oh, I forgot to mention, only delete um... archived threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all my Gmail threads that are from this or the last calendar month.' This is followed by an addition: 'only delete um, archived threads.' The addition correctly narrows the scope of the original request to only archived threads, which matches the original instruction. The final intent is a combined version of the initial and added constraints, and the control intent is to execute the task as revised.", "heard_transcript": "Delete all my uh, let me think, uh, Gmail threads that are from this or the last calendar month. Oh, I forgot to mention, only delete um, archived threads." }, "candidate_id": "wav_b8c32cf8d53286c6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-initiated pause ('let me think'), which is followed by a clear, coherent initial request. The addition of a new condition ('only delete um... archive threads') is introduced with a believable 'oh, I forgot to mention' phrase, and the correction is smoothly integrated into the ongoing utterance with a natural pause and filler ('um'). The prosody is consistent, and the flow reflects a realistic thought process." } ] }, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from this month. Make that 'this or the last calendar month' instead of 'this month'.", "long_pause_offsets": [ 63 ], "source_sha256": "70ca894169abc67ba12c74e2583588e6767d7f49d94dea8900aea88dbb404df8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the time filter to include both the current and previous calendar month, which matches the original instruction. The action, target app, and data type remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/revision/revision_03_time_filter/3030fce9c9283409/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "95b67cd043ba2177ad3cb2292eebcc98aa0d60aca4e55b5d332a65cd487b0ba8", "seconds": 14.24, "spoken_utterances": [ "Delete all my archived uh, let's see... gmail threads that are from um... this month. ", "Make that 'uh... this or the last calendar month' instead of 'this month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this month' but then corrects it to 'this or the last calendar month' after a pause. According to the revision type criterion, the correction spoken after a longer pause should be applied to the earlier instruction. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "delete all my archived uh let's see gmail threads that are from um this month. Make that uh this or the last calendar month instead of this month." }, "candidate_id": "wav_c23427695f40ed89", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and spontaneous speech pattern. It begins with a clear request, includes a realistic self-correction ('this month' to 'this or the last calendar month'), and uses natural fillers ('uh', 'let's see') that are well-placed and not overused. The intonation and pacing are consistent with someone formulating a task in real time, and the revision flows smoothly without sounding rehearsed or robotic." } ] }, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "3d32972090fc9cb893c27e93285a3d5ba47d6bf8a48003e23ad1fdf8997fa5a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/cancellation/cancel_034/c656f69e72dd6ae1/speech.wav", "wer": 0.08, "wav_sha256": "c67cf4b2e3f6313dd0904dd56c68bdce3aac4745f02b2f646fee5a726cd28406", "seconds": 13.32, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly revokes it with 'Hold on, let's drop the whole task,' which constitutes a cancellation. The final audible intent is to cancel, not to execute the original request.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "03413733556ed31b648dcbc09f8cbee2b63c29e04f574d35082ed41147485c32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/ambiguous_halt/halt_003/b30e1fb5dcaeca4b/speech.wav", "wer": 0.1, "wav_sha256": "284e06399268cfb9995b60e932bd41a45768022d237607cb03916ac08b379876", "seconds": 11.68, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month.", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection indicates a temporary hold on the task. The initial instruction is not canceled, and the speaker does not provide a new command or correction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "4d12842_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1113391fb259d84b3cd7c0fbaa0d1b63e141c4acff0877d606f48b28fd3ae5d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/filler/original/dd56c7dc0f82cdbf/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "04abbd14db47756a5b6dd940e82a5d5457573e842da3469ca1dd2b9e2d8dcfee", "seconds": 8.92, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('let's see', 'uh') but does not change the core task. The final intent matches the original instruction. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from last month—sorry, that should be this month.", "long_pause_offsets": [], "source_sha256": "c751fd60760cdb5c57d207e0a0febc76dd6d2b9a2fbcb83ed163f6fc134fe88b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/false_start/false_start_03_time_reference/a928af1c5bf86c4a/speech.wav", "wer": 0.15, "wav_sha256": "000063cc598f180667551f4ba160d36093d60c474af043c54af524d47b0e4ad1", "seconds": 9.76, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent or received money from last month—um... sorry, that should be this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'sorry, that should be this month.' This is a clear false start where the final intent is the corrected version. The correction is immediate and unambiguous, so the task should be executed with the corrected time frame.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from last month. Um, sorry, that should be this month." }, "candidate_id": "wav_44af174854d4b526", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complete false-start pattern. The speaker begins with 'last month', then self-corrects with 'sorry, that should be this month'—a clear, spontaneous revision with a believable tone of correction. The initial phrase is fully formed, and the correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody. The filler 'um' is used appropriately to signal hesitation, and the overall delivery sounds like a genuine moment of recollection and correction." } ] }, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month. I left out a detail: only befriend people I've sent or received money from.", "long_pause_offsets": [ 44 ], "source_sha256": "745c6a0fb1e1770dd8fa2c58fac2c1ba7653aca1827a83e3e1b037fabdd57c76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only befriend people I've sent or received money from', which refines the initial broad request. The time frame 'this month' is preserved from the first utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/addition/addition_02/b894284e9c1bbea0/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "3e9f47b275a667e962019e1fca9214e3270e1526d845efdb950f054321c475d2", "seconds": 12.48, "spoken_utterances": [ "Befriend on Venmo anyone I have uh... this month. ", "I left out a detail: only befriend people I've ... sent or received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Befriend on Venmo anyone I have uh this month' and then corrects with 'I left out a detail. Only befriend people I've sent or received money from.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. The addition type criterion is satisfied.", "heard_transcript": "Befriend on Venmo anyone I have uh this month. I left out a detail. Only befriend people I've sent or received money from." }, "candidate_id": "wav_1bfd6b93c6cb5dca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('bfriend on Venmo anyone I have uh... this month'), then naturally adds a clarifying detail ('I left out a detail. Only bfriend people I've sent or received money from'). The self-correction is embedded within the same utterance, with a believable pause and filler ('uh') before the initial request. The addition flows smoothly, maintaining consistent prosody and a natural conversational rhythm. The speaker sounds like they are formulating the request in real time, making the delivery highly realistic." } ] }, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money to this month. Make that 'sent or received money from', not 'sent or received money to'.", "long_pause_offsets": [ 70 ], "source_sha256": "322303f7af6db3617aa396f8e778092fed5a046d8bf7f9650c49e46116349274", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the preposition from 'to' to 'from', which matches the original instruction. The action, app, and time scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/revision/revision_02_transaction_relation/bee7f605c892d72c/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "dc5ba95400d76940f4c44885c7c2a95c13107992a1774fb7f1ccc5485306fe7e", "seconds": 14.8, "spoken_utterances": [ "Befriend on Venmo anyone I have sent or received money to ... this month. ", "Make that 'um... sent or received money from', not 'sent or received money to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent or received money to', then corrects it to 'sent or received money from' after a pause and self-correction. This matches the revision type criterion: the final intent is the corrected version, and the original incorrect clause is replaced. The rest of the instruction, including 'this month', remains unchanged and is preserved.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money to this month. Make that um, sent or received money from, not sent or received money to." }, "candidate_id": "wav_ee72a4121125e367", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with a natural-sounding self-correction. The pause before 'Make that' is brief and plausible, and the correction ('sent or received money from') is smoothly integrated into the same utterance with a slight pitch shift and emphasis, mimicking a spontaneous mental correction. The prosody is consistent and the pacing realistic." } ] }, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "ccb04f4a84f315dec4eb408e4aa3e36763e9b64b9947c137cbb012e660f961c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/cancellation/cancel_013/def44e422cc5862e/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "090b78c6f02f7994744b293f8e7bc95169f61031c2a344b770dc785419b74201", "seconds": 13.72, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month.", "Hold on, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then clearly says 'Hold on, I changed my mind. Please cancel the task.' This constitutes a direct and unambiguous cancellation of the entire task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month. Hold on, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "339270912248d5a19c7dd6c188e09aeaee4bc7d71c372bf6f2f6afda7817fae3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/ambiguous_halt/halt_001/9007ac3a22d13ae2/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "aa10e1aeec9960490e0c8f373a0be61b7a8145640494ff71c284cee4a805524f", "seconds": 11.44, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' which is a clear request to pause or hold. The initial instruction is fully stated before the pause command. Since the task is not canceled and the speaker does not resume or correct the instruction, the final control intent is to pause. This matches the 'ambiguous_halt' type criterion, which requires an audible request to hold/wait/pause with no permanent cancellation or permission to resume.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "628192f84aba661f56bec3004604790c82ac7d1141542dca23c76c8c2035b676", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/filler/original/0d1a4f88748ff1ff/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "89fb173a475e619f7eab9be293f6e85578513068f3585c75e6e0cf28b2608507", "seconds": 8.72, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um, let me think' which is a filler, but then clearly states the original instruction without any correction or change. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last week—sorry, make that this month.", "long_pause_offsets": [], "source_sha256": "4d56a468659b03d3cae060ca9fcbc9b3b227cacd13e0ad74c29f22031f8564b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'last week' with 'this month'; the action, app, and recipient scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/false_start/false_start_03_time_reference/312e789eef17b1ba/speech.wav", "wer": 0.0, "wav_sha256": "54a65f2f4618310acbbd3eee692ad29aff694f331f6de04b947d7b26ac5efe0b", "seconds": 12.96, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have sent money to last week—uh... sorry, make that ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last week' but immediately self-corrects to 'this month'. According to the 'false_start' type criterion, the final corrected value ('this month') overrides the initial mistaken value ('last week'), and the rest of the instruction remains unchanged. The final audible intent matches the original instruction.", "heard_transcript": "Befriend on uh let's see Venmo anyone I have sent money to last week uh sorry, make that this month." }, "candidate_id": "wav_d7e1fe75af8be90d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a filler ('uh') and a moment of thought ('let's see'), followed by a clear false start ('last week') and a self-correction ('this month'). The correction is delivered with a believable shift in intonation and pacing, and the overall prosody is consistent and human-like. The pauses and fillers are well-placed and contribute to the impression of real-time formulation." } ] }, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month. I left out a detail: only befriend people I have sent money to.", "long_pause_offsets": [ 44 ], "source_sha256": "f956d8dac6f84f093f3f1a4b0c05744e0174b02233038bf2d84cd8b8a9fb382d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'only befriend people I have sent money to', which narrows the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/addition/addition_02/bbddb8d3e64b5b0d/speech.wav", "wer": 0.3448275862068966, "wav_sha256": "4ee8028ee3844b530470eff157312860a3211483ab3042d5297ea406168bdb5b", "seconds": 15.92, "spoken_utterances": [ "Befriend on um, let's see... Venmo anyone I have uh, let me think... this month. ", "I left out a detail: only befriend people I have um... sent money to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Befriended on Venmo anyone I have sent money to this month' but then corrects with 'I left out a detail. Only befriend people I have send money to.' This is a clear addition that narrows the scope of the original instruction. The final intent is to befriend only those people to whom money has been sent, which matches the reference instruction. The addition refines the original request without contradicting it, satisfying the 'addition' type criterion.", "heard_transcript": "Befriended on um, let's see. Venmo anyone I have uh, let me think. This month. I left out a detail. Only befriend people I have um, send money to." }, "candidate_id": "wav_191d9cd813c64eed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with plausible fillers ('um', 'let me see', 'let me think') and a self-repair ('I left out a detail') that smoothly transitions into a more complete and specific instruction. The prosody reflects genuine hesitation and recollection, and the correction is integrated naturally into the flow without sounding rehearsed or robotic. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last month. That should be 'this month', not 'last month'.", "long_pause_offsets": [ 58 ], "source_sha256": "77ff8ce937338cdbb65e04d2d4f3aafc2d33fd03d802eaff55894a6073b619c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'last month' to 'this month', preserving the action, app, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/revision/revision_03_time_reference/bfc15d8476b8df3a/speech.wav", "wer": 0.15, "wav_sha256": "47ad3a7551785f901a25de789d23f225278a1dd08140eab426505468f728ca2b", "seconds": 12.44, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to um... last month. ", "That should be '... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' with 'That should be this month, not last month.' This is a clear revision where the final intent is the corrected version. The original instruction matches the final intent after the revision, so the task is executed as intended.", "heard_transcript": "Befriend on Venmo anyone I have sent money to last month. That should be this month, not last month." }, "candidate_id": "wav_63667e1a7d05ca46", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody. The self-correction ('That should be this month, not last month') is a believable, smoothly integrated revision with appropriate intonation and pacing. The pause before the correction is realistic, and the correction flows naturally within the same utterance, showing a plausible moment of realization." } ] }, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "44c222156ab9665c3c29f49b830ff6461e37c436a7721399ead9ac66f85cd12d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/cancellation/cancel_034/1e7de43e9e43129a/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "9e1b217cbde3ea3f57801929db803c0563a8df7c10701b94d26a9a45c31db1af", "seconds": 10.56, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month.", "Wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then clearly says 'Wait, let's drop the whole task,' which is an unambiguous cancellation of the entire request. The final interjection is a direct command to cancel, not a pause or hesitation. Therefore, the control intent is correctly identified as 'cancel' and the task is fully revoked.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month. Wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "845000bdbd0b67c21390918fb6d16f8813bb10955efc9604d275bcf23d2f77de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/ambiguous_halt/halt_028/aeb05b21f8128983/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "71d14ea37115f7f2e004e24dd0cab4c0cb7e18ad50a8c28805b8e360e1d4d82a", "seconds": 10.96, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The initial request is not executed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2929fd157401811b0e7601e3fff247fb29f3d2de0ccbc43345cc2859da077c93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/filler/original/e4af33a2c0fb9b78/speech.wav", "wer": 0.36363636363636365, "wav_sha256": "0e39632eba05f623c97ab90c4e9d1f09a5d35ba0c13e20d714501eda8d75c10c", "seconds": 6.56, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for such minor hesitations or pauses that do not alter the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to—wait, I meant received money from this month.", "long_pause_offsets": [], "source_sha256": "68a2423d7fd60fdc28658d4c7c14ef874e7ff7906e6e372f9da3ab9fae92b035", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately corrects to 'received money from', which matches the original instruction. The app (Venmo), action (befriend), and time condition (this month) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/false_start/false_start_02_transaction_relation/0a4a98cdc1e587db/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "e6612af679d89868ec6e440042b006ac4a69fbdfeda00333472ade4d4aa4e81f", "seconds": 9.48, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to—uh... wait, I meant um... received money from ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately self-corrects to 'received money from' with 'wait, I meant'. This is a false start where the final corrected value is the intended goal. The correction is immediate and unambiguous, and the rest of the instruction ('this month') remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Befriend on Venmo anyone I have sent money to uh wait I meant um received money from this month." }, "candidate_id": "wav_9df411672b6a2f16", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then self-corrects with a natural-sounding 'wait, I meant' followed by a smooth revision of the verb. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels spontaneous rather than rehearsed. The pause before 'wait' is brief and plausible, fitting a moment of reconsideration." } ] }, "is_default": true } }, { "task_id": "522e5e5_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have received money from. Oh, I forgot to mention, only include people from this month.", "long_pause_offsets": [ 53 ], "source_sha256": "c25077207fe7783e53d4c9d8e268aa537a2baff162708e5912b8f79833832b21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint, narrowing the initial broad request to only those people from 'this month'. The action, app, and recipient scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/addition/addition_01/b6f161caf9ca4435/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "cd1ee08381a8ca4a83d0507fa8bee613ad5ef55cc50fd080de4dcd31b81ddbe5", "seconds": 13.56, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from. ", "Oh, I forgot to mention, only include people from uh, let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to befriend everyone from whom money was received, then adds a constraint specifying 'this month' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final audible intent is clear and complete, matching the reference instruction exactly.", "heard_transcript": "Befriend on Venmo anyone I have received money from. Oh, I forgot to mention, only include people from uh, let's see, this month." }, "candidate_id": "wav_e2ea8e660a189018", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a believable self-correction. The initial phrase 'befriended on um Venmo anyone I have received money from' is followed by a plausible filler ('oh I forgot to mention') and a revision ('only include people from uh let's see this month'). The intonation, pacing, and use of 'uh' and 'let's see' sound spontaneous and conversational, with a clear change-of-mind flow. The addition of the time restriction is integrated naturally into the ongoing thought process." } ] }, "is_default": true } }, { "task_id": "522e5e5_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "cddfbba2ee7fc5c5118f74c07d3a065a823938734a2b7f77957933ab6aa4d270", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/cancellation/cancel_002/a5ae24c5bdc5ea31/speech.wav", "wer": 0.2, "wav_sha256": "22e5f67ae21027fff337e352b0a515b69e8c6b65691103d842eb583d9926167d", "seconds": 11.6, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month.", "Wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly and explicitly says 'cancel the whole task' as the final interjection, which constitutes a direct cancellation of the entire request. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Befriend on um Venmo anyone I have received money from uh let's see this month. Wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "1712f2f10d421239610e95af70c6287b7ca1707dbac42d8dae66c30a37e24f31", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/ambiguous_halt/halt_001/d0dc739b22244f54/speech.wav", "wer": 0.25, "wav_sha256": "8f87458be6e4142cf9c3ab29d3030d449a6bd32dfed6e03fed9d22657633da05", "seconds": 10.28, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'wait' after completing the full instruction. This constitutes a clear request to pause or hold the current task, with no indication of cancellation or permission to resume. The instruction is fully audible and understood, and the final interjection is a pause command. Therefore, the final intent matches the expected behavior for an 'ambiguous_halt' type.", "heard_transcript": "befriend on um Venmo anyone I have received money from uh let's see this month. wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "522e5e5_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3be549474f2bf811512be14ddf9911e597cd5f5a11bce3d7cdee2ccf3866f4c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/filler/original/e3e02160e1ae4292/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "6691de3727d70c759c1fb7f1b77433e3eebc9dc02a5d2d7aaef2cdd1a294ecbd", "seconds": 7.72, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not change the core request. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "Befriend on um Venmo anyone I have received money from uh let's see this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all treadmills—no, I meant weightlifting benches in my amazon cart.", "long_pause_offsets": [], "source_sha256": "e96dc248e418a8cd40deb1f110d12fd4d2b08b2183ce0514b8f9c8d6213cac50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'treadmills' to 'weightlifting benches' within the same utterance, preserving the action, scope ('all'), and destination ('amazon cart')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/false_start/false_start_01_product_category/f5c2ffed56c476e0/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "eaa6dcb0fbffc369e4e97544441c401238b20d4fa9ee16e63e6a63fa0a3eeb77", "seconds": 10.84, "spoken_utterances": [ "Place an order for all ... treadmills—uh... no, I meant hmmm... weightlifting benches in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'treadmills' but immediately self-corrects to 'weightlifting benches' after a hesitation. According to the 'false_start' type criterion, the final corrected value should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly, with no additional or missing constraints.", "heard_transcript": "Place an order for all treadmills, uh, no, I meant, uh, weightlifting benches in my Amazon cart." }, "candidate_id": "wav_5cc310ec0a210c81", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. The speaker begins with 'treadmills', then immediately self-corrects to 'weightlifting benches' with a believable 'no, I meant' structure. The correction is integrated smoothly into the ongoing sentence, and the prosody includes a natural hesitation and intonation shift. The flow feels like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for all weightlifting benches. Oh, I forgot to mention, place the order from my amazon cart.", "long_pause_offsets": [ 46 ], "source_sha256": "151b3a5b3a1373ffb1b0bab489087f0e400f30c03b75e431a4e368f9c5b58d7a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source context 'from my Amazon cart' without contradicting or removing any prior requirement. The action, item, and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/addition/addition_01/263a90a100777c70/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "309c215b9d7fbc423fc7a6c120f634f261c4a0a3764e2639766cd4c8366bfc31", "seconds": 13.08, "spoken_utterances": [ "Place an order for all uh, let me think... weightlifting benches. ", "Oh, I forgot to mention, place the order from ... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Place an order for all weightlifting benches' is followed by a longer pause and an addition: 'place the order from my Amazon cart'. This addition clarifies the source of the order, which is a relevant constraint. The final intent combines the original request with the added detail, matching the reference instruction. The hesitation and self-correction do not change the core task, and the addition is clearly intended to refine the original request.", "heard_transcript": "Place an order for all, uh, let me think, weightlifting benches. Oh, I forgot to mention, place the order from my Amazon cart." }, "candidate_id": "wav_2676213390faba67", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all treadmills in my amazon cart. Make that all weightlifting benches, not treadmills.", "long_pause_offsets": [ 53 ], "source_sha256": "0ca3202e7fdfd8671c6a2316a8a493ac287cf712ce52ad878a9b445aae086987", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial product category from treadmills to weightlifting benches, preserving the action, app, and scope of 'all' items." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/revision/revision_01_product_category/a1b3856776c5fa89/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "70a052f4fd3651b9b5c8446271b8e613741491f3f5732a5609ecb5598550fd11", "seconds": 11.88, "spoken_utterances": [ "Place an order for all treadmills in uh... my amazon cart. ", "Make that all ... weightlifting benches, not treadmills." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'treadmills' but immediately corrects to 'weightlifting benches' after a pause and the phrase 'Make that all'. This constitutes a revision where the final audible intent is to order weightlifting benches, matching the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Place an order for all treadmills and uh, my Amazon cart. Make that all, weightlifting benches, not treadmills." }, "candidate_id": "wav_7a2fc608329fd162", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('treadmills and uh...'), then smoothly corrects the item to 'weightlifting benches' with a natural-sounding self-repair ('Make that all...'). The prosody is consistent, and the correction flows logically within the utterance without awkward pauses or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "fcff850aab9d81701df2d5a227a448a71129645938c174f762f201b47b75663d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/cancellation/cancel_011/35877ae69cbb5797/speech.wav", "wer": 0.15, "wav_sha256": "936aa60cdb1611345c6e058682131d5a1341babeced04884378c85368d25c2d6", "seconds": 11.8, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart.", "Oh, wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then immediately corrects with 'Oh wait, I've changed my mind. Cancel that.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for all um weightlifting benches in my amazon cart. Oh wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "d1d331144ad80d399bd0df6630ed0327f12fe754cd6b716bca082479e9c7f57c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/ambiguous_halt/halt_024/336d7d99b319fc25/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "2b9962bd52b3c12454ed7db45208b8cbc2ef9278028f01c2b3873bc7f8235612", "seconds": 10.2, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not resume or complete the task. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Place an order for all um weightlifting benches in my amazon cart. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "12c667f4f24b47b06917011d17734da074fd7f71b2cff63be963e9f731d155eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/filler/original/bbfc2b16b4546000/speech.wav", "wer": 0.25, "wav_sha256": "0b696832fbb8a5ad65827677f3ff1b20a4cb6bacb3d62028b1d3bb8382a9d832", "seconds": 6.88, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task with a filler word ('um') but completes the original instruction without any correction or change in intent. The final audible goal matches the original instruction, so the task should be executed.", "heard_transcript": "Place an order for all um weightlifting benches in my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all beach towels—I meant beach balls in my amazon cart.", "long_pause_offsets": [], "source_sha256": "7ea67262c32e47106978f5fb1e0f071de7db338df30833b35737f002b38d844f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately corrects to 'beach balls' using 'I meant'. The final intended action is to order beach balls, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/false_start/false_start_01_product_category/757a8adb27cc888a/speech.wav", "wer": 0.0625, "wav_sha256": "3d7a75d2d24318c6a8cbe0ebf5f82d1adeca2cf10442e48dc69d61dc55a291ed", "seconds": 7.8, "spoken_utterances": [ "Place an order for all beach towels—uh... I meant ... beach balls in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately self-corrects to 'beach balls' with 'I meant'. According to the false_start type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Place an order for all beach towels. Uh, I meant beach balls in my Amazon cart." }, "candidate_id": "wav_b75334cc68e949a7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a clear false start ('beach towels') followed by a natural self-correction ('I meant beach balls') and a smooth continuation. The intonation and pacing are very natural, with a believable hesitation and a seamless transition. The correction is integrated into the flow of the sentence, making it sound like a spontaneous thought process rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for beach balls in my amazon cart. Oh, right, place an order for all beach balls in my amazon cart.", "long_pause_offsets": [ 50 ], "source_sha256": "e7284f0d97269c610b65be7925214f72cb7bf5eaadc58e179d25efa9f29d3473", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the word 'all', which was missing in the initial utterance, completing the original instruction's scope. The app, action, and item remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/addition/addition_01/42e41fd791120de3/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "77c69c78275851bd385ebc9e0b2fc86cf3f4df971ff78987c2c5c804d7891022", "seconds": 11.36, "spoken_utterances": [ "Place an order for beach balls ... in my amazon cart. ", "Oh, right, place an order for uh... all beach balls in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for beach balls in my Amazon cart.' This is followed by a self-correction where 'all' is added before 'beach balls' after a pause. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent is to place an order for 'all beach balls' in the Amazon cart, which matches the reference instruction. The addition of 'all' is a narrowing of the initial request, and the final combined goal is identical to the reference.", "heard_transcript": "Place an order for beach balls in my Amazon cart. Oh right, place an order for uh all beach balls in my Amazon cart." }, "candidate_id": "wav_7b762531d313723d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident request, then naturally self-corrects with a filler ('uh') and a plausible revision of the item ('all beach balls'), maintaining a smooth, continuous flow. The prosody is natural, with appropriate pauses and intonation that suggest spontaneous thought. The correction feels integrated into the utterance rather than tacked on." } ] }, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all beach towels in my amazon cart. Wait, the product category should be beach balls, not beach towels.", "long_pause_offsets": [ 55 ], "source_sha256": "1bdf9b95709689c4c2bf166988bb3d92ac82b45accfd9491aac8ce5e347b67d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the product category from beach towels to beach balls, while retaining the action, scope ('all'), and destination ('amazon cart')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/revision/revision_01_product_category/dff85e80a2cdab1d/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "1cd5b4b5f32f69a0b113da9f5b21a783017168cb177001792dd3c0f808c77219", "seconds": 11.64, "spoken_utterances": [ "Place an order for all beach towels ... in my amazon cart. ", "Wait, the product category should be uh... beach balls, not beach towels." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately corrects to 'beach balls' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to order beach balls, which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "Place an order for all beach towels in my Amazon cart. Wait, the product category should be uh beach balls, not beach towels." }, "candidate_id": "wav_85951f46e93d950f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and continuous prosody with a believable hesitation ('wait') and a clear, self-corrected revision. The speaker's intonation rises slightly on 'wait' and then drops on the correction, matching a genuine moment of realization. The correction is smoothly integrated into the same utterance without awkward phrasing or unnatural pauses, making it sound like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_036", "hide_variant_original": true, "source_sha256": "501ed6244f43c42e3923d9a467c9e020ce6fdb6d240925cc7c62e83fb62ecea8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/cancellation/cancel_036/3fbffc82c98c1a07/speech.wav", "wer": 0.16, "wav_sha256": "2a7dbf3f8115dd9b1dfb91d4e1bb211906d2bb98dcffe865aac55d27c1c87951", "seconds": 12.28, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart.", "Hang on, don't continue with that request. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to state the original task, then after a pause and hesitation, explicitly says 'Hang on, don't continue with that request. Cancel it.' This constitutes a clear cancellation of the task. The final audible intent is to cancel, not to execute the order. The cancellation is complete and unambiguous.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart. Hang on, don't continue with that request. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "7137395e3e059953155e9a3aad967e75aaaea4524b3c359325c094df7d5cb3d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/ambiguous_halt/halt_038/ef173eb10d81198d/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "37062185f7b3a608a300d1b0a8a93bb3c2ccc8f66664ec6f895fe8529a6c6845", "seconds": 11.48, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly requests to pause with 'Could you pause there for a second?'. This constitutes a clear, final request to hold the task, matching the 'ambiguous_halt' type criterion. The initial task is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "95baae1bee359e73121e863a6600a5264e0b75ec2accba24f828a2e0da40b5c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/filler/original/b12e26214005cf46/speech.wav", "wer": 0.25, "wav_sha256": "cd50beba53bfef8bd2ae8bbad635fffce1b5488153a9e28c4030d0f4331569b7", "seconds": 7.56, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let me think') but does not change the core request. The final intent remains to place an order for all beach balls in the Amazon cart, matching the original instruction. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all staplers in my ebay—sorry, I meant amazon cart.", "long_pause_offsets": [], "source_sha256": "4235f9e4fbe9a1386f0308ae58dde1b7eac913cb47ac811f69e914b3745ef94e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the action, item, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/false_start/false_start_02_app/3ea1075d346afc43/speech.wav", "wer": 0.0, "wav_sha256": "90051e4b8c61b69129701ee393e5d75575191596f6b0bbafd87375093f3409bd", "seconds": 7.52, "spoken_utterances": [ "Place an order for all staplers in my ebay—um... sorry, I meant ... amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon cart'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('eBay') is discarded. The final goal matches the original instruction with the corrected target.", "heard_transcript": "Place an order for all staplers in my eBay um, sorry, I meant Amazon cart." }, "candidate_id": "wav_bca9e8e347364f4b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for staplers in my amazon cart. Oh, right, this should apply only to all staplers.", "long_pause_offsets": [ 47 ], "source_sha256": "29d3c47149a69918a1924d88a13ac04e0fb71b2df28104176929758c51ab182b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the initial request should be restricted to 'all staplers', which matches the original instruction. The initial utterance was broad ('staplers'), but the addition refines it to the exact scope specified in the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/addition/addition_01/e8b17bc3afc0e6c9/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "0505d9a852ff5d21cce1e7d7c4a484f2502d766422910b4a9b8615b98b844105", "seconds": 11.84, "spoken_utterances": [ "Place an order for staplers in uh... my amazon cart. ", "Oh, right, this should apply only to uh, let me think... all staplers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for staplers in my Amazon cart.' The speaker then adds the clarification 'this should apply only to all staplers' after a pause and hesitation. This addition narrows the scope of the initial request to 'all staplers', which is consistent with the original instruction. The final intent is to place an order for all staplers, matching the original instruction. The addition refines the request without contradicting it, satisfying the addition type criterion.", "heard_transcript": "Place an order for staplers in uh my Amazon cart. All right, this should apply only to uh let me think, all staplers." }, "candidate_id": "wav_46b52b88ed2c33df", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all staplers in my amazon wishlist. Change the source to cart, not wishlist.", "long_pause_offsets": [ 55 ], "source_sha256": "97bc6c186ef129b52de24152e4e1bb1c5907fece9bd2c515376406cdfe1796f9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source from wishlist to cart, which aligns with the original instruction. The action, item scope, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/revision/revision_03_source_collection/51a6760a835b285e/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b7874aada24939ce788802079556b8528907423de8801bf62606f1dce4af907f", "seconds": 14.68, "spoken_utterances": [ "Place an order for um, let me think... all staplers in um... my amazon wishlist. ", "Change the source to uh... cart, not wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests placing an order for staplers in the wishlist, but then corrects the source to the cart. According to the revision type criterion, the correction spoken after a pause must be applied to the earlier instruction. The final audible intent is to order all staplers from the cart, which matches the original instruction's goal and target, despite the initial error.", "heard_transcript": "Place an order for um, let me think, all staplers in um, my Amazon wishlist. Change the source to uh, cart, not wishlist." }, "candidate_id": "wav_7a187ab5e5892e39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a self-initiated thought process ('let me think'), which is followed by a clear, coherent request. The revision ('change the source to cart not wishlist') is delivered with a smooth, conversational tone and a believable shift in focus. The prosody is consistent, and the correction flows naturally without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "44fd4b3867dd6adc6cd57aea46604c1cb0d1325daf6546de1fc48ed23175b286", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/cancellation/cancel_030/af0cee1894c9fec5/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "2920a69fff86da375678666262371f67762042562da3cae792f60e3d67f521b6", "seconds": 11.72, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart.", "Hold on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting to place an order for all staplers in the Amazon cart, but then clearly revokes the entire task by stating, 'Hold on, you can drop that task. I no longer need it.' This constitutes a direct cancellation of the original instruction. The final intent is to cancel, not to execute or pause, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for all staplers in um my Amazon cart. Hold on, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "ffc55789a11a26bdcaf52711f8e47152580a9493c7313753312ee82dbb6ecd56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/ambiguous_halt/halt_037/373c81935c183265/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "1e79b91a15549ebe097c1ee009aa8376a2dd177eed17c2f6daa05d23d9ba98a8", "seconds": 9.72, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to place an order is not executed, but the final control intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Place an order for all staplers in um my Amazon cart. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5238afc_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "037f095e0bebaf5107d7b1c51e946a6c3abebbe9401ec3e0404d78223433c1e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/filler/original/265c1ce484ce601b/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6851e5fbdcdcf9e13736374542e2e368831f2e5d453413e73b2a7eb6bc9b6aca", "seconds": 6.52, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. The task should be executed as requested.", "heard_transcript": "Place an order for all staplers in um my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill—sorry, I mean phone bill on venmo this year so far?", "long_pause_offsets": [], "source_sha256": "d48c0fa03874aaf2ade2fd102f3dcddddef56167e842015cd3369d9943b10ab5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'electricity bill' to 'phone bill' within the same utterance, preserving the app (Venmo), time frame ('this year so far'), and the action of querying the total amount paid." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/false_start/false_start_01_bill_category/535062db11addd11/speech.wav", "wer": 0.05, "wav_sha256": "01faa72b71acb50e8ec75b82f6d1b15543bc33953129674e81c045406f5921ff", "seconds": 9.2, "spoken_utterances": [ "How much have I paid in electricity bill—um... sorry, I mean ... phone bill on venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately self-corrects to 'phone bill'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the app (Venmo) and time frame ('this year so far'), remains unchanged and matches the reference. Therefore, the final audible intent matches the original instruction after the correction.", "heard_transcript": "How much have I paid in electricity bill um sorry I mean phone bill on Venmo this year so far?" }, "candidate_id": "wav_0687e50f20115cca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with a clear self-correction from 'electricity bill' to 'phone bill'. The filler 'um' is placed at a plausible point of hesitation, and the correction is integrated smoothly into the sentence without awkwardness. The prosody is consistent, and the overall flow mimics a real-time thought process." } ] }, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in phone bill on venmo so far? I didn't mention this, but it's so far this year.", "long_pause_offsets": [ 52 ], "source_sha256": "6712e6a07828b92362f95e9530c5685eff78f918c78cebbfc36fe8ea6c48e950", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time frame 'this year' to the initial question, refining the scope without changing the core request about the phone bill amount on Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/addition/addition_03/46f0b5c95443059f/speech.wav", "wer": 0.125, "wav_sha256": "fcd1b42b55951fa59951c7651fcc72f82ad814feedb88f7feb2dddcedee4c823", "seconds": 12.08, "spoken_utterances": [ "How much have I paid in uh... phone bill on hmmm... venmo so far? ", "I didn't mention this, but it's so far this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the phone bill paid on Venmo 'so far'. After a pause, they add the constraint 'this year', which is a clarifying addition to the original instruction. The final intent combines the original request with the added time frame. The addition is clear and does not contradict the original request, so the task is executed with the combined scope.", "heard_transcript": "How much have I paid in uh phone bill on Venmo so far? I didn't mention this, but it's so far this year." }, "candidate_id": "wav_4e53ab433c9b2430", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant question, then seamlessly adds a revision ('I didn't mention this but it's so far this year') that corrects the time frame. The intonation and pacing are very human, with a believable 'hmm' and a natural-sounding correction that integrates smoothly into the overall request." } ] }, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in phone bill on PayPal this year so far? Wait, use Venmo, not PayPal.", "long_pause_offsets": [ 63 ], "source_sha256": "7823b68ac7896c30801f081ddf1c57171b618c45dfb73510021e053cbddd1992", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, preserving the rest of the query including the expense category, time frame, and intent to retrieve a total amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/revision/revision_02_app/23c00eb14c17eb5f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3c5bcee66f0a9f0dfe37dd5a9140b94c082d4c46ba36f35bc3f105d5678514ff", "seconds": 11.64, "spoken_utterances": [ "How much have I paid in uh... phone bill on hmmm... PayPal this year so far? ", "Wait, use Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the phone bill amount on Venmo, which matches the original instruction. All other constraints (amount, time frame 'this year so far') remain unchanged.", "heard_transcript": "How much have I paid in uh phone bill on PayPal this year so far? Wait, use Venmo, not PayPal." }, "candidate_id": "wav_3d276454a79f722d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('uh') and a brief throat clear, which sounds like genuine speech preparation. The correction from 'PayPal' to 'Venmo' is delivered with a clear change in intonation and a self-correcting 'wait' that sounds spontaneous. The flow is continuous and the prosody remains consistent throughout, with no artificial pauses or robotic delivery." } ] }, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "48f79c86063c50eb103cfb6cc1002449afe7b2391e5f19aac00f0d480c7238e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/cancellation/cancel_039/9d39e59c0361ec71/speech.wav", "wer": 0.0, "wav_sha256": "aa83b83996572bdf90fbd4a02b233e2539555cf329bebc62aafe3f2c34ba4c39", "seconds": 10.96, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?", "Oh, wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately corrects themselves by saying 'Oh wait, don't go ahead with that. Cancel it.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute the original query.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far? Oh wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "b8e251aa925415912e844bb08911836f9bc45f3d87c4b4eb3fb7ac512cb85220", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/ambiguous_halt/halt_016/cb0fb4c79ae737ff/speech.wav", "wer": 0.0, "wav_sha256": "b61ba3e43bd7536f5028a4d555ac4000eb463a902947bbaed197df6370dfef2c", "seconds": 9.36, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not revoked or corrected, and the final interjection is a temporary hold command, not a cancellation. Therefore, the final control intent is correctly interpreted as 'pause'.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far? Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bed073b402dc6f826da094b5cb1dc3a48a0d42b51b17e458d877a28696de1e6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/filler/original/89c4096991a3ad8c/speech.wav", "wer": 0.0, "wav_sha256": "98a26c49b95a3553f08aad09547a953af8e87928d14ed53e42214691527ef27b", "seconds": 5.48, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker delivers the full original instruction without any hesitation, self-correction, or change in intent. The audio contains no fillers that alter the task goal, and the final utterance matches the original instruction exactly. Therefore, the task is executed as intended.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in electricity bill on cashapp—no, I mean venmo this year so far?", "long_pause_offsets": [], "source_sha256": "259ef0f5c39526fef5fb640a761b8fea1a56acd03af07f2aa8e771eb9d6f5ca7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the action, category, time frame, and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/false_start/false_start_02_app/c94948cc7f70aa6b/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "3f5b3701eb6e94d1ac1cd1e249c540bc0419db04a8100c57b2a9b6ff11308c08", "seconds": 9.96, "spoken_utterances": [ "How much have I paid in electricity bill on cashapp—hmmm... no, I mean venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, with only the corrected app name.", "heard_transcript": "How much have I paid in electricity bill on Cash App? Hmm. No, I mean Venmo. This year so far?" }, "candidate_id": "wav_a29437664a34cf9c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with a clear, slightly hesitant question, then self-corrects from 'cash app' to 'venmo' with a smooth, conversational tone and a natural pause. The intonation and rhythm are consistent with spontaneous speech, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "552869a_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in electricity bill on venmo so far? And just to be specific, it's so far this year.", "long_pause_offsets": [ 58 ], "source_sha256": "f59bf12dbd3b5843df7d36d72452375ca86bc50c436f1ac875022febbf915be7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'this year' to clarify the scope of 'so far', which is consistent with the original instruction's intent. The final goal combines the initial query with the added specificity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/addition/addition_03/15224bb20bd8c83b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "d3a4eb31733aa82a240bfa4cdb18fc93a2edd54c932d22c768a907102dce8344", "seconds": 14.44, "spoken_utterances": [ "How much have I paid in ... electricity bill uh, let's see... on venmo so far? ", "And just to be specific, it's so far uh... this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the core question and then adds the clarifying constraint 'this year' after a pause, which is the defining characteristic of an 'addition' type. The final intent combines the original request with the added temporal scope, matching the reference instruction exactly. The hesitation and filler words do not alter the final goal.", "heard_transcript": "How much have I paid in electricity bill uh let's see on Venmo so far and just to be specific it's so far uh this year" }, "candidate_id": "wav_92a745941753f403", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker's delivery is highly natural, with a realistic flow of thought. The initial question is followed by a plausible 'uh let's see' hesitation, and the speaker then adds clarifying details ('on Venmo so far') and a time frame ('this year') in a continuous, conversational manner. The prosody is varied and the pacing reflects genuine formulation of the request, making it sound like a spontaneous, real-time inquiry." } ] }, "is_default": true } }, { "task_id": "552869a_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "0333d10af265031e27f9746c6b99815891881f85f8c3abdd49da714c9d26fd07", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/cancellation/cancel_008/2d2ec9c0ed5daf22/speech.wav", "wer": 0.04, "wav_sha256": "99954a113e1bd0532d9b3d43885510e24af8e72e40ba4957bdb588884878f1cb", "seconds": 14.2, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question, then after a pause says 'Hang on. I want to cancel that task.' This is a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or new instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo? Let's see. This year, so far. Hang on. I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "646903618916c1f06bfdd433e71be9ed235b6660d9286cc0998616e51d2e9fa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/ambiguous_halt/halt_031/9a8587aeceba6b4f/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "6c500fa91f47d4850e51ca6f6795215b276aaf60383c190b6751d06d2ee00351", "seconds": 13.56, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on. I just', which is a clear request to pause the current task. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo? Let's see. This year, so far. Hang on. I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "21791b91e8079af3e169fcf207e31f644071a68bb95591005a1903889fa435bb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/filler/original/cee59151694622cd/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "a4f61b1707091ad6fb597a41d29c9787e2f5b4f9ea99a30f2caee41469dbe2ed", "seconds": 9.72, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core request. The final intent matches the original instruction exactly. The 'filler' type allows for hesitation and backchannels as long as the remaining task meaning is preserved.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo, let's see, this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill—I meant internet bill on venmo this year so far?", "long_pause_offsets": [], "source_sha256": "050cc27c63cd1e923bb5be8fdb67cd481bc7afb88a90f6bb98789515fa588d2d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'electricity bill' to 'internet bill' within the same utterance, preserving the app (Venmo), time frame (this year so far), and action (query total amount paid)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/false_start/false_start_01_bill_category/38b68689928fa8e3/speech.wav", "wer": 0.0, "wav_sha256": "eda67366a06a301b6ec1c43539ca982259343227b28cbe405a8084e6f07d43a0", "seconds": 8.04, "spoken_utterances": [ "How much have I paid in electricity bill—uh... I meant um... internet bill on venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately self-corrects to 'internet bill'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final request matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in electricity bill, uh, I meant um, internet bill on Venmo this year so far?" }, "candidate_id": "wav_c2fc5ac1dd729bcd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('electricity bill'), immediately self-corrects with 'I meant' to the correct term ('internet bill'), and smoothly continues with the rest of the request. The correction is natural and integrated into the flow, with a believable hesitation and filler ('uh') before the initial error. The prosody is consistent and the pacing realistic." } ] }, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in internet bill on venmo so far? And just to be specific, only include payments made this year.", "long_pause_offsets": [ 55 ], "source_sha256": "bce2ba051039ac243904840c95036f5f07a8c6dc89eb93af3b14937b81543926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this year') to the initial query, which is consistent with the original instruction's intent to ask about payments so far this year. The final goal combines the original scope with the added specificity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/addition/addition_03/43b98a5688153ffb/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "080f10d640a619d50aafbb37f5ca3f29ff609077e87601fb1595d9c5715e83e5", "seconds": 13.4, "spoken_utterances": [ "How much have I paid in ... internet bill on let me think... venmo so far? ", "And just to be specific, only include payments made um... this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How much have I paid in internet bill on venmo this year so far?'. The speaker begins with a similar request, then adds the constraint 'only include payments made this year' after a pause. This matches the 'addition' type, where the final goal combines the original request with the added constraint. The final audible intent is complete and matches the reference instruction with the added specificity.", "heard_transcript": "How much have I paid in internet bill on Venmo so far? And just to be specific, only include payments made um this year." }, "candidate_id": "wav_3e92d4304dfc2bdf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous speech patterns. The speaker uses a filler ('let me think') at a plausible point in the sentence, indicating genuine formulation. The addition of 'and just to be specific' followed by a clear, self-contained restriction ('only include payments made this year') flows naturally as a remembered detail. The prosody is varied and conversational, with appropriate pauses and intonation that suggest real-time thought processing." } ] }, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much have I paid in internet bill on venmo last month so far? Make that this year, not last month.", "long_pause_offsets": [ 66 ], "source_sha256": "305713e261eeab8155288b090c45f109825c07d3279a1f3f2bc9322dfabee051", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this year', preserving the action, app, and category. The phrase 'so far' is retained, aligning with the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/revision/revision_03_time_reference/0ebb4e961ea7161d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "1e996291a1a05fa1ce806f54f00cab812915a19a135991a66b21175f44d7ba52", "seconds": 12.28, "spoken_utterances": [ "How much have I paid in ... internet bill on ... venmo last month so far? ", "Make that ah... this year, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this year' with 'Make that uh this year, not last month.' This is a clear revision where the final intent is to query the amount paid this year. The correction is spoken after a short pause and directly modifies the time frame of the original request. The rest of the instruction remains unchanged, so the final goal matches the reference.", "heard_transcript": "How much have I paid in internet bill on Venmo last month so far? Make that uh this year, not last month." }, "candidate_id": "wav_2a3d475d7274c7e7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('uh') before correcting the time frame from 'last month' to 'this year'. The correction is smoothly integrated with a self-repair ('Make that uh...') and a believable change of mind. The prosody is consistent, and the pacing reflects genuine thought processing without robotic pauses or unnatural word boundaries." } ] }, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "940dec84dd4e7fd290e4650054e9ab7885f46b7c567aab13ed882a3e62f34c99", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/cancellation/cancel_030/c636876e8e0cd3f2/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "143d61b52423bfde927aa82bf11722ab8c3c3334efd45fdbe940537fcc42cd9e", "seconds": 13.32, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?", "Hold on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Hold on, you can drop that task. I no longer need it.' This is a clear and unambiguous cancellation of the initial request. The final interjection explicitly revokes the task, satisfying the cancellation type criterion.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far? Hold on, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "97bc1e1ff3256b5315eb046d62d55978b04f42e448a60be2a2e0cc4384af8714", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/ambiguous_halt/halt_001/80f348f3440e00d4/speech.wav", "wer": 0.0625, "wav_sha256": "2d531f4a92347da0ca309fdfb48e0ce5169644f39bb3095758079484e3314b33", "seconds": 11.16, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' which is an audible request to hold or pause. There is no cancellation of the task, nor is there a new instruction to resume. The final interjection clearly indicates a temporary halt, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far? Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "552869a_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e907e897cd459e11ddf4fda5012944c17e3a2611ea4463184e8ad5d2297b26a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/filler/original/25b550a003a43f32/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "48072c714f0caedc196d4e1a089e4d9b2872d6709d56549a15da598c1359d54e", "seconds": 8.4, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_artist_reference", "proposed_instruction": "Like all the songs from the bands I follow—I should say artists I follow on Spotify.", "long_pause_offsets": [], "source_sha256": "0160f12b7d6b4473fd7061ac3b103a4b8d5ed2efc95556c95b3a065abf89c40d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands' but immediately corrects to 'artists' within the same utterance. The final intended target is 'artists', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/false_start/false_start_01_artist_reference/cd6bb5401db57f1e/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "f07ba3982e7e8965921615dc6dd70db3f4030c599045c8b4261045102f9ec75b", "seconds": 8.04, "spoken_utterances": [ "Like ... all the songs from the um... bands I follow—uh... I should say artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands I follow' but immediately self-corrects to 'artists I follow'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final goal matches the reference instruction exactly.", "heard_transcript": "Like all the songs from the bands I follow uh, I should say artists I follow on Spotify." }, "candidate_id": "wav_525d56714b6025f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with 'bands' followed by a self-correction to 'artists' and a complete, coherent request. The intonation and pacing are very conversational, with a believable hesitation and a smooth, pragmatic flow. The correction feels spontaneous and integrated into the sentence, not mechanically inserted." } ] }, "is_default": true } }, { "task_id": "57c3486_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like all the songs from the artists I follow on Spotify.", "long_pause_offsets": [ 43 ], "source_sha256": "7f5b781ae1e46c5f6ff0ea7ac924bc7018311b394e51c548bd08d2b11091a9c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The second utterance adds the missing phrase 'all the songs from' to complete the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/addition/addition_01/193e97ea06659b0b/speech.wav", "wer": 0.0, "wav_sha256": "c2d4f82ae947c67f108b29b0e41eb5c82c46769f337120c9beaf5ee1bb263483", "seconds": 13.72, "spoken_utterances": [ "Like from the artists I follow on um... Spotify. ", "Oh, I forgot to mention, like let's see... all the songs from the artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance 'Like from the artists I follow on Spotify' is incomplete and grammatically awkward. The speaker then adds 'all the songs from the artists I follow on Spotify' after a pause and self-correction. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final intent is a complete and coherent request to like all songs from the artists the user follows on Spotify, which matches the original instruction. The addition clarifies and completes the initial incomplete thought, resulting in a valid and executable task.", "heard_transcript": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like, let's see, all the songs from the artists I follow on Spotify." }, "candidate_id": "wav_9c1615fb34d3750a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a natural, spontaneous flow with a believable 'oh, I forgot to mention' revision. The speaker starts with a partial request, hesitates with 'um', then corrects and completes the full task. The prosody is conversational, with appropriate pauses and intonation shifts that suggest genuine thought processing. The correction is integrated smoothly into the utterance, making it sound like a real-time formulation of a request." } ] }, "is_default": true } }, { "task_id": "57c3486_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "2a6a738f162b6548dc0a4767a7719186b5215edfb1ef61b8ca477c2d7fd635a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/cancellation/cancel_010/507846610848f644/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "750ea30746b733a3c72dee951cb895f03e4096eae360f382c9e57c710ef66400", "seconds": 11.76, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify.", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the user stating the original request, then immediately follows with a clear cancellation command: 'Oh wait, cancel the request I just made.' This constitutes a final, unambiguous cancellation of the task. The cancellation is not merely a pause or hesitation, but a direct revocation of the prior instruction. Therefore, the final control intent matches the expected 'cancel' behavior for this type.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "0f1307b9f7ad7ec917f6661f459c544fa116b719aa7169c326ce00c395ef943f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/ambiguous_halt/halt_010/f908cfe7bcc8e6cc/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "19722c61e8e2525b2d88c002b072f66cd379906bfc3d73c29e795028cee35b0d", "seconds": 10.56, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly says 'Hold on a second,' which is a standard verbal cue to pause or wait. There is no cancellation or instruction to resume, and the final interjection is a request to hold. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "da8240354f9ee4f9db7524dd6b80136d9cdcac1ef10076c4b1ccac82eed2bb71", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/filler/original/194de8afae53fb83/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "d780f708b5b62a7b09679c08ca7d4e59e9e872a26393046e1bb791762871bc22", "seconds": 7.2, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let's see') and a brief pause but does not change the core request. The final audible intent is to like all songs from the artists followed on Spotify, matching the original instruction. The hesitation is a filler and does not alter the task goal.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music—sorry, I meant Spotify.", "long_pause_offsets": [], "source_sha256": "56db02317ddfe8f0ed4fdd5a717ca6b95252724b0c3789835f135ef63852d4e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, scope, and target app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/false_start/false_start_02_app/b447fac66f02b24f/speech.wav", "wer": 0.05, "wav_sha256": "b8d03b09ad5a4b8c737a229e133aa35437849d9da6e17edc066cfc872eb0396e", "seconds": 8.52, "spoken_utterances": [ "Like all the songs from the artists I follow on uh, let's see... Apple Music—uh... sorry, I meant ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the reference instruction exactly.", "heard_transcript": "Like all the songs from the artists I follow on uh let's see, Apple Music uh sorry, I meant Spotify." }, "candidate_id": "wav_34dbd589933492dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start and self-correction. The speaker begins with 'Apple Music' and then immediately corrects to 'Spotify' with a believable 'sorry, I meant' phrase. The prosody includes a slight hesitation and a natural pitch drop on 'sorry,' followed by a smooth correction. The pacing and intonation are consistent with spontaneous speech, and the correction feels integrated into the flow rather than tacked on." } ] }, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Like all the songs from the on Spotify. I didn't mention this, but only like songs from the artists I follow.", "long_pause_offsets": [ 40 ], "source_sha256": "ceed268105e5489194f3dabe7415bd81d5d50df7f93bc8da9fa6f543a63a676d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'from the artists I follow' to the initial incomplete instruction, which already specified 'on Spotify'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/addition/addition_02/b33ff9d96f58820e/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "1052029569f710775606787f3f4e7e36e5717e91ca623a47e21f15673af40e11", "seconds": 13.4, "spoken_utterances": [ "Like ... all the songs from the on uh, let me think... Spotify. ", "I didn't mention this, but only like songs from the hmmm... artists I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all the songs from the um uh let me think. Spotify.' and then adds 'I didn't mention this, but only like songs from the hmm artists I follow.' The addition clarifies that the action should be limited to songs from the artists the user follows. This matches the 'addition' type criterion, where the final goal combines the original instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Like uh all the songs from the um uh let me think. Spotify. I didn't mention this, but only like songs from the hmm artists I follow." }, "candidate_id": "wav_57d5878881575982", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-correction ('let me think'), followed by a plausible addition of a restriction ('only like songs from the artists I follow'). The prosody is conversational, with a believable pause and filler ('hmm') before the final clarification. The flow from initial request to addition feels spontaneous and coherent." } ] }, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like on Spotify. I misspoke about the artist reference: it should be 'artists I follow', not 'artists I like'.", "long_pause_offsets": [ 55 ], "source_sha256": "0e808e173dcc56a3d589869a448485a6a890aaec0f914aad32f9ffdea68f12ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'artists I like' to 'artists I follow', which matches the original instruction. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/revision/revision_01_artist_reference/9facf2ed6d714bf3/speech.wav", "wer": 0.09375, "wav_sha256": "aa4a8c52988e777c33d50b6c50a1a4694e2b0fb6c7017a701bd2bd73437ac60c", "seconds": 15.96, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I like on ... Spotify. ", "I misspoke about the artist reference: it should be 'uh... artists I follow', not 'artists I like'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artists I like' but immediately corrects themselves to 'artists I follow' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to like all songs from the artists the user follows, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "Like um, let me think. All the songs from the artists I like on Spotify. I misspoke about the artist reference. It should be uh, artists I follow, not artists I like." }, "candidate_id": "wav_8157273a0de827da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic and natural sequence of spontaneous speech: it begins with a filler ('like um') and a hesitation ('let me think'), followed by a clear initial request. The speaker then self-corrects a specific term ('artists I like' to 'artists I follow') with a natural false start and repair, including a brief pause and a soft 'uh' before the correction. The prosody is conversational, with appropriate stress and pacing, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The flow is continuous and psychologically plausible." } ] }, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "06e3022a0f2b18d001dca911668a1ce285851f95cded48bb4c5ddbe066cbe4d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/cancellation/cancel_023/81181076d327a014/speech.wav", "wer": 0.0, "wav_sha256": "76d9ee81f1afe7b2a82809c1d3baa195f9a215ce99491168095d0d523fc74e07", "seconds": 12.76, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify.", "Hold on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to retrieve songs from followed artists on Spotify, then clearly revokes it with 'Hold on, never mind about that. Don't carry out the request.' This constitutes a direct cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify. Hold on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "22d5f76dc4d87f9428ff9c661722cdc197666dd649c8f4b32e743def7c009b1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/ambiguous_halt/halt_011/23cde13909b5ac92/speech.wav", "wer": 0.0, "wav_sha256": "05b4aaf268cc0b8756d251f21d31fd13f3bb5890d7cddfc2687e2c09e3a8d365", "seconds": 10.48, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request, then says 'Hang on a second' as a final interjection. This is a clear, audible request to pause or wait, with no indication of cancellation or resumption. The initial request is not executed, but the final intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "fc9b5e8a10ab4fba49a72d7ab42d97f180abd3c429e6815423b6f3cce9452432", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/filler/original/e8d2304c5618070c/speech.wav", "wer": 0.0, "wav_sha256": "5063d5e1503322cb469a3fda1bb416a40a1ec0284152a86201803d78ff82dc8b", "seconds": 7.36, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('um', 'let me think') but proceeds to state the original instruction without any correction or change in intent. The final utterance matches the original instruction, and the hesitation is a typical filler that does not alter the task goal. Therefore, the final intent is to execute the original request.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like—I should say artists I follow on Spotify.", "long_pause_offsets": [], "source_sha256": "2599adeed34efc126fa83283479e114489bd7570bc8e7ee7c75f4c668eab7d09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'artists I like' to 'artists I follow' within the same utterance, preserving the app (Spotify) and the action (like all songs). The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/false_start/false_start_01_artist_reference/34e76204e3cb7802/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "0d17efe0f42c0db84933c33125e74dff45b279f40f298f4eb7d0be71f3814ccf", "seconds": 8.88, "spoken_utterances": [ "Like hmmm... all the songs from the ... artists I like—uh... I should say artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artists I like' but immediately self-corrects to 'artists I follow on Spotify'. This is a false start, where the final corrected version matches the original instruction exactly. The correction is immediate and unambiguous, so the final intent is to execute the original task.", "heard_transcript": "Like, hmm, all the songs from the artists I like, uh, I should say artists I follow on Spotify." }, "candidate_id": "wav_e4db9da8c5e94833", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('hmm'), then formulates a request that is slightly off ('artists I like'), followed by a believable self-correction ('I should say artists I follow on Spotify'). The correction is smoothly integrated into the same utterance with a natural intonation shift and a plausible rephrasing. The pacing and prosody are consistent with spontaneous speech, and the false start feels authentic rather than forced." } ] }, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Like all the songs from the on Spotify. I left out a detail: only like songs from the artists I follow.", "long_pause_offsets": [ 40 ], "source_sha256": "4dedf076dfe6b349bc666d1c93476af677125ac81443f951adb325b107788249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'only like songs from the artists I follow', which clarifies the scope of the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/addition/addition_02/a4212fb9ce49b765/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "0e3882cb641d864f7d4aaad338a277f5b589d09e5b4492fad6f208c549b1863b", "seconds": 12.92, "spoken_utterances": [ "Like um... all the songs from the on Spotify. ", "I left out a detail: only like songs from the hmmm... artists I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all the songs from the on Spotify' which is a malformed version of the original instruction. They then correct this by adding the constraint 'only like songs from the hmm artists I follow'. This addition narrows the scope of the original request, which is the expected behavior for the 'addition' type. The final audible intent is to like only the songs from the artists the user follows, which matches the reference instruction after the addition.", "heard_transcript": "Like um all the songs from the on Spotify. I left out a detail. Only like songs from the hmm artists I follow." }, "candidate_id": "wav_7112fc3e9168200f", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 61 ], "source_sha256": "89dce863ff2ee65a3580dc2ccc186c56fd838efccd97942901bc3ae44b410d84", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the action and the scope of artists (those the user follows)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/revision/revision_02_app/0aeaaf6b22cbe4ba/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "983d353bf167f2c6897891b8b22fb6d049918a28287fa1d7891d139738e3c9de", "seconds": 13.72, "spoken_utterances": [ "Like um... all the songs from the artists I follow on um, let's see... Apple Music. ", "Make that ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify, not Apple Music' after a pause. This is a clear revision where the final intent is to use Spotify. The correction is unambiguous and the original instruction is fully preserved except for the app name, which is corrected. The final control intent is to execute the revised task.", "heard_transcript": "Like um all the songs from the artists I follow on um let's see Apple Music. Make that Spotify, not Apple Music." }, "candidate_id": "wav_970e7629cbe0e024", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('Like um...') and a plausible self-correction from 'Apple Music' to 'Spotify'. The correction is immediate and smoothly integrated into the sentence, with a clear shift in intonation and a brief pause before the new value. The flow feels spontaneous and unscripted, with no signs of artificiality or forced pacing." } ] }, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_035", "hide_variant_original": true, "source_sha256": "919d67324f548d19976bdf4c3f049f1cc9f7e4104888a10cbfc5ff26f01816d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/cancellation/cancel_035/a3680eaf8c015c67/speech.wav", "wer": 0.0, "wav_sha256": "277967b56d6dd0eff01d9580caddbf7f49736a6b4f3b1acd753d9f1497c15100", "seconds": 11.88, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify.", "Wait, stop working on that and cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction, followed by a clear self-correction and cancellation command: 'Wait, stop working on that and cancel the task.' This constitutes a final, unambiguous cancellation of the initial request. The cancellation is not merely a pause or hesitation, but a direct instruction to abort the task. Therefore, the final control intent matches the expected 'cancel' behavior.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify. Wait, stop working on that and cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "8713c535da949ef058097bf64fd3ffcdb89d6876f47e3ff9dc0c8df115730f3c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/ambiguous_halt/halt_022/159d2805d7c65a45/speech.wav", "wer": 0.0, "wav_sha256": "8f5e46a38ae6ea6664a20cc41f9fdf3e2480e50576c191b5b95efca3cff26ef1", "seconds": 10.8, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on. Let me think.', which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify. Hang on. Let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "57c3486_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4ea28eee411cf06f9440b7ea2d2929b3b6d6d72100c492d9083634d9533141d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/filler/original/1b19b81c0ab59330/speech.wav", "wer": 0.0, "wav_sha256": "bcb6affdb93163ee14adbf161de3e9378a8e90bd313b4cfbbe0c828d6e010d2e", "seconds": 6.76, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'um' and 'let me think' but does not change the core request. The final audible intent is to like all songs from the artists the user follows on Spotify, which matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_folder_reference", "proposed_instruction": "Delete all PDF files from the Documents folder—sorry, I meant downloads folder in my home directory.", "long_pause_offsets": [], "source_sha256": "7c90d4ddacfc11331b3d247a6569ad99f503d2d4c4ddae2539fb9b63e4db7820", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Documents folder' to 'downloads folder in my home directory', preserving the file type (PDF), action (delete), and scope (all files)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/false_start/false_start_02_folder_reference/93b51eca9526e270/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "7e6b11927e7af89a2017aad282e705ae8a20d60a821cf754a505e91d5f484570", "seconds": 9.4, "spoken_utterances": [ "Delete all let me think... PDF files from the Documents folder—um... sorry, I meant downloads folder in my home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'documents folder' but immediately self-corrects to 'downloads folder' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Delete all, let me think, PDF files from the documents folder. Um, sorry, I meant downloads folder in my home directory." }, "candidate_id": "wav_2def786313aa2cc5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic false start and self-correction. The speaker begins with 'delete all' and then hesitates with 'let me think', which is a natural filler. They then state an incorrect folder ('documents'), immediately self-correct with 'sorry, I meant downloads folder', and complete the request. The intonation, pacing, and natural flow of the correction, including the back-reference to the earlier mistake, make this sound like a spontaneous, real-time thought process." } ] }, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Delete all files from the downloads folder in my home directory. Oh, I forgot to mention, only delete PDF files.", "long_pause_offsets": [ 65 ], "source_sha256": "5d710589306d3a55e2507438a102d8d3ed241fd3cb62be82c98cf38676680c72", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only PDF files, preserving the folder and location." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/addition/addition_01/be8af8b6de7d7290/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "9384b76801adccd123a3aea8a33fee70f34b2c2e96110ad8ac987570d4654a0a", "seconds": 11.96, "spoken_utterances": [ "Delete all files from the uh... downloads folder in my um... home directory. ", "Oh, I forgot to mention, only delete ... PDF files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all files from the downloads folder. The speaker then added a constraint specifying that only PDF files should be deleted. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Delete all files from the uh downloads folder in my um home directory. Oh, I forgot to mention, only delete PDF files." }, "candidate_id": "wav_125ae414af9e2dc9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('delete all files from the uh downloads folder...') and then naturally adds a crucial restriction ('only delete PDF files') after a brief pause. The addition is grammatically and semantically coherent, with a plausible shift in focus. The prosody is natural, with appropriate fillers ('uh', 'um') and a believable conversational rhythm. The correction is integrated smoothly into the same utterance, making it sound like a spontaneous clarification rather than a rehearsed or disjointed edit." } ] }, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_folder_reference", "proposed_instruction": "Delete all PDF files from the Documents folder. Let me correct the folder: delete from the downloads folder in my home directory, not the Documents folder.", "long_pause_offsets": [ 48 ], "source_sha256": "0e1101ff7db18c1917d02d4f17c7ce33625ccc4c9ef59478e2787ee78dd65ddd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder from 'Documents' to 'downloads folder in my home directory' and negates the initial incorrect folder, preserving the file type and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/revision/revision_02_folder_reference/4cc4cd55a1a1c358/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "352c8ebcd9a9c9a4353911975327d1a056ab56a89ecd5d268de57bebb6bb4932", "seconds": 14.0, "spoken_utterances": [ "Delete all uh... PDF files from the uh, let's see... Documents folder. ", "Let me correct the folder: delete from the uh... downloads folder in my home directory, not the Documents folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the folder as 'documents folder' but immediately self-corrects with 'Let me correct that' and specifies the correct folder as 'downloads folder in my home directory'. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Delete all uh PDF files from the uh documents folder. Let me correct that. Delete from the uh downloads folder in my home directory, not the documents folder." }, "candidate_id": "wav_3e0171b37f8163da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('Delete all uh PDF files from the uh... documents folder'), then self-corrects with a natural-sounding revision ('Let me correct the folder: delete from the uh... downloads folder in my home directory not the documents folder'). The correction is integrated smoothly into the same utterance, with plausible filler words ('uh') and a realistic prosodic shift indicating a change of mind. The flow is continuous and the repair feels spontaneous." } ] }, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "cddb521ec352a8410f61eb3603c5005def931fca4ae04211ac6f66282b0c145c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/cancellation/cancel_010/a094884b938bfb30/speech.wav", "wer": 0.1, "wav_sha256": "5b018516cf3a9e53c01aa4ed306da28953728f538fb5585f6457dd652c463f07", "seconds": 9.68, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory.", "Wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a deletion request and then immediately follows with a clear cancellation command ('cancel the request I just made'). This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final audible intent is to cancel, not to execute the deletion.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory. Wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "e24dfe50657f205074ecc7f1146e2f4809d4ac3c41d6fec85ed709cb4e687c6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/ambiguous_halt/halt_014/97e876a8bd1fb429/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5646db6f2e9aebf5d4bff35a5c093c9491a8f8504cab739fa5289c816f176693", "seconds": 9.0, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the initial instruction and then clearly requests a pause with 'Can you hold on a second?'. This constitutes a final, audible request to pause, which matches the 'ambiguous_halt' type's expected behavior. No cancellation or execution is requested.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3f730ec818fe6dd56610ed5af779d032d577e4b30acf74a4d4219b5dd40e4924", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/filler/original/05727aafef9bae49/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "7e7c9719042aa0d3ff005da61c8414d0b6b13ce67e6e0946f1132a92ce009753", "seconds": 5.28, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler word 'uh' but the final intent is clear and matches the original instruction. The user does not change the goal or request a pause/cancellation. The task is executed as intended.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_file_filter", "proposed_instruction": "Delete all .png—no, that should be .jpg files from my file system ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "7b62303e728a28259ea966c5a44790082cd6d91358a479e581206d480e783309", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says .png but immediately corrects to .jpg; the action, scope, and destination folder remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/false_start/false_start_01_file_filter/6abae793e3ccc311/speech.wav", "wer": 0.375, "wav_sha256": "0961c42719e84bfbdba08ca534e03b05ca2e8cadbe6875b0bca94f4d63f356b3", "seconds": 9.08, "spoken_utterances": [ "Delete all .png—um... no, that should be ... .jpg files from my file system ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.png files' but immediately self-corrects to '.jpeg files'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The rest of the instruction, including the target folder, remains unchanged. The final intent matches the original instruction with the corrected file extension.", "heard_transcript": "Delete all .png files. No, that should be .jpeg files from my file system ~/downloads folder." }, "candidate_id": "wav_ecf1abd9914537da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident command ('Delete all .png files'), then immediately self-corrects with a natural-sounding hesitation ('No, that should be...') before smoothly continuing with the corrected instruction. The prosody includes a believable shift in intonation and pacing during the correction, and the final phrase is delivered at a steady, conversational pace. The repair feels spontaneous and integrated into the flow of speech." } ] }, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Delete all .jpg files from my file system folder. I left out a detail: do this only in the ~/downloads folder.", "long_pause_offsets": [ 50 ], "source_sha256": "8cec0f7d6753fbc063a8849360a821ed24502132fd9fbf0840fdab57a2f855ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint, narrowing the scope from the general file system folder to the specific ~/downloads folder. The action and file type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/addition/addition_02/50e3ae950cd4562b/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "98a201f7cf82519991dd2b6c3c730d9a42fc0b5e887b91ae04dd0206a44857f1", "seconds": 12.08, "spoken_utterances": [ "Delete all .jpg files from ... my file system folder. ", "I left out a detail: do this only in the ... ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all .jpg files from my file system ~/downloads folder.' The speaker then adds a clarifying constraint: 'Do this only in the ~/downloads folder.' This addition refines the scope of the original task, narrowing it to a specific folder. The final intent is to delete .jpg files only from the ~/downloads folder, which matches the original instruction's intent but with a more precise scope. Since the addition is a refinement and not a contradiction, the task is executed as specified.", "heard_transcript": "Delete all .jpg files from my file system ~/downloads folder. I left out a detail. Do this only in the ~/downloads folder." }, "candidate_id": "wav_9776e974237c7c65", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody and a believable pause before adding a restriction. The addition ('I left out a detail') is phrased conversationally, and the correction ('only in the Downloads folder') is smoothly integrated with a slight rise in pitch and emphasis, matching how a person would recall a missing detail. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_file_filter", "proposed_instruction": "Delete all .png files from my file system ~/downloads folder. Wait, delete all .png files, not .jpg files.", "long_pause_offsets": [ 62 ], "source_sha256": "1431762c17cd7370633f5c0cfa712918c8851ddd33ae0f11919cc6ae0b2e057b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the file extension from .jpg to .png, while retaining the action, location, and scope. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/revision/revision_01_file_filter/46545c4b0cac4be6/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "9ab9daa61946e2b3c850b9f9fa1bce0231d8ffb2fa316a995031a3927f06dbd7", "seconds": 15.56, "spoken_utterances": [ "Delete all hmmm... .png files from um, let's see... my file system ~/downloads folder. ", "Wait, delete all um... .png files, not .jpg files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.jpg files' but immediately corrects to '.png files' after a pause and the word 'wait'. The final instruction is a clear revision of the file type to delete. The folder and action remain consistent. The revision criterion is satisfied as the correction is applied to the earlier instruction.", "heard_transcript": "delete all um .png files from um let's see my file system uh downloads folder. wait, delete all um .png files, not .jpeg files." }, "candidate_id": "wav_e0a3e9f918490ae9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with natural fillers ('um'), then self-corrects the file type from '.jpeg' to '.png' in a spontaneous, believable manner. The correction is integrated into the same utterance with a natural pause and intonation shift, showing a plausible moment of realization. The prosody is consistent, and the delivery sounds like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "7f73ad04c2308a732b7c745edff2b0ad8729091d6339e56047be577ed4c25b8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/cancellation/cancel_028/75667405a738556f/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "882603de6e73754f26379a257e67405639d1f6a7ad3a1ec5487cb5890195d28a", "seconds": 14.36, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder.", "Hold on, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to delete .jpeg files, then immediately follows with 'hold on, disregard that instruction, don't carry it out,' which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder. hold on, disregard that instruction, don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_020", "hide_variant_original": true, "source_sha256": "48073c8dd9970f28af6dafd618a94365815b93e1522d4457c2076836b5bd4b4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/ambiguous_halt/halt_020/0fc4062910551b6b/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "2627c95f79843df179eca9f687590ece26f494ee3102cb4e48532fb8a68b40e5", "seconds": 13.16, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder.", "Hold on just a moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly says 'hold on just a moment please' as a final interjection, which is a standard request to pause or wait. There is no cancellation of the task, and the initial request is not completed. The final audible intent is to halt execution temporarily, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder. hold on just a moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ccf4e5ce1cfce68d47d9e7c517adbfbb827dc618ae7fd44385d37192eab91485", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/filler/original/32cf3d9f7d971ae4/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "df2ab298611e39b72906cc5e9970f52294c413ba6ff4b4c9485f81f75b58f913", "seconds": 9.04, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let me think') but does not change the core task. The final intent is to delete .jpeg files from the specified folder, matching the original instruction. The file extension is pronounced as '.jpeg' but is semantically equivalent to '.jpg'. No correction or change of intent is made.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_file_filter", "proposed_instruction": "Delete all .txt—sorry, that should be .pdf files from my file system ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "46ddebb860bce88a0e79f0a63e93704e09f05b696210ce0855c793b7878db805", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects .txt to .pdf within the same utterance, preserving the action, scope, and target directory." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/false_start/false_start_01_file_filter/9c0b72deb95b99c3/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "2a8f9a1e7fb87144c0f8b4541ea4a32ed2b7b3396b3210fb448ea045d7619573", "seconds": 12.08, "spoken_utterances": [ "Delete all .txt—um... sorry, that should be uh... .pdf files from my file system ... ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.txt' but immediately self-corrects to '.pdf'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction matches the reference, so the final task is to delete all .pdf files from the ~/downloads folder.", "heard_transcript": "Delete all .txt um, sorry, that should be uh .pdf files from my file system downloads folder." }, "candidate_id": "wav_3cfe500dfdee515e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('delete all .txt'), immediately self-corrects with a natural 'sorry' and 'that should be', then smoothly continues with the correct command ('pdf files') and specifies the location ('downloads folder'). The prosody includes a brief pause and filler ('um') at the start, and the correction flows without awkwardness or repetition. The delivery is continuous and sounds like a spontaneous, real-time correction." } ] }, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Delete all files from my file system ~/downloads folder. Oh, I forgot to mention, only delete .pdf files.", "long_pause_offsets": [ 57 ], "source_sha256": "565c2d99567bc008795bd328a53f52c18bc94bd77f80db182796fc440acfc773", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad deletion to only .pdf files, preserving the location and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/addition/addition_01/986d69cb586eb7fa/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "24cb742aa17566eb185bc8902a4746c324d77b01f7f237e9790683d1cd4a7953", "seconds": 13.72, "spoken_utterances": [ "Delete all files um... from my file system hmmm... ~/downloads folder. ", "Oh, I forgot to mention, only delete um... .pdf files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all files from the downloads folder. After a pause, the speaker added a constraint to delete only .pdf files. This is a clear addition that narrows the scope of the original request. The final intent is to delete only .pdf files, which matches the reference instruction. The addition is spoken after a longer pause, satisfying the revision/addition type criterion.", "heard_transcript": "delete all files um from my file system uh downloads folder. oh, I forgot to mention, only delete um pdf files." }, "candidate_id": "wav_14e826b6b658ef5a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_folder_reference", "proposed_instruction": "Delete all .pdf files from my file system ~/Documents folder. Let me correct the folder: delete from ~/downloads, not ~/Documents.", "long_pause_offsets": [ 62 ], "source_sha256": "43e60675b9ad1d7184e45acd00f0509373d8fb585cfbee51402807050dc9e78e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder path from ~/Documents to ~/downloads, while retaining the action and file type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/revision/revision_02_folder_reference/bb447d3d100c2f9e/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "cc6ceb9c88da29ae83f01acefae05ef1c754453e5dd62a1cff1d3aaaf931d654", "seconds": 18.32, "spoken_utterances": [ "Delete all .pdf files ... from my file system ... ~/Documents folder. ", "Let me correct the folder: delete from let's see... ~/downloads, not ~/Documents." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives an incorrect folder path ('~/documents') but immediately self-corrects to the intended folder ('~/downloads') after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction.", "heard_transcript": "Delete all .pdf files from my file system ~/documents folder. Let me correct the folder. Delete from, let's see, ~/downloads, not ~/documents." }, "candidate_id": "wav_8e21e2f7898782ea", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "131e160a27bab30e428ba0af2b0b5642e1a6ade39d907ae0d80486bb62b88c41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/cancellation/cancel_009/bc5a66e773ae550d/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "bb7612651c31eb699fb8a853a22159a2bf3c5488115516293995f9afad672d3e", "seconds": 13.48, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder.", "Wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to delete .pdf files from a specific folder, then immediately follows with 'wait, cancel that altogether,' which is a clear and final cancellation of the entire task. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "delete all .pdf files uh from my file system uh let me think tilde slash downloads folder. wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "9751a888897945a301e39506dccab4b8751fa243c2ac7c93976c147fb63cbacb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/ambiguous_halt/halt_025/618210feab81b63e/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "2a18c6fd3b0febd47826e1c84162e8a2393547cc873f843bd728cbcfb2d3cac3", "seconds": 14.0, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder.", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'wait, I need to think for a second,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request to delete PDF files is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "delete all let's see dot pdf files uh from my file system uh let me think tilde slash downloads folder. wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5a83b05_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ee2906fe09910f13c5d1b47b07ad4d07d867b26f8c130e7d87fbf2d153be7bd9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/filler/original/949ac18e17d2844c/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "b1e777b38b98942ce074db6dc50dd36105c93f4ecac39d41b40a049e34d627f9", "seconds": 9.36, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler phrase 'let's see' and a hesitation 'uh', then corrects the initial 'all' to 'all' again, and includes a self-pause 'let me think'. Despite these fillers and hesitations, the final, complete instruction matches the original task. The speaker does not change the action, target, or value. The 'filler' type allows for such hesitations and self-corrections as long as the final goal is clear and matches the original instruction.", "heard_transcript": "Delete all let's see dot pdf files uh from my file system uh let me think tilde slash downloads folder." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject—no, that should be empty subject and body.", "long_pause_offsets": [], "source_sha256": "bd0f788a30873acb9af8f09f82a268fad3372429f5fe8aedbbf3fc1a5b5ab7e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions only 'empty subject' but immediately corrects to 'empty subject and body', which matches the original instruction exactly. The final goal includes both conditions as required." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/false_start/false_start_03_empty_field_condition/51ab03b3747f4b1d/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a5d8d85b4acb0574366e589602809d2a21eae5da283a54182470a0651e11f002", "seconds": 12.28, "spoken_utterances": [ "Delete all my Gmail drafts that have um... empty subject—uh... no, that should be ... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but immediately self-corrects to 'empty subject and body' after a pause and 'no'. This is a false start where the final corrected instruction matches the original reference. The correction is clear and the final intent is fully specified.", "heard_transcript": "deletes all my Gmail drafts that have um empty subject uh no, that should be empty subject and body." }, "candidate_id": "wav_3aad08d04ce65cc5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'empty subject' to 'empty subject and body' with a plausible hesitation ('uh') and a smooth, continuous flow. The correction is integrated into the same utterance without awkwardness, and the prosody reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have empty subject and body.", "long_pause_offsets": [ 28 ], "source_sha256": "6e2e7f0920e89feb1f903dfc366472f1678f7ac71f1a7eb37b8010d612998c9c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'empty subject and body' to the initial request to delete drafts, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/addition/addition_01/8aacecc8c1837025/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "710e938f2a10f842a37ed3713d7850121c26fd4d8686614e2de70fe4e6ffbc78", "seconds": 11.88, "spoken_utterances": [ "Delete all my ... Gmail drafts. ", "Oh, I forgot to mention, only delete drafts that have ... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all Gmail drafts. After a pause and an interjection, the speaker added a constraint: only delete drafts that have empty subject and body. This matches the addition type criterion, where the final goal combines the initial instruction with the added condition. The final intent is fully specified and matches the original reference instruction.", "heard_transcript": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have empty subject and body." }, "candidate_id": "wav_f8eac92e12f96fd7", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Delete all my Outlook drafts that have empty subject and body. Wait, the app should be Gmail, not Outlook.", "long_pause_offsets": [ 63 ], "source_sha256": "4d9eeff6695299eef65003466f1f1107575e9b18a27817ee5c02fa31a623e52b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the rest of the instruction including the target (drafts), condition (empty subject and body), and scope (all my)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/revision/revision_01_app/96593b007d5a456d/speech.wav", "wer": 0.15, "wav_sha256": "99f99d8942c7663962dcdbe9a986810c41bc4c029dce8b855c33a041917990f2", "seconds": 12.44, "spoken_utterances": [ "Delete all my Outlook drafts that have empty subject ... and body. ", "Wait, the app should be hmmm... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail' after a pause, which is a revision. The final intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "to lead all my Outlook drafts that have empty subject and body. Wait, the app should be, Gmail, not Outlook." }, "candidate_id": "wav_be5af2c113082b50", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, coherent request, then pauses naturally with 'wait' to self-correct the application from 'Outlook' to 'Gmail'. The correction is delivered with a believable hesitation and a slight upward inflection on 'huh', followed by a smooth, self-contained revision. The prosody and pacing are consistent with spontaneous speech, and the revision feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "060fcee094fa43169629432f09785992f9364b13195cc02bc1e744dc1c5aa6b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/cancellation/cancel_030/9a1b3c6f645ef241/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "811c86799818f9db780bef7ecf7850a2cec6db3bb43ae5abe905a7ea699bdc68", "seconds": 12.76, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body.", "Wait, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the deletion task but then clearly revokes it by stating, 'Wait, you can drop that task. I no longer need it.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is a complete cancellation, not a pause or a revision.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body. Wait, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "72fe94f82523b48a7c99885ddca48064f1f22fcffa67b862a37f32acbe3e07a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/ambiguous_halt/halt_012/1bf6a6db6e83cbe3/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "07a4ddaf938dbbc2179323e0ff4724813f90626e72ae938cf1482c6cceec9ca1", "seconds": 10.6, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The recording ends with the explicit interjection 'Give me a second,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final audible intent is to pause, not to execute or cancel. The initial request is not completed, but the final control intent is to pause, which matches the expected behavior for this type.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "057e681fd43c97b6f5f3eeb175902e7134e1bedc54568b51478cadd7b8b7d90d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/filler/original/7c458b6a7d1984bb/speech.wav", "wer": 0.2, "wav_sha256": "d68d8652c0d1f24705c98c32cf6dc054f11f9be9c717fc8b81d50ded53bb5096", "seconds": 7.2, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') and a self-repair ('what's the') but does not change the core request. The final audible intent matches the original instruction. The filler words and hesitation do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject—sorry, I meant empty subject or body.", "long_pause_offsets": [], "source_sha256": "fe4fa502b711da9535d615d1508ebab0a516df39398d69fa2306579f002050ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but immediately corrects to 'empty subject or body'. The final intended condition includes both criteria, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/false_start/false_start_03_empty_field_condition/613a902557980793/speech.wav", "wer": 0.3888888888888889, "wav_sha256": "5b3a4502921c8a01ce274f2a247d0f65352c4df4082374b5ba9d70ae52188dfa", "seconds": 7.16, "spoken_utterances": [ "Delete all my um... Gmail drafts that have empty subject—uh... sorry, I meant ... empty subject or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but then self-corrects to 'empty subject or body' after a pause and an apology. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is correctly executed.", "heard_transcript": "Delete all my um Gmail drafts that have empty subject uh uh sorry, I meant empty subject or body." }, "candidate_id": "wav_cc508beaca8adbbb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Delete all my um Gmail drafts that have empty subject' and then self-corrects with 'sorry, I meant empty subject or body' in a smooth, conversational manner. The repair is immediate, the intonation reflects a genuine moment of correction, and the pacing feels spontaneous rather than rehearsed. The filler 'um' and the self-correction are well-integrated into the flow." } ] }, "is_default": true } }, { "task_id": "5e27cd7_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Delete all my Outlook drafts that have empty subject or body. Wait, delete drafts in Gmail, not Outlook.", "long_pause_offsets": [ 62 ], "source_sha256": "9c3af6fc0e4aed8847a1d436922a5ada5be650e652226467e11ce4a3e0a5e048", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the rest of the instruction: deleting drafts with empty subject or body." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/revision/revision_01_app/ee3e962854fd1023/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "0859d69473d928d3888b2504f6cc70dd4750d57b2e92f49713b83dbcc538e60c", "seconds": 12.04, "spoken_utterances": [ "Delete all my ... Outlook drafts that have empty subject uh... or body. ", "Wait, delete drafts in uh, let's see... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the email service as 'Outlook' but immediately corrects it to 'Gmail' after a pause and self-correction. The revision applies the correction to the earlier instruction, preserving the action (delete), the condition (drafts with empty subject or body), and the target (Gmail). The final intent matches the original instruction exactly.", "heard_transcript": "Delete all my Outlook drafts that have empty subject uh or body. Wait, delete drafts in uh let's see, Gmail, not Outlook." }, "candidate_id": "wav_74949f50177104db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then self-corrects with a natural hesitation ('uh') and a plausible revision of the email service from 'Outlook' to 'Gmail'. The correction is integrated smoothly into the same utterance, with a believable shift in intonation and pacing. The pause and filler ('let's see') before the correction sound like genuine recollection, not a rehearsed or robotic edit. The prosody remains consistent and the delivery is continuous, making it sound like a spontaneous, real-time correction." } ] }, "is_default": true } }, { "task_id": "5e27cd7_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "151237c26ff7771b3a02291062552f604476e03b476ae377f0e9d95aa27efa2b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/cancellation/cancel_019/bff7412e722971ee/speech.wav", "wer": 0.0, "wav_sha256": "4ea0800af5d5f2757202973c66ec1c717d2d58fa8fc289ee1bf7616314098c8e", "seconds": 10.6, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body.", "Oh, wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Oh wait, never mind, don't do that.' This constitutes a cancellation of the entire task as per the type-specific criterion. The final interjection is a direct cancellation, not a pause or ambiguous halt.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body. Oh wait, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "84b895f1da36b8bad25576646ee75ef1749d5cbf85b6f39d913fda8148787897", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/ambiguous_halt/halt_014/ad65773cea4711a2/speech.wav", "wer": 0.0, "wav_sha256": "dcc6ed6c6f655b39318af516ea39fdd0680575a947ba00c66b8387f9ab286fc5", "seconds": 9.76, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Can you hold on a second?' as the final interjection, which is a clear request to pause. The initial part of the sentence is a filler and hesitation, not a final directive to execute. The final audible intent is to pause, not to execute or cancel. Therefore, the audio matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c6faa857ec79608c7f0a751ac47db3b72b35eb33ccf4ebfe8ed5648c73624f12", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/filler/original/0dbaa977a5f8748b/speech.wav", "wer": 0.0, "wav_sha256": "4657ec0cd54eb87b63980501e339174d18ab7b77b9044054060de5ca793f4f03", "seconds": 6.32, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let me think') and a pause but does not change the core task. The final intent remains to delete Gmail drafts with empty subject or body, matching the original instruction. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Delete all my Outlook—I should say Gmail drafts that have empty subject and body.", "long_pause_offsets": [], "source_sha256": "3bf3ec0e780a16abaf99e8fd7399a30677f05a4a117722faa1ed33b45561c823", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, action, and all other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/false_start/false_start_01_app/9bcf97a754f3ead6/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "70860c1e0a457bda3fe84b5a9ca1930694430264743ab31f22f21f079ba6e128", "seconds": 10.2, "spoken_utterances": [ "Delete all my uh... Outlook—um... I should say Gmail drafts that have empty subject hmmm... and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Delete all my uh Outlook um, I should say Gmail drafts that have empty subject and body." }, "candidate_id": "wav_ddd6157dbdd2ac71", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('Outlook'), self-corrects to 'Gmail drafts' with a natural hesitation ('I should say'), and continues with a smooth, coherent request. The pacing is conversational, with a believable pause and filler ('hmm') before the final clause. The correction feels spontaneous and integrated into the flow." } ] }, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have an empty subject and body.", "long_pause_offsets": [ 28 ], "source_sha256": "3dd8cd3c7b6d7e48c38a72dc945100fcdaaa5e4cce849f3ede3717395f938249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition about empty subject and body, which refines the initial broad request. The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/addition/addition_01/64f75c2e4a881d69/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "0c5eae76d8d27d0b4c519a392f8c3a80679e044c834c64618a84a64783e50f98", "seconds": 13.24, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts. ", "Oh, I forgot to mention, only delete drafts that have an let me think... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Delete all my Gmail drafts' and then adds a constraint 'only delete drafts that haven't... empty subject and body'. The phrase 'haven't' is likely a self-correction or hesitation for 'have', but the final audible intent is to delete drafts with empty subject and body. The addition refines the initial request, and the final goal matches the original instruction. The hesitation and self-correction do not change the final intent.", "heard_transcript": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that haven't, let me think, empty subject and body." }, "candidate_id": "wav_32d2a56f21c01d3e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, direct command. After a natural pause, they add a clarifying detail ('only delete drafts that have...') with a believable hesitation ('let me think') and a smooth, self-corrected continuation. The prosody is consistent, and the addition feels like a spontaneous afterthought rather than a rehearsed or robotic insertion." } ] }, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_email_state", "proposed_instruction": "Delete all my Gmail sent that have empty subject and body. Change the email state to drafts, not sent.", "long_pause_offsets": [ 59 ], "source_sha256": "befba26b9898734cdee46386f66eda36c0429c13441b0f3fd54203e9fb994dd1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email state from 'sent' to 'drafts', preserving the rest of the instruction including the condition of empty subject and body." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/revision/revision_02_email_state/b33fc32a079b1a29/speech.wav", "wer": 0.05, "wav_sha256": "1ee7d017b1bea45372377ef807caf784bfabed4b0004efd9169703eef346a80f", "seconds": 11.52, "spoken_utterances": [ "Delete all my um... Gmail sent that have ... empty subject ... and body. ", "Change the email state to drafts, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail sent' but immediately corrects to 'Gmail drafts' in the revision. The final instruction is a revision of the original request, changing the target folder from 'sent' to 'drafts'. The condition of empty subject and body is preserved. The additional instruction to 'change the email state to drafts, not sent' is consistent with the correction and clarifies the desired outcome. The final intent matches the revised goal.", "heard_transcript": "Delete all my um Gmail sent that have empty subject and body. Change the email state to drafts, not sent." }, "candidate_id": "wav_7bd1724d4f7f18e5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a filled pause ('um') and a slightly hesitant delivery. The correction ('change the email state to drafts, not sent') is abrupt and delivered with a flat, declarative tone, lacking the natural intonation and flow of a spontaneous thought process. The transition between the initial request and the correction feels more like a post-hoc edit than a genuine self-repair, reducing its realism." } ] }, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "d3325902550232cfae05906c3249cc8715bb1ec67c7cab1c206675b82259761a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/cancellation/cancel_019/f772be97ed91b88a/speech.wav", "wer": 0.1, "wav_sha256": "cdf7bd94c784c8bf86bb256bc581db937b694104da975f9b09680b5068734abe", "seconds": 12.4, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body.", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to delete drafts with empty subject and body, then immediately follows with 'Hang on, never mind, don't do that,' which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body. Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "531eb1e908612a2af25da05bc8e16410b900c91f38941176377a359cb6c54703", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/ambiguous_halt/halt_011/86fe3617dd84d2ce/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "b8600c04e21ec091b0c1ecded41a7018279441ba847cc2153f1ddae7d11f88ab", "seconds": 10.76, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then makes a final interjection 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final audible intent is to pause, not to execute or cancel. No correction or new instruction is given after the pause request.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "5e27cd7_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "133f3b49de7b50eeb508ebe85b15cb02816ad258fe15685c677c7a3a18c77a49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/filler/original/28ffe785b9e1e44d/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "1fbdc6e009f47bef80d4099f44ca71df68fa6cf1bdbd109d3d4f98d4a9bfa0b4", "seconds": 7.36, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Delete all my Gmail drafts that have uh empty subject and body.' The 'uh' is a filler sound that does not alter the task goal. The final intent matches the original instruction, and the filler does not constitute a change in control intent. Therefore, the task is executed as intended.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_email_state", "proposed_instruction": "Send all my scheduled—I meant future-scheduled emails on Gmail right away.", "long_pause_offsets": [], "source_sha256": "4a67e7799688d31814425e9abd555e1da658a919251730b8a316dbc3c54daf3c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'scheduled' to 'future-scheduled' within the same utterance, clarifying the intended scope of emails to be sent. The app (Gmail) and action (send right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/false_start/false_start_01_email_state/34361990f2f331d9/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "6ed6a665684eb7c553e101508dbe89dff5318164e6c0f3dce0262026d775d734", "seconds": 9.52, "spoken_utterances": [ "Send all my scheduled—um... I meant hmmm... future-scheduled emails on Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'scheduled' but immediately self-corrects to 'future scheduled' with 'I meant'. This is a false start where the final corrected value replaces the initial mistaken one. The final intent matches the original instruction, and the correction is immediate and clear. Therefore, the task should be executed as per the final corrected intent.", "heard_transcript": "Send all my scheduled um, I meant future scheduled emails on Gmail right away." }, "candidate_id": "wav_d876dac4e7d4f6a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('send all my scheduled um...'), followed by a natural self-correction ('I meant...') and a smooth continuation. The prosody is consistent, the repair is integrated into the flow, and the intonation reflects genuine hesitation and correction. The pacing and rhythm are very natural, with a believable pause and filler ('um') before the correction." } ] }, "is_default": true } }, { "task_id": "6a5e690_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Send all my future-scheduled emails on Gmail. Oh, and one more thing: send them right away.", "long_pause_offsets": [ 46 ], "source_sha256": "7761cab5df22c696df33843ffe5be85a41746b923803ca44b67d3cc62f5569a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the urgency condition 'right away' to the initial instruction, which is consistent with the original goal and does not contradict or remove any existing requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/addition/addition_03/e34eb3d023366eed/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4b43b684c6e31949f1499ee9dfef5b67ef74b3aca8ef478be6867be888ddaacd", "seconds": 13.12, "spoken_utterances": [ "Send all my uh... future-scheduled emails on let's see... Gmail. ", "Oh, and one more thing: send them uh... right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'future scheduled emails on let's see, Gmail' and then adds 'send them uh right away' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final intent is to send the emails immediately, which is consistent with the reference instruction.", "heard_transcript": "Send all my uh future scheduled emails on let's see, Gmail. Oh and one more thing, send them uh right away." }, "candidate_id": "wav_18de1b7cae2172a9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear initial request, includes a natural hesitation ('uh') and a plausible self-correction ('let's see'), then adds a new detail ('send them uh right away') after a short pause. The prosody is conversational, with appropriate stress and pacing. The addition feels spontaneous and integrated into the flow, not tacked on." } ] }, "is_default": true } }, { "task_id": "6a5e690_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_016", "hide_variant_original": true, "source_sha256": "0fdd17f481aafaa033dbb5761502bf9ed1a2bc82a5752527a7a0a2fc45eeb0cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/cancellation/cancel_016/fcd5145c3acc0d60/speech.wav", "wer": 0.0, "wav_sha256": "afd22bded9b1685aa351ea96403dcacbae398a19c22c11d6008c7c77bdbc9d44", "seconds": 10.68, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away.", "Hold on, you know what? Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues the task to send all future-scheduled emails on Gmail, then immediately corrects themselves by saying 'Hold on. You know what? Cancel that.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. Therefore, the final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Send all my future scheduled emails on Gmail right away. Hold on. You know what? Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "871ff112f63674c12fa516cd7adf01fbcef86589f628b8f74109d126893198d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/ambiguous_halt/halt_031/c477fcc29f56a996/speech.wav", "wer": 0.0, "wav_sha256": "37ba48f6160ce242fc33ffa7ac9c4d79cbe9a3723f8c967bb9f16af3c1aec571", "seconds": 9.44, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or hold. The final interjection indicates a temporary halt, not a cancellation or a new instruction. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause.", "heard_transcript": "Send all my future scheduled emails on Gmail right away. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3ecef42de2604667d6406be2709ad5fa20a9072b7b6c5cefc7536d4be7f1276c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/filler/original/d524167b16f5877e/speech.wav", "wer": 0.0, "wav_sha256": "aaf98f057adb4cf0c4edd14311deaeff3a8873749318226c3e59d3b225828e31", "seconds": 5.76, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Send all my future scheduled emails on Gmail right away.' There are no hesitations, fillers, or self-corrections that alter the task. The final intent matches the original instruction, and the 'filler' type allows for minor hesitations or backchannels that do not change the goal. Since no such elements are present, the task is executed as stated.", "heard_transcript": "Send all my future scheduled emails on Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook—sorry, I should have said Gmail right away.", "long_pause_offsets": [], "source_sha256": "40b3935b3f45195f71e4b5b0cd8097de2e83e61afa8cd3ded61f0b7bb0d1ccc2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, scope, and urgency. The final intended app is Gmail." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/false_start/false_start_02_app/4a88ca0fa2b70803/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "9093304fb2120ca2a1054dfb181612754ff7faf49f95eecc7ec72e55d4202e69", "seconds": 8.56, "spoken_utterances": [ "Send all my ... future-scheduled emails on Outlook—hmmm... sorry, I should have said ... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Send all my future scheduled emails on Outlook. Sorry, I should've said Gmail right away." }, "candidate_id": "wav_e592307641e8ef8e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request, then pauses with a thoughtful 'hmm' before a natural self-correction. The correction is smoothly integrated with 'sorry I should have said' and a clear rephrasing of the intended task. The prosody and pacing feel spontaneous and conversational, with a believable moment of reconsideration." } ] }, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future-scheduled.", "long_pause_offsets": [ 40 ], "source_sha256": "8c8860353e97f886a01a020700735667e43fdb92c5c3a401083edea1ad740c17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'future-scheduled', which narrows the initial broad request to only those emails. The app (Gmail), action (send), and timing (right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/addition/addition_01/717d53e3c09c09e5/speech.wav", "wer": 0.0, "wav_sha256": "c93974f48fbe3746da596ea5963c7aef4c3edadd15b9b97d27bfe6fb922cd78d", "seconds": 10.96, "spoken_utterances": [ "Send all my emails on ... Gmail right away. ", "Oh, I forgot to mention, only send emails that are ... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to send all emails on Gmail right away. After a pause and an interjection, the speaker added a constraint to only send 'future scheduled' emails. This is a clear addition that narrows the scope of the original request. The final intent is to execute the original task with the added constraint, matching the expected behavior for an 'addition' type.", "heard_transcript": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future scheduled." }, "candidate_id": "wav_5f2a56cf9d080fe2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial command with natural prosody. After a brief pause, they add a restriction ('only send emails that are future scheduled') in a way that sounds like a spontaneous afterthought, with a slight shift in intonation and a natural, conversational tone. The flow feels like a real person remembering a detail mid-sentence, making it highly realistic." } ] }, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later. That should be 'right away', not 'later'.", "long_pause_offsets": [ 52 ], "source_sha256": "5935e45afbd045ba484e1be094f8ee5c5c254f25e263a24ef3ecc46cbf3e09bb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time instruction from 'later' to 'right away', while retaining the app (Gmail), action (send), and target (all future-scheduled emails)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/revision/revision_03_execution_timing/31d9bdb72f64c983/speech.wav", "wer": 0.0, "wav_sha256": "fd26ff4ab490dc317ec0779d6df27da11620050b21cdc7959d75050de036b9d4", "seconds": 12.04, "spoken_utterances": [ "Send all my future-scheduled emails on um, let's see... Gmail later. ", "That should be '... right away', not 'later'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail later' but immediately corrects themselves with 'That should be right away, not later.' This constitutes a revision where the final intent is to send the emails right away, which matches the original instruction. The correction is clear and unambiguous, and the final control intent is to execute the revised task.", "heard_transcript": "Send all my future scheduled emails on um, let's see, Gmail later. That should be right away, not later." }, "candidate_id": "wav_5727b6f4646ccee0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a clear intent, uses a plausible filler ('um') and a self-correction ('let's see'), then smoothly revises the original request from 'Gmail later' to 'right away'. The correction is integrated into the same utterance with natural intonation and pacing, showing a believable moment of reconsideration without any robotic or disjointed transitions." } ] }, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "86ddf384a238caa83338c9775e030a23841962c235f68f74d01e30f59d9e637a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/cancellation/cancel_024/4589b7deb9ee15ab/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "0bf42247d69871c481c5cbe2401b954af67ccb0238b8ad2fd97b0bb3f33b0efe", "seconds": 13.16, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away.", "Oh, wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send scheduled emails, but immediately corrects themselves with 'Oh wait, oh never mind, cancel the whole thing.' This constitutes a clear cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away. Oh wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "12627eafac1b42f7308b386d51b0771619bcffe1ae6db281c36a237f8bfe5f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/ambiguous_halt/halt_036/c760b059785a389c/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d7e37477c9624e825a51a29daaac7468c0b7cd312bf692a5590c34298d59b70d", "seconds": 12.4, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second, before you go ahead,' which is a clear request to pause the task. This matches the 'ambiguous_halt' type, where the final intent is to halt execution temporarily. The initial instruction to send emails is not canceled, but the speaker explicitly requests a pause before proceeding.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "519e494ae9064ebcdb5613f44d491dddd5ed1e60eb817442dbc3c75f25b32e35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/filler/original/529301c9ecc21dfe/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "f3283c3efd59aad61ef793589be10adc5fff87908342635dd364328c89a66618", "seconds": 8.08, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core task. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored, and the remaining task meaning must match the original instruction. The final intent is clear and matches the reference.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook—sorry, I meant Gmail right away.", "long_pause_offsets": [], "source_sha256": "3a7cec03f8df2405b2334e2fad0c88b06f9429f9c31a2637f9ce4b9e57c32f81", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, scope, and urgency." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/false_start/false_start_02_app/e3f23fc2c2011317/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a9dc60ee3b0e20ebcf01365d18deb28ed5a7f8a36357aedd3b1af5abb2e33120", "seconds": 9.88, "spoken_utterances": [ "Send all my um, let me think... future-scheduled emails on Outlook—um... sorry, I meant Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail' after a pause and apology. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Send all my um, let me think, future scheduled emails on Outlook um, sorry, I meant Gmail right away." }, "candidate_id": "wav_4385a5b86a16c020", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. It begins with a filler ('um') and hesitation ('let me think'), then makes a clear false start ('Outlook'), immediately self-corrects to 'Gmail', and finally adds 'right away' as a follow-up clarification. The flow is continuous, with realistic pauses and intonation shifts that mimic genuine thought processing. The correction is embedded within the utterance, not a separate sentence, making it sound spontaneous and credible." } ] }, "is_default": true } }, { "task_id": "6a5e690_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Send all my emails on Gmail right away. Oh, right, these are only the emails that are future-scheduled.", "long_pause_offsets": [ 40 ], "source_sha256": "5515efc6be5c9fd57dbf96a57e7f33447e00edffb46769fb9e80f0e28b88655f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'future-scheduled' to the initial broad request 'all my emails', narrowing the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/addition/addition_01/f7ea983b8c0a865b/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "6e4787f913ebbf4514d71dbdcae7e0757986628a70170ab37fb4823f4f08d64e", "seconds": 14.52, "spoken_utterances": [ "Send all my emails on uh, let me think... Gmail right away. ", "Oh, right, these are only the emails that are uh, let's see... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send all my emails on Gmail right away' but then corrects themselves with 'Oh right, these are only the emails that are future scheduled.' This is a clear addition that narrows the scope of the original instruction. The final intent is to send only the future-scheduled emails, which matches the original instruction with the added constraint. Therefore, the task is executed with the corrected scope.", "heard_transcript": "Send all my emails on uh let me think Gmail right away. Oh right, these are only the emails that are uh let's see future scheduled." }, "candidate_id": "wav_ec5a7a14fe596663", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous thought process with a believable 'let me think' pause, followed by a self-correction ('Oh right') and a clear, additive clarification ('These are only the emails that are... future scheduled'). The prosody is conversational, with appropriate intonation and pacing. The addition of the detail about 'future scheduled' emails feels like a remembered restriction, not a forced insertion, and the flow is continuous and pragmatic." } ] }, "is_default": true } }, { "task_id": "6a5e690_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_email_state", "proposed_instruction": "Send all my draft emails on Gmail right away. Wait, send all my draft emails on Gmail right away, not future-scheduled.", "long_pause_offsets": [ 46 ], "source_sha256": "1ac2be659e498a73bb555ff37ea91bfd094a62ef0b50f4bbd0f48922837870c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the action should apply to draft emails, not future-scheduled ones, which removes the original ambiguity and aligns with the final instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/revision/revision_01_email_state/2467679f8e593208/speech.wav", "wer": 0.0, "wav_sha256": "f8a12d00def460b3e39d66799d6a082a5ea20f1468fb1c9d04eb47d9d95b4158", "seconds": 15.96, "spoken_utterances": [ "Send all my uh... draft emails on um... Gmail right away. ", "Wait, send all my draft emails on Gmail right away, not ... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'draft emails' but then corrects to 'not future scheduled', which clarifies the intended scope. The revision narrows the original instruction from 'future-scheduled emails' to 'draft emails' while excluding future-scheduled ones. The final intent is a clear revision of the original instruction, and the correction is applied as per the revision type criterion.", "heard_transcript": "Send all my uh draft emails on um Gmail right away. Wait, send all my draft emails on Gmail right away, not future scheduled." }, "candidate_id": "wav_03a119f8df112a33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a slightly mumbled phrase ('send all my uh'), then corrects to 'send all my draft emails on Gmail right away'. The correction is immediate and smooth, with a clear shift in intonation and pacing. The speaker then adds a clarifying revision ('not future scheduled') with a natural pause and a slightly lower pitch, indicating a self-correction. The prosody is consistent, and the flow feels spontaneous and conversational." } ] }, "is_default": true } }, { "task_id": "6a5e690_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "736a85264a13200cf9561ac3fbac4f3409e7a975b1e8d75e0c5194170ca3a36d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/cancellation/cancel_024/ea627da849fe6319/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "bb0c6e0d530400ef18bb7f355ea48503ececb28882116dfdd85d70b2e99c374b", "seconds": 16.2, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on um, let's see... Gmail right away.", "Wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a request to send scheduled emails, but immediately corrects and cancels the entire task with 'Wait, oh never mind. Cancel the whole thing.' This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send all my future scheduled emails on emails on um, let's see, Gmail right away. Wait, oh never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6a5e690_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "88503b9932cdbc06f7ca16996b7b4e138423bddc4764fd6379b845489a039da8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/filler/original/201f373f1592f364/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "068cb34f9bbffd7b79ed896ad00930eaa1e9bfbcf762c6a0656bd8b42543dfc7", "seconds": 11.16, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on um, let's see... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the platform as 'emails' but immediately self-corrects to 'Gmail'. The phrase 'let's see' is a filler and does not alter the final goal. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Send all my future scheduled emails on emails on um, let's see, Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_reference", "proposed_instruction": "How many people follow the artist of the last played song—sorry, I meant currently playing song on Spotify?", "long_pause_offsets": [], "source_sha256": "f35346b96a952f048f28da97aa42341519decb15038514885f3b81efeb52a734", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately corrects to 'currently playing song'. The final intended query matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/false_start/false_start_01_song_reference/35dab78ab6a62a4d/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "da6e4a713efd73a605cdd8e529d04705d93728ab26e51a23ce38a59cb07b9976", "seconds": 10.2, "spoken_utterances": [ "How many people follow uh... the artist of the um... last played song—ah... sorry, I meant currently playing song on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately self-corrects to 'currently playing song'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved otherwise. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow uh the artist of the um last played song that, sorry, I meant currently playing song on Spotify?" }, "candidate_id": "wav_937d012999e6e825", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous false start with 'last played song' followed by a self-correction to 'currently playing song'. The filler words 'uh' and 'um' are placed at plausible points of hesitation, and the correction flows smoothly without awkwardness. The prosody is consistent, and the overall delivery sounds like a genuine, unscripted attempt to formulate the request." } ] }, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to mention, it's about the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "f1734efda5df0b51f6412b1336bd06d73b6924db7eb0487a9c5ef1fe95b6b87f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context 'of the currently playing song' to the initial question, completing the reference to the correct artist. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/addition/addition_01/56385a454ad0d202/speech.wav", "wer": 0.04, "wav_sha256": "497896e09aef2c147737f81a588e1bedd45d23fbfb5e7cc6d262dcdea54e4c64", "seconds": 13.8, "spoken_utterances": [ "How many people follow let's see... the artist on Spotify? ", "Oh, I forgot to mention, it's about the artist of the ah... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the artist on Spotify, then adds a clarifying detail about the 'currently playing song' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added information. The final intent is fully aligned with the original instruction.", "heard_transcript": "How many people follow, let's see, the artist on Spotify? Oh, I forgot to mention, it's about the artist of the uh currently playing song." }, "candidate_id": "wav_dfb0e4a71fd83dda", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_song_reference", "proposed_instruction": "How many people follow the artist of the latest released song on Spotify? Wait, I meant the artist of the currently playing song, not the latest released song.", "long_pause_offsets": [ 74 ], "source_sha256": "b57e284e4313fc6138f1733a7cf24376935da64cd5ff7ed8f4f65369f2851b50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'latest released song' to 'currently playing song', preserving the rest of the query including the action and app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/revision/revision_01_song_reference/0f348da374122483/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "9860ec223f6c8be0115972c194ec20d8e94c7afe7d7bee021660583802cc27dd", "seconds": 14.68, "spoken_utterances": [ "How many people follow uh, let me think... the artist of the latest released song um... on Spotify? ", "Wait, I meant the artist of the currently playing song, not the latest released song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'latest release song' but immediately corrects themselves with 'Wait, I meant the artist of the currently playing song, not the latest release song.' This constitutes a revision where the final intent is the corrected version. The original instruction matches this final intent, so the task is executed as revised.", "heard_transcript": "How many people follow uh let me think the artist of the latest release song um on Spotify? Wait, I meant the artist of the currently playing song, not the latest release song." }, "candidate_id": "wav_1b8fc9b7ffc51537", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh, let me think'), formulates a slightly incorrect initial request ('latest release song'), then self-corrects with a plausible revision ('Wait, I meant the artist of the currently playing song, not the latest release song'). The correction is integrated smoothly into the same utterance with a believable shift in intonation and emphasis, demonstrating a realistic thought process and repair." } ] }, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "003c3a608518a0a331d084f7d1a3de7256d6338dcd80ecc2458ee48d987d1434", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/cancellation/cancel_026/e88a2f69855ef378/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "5ba3d3b2af13d879b138483e537b794768c8eafffdcaf7628649b3bbdbda5dc1", "seconds": 11.56, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?", "Hold on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial request. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify? Hold on, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "f703e1c35d678752a396dc125a5a5a3536f3d614bd5075b630c084abf205e5a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/ambiguous_halt/halt_034/66f011b93e69cd83/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d0095f6503bde23b677d58245da046c38e45c4f1944497050ca7c40c021afb92", "seconds": 10.44, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or resumption. The initial request is not executed, but the intent to pause is audible and final.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify? Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e44e996850acbcf17a821dafc35733a1aba600bc27a648a90ab85b4d347582fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/filler/original/07b30797f8b164f1/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "5118bfef84536ac5b5561a2a5c11b58ee30dcd8fb0564c97c53f2483a87cb527", "seconds": 6.4, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker hesitated with 'uh' but did not change the core request. The final intent matches the original instruction, and the hesitation is a filler, not a change in control. The task is to execute the original request.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music—no, that should be Spotify?", "long_pause_offsets": [], "source_sha256": "2b21fadb87c72b9cda40c0e8807133b2e608a9bdbc6213df5969f1732a817b2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/false_start/false_start_02_app/9eae7d234ec4c6bc/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "54437956d65fa25b074c6208c7b6c15df950e8be5b085c1c49469aa3b7997acd", "seconds": 9.16, "spoken_utterances": [ "How many people follow uh... the artist of the currently playing song on Apple Music—um... no, that should be uh... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow uh the artist of the currently playing song on Apple Music? Um, no, that should be uh Spotify." }, "candidate_id": "wav_01d0d93cc226e46d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then self-corrects 'Apple Music' to 'Spotify' with a plausible false start and immediate repair. The intonation and pacing are consistent with spontaneous speech, and the correction flows smoothly within the same utterance without awkwardness or hesitation." } ] }, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to say, it's the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "9727f2c5b16ae95768910e3d2ea734c95ef89a6d2a54743ba0fcc95ab1b906d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specification 'of the currently playing song' to the initial broad request about the artist, completing the intended target." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/addition/addition_01/ac10c0f368885605/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "1f68bfad70c352f630f5ab1c23ed36761ed001334c81319007173ca23906f66f", "seconds": 11.52, "spoken_utterances": [ "How many people follow the artist uh... on Spotify? ", "Oh, I forgot to say, it's the artist of the ... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, missing the crucial qualifier 'of the currently playing song'. The speaker then adds this missing constraint after a pause, completing the original instruction. The addition is clear and directly resolves the ambiguity, resulting in a complete and executable task that matches the reference. The type is 'addition', and the final intent is fully aligned with the original instruction.", "heard_transcript": "How many people follow the artist uh on Spotify? Oh, I forgot to say, it's the artist of the currently playing song." }, "candidate_id": "wav_91a2b83002b33cbc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then hesitates with 'uh' before continuing. The addition of 'Oh I forgot to say it's the artist of the currently playing song' is delivered with a believable, slightly self-correcting intonation and a natural pause before the clarification. The flow feels spontaneous and conversational, with the filler and follow-up fitting seamlessly into the task formulation." } ] }, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_song_reference", "proposed_instruction": "How many people follow the artist of the latest played song on Spotify? Wait, I meant the artist of the currently playing song, not the latest played song.", "long_pause_offsets": [ 72 ], "source_sha256": "c1b8758e5f734e55359b62ab59e2de1b9721ade8a7b72eaa30ed46501dbfb70f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'latest played song' to 'currently playing song', which is the same as the original instruction. The correction clarifies the intended reference, and the final goal matches the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/revision/revision_01_song_reference/5e6231a0952dc391/speech.wav", "wer": 0.0625, "wav_sha256": "0bb5612e3ed25a2d62d22b9d65054e4e1dc9e9fe379900dca631f53956a0d120", "seconds": 14.4, "spoken_utterances": [ "How many people follow hmmm... the artist of the latest played song on Spotify? ", "Wait, I meant the artist of the let me think... currently playing song, not the latest played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'latest played song' but immediately corrects themselves to 'currently playing song' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to ask about the currently playing song, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "How many people follow, um, the artist of the latest played song on Spotify? Wait, I meant the artist of the, let me think, currently playing song, not the latest played song." }, "candidate_id": "wav_07013d997d225ad3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' and 'wait,' followed by a self-correction that is smoothly integrated into the flow. The revision from 'latest played song' to 'currently playing song' is delivered with a believable change-of-mind intonation and a brief pause ('let me think') that sounds spontaneous. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "0827f2d32123c9ad4b81f335a8535bd15c1e1465c5980912a602c1662cd3ed58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/cancellation/cancel_009/7679f91f7d33b552/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "4369f2b771b5cd20f7b017fcaad2d8f757db3f130ee7ba293b63035d587de01f", "seconds": 9.84, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately follows with 'Oh wait, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify? Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "08175044958ad991186731486f312dc6743eb258e67daaac8f8dcdb9f37f24e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/ambiguous_halt/halt_039/a36d947bd5bf6195/speech.wav", "wer": 0.05, "wav_sha256": "7bd905f04512a76725322850c074ca7ab6840036cc279b17d568b1e7c4ea9113", "seconds": 9.4, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, give me a second first,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial request is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "58cfa44345fee02b95be3551e1b2976cba6c6ec4d9b892305ec7df5068adec07", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/filler/original/4b845533ea363ba7/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "30ee3b0aa7819a0bc952d0ef74126d2ff688f68b15ee42a7d3962bcbfb622984", "seconds": 5.36, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'How many people follow' and hesitates with 'um' before continuing the original instruction. The hesitation and filler word do not alter the task goal. The final intent matches the original instruction, and the control intent is to execute the task.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music—sorry, that should be Spotify?", "long_pause_offsets": [], "source_sha256": "168d93cc7ea2a47663e0e06f07d9172ecf4c56b2ceb5070261e69fc2f7f9e603", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/false_start/false_start_02_app/54c40f2477776c74/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "2f53cde40c13587e990d15c294871620ef96ba8203d23599fbdae715dadd5463", "seconds": 9.92, "spoken_utterances": [ "How many people follow the artist of the currently playing song uh... on Apple Music—um... sorry, that should be Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Apple Music um sorry, that should be Spotify." }, "candidate_id": "wav_ea4e83ff0e53134d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a clear and natural false start with 'uh' and 'um' fillers, a plausible self-correction ('sorry that should be Spotify'), and a smooth, continuous flow. The prosody is conversational, with appropriate pauses and intonation that suggest spontaneous speech rather than a rehearsed or robotic delivery. The correction is integrated naturally into the utterance, making it highly realistic." } ] }, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to mention, it's the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "c4c1a0a30db3517dc0c5282972744737025f8f3647dc68c5ec2b5b2a32691d13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specification 'of the currently playing song' to the initial question, completing the reference to the correct artist. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/addition/addition_01/c2444c45ade2c956/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "f2ef2bd852db2d83601c433eb3565aab20715ded2b399981c2893625c5a18b48", "seconds": 15.36, "spoken_utterances": [ "How many people follow let's see... the artist uh... on Spotify? ", "Oh, I forgot to mention, it's the artist of the ... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the number of followers of 'the artist' on Spotify, then adds the clarifying detail 'it's the artist of the currently playing song' after a longer pause. This is a clear addition that narrows the scope of the initial request. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many people follow, let's see, the artist uh, on Spotify? Oh, I forgot to mention, it's the artist of the currently playing song." }, "candidate_id": "wav_45a0855307f83c78", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music? Let me correct the app: the question should be about Spotify, not Apple Music.", "long_pause_offsets": [ 80 ], "source_sha256": "4b030d7c037d55342f4c74d12286c1e218a573c41c743dbc58ca877a98e95b37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, aligning with the original instruction. All other components of the query remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/revision/revision_02_app/fd97a8c1588792a6/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "e0d8485200922d740716a2ee8159d9b10614a2475be94d87187e60c3c140d6c8", "seconds": 13.68, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song uh... on Apple Music? ", "Let me correct the app: the question should be about Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects the app to Spotify. This is a revision where the final intent is to query Spotify, matching the original instruction. The correction is clear and the final goal is fully specified.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Apple Music. Let me correct the app. The question should be about Spotify, not Apple Music." }, "candidate_id": "wav_a1844f550f4d652d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects the app name from 'Apple Music' to 'Spotify' in a natural, conversational manner. The correction is smoothly integrated with a filler word ('uh') and a clear, audible rephrasing ('Let me correct the app'). The prosody is consistent, and the pacing reflects genuine thought processing, making the revision sound spontaneous and realistic." } ] }, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "9320583425994037af95ea31be1de123f6f05b589eac1e44de5a5765aac9418a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/cancellation/cancel_025/a3306c27fc13f000/speech.wav", "wer": 0.0, "wav_sha256": "baf69b5bae807009196705a07bfa281e508b6ad5d302682d784122e3f4daa088", "seconds": 14.6, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?", "Oh, wait, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question but then clearly revokes the entire request with 'Oh wait, forget that request, don't do it.' This constitutes a direct cancellation of the task. The final interjection is a complete and unambiguous cancellation, not a pause or a revision. Therefore, the control intent is correctly identified as 'cancel' and the verdict is 'pass'.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify? Oh wait, forget that request, don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "5bade18a62f260e91af90dce0f1af07d466f41f0b4d7ed80425c90d599cfaf46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/ambiguous_halt/halt_016/5c20ff28cecf3727/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "a2e83e80a513f3a32c3498a2a2272d42538d4929ab9e231c012f42447fcb26b9", "seconds": 12.88, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a second jeez', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify just a second jeez" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6bdbc26_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1836c7063558f7f5129c7e02174520f444a309321684d34ef7c3b263b8143156", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/filler/original/fe280ec82867431c/speech.wav", "wer": 0.0, "wav_sha256": "a649383334417c13d645268a564ad2e5cfb6c0535e902ed5f8337675c5f3bca8", "seconds": 9.48, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see', 'let me think') but does not change the core request. The final audible intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored, and the remaining task meaning must match the original instruction.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family and friends—no, I mean roommates and coworkers.", "long_pause_offsets": [], "source_sha256": "2869d11a62d3cef038a6d8803fc66cd17b84a7d874b8498da589246edfc2b419", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'family and friends' with 'roommates and coworkers'; the action, app, and request type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/false_start/false_start_04_requester_group/cf271c8a77dc40f0/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "9caff59366e50c5fc9558ecbd85f042182cb5e19c2ea6a9c3bac4d1358274f03", "seconds": 10.52, "spoken_utterances": [ "Accept all pending ... Venmo payment requests from um... my family and friends—uh... no, I mean roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family and friends' but immediately self-corrects to 'roommates and coworkers'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Accept all pending Venmo payment requests from um my family and friends uh no, I mean roommates and coworkers." }, "candidate_id": "wav_a823d5558a10e6ae", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural false-start pattern. The speaker begins with 'family and friends', then self-corrects with 'no, I mean roommates and coworkers' using a believable shift in intonation and a clear, spontaneous-sounding revision. The initial phrase is fully formed, the pause before the correction is natural, and the correction flows smoothly without awkwardness. The prosody and rhythm are consistent with a person recalling and adjusting a detail in real time." } ] }, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Accept all pending Venmo payment requests. I left out a detail: only accept requests from my roommates and coworkers.", "long_pause_offsets": [ 43 ], "source_sha256": "2b67cdbac427b0332b2d92e560ff31ee341de616c85f6aaad2b8184c80fca266", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the initial broad request to only include requests from roommates and coworkers, which matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/addition/addition_02/40547f8e7cbd5988/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "0a81d34b22f12b41808fc62decd851d3a141071521e7f8b3f8769511f3a1b6c1", "seconds": 11.8, "spoken_utterances": [ "Accept all uh... pending um... Venmo payment requests. ", "I left out a detail: only accept requests from my roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Accept all pending Venmo payment requests'. The speaker then adds a clarifying constraint: 'Only accept requests from my roommates and coworkers.' This addition narrows the scope of the original request, and the final intent is to execute the original task with the added constraint. The final control intent is to execute the combined task.", "heard_transcript": "Accept all uh pending um Venmo payment requests. I left out a detail. Only accept requests from my roommates and coworkers." }, "candidate_id": "wav_0fb8dab50fea7826", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding false start ('accept all uh pending um'), followed by a clear self-correction ('I left out a detail') and a smooth, additive revision that specifies the restriction ('only accept requests for my roommates and coworkers'). The prosody is conversational, with appropriate pauses and filler words that reflect genuine thought processing. The correction is integrated into the same utterance, making the flow feel spontaneous and realistic." } ] }, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family. I misspoke about the requester group: it should be roommates and coworkers, not family.", "long_pause_offsets": [ 58 ], "source_sha256": "3e75983dfcb37c004c74bea4751a7ce972db8694d1ab047ec81afee148238c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the requester group from 'family' to 'roommates and coworkers', while retaining the action, app, and request status. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/revision/revision_04_requester_group/cfb5ee89520f5c7b/speech.wav", "wer": 0.1, "wav_sha256": "5af535e07530be7c80b11ad89f81bda4d030e93fc831af054167c9b523448d4e", "seconds": 15.44, "spoken_utterances": [ "Accept all uh, let's see... pending Venmo payment requests from uh, let me think... my family. ", "I misspoke about the requester group: it should be roommates and coworkers, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately corrects themselves with 'I misspoke about the requester group. It should be roommates and coworkers, not family.' This is a clear revision where the final intent is to accept requests from 'roommates and coworkers,' matching the original instruction. The revision is spoken after a pause and explicitly corrects the earlier value, satisfying the revision type criterion.", "heard_transcript": "Accept all uh let's see pending Venmo payment requests from uh let me think my family. I misspoke about the requester group. It should be roommates and coworkers, not family." }, "candidate_id": "wav_2ecd656115ebe846", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker starts with a clear initial request, then uses a believable 'let me think' filler pause before stating 'my family'. The self-correction is smooth and pragmatic: 'I misspoke about the requester group, it should be roommates and coworkers, not family.' The intonation and pacing of the correction sound spontaneous and not rehearsed, with a clear reference back to the earlier mistake. The prosody is consistent and the delivery is continuous without robotic or stilted qualities." } ] }, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_035", "hide_variant_original": true, "source_sha256": "cc3398e891a9af7e9f7d5b3a3eafb3c1d541440d59ad7a89f9c2402a05414977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/cancellation/cancel_035/29e71453de3a9e38/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "fab3c8733b08976364625efe8aa2b079b2fec6d9413adcd23663d1fc6bd1efdc", "seconds": 17.4, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers.", "Hold on, stop working on that and cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hold on, stop working on that and cancel the task.' This constitutes a direct cancellation of the entire task, satisfying the 'cancellation' type criterion. The initial hesitation and self-correction do not alter the final intent.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers. Hold on, stop working on that and cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "69d8556f6604c5c364e9886bf9ba3c7c807759afc78aa0e97b3d52ae549d59cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/ambiguous_halt/halt_040/7c94d6cde9974463/speech.wav", "wer": 0.125, "wav_sha256": "2bbb63f4597028592f20532e33b507f7182617dd53b65a925138a99a5863996e", "seconds": 15.36, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request with 'Accept all pending Venmo payment requests from my roommates and coworkers' but then immediately interjects 'Hang on, just a minute,' which is a clear request to pause. The final interjection is not a cancellation or a revision, but a temporary halt. The type is 'ambiguous_halt', and the final audible intent matches the expected control intent of pausing.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bc966380685f328a80c04390eefdfeb0d2d1f60eff132bbf4cdcc9ff16f3d36c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/filler/original/e479d91bd84ccc41/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "a91baf429294a7bf3e55d9feb0849844f4da2640d80d5a3c6d17e7228201b350", "seconds": 11.64, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitations ('um', 'let me think') and pauses but does not change the core request. The final intent matches the original instruction exactly, with no self-correction or modification. The hesitations are filler and do not alter the task goal.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Reject all pending Venmo payment requests from my family—no, I mean friends and roommates.", "long_pause_offsets": [], "source_sha256": "8f4a137c80203c2dabda147d9c2ad68d4e73cb353969df1c7cabaeca4917518b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'friends and roommates', which matches the original instruction. The app and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/false_start/false_start_04_requester_group/a65eb1d3f46dd032/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1ae17a077160da011fcea7afb1e0f8f310a05aeb8f7bc67aaae41bbcb7170b19", "seconds": 9.04, "spoken_utterances": [ "Reject all pending Venmo payment requests from uh... my family—um... no, I mean friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately self-corrects to 'friends and roommates'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Reject all pending Venmo payment requests from uh my family. Um, no, I mean friends and roommates." }, "candidate_id": "wav_481d11b862ec55c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'my family', hesitates with 'um', then clearly self-corrects to 'no, I mean friends and roommates'. The intonation and pacing of the correction sound spontaneous and conversational, with a believable shift in focus. The pause and filler are used appropriately to signal a change of mind, and the overall flow is continuous and realistic." } ] }, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Reject all Venmo payment requests from my friends and roommates. Oh, I forgot to mention, only reject pending requests.", "long_pause_offsets": [ 65 ], "source_sha256": "e679c1f9f391358c7d9dfc8d0c58cda271851d08d17556782afe64db41eb039d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'pending' to the initial request, refining the scope without changing the app, action, or recipients. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/addition/addition_01/04bf816217bfd324/speech.wav", "wer": 0.05, "wav_sha256": "bfd084aca6b63f44e427916c202634d74bfb1bcf3e76a46690ff29861bb2dd6b", "seconds": 11.04, "spoken_utterances": [ "Reject all Venmo payment requests from ... my friends and roommates. ", "Oh, I forgot to mention, only reject um... pending requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Reject all Venmo payment requests from my friends and roommates.' This is followed by a longer pause and an addition: 'only reject pending requests.' This addition narrows the scope of the original instruction to only 'pending' requests. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type is correctly applied and the final goal is fully satisfied.", "heard_transcript": "Reject all Venmo payment requests from my friends and roommates. Oh, I forgot to mention, only reject um pending requests." }, "candidate_id": "wav_f8bea58fa9deaf60", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural pacing and a believable, additive follow-up. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to introduce a restriction, and the correction 'only reject pending requests' is smoothly integrated with a natural pause and filler ('um'). The prosody is consistent, and the addition feels spontaneous rather than forced." } ] }, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_requester_group", "proposed_instruction": "Reject all pending Venmo payment requests from my my family. Change the requester group to my friends and roommates, not my family.", "long_pause_offsets": [ 61 ], "source_sha256": "6a6a343e01b916bf55a57998a786bea646d398dac7d2e918c8caf286094ca4ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the requester group from 'my family' to 'my friends and roommates', which matches the original instruction. The action and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/revision/revision_04_requester_group/c09536f1cdd850a6/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "337436db9b44b508052cf7b9b3ef67ebbbf3a8d275e8063de9fff2665b16d120", "seconds": 14.68, "spoken_utterances": [ "Reject all um, let me think... pending Venmo payment requests from um... my my family. ", "Change the requester group to my ... friends and roommates, not my family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'from my family' but immediately corrects to 'Change the requester group to my friends and roommates, not my family.' This is a clear revision where the final intent overrides the initial incorrect value. The final instruction matches the original reference exactly, so the revision criterion is satisfied.", "heard_transcript": "Reject all um, let me think, pending Venmo payment requests from um, my family. Change the requester group to my, friends and roommates, not my family." }, "candidate_id": "wav_a43c56ecdbef2c33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic flow. The speaker begins with a clear request, uses a believable filler ('let me think'), hesitates with a pause and 'um', then corrects the group from 'my family' to 'my friends and roommates' with a smooth, self-correcting intonation. The revision is pragmatic and the prosody reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "1becefa59432356162028c58d6984707677906941c4b2469803cecc7001455e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/cancellation/cancel_021/81fe76edecf70316/speech.wav", "wer": 0.08, "wav_sha256": "6e520aab6e3550f27d223fe265da3a7feb6d031413408e978c6dcd5b71148257", "seconds": 12.44, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates.", "Oh, wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task but then clearly revokes it with 'Oh wait, never mind. I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion. The final interjection is a complete and unambiguous cancellation, not a pause or a revision.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates. Oh wait, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "d6b94be10f09c1b0540f41c27cb718716271ef8c5c7d92592be595e0dbe04b47", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/ambiguous_halt/halt_001/ad428fca460bdc05/speech.wav", "wer": 0.125, "wav_sha256": "daa13c2cafade1c374589a6cc8f8c6d9cce5f5b2304d5a5cf81aa01f433f4ace", "seconds": 10.48, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait' after a brief hesitation, which constitutes a clear request to pause or hold the current task. The initial instruction to reject payment requests is not canceled, but the speaker explicitly halts execution to reconsider. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "83f3d47e43555e3a2d42ce2ad10ae6f085b22c1f5add44c591edd1462f357e0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/filler/original/3ab1e684dbce76be/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "149a234fdabbe9cdad99ae24ab81ff0c3791cb05ce2e7c893806dcfa24a782a3", "seconds": 7.72, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core request. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family—sorry, I mean coworkers and friends.", "long_pause_offsets": [], "source_sha256": "daedf01420f56fb165f1095b554024274361420fd4a9728f17535974ea0d7839", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'coworkers and friends'. The final intended group matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/false_start/false_start_04_requester_group/cd82dd943de36c2f/speech.wav", "wer": 0.1875, "wav_sha256": "eb369dd6f8f3914cd06eae214c2c95a1d47d06de3a8c3fa80faa3e43a04db141", "seconds": 10.04, "spoken_utterances": [ "Accept all pending ... Venmo payment requests from my family—hmmm... sorry, I mean coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'coworkers and friends'. This is a false start, and the final intent matches the original instruction. The correction is clear and immediate, so the task should be executed as stated in the correction.", "heard_transcript": "Accept all pending Venmo payment requests from my family, whom, sorry, I mean coworkers and friends." }, "candidate_id": "wav_30c2d1eb12fc737a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'family' followed by a self-correction to 'co-workers and friends'. The pause and filler 'hmm' before the correction sound spontaneous, and the correction is delivered with a slight shift in intonation that mimics genuine recollection. The overall prosody is smooth and conversational, with no signs of artificiality or forced pacing." } ] }, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Accept all pending payment requests from my coworkers and friends. Oh, and one more thing: do this on Venmo.", "long_pause_offsets": [ 67 ], "source_sha256": "bedeb3f4624e81757b9c61ba1d1980dbda8183a057026ae795fe17538991e27f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Venmo' without changing any other requirement; the initial instruction already specifies the action, scope, and recipients." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/addition/addition_02/90cc2e6c326ec652/speech.wav", "wer": 0.125, "wav_sha256": "ce998fb3007878dab7c5b237429a4ba47bd144a669140d63851a56b43129f9bf", "seconds": 17.28, "spoken_utterances": [ "Accept all pending payment requests from my hmmm... coworkers and uh, let's see... friends. ", "Oh, and one more thing: do this on um... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to accept all pending payment requests from coworkers and friends. The speaker initially omits the app name, then adds 'Do this on um Venmo' after a pause, which specifies the platform. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, with the addition of the app name.", "heard_transcript": "Accept all pending payment requests from my coworkers and uh let's see, friends. Oh, and one more thing. Do this on um Venmo." }, "candidate_id": "wav_f960bcb675e497c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'uh', 'let's see') at plausible points during formulation. The addition of 'Oh and one more thing' followed by a new instruction ('Do this on Venmo') is a believable conversational addition with a clear shift in focus. The prosody is continuous and the pacing reflects spontaneous thought, including a slight hesitation before 'Venmo'." } ] }, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_request_status", "proposed_instruction": "Accept all completed Venmo payment requests from my coworkers and friends. Make that pending payment requests, not completed ones.", "long_pause_offsets": [ 75 ], "source_sha256": "5d5672f2091c9b78a60a02c0fa639b38e3d6969a5ab1986abdc5ccfcce30128d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial request by replacing 'completed' with 'pending' and explicitly excludes 'completed ones', restoring the original condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/revision/revision_02_request_status/1953ec9c5ec18c3c/speech.wav", "wer": 0.38095238095238093, "wav_sha256": "a9898a2f1bad7303272feda2e2db0e8768e832199d69087a9a49fc34fb67c5aa", "seconds": 13.16, "spoken_utterances": [ "Accept all completed uh... Venmo payment requests from my coworkers and um... friends. ", "Make that pending payment requests, not completed ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'completed' but immediately corrects to 'pending' after a pause, which is the revision type. The final audible intent is to accept pending requests, matching the original instruction. The correction is clear and unambiguous.", "heard_transcript": "Accept all completed uh Venmo payment requests from my coworkers and um friends. Make that pending payment requests, not completed ones." }, "candidate_id": "wav_ec915012759dd689", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction. The initial 'completed' is followed by a smooth, plausible revision to 'pending' without awkward phrasing. The prosody is consistent, and the correction feels spontaneous rather than forced. The speaker maintains a steady pace and tone throughout, with only a brief, natural hesitation before the correction." } ] }, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "35de8e3d5cc5974c750c485382476bd102329cd19027f9bc926105ed78505ac8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/cancellation/cancel_001/a5334f81194f1048/speech.wav", "wer": 0.2, "wav_sha256": "4db760db37334a41381e33d18d61419ba68838db227ea16e5db2786f9bbc2137", "seconds": 11.6, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends.", "Hang on, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then immediately follows with 'Hang on, cancel that,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends. Hang on, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "7e948cfbeefd4a087ec9e3f8f58d6870113c8df859e5fb375a7dce608bc74af1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/ambiguous_halt/halt_022/ada2d851d6b53707/speech.wav", "wer": 0.1875, "wav_sha256": "7350c5c43bf04ea8c14d8d96c612e981d1d023ad51b627f926e693654e58339c", "seconds": 11.84, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The initial instruction is not canceled, and the speaker does not provide a new command or correction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "6ea6792_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "84d5b49f935d18e923503c1030fd7564560eebbbebb1b3e62715aa60ed0ab1f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/filler/original/3109fbe8826ec23a/speech.wav", "wer": 0.2727272727272727, "wav_sha256": "c0b94bb29bd1a5bdc725b84b163a1b51f3ab82fdd659c83e076ed4ca1a256e3d", "seconds": 7.88, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not alter the goal. Since the goal is fully and correctly stated, the task should be executed.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to watermelon—sorry, make that sunshine, all lowercase.", "long_pause_offsets": [], "source_sha256": "7365b608ba015b1046c446e94bebd08d1f28e37c7ac4136768643a2505709fca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'watermelon' with 'sunshine'; the app, action, and case requirement remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/false_start/false_start_02_password/751fcc92a014712b/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "e955a71fbef31f085529f0e1d06acc073a629e96da7c4c98daa81e13b9c1efd9", "seconds": 9.72, "spoken_utterances": [ "Change my um... venmo password to watermelon—hmmm... sorry, make that uh... sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'watermelon' but immediately self-corrects to 'sunshine' after a pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value 'sunshine' should be applied, and the original incorrect value 'watermelon' should be discarded. The final instruction matches the original reference instruction exactly, including the lowercase requirement.", "heard_transcript": "Change my um Venmo password to watermelon. Hmm. Sorry, make that uh sunshine, all lowercase." }, "candidate_id": "wav_977a96d76c9943f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'Venmo' followed by a self-correction to 'PayPal', then a pause and a clear, spontaneous correction to 'sunshine'. The intonation and rhythm are consistent with someone formulating a request in real time, and the correction flows smoothly without robotic repetition. The filler 'um' and the thoughtful 'hmm' are placed naturally, enhancing realism." } ] }, "is_default": true } }, { "task_id": "7574325_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Change my PayPal password to sunshine, all lowercase. Let me correct the app: change your Venmo password to sunshine, all lowercase.", "long_pause_offsets": [ 54 ], "source_sha256": "9e6bc0bfe297ec24671c236b4c1085d63fa0a3764f4ca325e1bc8a4da0d0fa63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from PayPal to Venmo, while retaining the password value and case requirement. The final instruction matches the original instruction in all meaningful aspects." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/revision/revision_01_app/88fd55494750cd00/speech.wav", "wer": 0.04, "wav_sha256": "aa2d4e794a2cd5b05ec057803254d0dc3c0e354f0319c27526b39e7d03fc7687", "seconds": 13.8, "spoken_utterances": [ "Change my uh... PayPal password to sunshine, all lowercase. ", "Let me correct the app: change your uh, let's see... Venmo password to sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a self-initiated repair. The final instruction matches the original reference exactly, including the password and case requirement. The revision type criterion is satisfied as the correction is spoken after a pause and applies to the earlier instruction.", "heard_transcript": "Change my uh PayPal password to sunshine, all lowercase. Let me correct the app. Change your uh, let's see, Venmo password to sunshine, all lowercase." }, "candidate_id": "wav_13509096202aafe3", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "7574325_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "52d7501bcfbbc7411e0e5d9ef16784e1649ef5a7c39459089ca78b061e0b3129", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/cancellation/cancel_032/9130d96425d7f51e/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "83e0384cc536bb851ff6dfc4d56c7950034bdf581656caa88a4f0b3b52b782db", "seconds": 11.92, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase.", "Wait, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially attempts to execute the task but clearly revokes it at the end by stating, 'Wait, I don't need this anymore. Please cancel the request.' This constitutes a direct cancellation of the entire task, satisfying the cancellation type criterion.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine, all lowercase. Wait, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "8d35d58ce041b194cb8db100aac9f166591c88a0320c1ee89eec342deae25dee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/ambiguous_halt/halt_017/99bea85cee5cfc84/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "bafdf8a600dcb8508140d21d03a8e4ec608aa58f040bb4ac9f9c7798782afb16", "seconds": 9.84, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to change the password is not canceled or revised, but the user explicitly requests a hold on execution. Therefore, the final control intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine all lowercase. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5d5a7526da40f1cf8b7307f66b52789090e9c5ab26a694c24770f7a452304fbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/filler/original/41e895cda44a2c80/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "f22b95d87ebe03501d608dd0bbe4dec23f2525f440bf3bde3decf4b0a16157bb", "seconds": 6.48, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not alter the core request. The final instruction clearly matches the original task: changing the Venmo password to 'sunshine' in all lowercase. The fillers do not constitute a change in intent, so the execution criterion is satisfied.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to strawberry—sorry, I meant blueberry, all lowercase.", "long_pause_offsets": [], "source_sha256": "48d80305554af082b56f6506d11e66dbdb38f1bc2e6304435ca79cd7fc58a2df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'strawberry' but immediately corrects to 'blueberry' with 'I meant'. The final password value is blueberry, and the lowercase requirement is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/false_start/false_start_02_password/ac0ab046cc9aa661/speech.wav", "wer": 0.0, "wav_sha256": "afc9c6a36920706c924e63ec9d316ff29df7af17ad4d462712337f62d4e38453", "seconds": 8.12, "spoken_utterances": [ "Change my um... venmo password to strawberry—uh... sorry, I meant blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'strawberry' but immediately self-corrects to 'blueberry'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, including the lowercase requirement.", "heard_transcript": "Change my um Venmo password to strawberry uh sorry, I meant blueberry, all lowercase." }, "candidate_id": "wav_67e8bb0d6e3ff7b0", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Change my venmo password to blueberry,. I left out a detail: make sure the password is all lowercase.", "long_pause_offsets": [ 40 ], "source_sha256": "4203fae4e04b8c9f0a2b7b19d562b10a46523b8615a04eaa6794fa2ba43f0c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the lowercase requirement, which is consistent with the original instruction's specification of 'all lowercase'. The password value 'blueberry' is preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/addition/addition_02/c91ff68817d4ffdf/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "c692b300b8a4c11c41e5a13e5e1481a7647bb5818f19009804e70cabc2ac4609", "seconds": 14.36, "spoken_utterances": [ "Change my venmo password to uh, let's see... blueberry,. ", "I left out a detail: make sure the password is hmmm... all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the password change request, then adds a constraint ('all lowercase') after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final audible intent is to change the password to 'blueberry' and ensure it is all lowercase, which matches the original instruction.", "heard_transcript": "Change my Venmo password to uh let's see blueberry. I left out a detail. Make sure the password is um all lowercase." }, "candidate_id": "wav_0acff4d5c29a29a7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent initial request with a natural pause and filler ('uh, let's see') before stating the password. The follow-up addition ('I left out a detail') is smoothly integrated with a believable self-correction and a pause before completing the thought. The prosody is consistent, and the flow reflects a person recalling and adding information in real time." } ] }, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Change my PayPal password to blueberry, all lowercase. Wait, change my Venmo password to blueberry, all lowercase.", "long_pause_offsets": [ 55 ], "source_sha256": "8b48973fe17383927673107e60516f6d3fc39309b72b0c8e7b08f078b491f8b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the password value and case requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/revision/revision_01_app/571faa2ea19d7aec/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "4fdf8df1fe6c62a6276b24dd6fe2d3510bd5c846eb77728f377bfb478d8e7750", "seconds": 14.64, "spoken_utterances": [ "Change my um... PayPal password to ... blueberry, all lowercase. ", "Wait, change my hmmm... Venmo password to blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and the word 'Wait'. This is a clear revision where the final target app is 'Venmo'. All other constraints (password value 'blueberry', all lowercase) remain consistent. The revision criterion is satisfied.", "heard_transcript": "Change my um PayPal password to blueberry, all lowercase. Wait, change my Venmo password to blueberry, all lowercase." }, "candidate_id": "wav_9eaa2f5b018cf455", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request to change a PayPal password, then self-corrects to Venmo with a natural-sounding 'wait' and a brief pause. The correction is smoothly integrated, and the prosody reflects genuine thought processing. The second part of the utterance is delivered with consistent pacing and intonation, making the revision sound spontaneous and credible." } ] }, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "ff6d25673391d825d64cccbbe97405273b0cab6ae1e4c9ba08a6b132920b7d97", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/cancellation/cancel_025/ecd5590fcc600432/speech.wav", "wer": 0.0625, "wav_sha256": "fe93a588e91fc639791ecf3eefa68a7bf0572e2a35b40bd49d1c2079279cb965", "seconds": 11.96, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase.", "Wait, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to change the Venmo password, but immediately follows with 'Wait, forget that request. Don't do it,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Change my um Venmo password to blueberry, all lowercase. Wait, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "2da7ae0a2ded091665aef4876c301f574c36e506e84371ad42e39d67350bfa42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/ambiguous_halt/halt_039/cab73c78a337d6eb/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "3442592b1a3b6e9df6e19e6571b1342442d203ca699e0d07cb4b389c9475edcd", "seconds": 11.16, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase.", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, give me a second first,' which is a clear request to pause or hold. The initial instruction to change the password is not canceled, but the final interjection indicates a temporary halt. This matches the 'ambiguous_halt' type, where the final intent is to pause, not execute or cancel.", "heard_transcript": "Change my um Venmo password to blueberry, all lowercase. Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c7340b1672dd590775e342451868dd6d1cd7bac146670a5f56d031e7934d0e56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/filler/original/9a0d08095e703b3b/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "9bd3c0ce552a6cc0fa126ad4e3cd008b8de17c42c07da7d163faadd9baa33900", "seconds": 7.2, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler ('um') but continues with the full original instruction without any correction or change in intent. The final audible task matches the original instruction exactly, including the password value and formatting requirement. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Change my um, Venmo password to blueberry, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to watermelon—scratch that, make it pineapple, all lowercase.", "long_pause_offsets": [], "source_sha256": "bea1579250b80e72bb07203ad4a4181ea662df0527e0308606691652865e7128", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces watermelon with pineapple; the app, action, and case requirement remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/false_start/false_start_02_password/11d5e7e5600c37be/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "01317ed9c987afe1358c6286df5eddcafb196dcb9802112c3f643ad1eaef266b", "seconds": 7.76, "spoken_utterances": [ "Change my venmo password to watermelon—um... scratch that, make it ... pineapple, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a wrong password ('watermelon') but immediately self-corrects with 'scratch that' and provides the correct value ('pineapple, all lowercase'). According to the 'false_start' type criterion, the final corrected value must be executed, and the initial incorrect value is discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Change my Venmo password to watermelon. Um, scratch that, make it pineapple, all lowercase." }, "candidate_id": "wav_fff1a9dc81a7471b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly natural revision pattern. The speaker first makes a complete request ('Change my Venmo password to watermelon'), then uses a believable self-correction ('Scratch that, make it pineapple all lowercase'). The intonation and pacing of the correction sound spontaneous, with a clear shift in focus and a pragmatic back-reference to the initial request. The flow is continuous and the change-of-mind is expressed in a way that mimics real human hesitation and reformulation." } ] }, "is_default": true } }, { "task_id": "7574325_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "2d951c8fc625df19754ccc5dfd1fce3e3f03e3e5ceae04896e503ccf439b8abb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/cancellation/cancel_029/6533e910e9037beb/speech.wav", "wer": 0.0, "wav_sha256": "4a9fe401f702d4be1373a1f7a7f969f4c085f84f04beeaa9390a4a491832d7e1", "seconds": 10.96, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to change the Venmo password, but immediately follows with 'Hang on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection confirms the cancellation intent, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "49be5d08a04b9dc91de19558bb2addbc494916cff81086dbf6d265574c433a16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/ambiguous_halt/halt_031/67f907be70e7ead3/speech.wav", "wer": 0.0, "wav_sha256": "91070b7d6002d4ae36095428914a4eea5aa16b28a3517df0c27771b5e5568c9c", "seconds": 10.04, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. This constitutes an ambiguous_halt as per the rubric, where the speaker indicates a need to stop and think without canceling the task. The final interjection is a pause request, not a cancellation or a new instruction.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7574325_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "58c8eee4eddc6b55e5cb90928a50c4dca4a19fb88748b8063b8c31c5616e2694", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/filler/original/1cccb2c89ab55ebf/speech.wav", "wer": 0.0, "wav_sha256": "ba9500df1e7a7fd06384ac32e22ae69a9a919aa71f3a88a777dded8c5ce505f6", "seconds": 6.4, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker clearly states the full instruction to change the Venmo password to 'pineapple' in all lowercase. The 'filler' type allows for minor hesitations or fillers, but the final intent is clear and matches the original instruction. No errors or omissions are present in the final utterance.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 3654328627—sorry, that should be 3654328626 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "cd03732f16fcc735128f9e99d0949fd7ab58b3275d637d5742776d6263569523", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in seven, then corrects it to end in six. The final number matches the original instruction, and all other conditions (message types, action, and intent) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/false_start/false_start_02_phone_number/0cde7f08af45041c/speech.wav", "wer": 0.10256410256410256, "wav_sha256": "1a3ebe0d5b611d005a92cc2b59e62e057b4b76b4aaac48be57f5a5575e1214ca", "seconds": 16.88, "spoken_utterances": [ "All phone text messages and voice messages from uh... three six five, four three two, eight six two seven—ah... sorry, that should be three six five, four three two, eight six two six uh... are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the number as 3654328627, then immediately self-corrects to 3654328626. According to the 'false_start' type criterion, the final intent is the corrected version. The action 'delete them' remains unchanged. The final audible task matches the original instruction after the correction.", "heard_transcript": "All phone text messages and voice messages from uh three six five four three two eight six two seven uh sorry, that should be three six five four three two eight six two six uh are spam, delete them." }, "candidate_id": "wav_852a6be7d6b79118", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects the phone number from '8627' to '8626' with a natural-sounding 'uh, sorry, that should be...' followed by a smooth continuation. The prosody includes a believable pause and filler, and the correction is integrated seamlessly into the flow. The delivery sounds like a real person recalling and fixing a detail mid-sentence." } ] }, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "All phone text messages and voice messages from are spam, delete them. One more detail: these are from the number 3654328626.", "long_pause_offsets": [ 71 ], "source_sha256": "eaae1336a73cf63052f5b11a96821cb08605b54ee9d6206947f952b93534e4a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing phone number, which matches the original instruction's number. The initial utterance omits the number but otherwise matches the final goal. The addition fills the missing value without altering other requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/addition/addition_02/4840821def2a04e9/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "bcc343e83180bdcd99fad618d985a788bded548051db337b630400e8d64e4cad", "seconds": 16.12, "spoken_utterances": [ "All ... phone text messages and uh... voice messages from ... are spam, delete them. ", "One more detail: these are from the number three six five, four three two, eight six two six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to delete spam messages from a specific number. The speaker initially omits the number but then adds it clearly after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added detail. The number is correctly stated as 'three six five, four three two, eight six two six', which matches the original reference number. No other constraints are changed or removed.", "heard_transcript": "All phone text messages and uh voice messages from are spam, delete them. One more detail, these are from the number three six five, four three two, eight six two six." }, "candidate_id": "wav_9c6c78789c1bd92a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to delete spam messages, then adds a specific detail about the sender's number. The addition is delivered with a slight pause and a conversational tone, sounding like a spontaneous recollection rather than a rehearsed or robotic insertion. The prosody is consistent, and the number is pronounced clearly without hesitation." } ] }, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_message_channel_scope", "proposed_instruction": "All phone text messages from 3654328626 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages.", "long_pause_offsets": [ 63 ], "source_sha256": "60989f3298609ddc99d7f40497755150db69d2d469962043f991205d8395d167", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type scope from only phone text messages to include both phone text and voice messages, while retaining the sender number and the delete action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/revision/revision_01_message_channel_scope/ecbe980b1a2b7993/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "2eb9ac09f065563b952db0e00350918739b0f94276f48d51b7491f59f445d046", "seconds": 19.36, "spoken_utterances": [ "All phone text messages from um... three six five, four three two, eight six two six uh... are spam, delete them. ", "Wait, the messages should be um... phone text messages and voice messages, not just phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests deletion of only 'phone text messages' from the number, but then corrects this to include 'phone text messages and voice messages'. This is a revision of the message type to be deleted. The final intent is to delete both types of messages, which matches the original instruction. The correction is clear and applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "All phone text messages from um, 3654328626 uh, are spam, delete them. Wait, the messages should be um, phone text messages and voice messages, not just phone text messages." }, "candidate_id": "wav_88f94dab86b60ffb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'wait' before revising the message type from 'phone text messages' to include 'voice messages'. The correction is smoothly integrated into the same utterance with a believable self-repair ('should be... not just...'), and the prosody reflects a genuine moment of reconsideration. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "9ceef6b0db06463cfd8004dad3cb82795543d50d4a5f37cdfd813ef62ebd18a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/cancellation/cancel_009/c338d42ab1b1862e/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "0300bc2671e163fac5bd8a69b920d37cf36809895c0997fe96429eb131516dde", "seconds": 16.04, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a deletion task, then immediately follows with 'Hang on, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "All phone text messages and voice messages from 3654328626 are spam, um, delete them. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "a43005343635c827df2f7517d0c95c72717778c12c0911af1dae25f3f5722881", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/ambiguous_halt/halt_001/8d2ff83133b3069c/speech.wav", "wer": 0.12, "wav_sha256": "936ba5f35679181d47be590be6b5c8e941efea9e1b41447ca839e1366f031777", "seconds": 15.08, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' after completing the instruction to delete messages. This constitutes a clear request to pause or hold the current action, consistent with the 'ambiguous_halt' type. The final interjection is not a cancellation or a correction, but a directive to stop and reconsider, which matches the expected behavior for this type.", "heard_transcript": "All phone text messages and voice messages from 3654328626 uh are spam, um delete them. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ac81a0cb1a1462fcdb8eea20da8b48aa1a8a15b805453f5bd46672b150765c1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/filler/original/b9cfd095475a6bd7/speech.wav", "wer": 0.125, "wav_sha256": "8bb675f766eaf46e946efd0719df2228192fe58906610e0f5725e986619fa639", "seconds": 12.32, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a brief hesitation ('um') but does not alter the core instruction. The final control request is 'delete them', which matches the original task. The hesitation is a filler and does not change the goal. The final intent is to execute the deletion of messages from the specified number.", "heard_transcript": "All phone text messages and voice messages from 3654328626 are spam, um, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_message_channel_scope", "proposed_instruction": "All phone calls and voice messages—sorry, I meant phone text messages and voice messages from 9294880327 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "96b024a94691ad302cf6338b409e082e5ab469691fc80db6f9f021db223e3a27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone calls' but immediately corrects to 'phone text messages' within the same utterance. The final intended message type is 'phone text messages and voice messages', and the phone number is correctly specified as 9294880327. The action and target remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/false_start/false_start_01_message_channel_scope/480e66cb753ce259/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "38298e3b693177b877a01f57618fbae269835cbb1723ab7919ebf4c4e5cacd50", "seconds": 12.2, "spoken_utterances": [ "All um... phone calls and voice messages—uh... sorry, I meant phone text messages and voice messages from nine two nine, four eight eight, zero three two seven are spam, hmmm... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone calls' but immediately self-corrects to 'phone text messages'. The final intent is clear and matches the original instruction with the correction applied. The type is 'false_start', so the immediate self-correction is preserved and the mistaken value is discarded. All other constraints, including the phone number and action, are consistent.", "heard_transcript": "All um phone calls and voice messages, uh sorry, I meant phone text messages and voice messages from nine two nine four eight eight zero three two seven are spam, delete them." }, "candidate_id": "wav_6f35c977d97706e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural false start ('All um phone calls...'), immediately self-corrects with 'sorry, I meant...', and smoothly continues with the correct request. The correction is integrated into the flow without awkwardness, and the delivery includes a believable hesitation and a natural pause before the final command. The prosody and rhythm are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "All phone text messages and from 9294880327 are spam, delete them. And make sure to include voice messages too.", "long_pause_offsets": [ 67 ], "source_sha256": "0550fcf8b52ba373e57f2ffa2b283ea2b14554a7d0a3a71c8b78968b476bbc1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include voice messages, which was not in the initial utterance but is consistent with the original instruction. The phone number and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/addition/addition_02/d986c6fb8d92e6a1/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "a68c2f43b5e5fb61f4da45ef9a84701ee2bcff47425c1a902cc3b01d72bfaf98", "seconds": 17.36, "spoken_utterances": [ "All um... phone text messages and um, let me think... from nine two nine, four eight eight, zero three two seven are spam, delete them. ", "And make sure to include ... voice messages too." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the reference. The speaker then adds 'And make sure to include voice messages too' after a pause, which is a clear addition to the original request. The final intent combines both the original and added constraints, fully satisfying the 'addition' type criterion.", "heard_transcript": "All um phone text messages and um let me think from nine two nine four eight eight zero three two seven are spam, delete them. And make sure to include voice messages too." }, "candidate_id": "wav_1954a690f4a8eafa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural, spontaneous fillers ('um', 'let me think') that are well-placed and not overused. The delivery has a realistic, hesitant rhythm, and the addition of 'voice messages too' at the end sounds like a remembered detail rather than a rehearsed command. The prosody is varied and the pacing reflects genuine thought process." } ] }, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_message_channel_scope", "proposed_instruction": "All phone text messages from 9294880327 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages.", "long_pause_offsets": [ 63 ], "source_sha256": "5a6d0470d14c9c0be2e34b924bfd5a8bd96f3e094f5413d87bd5e42b7b28d9ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the message type from 'phone text messages' to include 'voice messages', matching the original instruction's scope. The phone number and action (delete spam) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/revision/revision_01_message_channel_scope/19ef18dd0b008b45/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "23367f4f8b9304c035ce4369e07af048980520cfea81a08c10dcfc2af9ee03f1", "seconds": 19.92, "spoken_utterances": [ "All phone text messages uh, let me think... from nine two nine, four eight eight, zero three two seven are spam, ... delete them. ", "Wait, the messages should be ... phone text messages and voice messages, not just phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' but immediately corrects to 'phone text messages and voice messages' after a pause. This is a revision of the message type. The final intent includes both message types, which matches the original instruction. The phone number and action (delete spam) remain unchanged. Therefore, the revision is correctly applied and the task is fully executable.", "heard_transcript": "All phone text messages, let me think, from 9294880327 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages." }, "candidate_id": "wav_ddaf3518202f85ed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), delivers a clear initial request, then self-corrects with a plausible revision ('wait, the messages should be...') that smoothly integrates the missing 'voice messages' and corrects the phone number. The prosody is consistent, the correction is embedded within the same utterance, and the pacing reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "8668665d70f7b6c4e8f1fd9bf0e72c0afc51361eabf54289945f6c8553fd11d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/cancellation/cancel_006/2d6ec7cf264ca9a4/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "c0d5206bd280a5f2960c32f4a136dc6fa8fe9d07b2f8eb529cf1e8c47cacf7b4", "seconds": 16.0, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them.", "Oh, wait, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a complete instruction to delete spam messages from a specific number. However, the speaker immediately follows with 'Oh, wait, please cancel the entire task,' which is a clear and unambiguous cancellation of the prior request. According to the cancellation criterion, the recording must clearly revoke/cancel the task, which it does. The final interjection is a direct cancellation command, not a pause or a request to wait. Therefore, the final intent is cancellation, and the task should not be executed.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam. Delete them. Oh, wait, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "24de9fc93801299518a2ed0f030e69d80648db632370c572ffd73038f0f971f6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/ambiguous_halt/halt_037/590d99201e7aa1d4/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "e349ff664a9ffe1a2f86271c7e392113724581b6a855731aa8c28f739ff39e0d", "seconds": 14.72, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the original instruction to delete spam messages from 9294880327, followed by the explicit final interjection 'Pause for a moment.' This constitutes a clear request to hold or wait, with no indication of cancellation or resumption. According to the rubric, such a final interjection qualifies as an 'ambiguous_halt' where the agent should pause. The control intent is correctly identified as 'pause'.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam. Delete them. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a5da11abfbd8f94d05140473405402224aa9881d640a6e5a9ee87d8b4fbed6a0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/filler/original/f824dab3c492049a/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "7b0114b68d3f599bbe7bf225dddf5a3faf54bd8a883d53395371af5ece6e8624", "seconds": 11.08, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'uh' but continues the original instruction without changing the action, target, or value. The phone number is clearly stated and matches the reference. The final control intent is to execute the deletion task as originally requested.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 5708520673—sorry, that should be 5708520672 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "705b7ea6ec572785efa2d250bfa55e093a9cba4bbf304650e8bc78c592536c99", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in three, then immediately corrects it to end in two. The final number matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/false_start/false_start_02_phone_number/2674626f4c44cd0c/speech.wav", "wer": 0.0, "wav_sha256": "0b662333ecd0a75957c6e9aebd8e70a01a34a121827a94e0ce9d9ebce9724df2", "seconds": 18.36, "spoken_utterances": [ "All phone text messages and voice messages from um, let me think... five seven zero, eight five two, zero six seven three—um... sorry, that should be five seven zero, eight five two, zero six seven two are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in 'three' but immediately self-corrects to 'two'. The final number spoken is 5708520672, which matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion. The task intent remains consistent and executable.", "heard_transcript": "All phone text messages and voice messages from um, let me think, five seven zero eight five two zero six seven three um, sorry, that should be five seven zero eight five two zero six seven two are spam, delete them." }, "candidate_id": "wav_1c6cf6eef3cdb699", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'um' followed by a plausible hesitation ('let me think'), then delivers a clear initial number. The self-correction ('sorry, that should be...') is spoken with a slightly lower pitch and faster pace, typical of a spontaneous correction. The number is corrected from 0673 to 0672, and the sentence flows smoothly into the final command. The prosody and rhythm are consistent with a person recalling and correcting a detail in real time." } ] }, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "All from 5708520672 are spam, delete them. I left out a detail: only the phone text messages and voice messages.", "long_pause_offsets": [ 43 ], "source_sha256": "bd6d000e7dc742a66a94e1f94208b537d304eb91b7d03030cb8459da33a50b21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that limits the deletion to only phone text and voice messages, which is consistent with the original instruction. The original instruction already specified the sender and action, and the followup refines the message type without contradicting any other part." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/addition/addition_02/64d258bcd31093f7/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "f60749cae03a323b157b341c68e380b365b9228b7fb767b88a5c04a0a178315b", "seconds": 17.16, "spoken_utterances": [ "All ... from five seven zero, eight five two, zero six seven two um... are spam, delete them. ", "I left out a detail: only the uh... phone text messages and voice messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the full instruction to delete spam messages from the number. After a pause, they add a clarifying detail that the action should apply only to 'phone text messages and voice messages', which is a narrowing of the original instruction. Since the original instruction already specified 'phone text messages and voice messages', the addition does not change the scope. The final intent remains to delete spam messages from the specified number, and the addition is consistent with the original. Therefore, the task is executed as intended.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, delete them. I left out a detail, only the uh phone text messages and voice messages." }, "candidate_id": "wav_0bd870ac1e8e01c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to delete spam messages from a specific number. After a brief pause and filler ('um'), they add a detail about only deleting phone, text, and voice messages. The addition is smoothly integrated with a natural-sounding 'I left out a detail' self-correction, and the prosody and pacing are consistent with spontaneous speech. The flow feels like a real-time recollection and refinement of a task." } ] }, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 5708520673 are spam, delete them. Let me correct the number: messages from 5708520672, not 5708520673.", "long_pause_offsets": [ 82 ], "source_sha256": "0ef64fd4d8dea5d74550582487cdffc2a5cca8e5865ba048f5ed2c71cb0482b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up revision corrects the phone number from 5708520673 to 5708520672, which matches the original instruction. The action, message type, and deletion directive remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/revision/revision_02_phone_number/c39996c591e87e24/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "67105c80d782d4522201b41082e168d6db4444afd9d0a5b2d9d9b48789c6d606", "seconds": 24.24, "spoken_utterances": [ "All uh... phone text messages and voice messages from five seven zero, eight five two, zero six seven three let me think... are spam, delete them. ", "Let me correct the number: messages from five seven zero, eight five two, zero six seven two, not five seven zero, eight five two, zero six seven three." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong number (5708520673), then immediately corrects it to 5708520672 after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "All uh phone text messages and voice messages from 5708520673, let me think. are spam, delete them. Let me correct the number. Messages from 5708520672, not 5708520673." }, "candidate_id": "wav_5929b561183d0df9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural hesitation ('let me think') and a self-repair that flows smoothly within the same utterance. The correction of the phone number is immediate and integrated without restarting the sentence, showing a realistic thought process. The prosody is consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "e2f7f48a2bdadeea4aa246bb5f18f616ecac858c57b045dbc1b240ef4dd277fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/cancellation/cancel_029/1e10a5ea04ecda8e/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "53dacb82d1fb2875ec6e8a0a9cc98ad4ed43d3ba6c557ee8fb76d3939a6329b8", "seconds": 15.56, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them.", "Hold on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a task to delete messages from a specific number, then immediately follows with 'Hold on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "All phone text messages and voice messages from five seven zero eight five two zero six seven two are spam, uh, delete them. Hold on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "9232d9ceb9214adb09b1b084b14a3f47bc25b0001bb33167179f095e018401ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/ambiguous_halt/halt_040/3b93d7c07de7b475/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "b29606eec566f92eafeae4737addf9a7ab1f16bd1d855321e08c918ee8f77f3e", "seconds": 14.84, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, just a minute,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or provide a new instruction. The initial request to delete spam messages is not revoked, but the speaker explicitly halts execution. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, uh, delete them. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "771d8fc_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ada2b046166947299ba8f7656c381f25f26c258666bd9513547d433cccd24a09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/filler/original/97b83febd2ea4a75/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "845850b4ac20f319abbd135e38e5d390367488cd5e3a820bbf1e168f9751d54c", "seconds": 11.32, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task with a filler word 'uh' but immediately proceeds to complete the original instruction without any correction or change in intent. The final audible command is to delete the messages, matching the original instruction. The filler does not alter the goal, so the task is executed as intended.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, uh, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Place an order for everything in my eBay—sorry, I meant amazon cart and wishlist for my home address.", "long_pause_offsets": [], "source_sha256": "760052ddfd6ec22b02c9ccea195900a6f1c0c46f55f13d66e507b63a16735f5f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'eBay' to 'Amazon' within the same utterance, preserving the action, scope (cart and wishlist), and destination (home address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/false_start/false_start_01_app/a94ba26e59fc2bba/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "cac51569bb73e603c610d42e56aacfb01fc0763ebc469a629c82251c091d1906", "seconds": 12.92, "spoken_utterances": [ "Place an order for everything in my eBay—ah... sorry, I meant um... amazon cart and ... wishlist for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' after a brief hesitation. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('eBay') is discarded. The final instruction matches the original reference instruction in both action and target, with no other changes.", "heard_transcript": "Place an order for everything in my eBay uh uh sorry, I meant um Amazon cart and wishlist for my home address." }, "candidate_id": "wav_75011ad82f143d17", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('eBay'), self-corrects smoothly to 'Amazon', and continues with a natural, flowing delivery. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause and filler 'uh uh' are placed at a believable point of hesitation, and the overall rhythm is that of someone thinking aloud while speaking." } ] }, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Place an order for everything in my amazon for my home address. I left out a detail: include both my cart and wishlist.", "long_pause_offsets": [ 64 ], "source_sha256": "496c7bfa60580f243602f09a0a149101263acdc8f71bb04388e5dfb68654f848", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include both the cart and wishlist, completing the initial instruction which only mentioned Amazon. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/addition/addition_02/a269e270f41245cf/speech.wav", "wer": 0.04, "wav_sha256": "f970338095f6911aa3b3a6db483845d993ed2b583c6e8e4889226dea1c572605", "seconds": 14.08, "spoken_utterances": [ "Place an order for everything in my um... amazon for my home address. ", "I left out a detail: include both my hmmm... cart and wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'everything in my Amazon' but then corrects themselves by adding 'Include both my cart and wishlist.' This is a clear addition that narrows the initial broad request. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Place an order for everything in my um Amazon for my home address. I left out a detail. Include both my cart and wishlist." }, "candidate_id": "wav_e1a634d8bfcc4b36", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'um' before stating the core request. The addition of 'I left out a detail' followed by a clear, smoothly integrated follow-up about including both cart and wishlist sounds like a genuine, spontaneous recollection. The prosody is consistent, and the flow from initial request to addition feels organic and unforced." } ] }, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address. Change the destination to home address, not work address.", "long_pause_offsets": [ 82 ], "source_sha256": "690aa102c2bac723f78836c727b532ae53c8c1e844901b15bbd84f5dc4ffa90b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination from work address to home address, preserving the action, items, and scope of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/revision/revision_03_destination_address/592ee1ed9ea72311/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "0d20c73ca2297dda643067006cc037a10bbdb2cf7123894091ce599446339ec9", "seconds": 15.64, "spoken_utterances": [ "Place an order for everything in my amazon cart and wishlist um... for my work address. ", "Change the destination to hmmm... home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the task for the 'work address' but immediately corrects it by saying 'Change the destination to, um, home address, not work address.' This constitutes a revision where the final intent is to use the 'home address.' The correction is clear and unambiguous, and all other conditions (ordering everything from cart and wishlist) remain unchanged. Therefore, the final audible intent matches the original instruction.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, um, for my work address. Change the destination to, um, home address, not work address." }, "candidate_id": "wav_b194633dd8dd315d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a coherent, multi-part request with natural hesitation and self-correction. The initial request is clear, followed by a plausible pause and filler ('um'), then a revision of the destination. The correction ('change the destination to...') is smoothly integrated with a natural shift in intonation, and the final phrase is spoken with appropriate emphasis. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "922259422f8c8848e68d306bafa03db71670c1580b564ab6fbd294bd112445d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/cancellation/cancel_030/7b5529475493e486/speech.wav", "wer": 0.0625, "wav_sha256": "ebca1a5f72293c0b0d9ec736d12667483a19d6962ad96227fef38a9aea6148bd", "seconds": 19.12, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address.", "Hang on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to place an order, then after a pause and hesitation, explicitly says 'Hang on. You can drop that task. I no longer need it.' This is a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist. Uh, let me think. For my home address. Hang on. You can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "8015757d4ca1f709eea85a472c276721cee5798c0d87b7175275de526d3bc0e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/ambiguous_halt/halt_032/cedec8499a544653/speech.wav", "wer": 0.08, "wav_sha256": "99caeb152b43493c6dfa31283160aeafb07e2cf0a236830947472c0b778e5532", "seconds": 16.8, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Wait, before you do that,' which is a clear request to pause or halt the current action. The final interjection is a pause command, not a cancellation or a new instruction. The initial request is not executed due to the halt, which is consistent with the 'ambiguous_halt' type. No issues are present as the final intent is a valid pause.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think, for my home address. Wait, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "857496a94eac34c46365edba9ea7d54a3498afbd4a7b7bc803ade9906d6e42bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/filler/original/9c26c61d18565efc/speech.wav", "wer": 0.1, "wav_sha256": "780a97c77fc42dda8a3557467bf01e75688aecc43e5f9a9a7eeefd6ed64facb3", "seconds": 12.8, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the correct instruction, hesitates with filler words ('uh', 'let me think'), but completes the original task without changing the action, target, or value. The hesitation is a filler, not a final directive to pause or cancel. The final intent matches the original instruction.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think, for my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my home address—no, that should be work address.", "long_pause_offsets": [], "source_sha256": "4892b9c8e0cad2fec99a8527fc08c834f76433fb50648487905ac2fc44144cd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects it to 'work address'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/false_start/false_start_03_destination_address/38a81afc2cd12661/speech.wav", "wer": 0.0, "wav_sha256": "84f91d23389ca52e64e45f53a1dc3f76cf240d092dec56a77efc8ef32b9911fc", "seconds": 11.48, "spoken_utterances": [ "Place an order for everything in my uh, let me think... amazon cart and wishlist uh... for my home address—um... no, that should be work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately self-corrects to 'work address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for everything in my uh let me think Amazon cart and wishlist uh for my home address um no, that should be work address." }, "candidate_id": "wav_fe9faaa115b48a47", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. It begins with a clear false start ('uh, let me think'), followed by a complete but incorrect initial request ('Amazon cart and wishlist for my home address'), then a believable change-of-mind with a self-repair ('No, that should be work address'). The intonation and pacing of the correction sound spontaneous and conversational, with a natural pause and shift in tone. The filler 'uh' and hesitation are well-placed and contribute to the realism." } ] }, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist. Oh, I forgot to mention, use my work address.", "long_pause_offsets": [ 62 ], "source_sha256": "451140102b09884eb858f5b9cd68fee2353d9fd8d65745112d407481b102f7b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing address constraint, which was omitted in the initial instruction but is present in the original. The action, scope (cart and wishlist), and app (Amazon) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/addition/addition_01/2cd02d4b2f9bbd97/speech.wav", "wer": 0.0, "wav_sha256": "36fbe36e12bbe0a32e0f8402ac6c1fcb26e96f0937bbd8259a6c1fb6676afd38", "seconds": 9.96, "spoken_utterances": [ "Place an order for everything in my um... amazon cart and ... wishlist. ", "Oh, I forgot to mention, use my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. The speaker then adds a new constraint, 'use my work address,' which is a valid addition. The final intent combines the original task with the added detail, fully satisfying the 'addition' type criterion.", "heard_transcript": "Place an order for everything in my um Amazon cart and wishlist. Oh, I forgot to mention, use my work address." }, "candidate_id": "wav_ff54d6c2134163ea", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural, spontaneous flow with a clear addition after an initial request. The speaker begins with a complete instruction, then pauses and adds a missing detail ('Oh, I forgot to mention...') with a believable intonation shift and pragmatic back-reference. The prosody is smooth, the pace is conversational, and the filler 'um' is used naturally. The addition feels like a genuine afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my home address. Let me correct the destination: ship to my work address, not home address.", "long_pause_offsets": [ 82 ], "source_sha256": "9a522cbab2847c54768abf63fbcece3f46f84e4c12b47865d37cd6e2494e34a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the destination from 'home address' to 'work address', replacing the initial incorrect value while preserving the action, items, and the requirement to order everything from both cart and wishlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/revision/revision_03_destination_address/5ec52c8faf340563/speech.wav", "wer": 0.0625, "wav_sha256": "242a6c5e9deea3a94f74695b5ec4ea4ea7627eedf98893dd23dfc226301421a5", "seconds": 15.08, "spoken_utterances": [ "Place an order for everything in my um... amazon cart and let me think... wishlist for my home address. ", "Let me correct the destination: ship to my work address, not home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests shipping to the 'home address' but immediately corrects this by stating 'Let me correct the destination, ship to my work address not home address.' This constitutes a revision where the final, corrected instruction overrides the initial incorrect one. The final intent matches the original reference instruction, so the task is considered passed.", "heard_transcript": "Place an order for everything in my um, Amazon cart and let me think, wishlist for my home address. Let me correct the destination, ship to my work address not home address." }, "candidate_id": "wav_7f0db28b3a4894e1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, inserts a plausible 'let me think' pause, and then smoothly corrects the destination with a believable 'let me correct' phrase. The intonation and pacing are consistent with spontaneous speech, and the correction is integrated into the same utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "c0b1466c7a4c8edc0231dd29f38f5e58511f04284a29e48d6e7e0acf018f471c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/cancellation/cancel_010/869bf9937957beb9/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "d3b3e4c8b437a48b386d9ce4001cd0cdf4a5299da7c6aafd5b9e23879fdd8a3d", "seconds": 10.64, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address.", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a cancellation command ('cancel the request I just made') following the initial instruction. The final interjection is a direct revocation of the task, satisfying the cancellation type criterion. The original instruction is irrelevant after the explicit cancellation.", "heard_transcript": "Place an order for everything in my amazon cart and wishlist for my work address. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "8d2cb7d5033cb033531f5ca85ddb52aa1e0c42e30c29f7ad4269e6161afb4eb4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/ambiguous_halt/halt_023/753158371f91b0d5/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "f5faa03240fe5fd2de120b5c55d3c397ea5f46ecf3aad041aee6e1fa207e32b5", "seconds": 9.68, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the evaluation passes.", "heard_transcript": "Place an order for everything in my amazon cart and wishlist for my work address. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ce38099781cceadc158824f9d8cd548c88736204d8cc942ce483321ff8931e88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/filler/original/2b187ae6851b22e9/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "dbcbe05521d11808432143dd6025aae1c9dfef232d3005f0107537c8519c7b3e", "seconds": 5.92, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without modification. The final intent matches the original instruction exactly, with no missing or extra constraints. The type is 'filler', which permits minor hesitations or fillers but requires the final task to be executed as stated. Since the final utterance is a complete and unmodified execution of the original request, the task passes.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist for my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address—sorry, make that home address.", "long_pause_offsets": [], "source_sha256": "550bca5ab554ad5c9c7879541045ae0309c89b0588d8bd1f956ee68ade5352b8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'work address' with 'home address'; the action, source items, and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/false_start/false_start_03_destination_address/6f74faf115fff669/speech.wav", "wer": 0.12, "wav_sha256": "5f0dcb2f5b397e5772db72c6bf392ba1162cb60ff15c01e22f3e72175263c3b5", "seconds": 10.6, "spoken_utterances": [ "Place an order for everything in my amazon cart and wishlist uh, let me think... for my work address—uh... sorry, make that home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately self-corrects to 'home address' after a pause. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think. For my work address, uh, sorry, make that home address." }, "candidate_id": "wav_f2f4357f98ff5a70", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with a clear request, then hesitates with 'uh let me think' before correcting the delivery address from 'work address' to 'home address' with a spontaneous-sounding 'sorry, make that home address.' The self-correction is integrated smoothly into the utterance, with natural intonation and pacing. The prosody is consistent, and the repair feels like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist. Oh, I forgot to mention, deliver it to my home address.", "long_pause_offsets": [ 62 ], "source_sha256": "571688afa8a20a01add8378a5701fc19fd69ff2f3b3f317f8407efe0bc38728a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement, which was not in the initial utterance but is present in the original instruction. The final goal combines both the initial order request and the added delivery condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/addition/addition_01/e668da694fcfb1e8/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "e08d9fb70108d868fcf04af907c679e1be63761c885fc38e7b14941c9594b80a", "seconds": 16.56, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and uh, let's see... wishlist. ", "Oh, I forgot to mention, deliver it to my um... home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request to place an order for everything in the Amazon cart and wishlist. After a pause and a self-correction ('Oh, I forgot to mention'), the speaker adds the delivery address as 'home address'. This is a clear addition of a constraint to the original task. The final intent combines the original order request with the new delivery address, satisfying the 'addition' type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and uh let's see, wishlist. Oh, I forgot to mention, deliver it to my um home address." }, "candidate_id": "wav_f45a97519589311a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address. Change the destination to home address instead of work address.", "long_pause_offsets": [ 82 ], "source_sha256": "53c4abb74d35addeedff34399f31a07dadc75512ca5ad93165053edc6cb60c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination from work address to home address, preserving the action, scope (cart and wishlist), and the intent to place an order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/revision/revision_03_destination_address/96cf9069bdd262b7/speech.wav", "wer": 0.0, "wav_sha256": "8ba1d0ac6385e892ec0e56566fde8d0da47b194f8b6cbfe518236496adf9f355", "seconds": 14.0, "spoken_utterances": [ "Place an order for everything in my amazon cart and uh... wishlist for my work address. ", "Change the destination to let me think... home address instead of work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the order for the 'work address' but immediately corrects it with 'Change the destination to... home address instead of work address.' This constitutes a revision where the final intent is to use the 'home address.' The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and uh wishlist for my work address. Change the destination to let me think, home address instead of work address." }, "candidate_id": "wav_f9f835fc9b1bab39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural hesitation ('and uh') and a clear, self-contained revision. The correction ('Change the destination to...') is smoothly integrated, with a believable 'let me think' filler and a pragmatic shift from 'work address' to 'home address'. The prosody is consistent, and the flow reflects a realistic thought process." } ] }, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "ee0385db25c54ee2f8c0ef76e07007e4067651f52dd3af19a221b81b4fffa277", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/cancellation/cancel_015/383c8b9e4cac3143/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "d975109f604119c5abd9ed45d6e3cf1ac64e88b01fc0258a9aa767d6c81b1db7", "seconds": 14.12, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address.", "Oh, wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to place an order, but after a pause and hesitation, clearly revokes the entire request by saying 'cancel the whole thing.' This constitutes a final cancellation of the original task, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address. Oh wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "4acc1b65a01739aa084ef54e22e20845b0a4688af9026d6ae66993eb7267e43e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/ambiguous_halt/halt_036/f066d5c94e088009/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "1992bfeb0915b1ec8c9214dd0ac35d475f284930280753fa14448ae56900e59a", "seconds": 13.88, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Wait a second before you go ahead,' which is a clear request to pause or hold the action. This is a final interjection that does not cancel the task but explicitly asks for a delay. The initial request to place an order is not revoked, and the speaker does not say 'cancel' or 'never mind.' Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address. Wait a second before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "77bcb81_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f6e51dbf9a82beaf552221c7565e1d1a4a2b6feb4c0481a54525abf5673f5058", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/filler/original/3272d17a407e43eb/speech.wav", "wer": 0.25, "wav_sha256": "135ace661bc7e5f622a2d5e9e4fc32188c28e8d85d048f0867c8de246d9129fe", "seconds": 9.72, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let me think', 'hmm') but does not change the core request. The final intent remains to place an order for everything in the Amazon cart and wishlist for the home address, matching the original instruction. The hesitation is a filler and does not alter the task goal.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_completion_state", "proposed_instruction": "How many activities are incomplete—sorry, I meant completed in my bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "0d365ac03c8e060027f850343f4aa23e5e2add4237e5235a27db91baffc768d2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'incomplete' to 'completed' within the same utterance, preserving the app (Simple Note), note title ('bucket list'), and the action of counting activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/false_start/false_start_01_completion_state/d0eb25e97fd427ec/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "26f2fdc121ea2525161a3983b68019d9e9a0fd27b83ac710198e7fdd1cb42d98", "seconds": 9.8, "spoken_utterances": [ "How many activities are incomplete—um... sorry, I meant completed in my bucket list note in ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'incomplete' but immediately self-corrects to 'completed'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is not executed. The final intent matches the original instruction exactly.", "heard_transcript": "How many activities are incomplete? Sorry, I meant completed in my bucket list note in Simple Note." }, "candidate_id": "wav_a0377a6fd8841585", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and credible false-start pattern. The speaker initially says 'incompletem' with a clear mispronunciation, then immediately self-corrects with 'sorry, I meant completed' in a spontaneous, conversational tone. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and recovery. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many activities are in my bucket list note in Simple Note? Oh, I forgot to mention, only count the activities that are completed.", "long_pause_offsets": [ 63 ], "source_sha256": "1bd95b8f68a6ba416adba26220475fb0922c92ee8007d8ee6f5d3401b8cc98be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a constraint to only count completed activities, which refines the initial broad request to count all activities. The final goal is fully specified and matches the intent of the original instruction with a clarifying restriction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/addition/addition_01/a56aaf659c9c6d52/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "82ba6bf692d21eb04aa08d9a206e41c15e16392eccc6579ca584a59c69b35f02", "seconds": 13.84, "spoken_utterances": [ "How many activities are in ah... my bucket list note in let's see... Simple Note? ", "Oh, I forgot to mention, only count the activities that are ... completed." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many activities are in my bucket list note in Simple Note?'. After a pause and a self-correction, the speaker adds the constraint 'only count the activities that are completed'. This is a clear addition of a new condition to the original request. The final intent combines the original task with the added constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "How many activities are in uh my bucket list note in let's see, Simple Note. Oh, I forgot to mention, only count the activities that are completed." }, "candidate_id": "wav_374b960f0247e956", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, uses a plausible filler ('uh') and hesitation ('let's see'), and then adds a relevant restriction ('only count the activities that are completed') in a way that sounds like a spontaneous afterthought. The prosody is consistent, and the addition flows naturally without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_note_app", "proposed_instruction": "How many activities are completed in my bucket list note in Evernote? Let me correct the app: look for the note in Simple Note, not Evernote.", "long_pause_offsets": [ 70 ], "source_sha256": "7c83452b4852caf47b62c81b5670dce296ed79daeb81f668349fbf0897210986", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Evernote to Simple Note, preserving the note title and the query about completed activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/revision/revision_03_note_app/8d1f9d28b7300ae8/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "61c4e7626c857f9833604bf18172a6c3491044ab0ad8f360763cd4204c67d896", "seconds": 17.84, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in um, let's see... Evernote? ", "Let me correct the app: look for the note in um... Simple Note, not Evernote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Evernote but immediately self-corrects, specifying 'Simple Note, not Evernote' as the correct app. This constitutes a revision where the final, corrected target app is Simple Note, matching the original instruction. The rest of the task (counting completed activities in the 'bucket list' note) remains unchanged and is executed as intended.", "heard_transcript": "How many activities are completed in my bucket list note in um, let me see, Evernote? Let me correct the app. Look for the note in um, Simple Note, not Evernote." }, "candidate_id": "wav_76f8bdb29ac5a720", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complex, multi-part request with natural hesitation ('uh', 'um'), a clear false start ('Evernote'), and a self-correction ('Let me correct the app') that flows into a believable revision ('SimpleNote'). The prosody is varied, with appropriate stress and pacing, and the correction is integrated smoothly into the ongoing utterance rather than sounding like a separate sentence. The speaker sounds like they are actively thinking through the task." } ] }, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "84329a448869c4cd235eca03e1e3319962253a4430dabd8ddec3f22c0b5453b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/cancellation/cancel_031/087ddea75ee7b4e9/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "8f26a3a121fd2c9c5f8d050758351eedc443b8bcc802895d47245126dfff9efc", "seconds": 11.12, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?", "Wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the entire task by stating 'Wait, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the original request. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note? Wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "5dc14754cc5f601e61fd9efc209ee8bcaff037319f1b2ef27131eb72a1d82dfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/ambiguous_halt/halt_025/b8cd3df571e7a85a/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "ad12bcfbe0a7afc621e11d344e05274b37d91ebcdaf7e8e8de9cb5ea21aad55e", "seconds": 9.92, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I need to think for a second,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note? Wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ffbdcfda2c5e94d5c1f1fd0a6dd44581f5c086c480472860eaedbeab3b541bfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/filler/original/5fa03f6b0c78a9ff/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "768be97ba0850ab6f4519896fd6691bf1b3ef72e2e2685fe263ac2159b441dcf", "seconds": 5.68, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'uh' as a filler, then proceeds to complete the original instruction without any correction, cancellation, or addition. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_note_app", "proposed_instruction": "How many activities are left to do in my bucket list as per my Google Keep—I should say SimpleNote note?", "long_pause_offsets": [], "source_sha256": "48e3ae66eb605cf79d3ffe86d5e7fb0a46bcbf5d6330b0769e5d7e3fb7e06a73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Google Keep but immediately corrects to SimpleNote, which is the app specified in the original instruction. The rest of the query remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/false_start/false_start_03_note_app/c34fbbb3064afaeb/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "36acf89de8eb08c172620d63d24a3595eaa82e21cb04a508ddadb7535e5e44bd", "seconds": 8.24, "spoken_utterances": [ "How many activities are um... left to do in my ... bucket list as per my Google Keep—uh... I should say SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Keep' but immediately self-corrects to 'SimpleNote'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved otherwise. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many activities are uh left to do in my bucket list as per my Google Keep uh, I should say SimpleNote." }, "candidate_id": "wav_a54916fcad529973", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('uhm') and smoothly self-corrects from 'Google Keep' to 'Simple Note' without any awkwardness. The prosody is consistent, the pace is conversational, and the correction feels spontaneous and integrated into the flow of the sentence." } ] }, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many activities are bucket list as per my SimpleNote note? Oh, I forgot to mention, it's only the activities left to do in my bucket list.", "long_pause_offsets": [ 63 ], "source_sha256": "7cb5f38a5edfba7b84538648d0bf9a5006a59214a111d33944fc9b4116fcae61", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the initial ambiguous phrasing 'activities are bucket list' by specifying 'only the activities left to do in my bucket list', which matches the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/addition/addition_01/7c35f75797d685ae/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "3fdd651289b1ebff93b3d65934250effc0200eaad7a8315ac2cd76375b989c0f", "seconds": 13.24, "spoken_utterances": [ "How many activities are uh... bucket list as per my ... SimpleNote note? ", "Oh, I forgot to mention, it's only the activities um... left to do in my bucket list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('activities are a bucket list'), but the speaker adds the clarifying phrase 'only the activities left to do in my bucket list' after a pause. This addition completes the original request by specifying the subset of activities to count. The final intent matches the reference instruction, which is to count activities left to do in the bucket list as per the note.", "heard_transcript": "How many activities are a bucket list as per my SimpleNote? Oh, I forgot to mention it's only the activities um left to do in my bucket list." }, "candidate_id": "wav_1e54c0ceff52f5d3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then naturally adds a clarifying detail ('Oh, I forgot to mention...') with a believable self-correction. The prosody is smooth, the pause before the addition is realistic, and the correction flows naturally without sounding forced. The speaker's voice is consistent and the delivery sounds like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_completion_state", "proposed_instruction": "How many activities are completed in my bucket list as per my SimpleNote note? Wait, it should be 'left to do' in my bucket list, not 'completed'.", "long_pause_offsets": [ 79 ], "source_sha256": "6169b83fc488d3e797def0ede10ab8da952080f645004bedf4c69a9ce39d0ce8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial phrasing from 'completed' to 'left to do', preserving the app (SimpleNote), list (bucket list), and the intent to count remaining activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/revision/revision_01_completion_state/0bc599429ecb867e/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "14ae32aca74830d5471b31331aa6be96d26c8a8eb53a97733f64167ce4bf09aa", "seconds": 16.12, "spoken_utterances": [ "How many activities are uh... completed in my ah... bucket list as per my ... SimpleNote note? ", "Wait, it should be 'left to do' in my bucket list, not 'completed'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'completed' activities but immediately corrects themselves to 'left to do' after a pause. This is a clear revision where the final intent is the corrected version. The original instruction is revised to match the final audible request, so the task is executed as revised.", "heard_transcript": "How many activities are uh completed in my uh bucket list as per my SimpleNote note? Wait, it should be left to do in my bucket list, not completed." }, "candidate_id": "wav_aef6cc5d6d3940fc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('completed') and immediately self-corrects with a natural-sounding 'wait, it should be left to do'. The correction is integrated smoothly into the sentence, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit. The pacing and filler ('uh') are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "fcff27053f79295fda80e68b1fd6956cba512f95f120cade0e8a00a21aa9df0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/cancellation/cancel_031/e31c39c447d51200/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "305aa14a659cb2e2c92d7a14080a5a2a458e5c2cf17159bf8ecb8a642f092597", "seconds": 11.96, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?", "Oh, wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the entire task by stating 'Oh wait, there's no need to do that anymore. Cancel it.' This constitutes a final, explicit cancellation of the original request. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "How many activities are left to do in my bucket list as per my um, let's see, SimpleNote? Oh wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "feb6cd2e2bb698c7270e5afa7be03bdec428b2e1f67ede6630349a03901189f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/ambiguous_halt/halt_017/6c4ea418167a88d3/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "d5f07dd50649ac3f2c0bbf0368fde81bc868169a7a7fa17093541ac5a6ee2653", "seconds": 10.32, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment please,' which is a clear request to pause or wait. The user is in the process of checking their SimpleNote note, as indicated by 'as per my um let's see. SimpleNote.' This constitutes an ambiguous_halt, where the user asks for a pause without canceling the task. The final interjection is a request to wait, not to cancel or execute, satisfying the ambiguous_halt criterion.", "heard_transcript": "How many activities are left to do in my bucket list as per my um let's see. SimpleNote. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e213d8856d287fb2c66ee7f89091c28541dbba0242322bb13c13425e4640dac7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/filler/original/721327583ae21e5d/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "02ad8b925d250d87fdd567320c3c81232f12d7ba318c019457c68dc4b6672009", "seconds": 7.2, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'let's see') and a brief hesitation but does not change the core request. The final intent remains to count bucket list activities from the SimpleNote note, matching the original instruction. The hesitation is a natural speech pattern and does not alter the task goal.", "heard_transcript": "How many activities are left to do in my bucket list as per my um let's see SimpleNote?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_note_description", "proposed_instruction": "How many activities are done in my daily goals—wait, I meant bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "668fa89bcfc47b8589e6c64e7843c8099b71c50a7aafc19f5f623bb630024e34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'daily goals' to 'bucket list note' within the same utterance, preserving the app (Simple Note) and the action (counting done activities)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/false_start/false_start_02_note_description/b16783e73f18cee6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "5954a4705c38bccf031125d85d8f1167f9cf8b69fc4f600818b96dbb1fbef58d", "seconds": 9.96, "spoken_utterances": [ "How many activities are done in my um... daily goals—uh... wait, I meant hmmm... bucket list note in Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily goals' but immediately self-corrects to 'bucket list note'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many activities are done in my um daily goals uh wait, I meant bucket list note in Simple Note?" }, "candidate_id": "wav_20ecf1f869bbae1f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start and self-correction. The speaker begins with 'daily goals', hesitates with 'uh', then clearly self-corrects to 'bucket list note' with a believable 'wait, I meant' phrase. The prosody includes a natural upward inflection on 'uh' and a downward correction, and the pacing feels like a real-time thought process. The correction is integrated smoothly into the same utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many activities are done in Simple Note? Oh, I forgot to mention, it's in my bucket list note.", "long_pause_offsets": [ 45 ], "source_sha256": "644e27e46dc0b0f91a256af068951fe72f9bd578f4a22dc1f7eb2d2f8358b8ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context that the note is the 'bucket list note', which specifies the exact note to query. The app and the action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/addition/addition_01/9b19f0387bb9350c/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "1f580f7de7f08849965f56c944c7309c49cbe8d876fb77fad0043cf2e63540db", "seconds": 14.0, "spoken_utterances": [ "How many activities are done in um, let me think... Simple Note? ", "Oh, I forgot to mention, it's in my uh, let's see... bucket list note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many activities are done in Simple Note?'. After a pause and self-correction, the speaker adds the constraint 'it's in my bucket list note'. This is a clear addition that narrows the scope of the original request. The final intent combines both the original and added information, matching the reference instruction. The addition type criterion is satisfied.", "heard_transcript": "How many activities are done in um, let me think, Simple Note? Oh, I forgot to mention, it's in my uh, let's see, bucket list note." }, "candidate_id": "wav_ae6011a023cf19a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural, spontaneous fillers like 'um' and 'let me think' at the beginning, followed by a believable self-correction ('Oh, I forgot to mention...') and a clear addition of the missing detail ('bucket list note'). The prosody is conversational, with a natural rise and fall in intonation, and the pacing includes plausible pauses that suggest genuine thought processing. The addition feels integrated into the flow rather than tacked on." } ] }, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_completion_state", "proposed_instruction": "How many activities are all in my bucket list note in Simple Note? Wait, I meant 'done' activities, not 'all'.", "long_pause_offsets": [ 67 ], "source_sha256": "1500d2ca547d43f344648a398b8ee6edf12dc89c0d5b7f6383d21b9b9c13421f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects 'all' to 'done', preserving the app (Simple Note), note (bucket list), and the action of counting activities. The final goal is to count only the 'done' activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/revision/revision_01_completion_state/c0d1293db0444cc9/speech.wav", "wer": 0.12, "wav_sha256": "657201d0ecaf63348dc299598bb92853172514a5f35856a8af67dbcaf367c1c0", "seconds": 16.32, "spoken_utterances": [ "How many activities are all in um... my bucket list note in uh, let's see... Simple Note? ", "Wait, I meant 'uh... done' activities, not 'all'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'all' activities but immediately self-corrects to 'done activities' after a pause and repair. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to count 'done activities', which matches the original instruction's structure and target ('my bucket list note in Simple Note') but with the correct value ('done' instead of 'all'). No other constraints were changed, so the revision is correctly applied.", "heard_transcript": "How many activities are all in um my bucket list note in uh let's see Simple Note? Wait, I meant uh done activities, not all." }, "candidate_id": "wav_35478cef19152d88", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous flow with a clear false start ('all in um... my bucket list note') followed by a self-correction ('wait, I meant uh done activities not all'). The intonation and pacing reflect genuine hesitation and recollection, with a believable shift in focus. The pauses and filler words ('um', 'uh', 'let's see') are well-placed and contribute to the impression of someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "41fe8ab1e93dfa0c4b3f25ed2f60bcd763d1884f09d9771a73983f71e475a798", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/cancellation/cancel_024/826bf35b093e2dc3/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "9cc5941d4f781f01fbb0bf31c7fcf04232faff2910eab0b07ca0568f25ed803c", "seconds": 13.48, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Hang on, no, never mind, cancel the whole thing.' This constitutes a final, unambiguous cancellation of the entire task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "How many activities are done in um my bucket list note in let me think Simple Note? Hang on, no, never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "90bb7e10da93fe42c1706e6c8d7b9ca8f921b59f932e79dac32511445f346517", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/ambiguous_halt/halt_031/da2989f3a085fc8a/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "ae172a0673b59c6757de90ccba098a3a603997ce15e27ed72e246d3e86bee0c2", "seconds": 12.56, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present as the final intent is to pause, not execute or cancel.", "heard_transcript": "How many activities are done in um my bucket list note in let me think, Simple Note? Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7847649_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "46feb7a44c720d62c86f64010741f1b447fd93d1857696a0642e3944765d4309", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/filler/original/8c263be8b55fac0e/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8e774c708934895bbc8a1ccc89f620b3f26ba73d7924115504b2085f6656982d", "seconds": 8.64, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and a pause but does not change the core request. The final audible intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "How many activities are done in um my bucket list note in let me think Simple Note?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I mean as many things as possible with $400 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "32161cb5bdea80c1140c5538e6045b66758eacb64077771c3735c7dd5a303501", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and scope, with only the initial mistaken phrasing being replaced." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/false_start/false_start_01_selection_objective/a5cdb506e3221b3e/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "6d4226cfb156e4766c14980aafa98dcc1ac6dadf47280467002dd9fa1779c640", "seconds": 10.24, "spoken_utterances": [ "Make an order for uh... as few things as possible—ah... no, I mean as many things as possible with four hundred dollars um... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately corrects themselves to 'as many things as possible' with a self-repair. According to the 'false_start' type criterion, the final corrected intent should be executed. The monetary value and target (Amazon wishlist) remain consistent and correctly stated. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Make an order for uh as few things as possible. I no, I mean as many things as possible with four hundred dollars um for my Amazon wishlist." }, "candidate_id": "wav_cdf975761504df49", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'as few things as possible' immediately corrected to 'as many things as possible'. The self-repair is smooth, the intonation reflects genuine reconsideration, and the pause before the correction feels like a real moment of thought. The speaker's voice and prosody remain consistent throughout, with no signs of artificiality or disjointedness." } ] }, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $400 from my amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "ab8a6236890e542a04709018ee8af40662e9182097ee2fa10afe94694962c34d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial instruction, which already specified the amount and source. The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/addition/addition_01/fa451e2c6b49479f/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "85563bd168be3cb185514896c6a4cfbccf702218369aa701a37a661ff1abfd5b", "seconds": 13.8, "spoken_utterances": [ "Make an order for with ... four hundred dollars uh... from my amazon wishlist. ", "Oh, I forgot to mention, make the order for ... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was incomplete ('Make an order for with four hundred dollars...'), but the speaker immediately corrected and completed it with 'make the order for as many things as possible.' This constitutes a clear addition that combines with the initial instruction to form the complete goal matching the reference. The final intent is fully aligned with the original instruction.", "heard_transcript": "Make an order for with four hundred dollars uh from my Amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible." }, "candidate_id": "wav_043c75e9f018b847", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Make an order for as many things as possible with $400 from my ebay wishlist. Let me correct the app: use 'amazon', not 'ebay'.", "long_pause_offsets": [ 78 ], "source_sha256": "3f8360902b51bc03702f0300842695f57d03e432b0622301e02f1b1a88d9234b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from 'ebay' to 'amazon' while preserving the amount ($400) and the goal of ordering as many items as possible from the wishlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/revision/revision_03_app/3407e8f5a569dc7a/speech.wav", "wer": 0.0, "wav_sha256": "d309f32185453f1d4e9195ac83010960a8a03e93658da14c49dad4f1e8ae159c", "seconds": 14.68, "spoken_utterances": [ "Make an order for as many things as possible with four hundred dollars let's see... from my ebay wishlist. ", "Let me correct the app: use 'ah... amazon', not 'ebay'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects it to 'Amazon, not eBay' after a short pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to use Amazon, which matches the original reference instruction. The monetary value and goal of ordering as many items as possible remain unchanged.", "heard_transcript": "Make an order for as many things as possible with four hundred dollars, let's see, from my eBay wishlist. Let me correct the app, use Amazon, not eBay." }, "candidate_id": "wav_26e3147e69032be1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a clear, confident request, then hesitates with 'let me see' before correcting the app from 'eBay' to 'Amazon'. The self-correction is smooth, with a natural pause and a clear, unforced intonation shift. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "aadf2b9abcb07cd32a4a336856c628d11ffe58f345a90224f973fc0ffca223b4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/cancellation/cancel_037/1ede6da09ffd79e2/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "9b8d904dc1c12acde1a20489cbeb7ca9bfa3359718d32190da623ae89951dd5d", "seconds": 14.56, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist.", "Hang on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly and unambiguously cancels the entire task with 'Hang on, stop there, cancel the whole task.' This meets the criterion for a cancellation, as the final audible intent is to revoke the task, not to proceed with it.", "heard_transcript": "Make an order for as many things as possible with um four hundred dollars um from my amazon wishlist. Hang on, stop there, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "e8face98b0da38d3d3d57d3eda363bac36be42a249eede7584b4b120a31592fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/ambiguous_halt/halt_030/4ee0763e655a1cc1/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "498e132326d3c5062685988e67870e1f773e0d5b808f9ff5b38b84206a0c2b9c", "seconds": 12.96, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist.", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause the current task. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present as the final intent is to pause, not execute or cancel.", "heard_transcript": "Make an order for as many things as possible with um four hundred dollars um from my amazon wishlist. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "36586c81255f71baf86bdb5b4b9f8ebc3fb143c3b0248693fc74c204a3807aa9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/filler/original/1445fef377cc34d2/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "d85baca40827329f4eda8ba315215093a5fcbb938e1c670e4ad538bd7cfd4acd", "seconds": 9.44, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um') but does not change the core request. The final intent matches the original instruction, with 'four hundred dollars' being the audible equivalent of '$400'. The task is executed as requested.", "heard_transcript": "Make an order for as many things as possible with um, four hundred dollars um, from my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I meant as many things as possible with $500 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "e293ccadce001b47e4c46d6c4011e7ed303f71e86687850216cb34818e56ac25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/false_start/false_start_01_selection_objective/247efd372e790d12/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "5723bf0ecfb6637e97e0650dcd97650e412e95a1fc7ad653c70b63f9cf9aa56e", "seconds": 12.2, "spoken_utterances": [ "Make an order for as few things as possible—uh... no, I meant hmmm... as many things as possible with ... five hundred dollars from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately self-corrects to 'as many things as possible'. This is a false start, and the final intent is the corrected version. The monetary value and source are consistent with the original instruction. The final control intent is to execute the corrected task.", "heard_transcript": "Make an order for as few things as possible. Um, no, I meant, as many things as possible with five hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_df11239ca6ab605b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous flow. The speaker begins with a clear false start ('as few things as possible'), immediately self-corrects with a believable 'no, I meant' and a brief throat clear, then smoothly continues with the corrected request. The prosody is consistent, the correction is integrated into the utterance, and the pacing feels like genuine real-time thought processing." } ] }, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $500 from my amazon wishlist. Oh, I forgot to say, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "8782a0e50e31677804aec059c395e4e9a1c63c9f2c39b5a98028eb96e0bcd93e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial order request, which already included the amount and source. The final goal combines both utterances without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/addition/addition_01/37a48c39b0cfd443/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "d4f58e1efa677448230be62a58703361f3a86091e5ba51bfca88258b8d5a19ca", "seconds": 13.48, "spoken_utterances": [ "Make an order for with five hundred dollars let's see... from my amazon wishlist. ", "Oh, I forgot to say, make the order for hmmm... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make an order for with five hundred dollars' and then adds 'as many things as possible' after a pause and self-correction. This matches the 'addition' type, where the final goal combines the initial instruction with the added constraint. The final audible intent is to make an order for as many items as possible within the $500 budget from the Amazon wishlist, which aligns with the original instruction.", "heard_transcript": "Make an order for with five hundred dollars, let's see. From my Amazon wishlist. Oh, I forgot to say, make the order for, hmm, as many things as possible." }, "candidate_id": "wav_cc54df14955377af", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' and a pause, which is a believable moment of thought. The correction 'Oh, I forgot to say' is a plausible self-aware interjection, followed by a smooth, additive continuation. The prosody is consistent, and the flow from initial request to addition feels spontaneous and human." } ] }, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $300 from my amazon wishlist. Change the budget to $500, not $300.", "long_pause_offsets": [ 80 ], "source_sha256": "53b9079b63b1adbdcce661b84f624931b43a8f7758b2e6328341290e8d01a17e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the budget from three hundred to five hundred dollars, preserving the action, scope, and source from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/revision/revision_02_budget_amount/dc8a2e3c6a4aed95/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "2eca55d7e10f69a99aa579722f5beb05b1c28f463fb374664d5126d56e9636ed", "seconds": 13.48, "spoken_utterances": [ "Make an order for as many things as possible with um... three hundred dollars from my amazon wishlist. ", "Change the budget to uh... five hundred dollars, not three hundred dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a budget of $300, then clearly revises it to $500 after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the action, target, and budget value.", "heard_transcript": "Make an order for as many things as possible with um, three hundred dollars from my Amazon wishlist. Change the budget to uh, five hundred dollars, not three hundred dollars." }, "candidate_id": "wav_19242ea9f710fa2a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural revision flow. The speaker starts with a clear request, then self-corrects the budget amount from $300 to $500 with a smooth, spontaneous-sounding 'change the budget to uh $500 not $300'. The prosody is consistent, the correction is integrated into the same utterance, and the intonation reflects a genuine moment of reconsideration. The delivery is fluent and lacks any robotic or disjointed qualities." } ] }, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "0cbe445c6425498a2212e252cd4388cd0ea389f1aed7890215e247ff5e6c9e40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/cancellation/cancel_003/d9447eba04d00ef0/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "cdeb907a1669726e39037ef46c5c810b993488ef6dccf55799d93d895ef1a854", "seconds": 13.04, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist.", "Wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to make an order with $500 from the Amazon wishlist, then immediately corrects themselves by saying 'Wait, please cancel that.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist. Wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "27844723f93dcbaeb077ec7f4ad56ba2517d7f22229f24fa7a5fe465a8952887", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/ambiguous_halt/halt_025/b29649b2c29e8b69/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "f57c74b097470281f384e90a2cd518f6be6433c238aeaf4aaa68b3cbdb01d9d6", "seconds": 13.04, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist.", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I need to think for a second,' which is a clear request to pause the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist. Wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9d562cd28ab39be50346aac3eb2047cb55b1fba7cdb848956df5b495c7b77588", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/filler/original/b0594a5b26b6cdb0/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "07c0444c2535ef6194665e17415cf97c86e2a920bd8fa2cb100ea2d74c7731d9", "seconds": 9.12, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler ('um') and a pause ('Hmm'), but the final instruction clearly matches the original request. The speaker corrects the initial hesitation and completes the task with the same intent. The monetary value is stated as 'five hundred dollars', which is equivalent to '$500'. The target app ('Amazon wishlist') and the goal ('order as many things as possible') are preserved. No correction or change to the core task is made after the initial hesitation.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I mean as many things as possible with $600 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "d7626b298220f06db32c34515f8461068ffc118f306e9cca38624db280f5505f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/false_start/false_start_01_selection_objective/8f8d4b3d12a66bc2/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "1bb0c160e770458fd583795cc471402b01ac16fffea481c2f5119bbf75a2b0aa", "seconds": 12.08, "spoken_utterances": [ "Make an order for as few things as possible—um... no, I mean hmmm... as many things as possible with ... six hundred dollars from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately self-corrects to 'as many things as possible'. According to the 'false_start' type criterion, the final corrected intent should be executed. The monetary value and source (Amazon wishlist) remain consistent. The correction is immediate and unambiguous, satisfying the type-specific requirement.", "heard_transcript": "Make an order for as few things as possible. Um, no, I mean, as many things as possible with six hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_6497ed878cb0cd23", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and fluent self-correction. The speaker initially says 'as few things as possible' and immediately corrects to 'as many things as possible' with a clear, spontaneous 'no, I mean' transition. The prosody is consistent, the pace is conversational, and the correction feels like a genuine moment of reconsideration rather than a rehearsed edit. The sentence flows smoothly without awkward pauses or unnatural phrasing." } ] }, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $600 from my amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "0027efb19a80eb1addc0c6f446873781c683f7d354f28eba3f6ae3369d92a175", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial order amount, combining both utterances into a complete goal that matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/addition/addition_01/93a07eaf19a6281e/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "fc5c11bf8626271fe0242e095b406faa4647d06b915a67d36e17042d8346e601", "seconds": 15.0, "spoken_utterances": [ "Make an order for with ... six hundred dollars uh... from my amazon wishlist. ", "Oh, I forgot to mention, make the order for ... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make an order for with six hundred dollars from my Amazon wishlist' and then adds 'make the order for as many things as possible' after a pause. This is a clear addition that narrows the original broad request to a specific constraint. The final intent combines the original instruction with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "Make an order for with six hundred dollars uh from my Amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible." }, "candidate_id": "wav_90bece520d3ba353", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Make an order for...') and then adds a detail ('Oh, I forgot to mention...') with a believable hesitation and self-correction. The prosody is consistent, the pacing includes a realistic pause and filler ('uh'), and the addition flows smoothly without sounding forced or robotic. The speaker's intonation reflects genuine recollection, making the addition feel spontaneous and credible." } ] }, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $600 from my amazon shopping list. Change the source collection to 'wishlist', not 'shopping list'.", "long_pause_offsets": [ 85 ], "source_sha256": "56cb2e3694e154d62cd5905832495a43dcdc38649902bb891d0b79d2195eed34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source collection from 'shopping list' to 'wishlist' while preserving the monetary limit and the goal of maximizing items. The original instruction and final spoken goal match in all meaningful requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/revision/revision_04_source_collection/e34c657078b76c18/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "1a22d0738c3162d2067124c3aff71038fee22528c0e3fc41dc3dffaa0e33a6cb", "seconds": 17.64, "spoken_utterances": [ "Make an order for as many things as possible with let me think... six hundred dollars let's see... from my amazon shopping list. ", "Change the source collection to 'um... wishlist', not 'shopping list'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the amount and the target collection.", "heard_transcript": "Make an order for as many things as possible with, let me think, six hundred dollars, let's see, from my Amazon shopping list. Change the source collection to, um, wishlist, not shopping list." }, "candidate_id": "wav_aed8d73528b0fe89", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a realistic 'let me think' filler, then states a value ('six hundred dollars'), hesitates with 'let's see', and then clearly revises the source collection from 'shopping list' to 'wishlist'. The intonation and pacing are consistent with someone formulating a request in real time, and the revision is smoothly integrated into the ongoing utterance without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "e9facae1b90eda74c99dda02ad60949de2cf95deddf0d786b72ccf3a4836c953", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/cancellation/cancel_029/ea6b18694b3a4411/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "b37bac35077d68e2e2f3727d7fc995d28249e4d960f1124c5c00894103efed52", "seconds": 11.76, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a complete order request, then immediately follows with 'Hang on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "Make an order for as many things as possible with six hundred dollars uh from my Amazon wishlist. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "3faf3aa15755c66b65306093ccc88b583c74ad3d1a57e91c0d571ce61e7c4970", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/ambiguous_halt/halt_026/7da286241974a4fc/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "5dbb85745dd5fb0d221890faaf199bf2d54bce33160065b9c5ccb694a9a55437", "seconds": 12.0, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation. The initial instruction is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Make an order for as many things as possible with six hundred dollars uh from my Amazon wishlist. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7b97e6d_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "68b646f5398daefe0c306b80f33bada73f496f0d7353a6f44b78b1447e1c0681", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/filler/original/f061025cc20f96aa/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "287232743dcbdc607b002a70340d8912f8856e824cebe0ab77159ac041e1a8a8", "seconds": 7.48, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh') but does not change the core request. The final intent matches the original instruction in action, target, and value. The filler nature of the pauses and hesitations does not alter the goal, so the task should be executed.", "heard_transcript": "Make an order for as many things as possible with uh six hundred dollars uh from my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Buy me a kitchen scale—I meant food scale as David recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "61fe4fd5cd507fb232861f55b26ccb5d0bb146d287575ce0241707a143748af5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'kitchen scale' to 'food scale' within the same utterance, preserving the rest of the instruction including the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/false_start/false_start_01_product_category/08d5daf691ce2e27/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1a718cf22b46cd4fefeeda69d6e0eedf63e70120373bb26c10fcc4694e8a13f4", "seconds": 7.52, "spoken_utterances": [ "Buy me a kitchen scale—um... I meant ... food scale as David recommended in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'kitchen scale' but immediately corrects to 'food scale' with 'I meant'. This is a false start where the final corrected value is the intended target. The rest of the instruction, including 'as David recommended in his phone message', is preserved and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Buy me a kitchen scale. Um, I meant, food scale as David recommended in his phone message." }, "candidate_id": "wav_082a9d224959d078", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'kitchen scale', self-corrects to 'food scale' with a clear, spontaneous 'I meant' repair, and smoothly continues with the full context 'as David recommended in his phone message'. The prosody is consistent, the pause before the correction is brief and realistic, and the entire utterance flows as a single, coherent thought process without any awkwardness or unnatural phrasing." } ] }, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Buy me a food scale in his phone message. Oh, I forgot to say, it should be as David recommended.", "long_pause_offsets": [ 42 ], "source_sha256": "13184b8ae78a1b9aa6378fb399b4150a0c4727a6d570d42088a71d3ed07a4f0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'as David recommended', which was present in the original instruction. The initial utterance incorrectly placed 'in his phone message' as the location or context, but the correction clarifies the recommendation source. The final goal matches the original by specifying the item and the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/addition/addition_01/d8fba02856b57b51/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "613dceffa90e5176c7f1cfef71340270e131d9692eecc4a830eac28efc2029dc", "seconds": 10.0, "spoken_utterances": [ "Buy me a um... food scale ... in his phone message. ", "Oh, I forgot to say, it should be as David recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Buy me a food scale in his phone message.' The speaker then adds 'Oh, I forgot to say, it should be as David recommended.' This addition clarifies the condition under which the scale should be bought, modifying the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction. The addition is clear and directly relevant, satisfying the addition type criterion.", "heard_transcript": "Buy me a um, food scale in his phone message. Oh, I forgot to say, it should be as David recommended." }, "candidate_id": "wav_4bcf3c6c73de073d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a slightly off-the-cuff phrasing ('in his phone message'), which sounds like a spontaneous thought. The addition of the constraint ('it should be as David recommended') is delivered with a believable, slightly apologetic tone and a natural pause before the correction, fitting the 'addition' pattern well. The prosody is consistent, and the flow feels like a real-time recollection." } ] }, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_person_name", "proposed_instruction": "Buy me a food scale as Sarah recommended in his phone message. I misspoke about the person: it should be David, not Sarah.", "long_pause_offsets": [ 63 ], "source_sha256": "339a1bb2dac4d67aa35668bf8a49b488a51611d81390330fd4a416ba3affd6da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recommender from Sarah to David, preserving the action, item, and context of the phone message." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/revision/revision_02_person_name/904e4025dfab4251/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "6c65a4820e3398f959d580e744b4adbc49fed94cde325b57a1a53c69e785518f", "seconds": 14.08, "spoken_utterances": [ "Buy me a uh, let's see... food scale as ... Sarah recommended um... in his phone message. ", "I misspoke about the person: it should be David, not Sarah." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the recommender as 'Sarah' but immediately corrects themselves with 'I misspoke about the person, it should be David, not Sarah.' This constitutes a revision where the final intent is the corrected version. The original instruction is fully preserved except for the name, which is corrected. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Buy me a uh let's see food scale as Sarah recommended um in his phone message. I misspoke about the person, it should be David, not Sarah." }, "candidate_id": "wav_7388041a588b08ab", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous revision flow. The speaker begins with a filler ('let's see'), makes a clear mistake ('Sarah'), immediately self-corrects with 'I misspoke about the person, it should be David, not Sarah,' and maintains a consistent, conversational pace. The correction is integrated smoothly into the utterance, and the prosody reflects genuine recollection and correction rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "ab5484cf22b4718c62a3069dd2cfbb7f450c5a4c1e9b8d0a4b6567ad7163c926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/cancellation/cancel_026/74055ddf60016657/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "338a0fcbe3dd7166089132842d6afc8fd907dff32f7db94b79b33a9016b1a1b7", "seconds": 11.04, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message.", "Hang on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to buy a food scale, then immediately follows with 'Hang on, forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and not followed by any resumption or new instruction, satisfying the criterion for a 'cancellation' type task.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message. Hang on, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "ea23d2961e9d86ccbda037ad14f084e20a56d73beb74156d3a988d6f70a31862", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/ambiguous_halt/halt_040/03fdef4a4a712f0d/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "842ca29cc69fa60d2b49fa4c2d60808068f4710f336866b404b2a82e85ed71f7", "seconds": 9.76, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly requests to hold or wait with 'Hang on, just a minute.' This constitutes a final, audible request to pause, which matches the 'ambiguous_halt' type criterion. No cancellation or further instruction is given, so the task is not executed or revoked.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6dfa76d0e5cab002da6232c82a0538fcaeb626bb53c23f471e99aa93df19e476", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/filler/original/c6315b836a9132f4/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "e0f16af90b0936a825878c203c5b25229beffc1c9e26c8c4a2353338933418c7", "seconds": 6.2, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker hesitated with 'uh' but did not change the core request. The final intent matches the original instruction, and the hesitation is a filler that does not alter the task goal.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Buy me a blender—sorry, I meant stand mixer as Connor recommended in their phone message.", "long_pause_offsets": [], "source_sha256": "3cdb2643100dd9ebd6fcbe38f8061f926654843a7cdd92187bbc014b9e3d46b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'blender' to 'stand mixer' within the same utterance, preserving the rest of the instruction including the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/false_start/false_start_01_product_category/24dd5c9de74aa45d/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "911bcae7432aa850e93a7b6c8d175b47cf253a8264c013dccb0b126191891bcc", "seconds": 7.0, "spoken_utterances": [ "Buy me a ah... blender—um... sorry, I meant stand mixer as Connor recommended ... in their phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blender' but immediately self-corrects to 'stand mixer'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including the condition 'as Connor recommended in their phone message', is preserved. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Buy me a a blender um sorry I meant stand mixer as Connor recommended in their phone message." }, "candidate_id": "wav_7d27948170295caf", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Buy me a stand mixer in their phone message. Oh, I forgot to mention, it should be as Connor recommended.", "long_pause_offsets": [ 45 ], "source_sha256": "2c5843f98a2e66b0c2d30c596af1afa656abcda5cda140f48f69217d637ab523", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'as Connor recommended', which clarifies the source of the recommendation. The initial phrase 'in their phone message' is replaced by the more specific and correct reference to Connor's recommendation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/addition/addition_01/d90bfeb2af8506c6/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f9654dc7dcdcd782248e88a5b94b68a661d9fa4fd714816c8b4dc0278a784d69", "seconds": 11.76, "spoken_utterances": [ "Buy me a stand mixer um... in their phone message. ", "Oh, I forgot to mention, it should be as hmmm... Connor recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Buy me a stand mixer, um, in their phone message.' This is followed by a longer pause and an addition: 'Oh, I forgot to mention, it should be as, hmm, Connor recommended.' The addition clarifies the condition under which the mixer should be bought, completing the original request. The final intent combines the initial request with the added constraint, matching the original instruction. The hesitation and filler words do not change the final goal.", "heard_transcript": "Buy me a stand mixer, um, in their phone message. Oh, I forgot to mention, it should be as, hmm, Connor recommended." }, "candidate_id": "wav_46773ed6a166ff52", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_product_category", "proposed_instruction": "Buy me a blender as Connor recommended in their phone message. Change the product to a stand mixer, not a blender.", "long_pause_offsets": [ 63 ], "source_sha256": "19a18b1dcb0fa44013cd6f46a3cc853ef22715477f5749c7f39e97ae00dda59a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the product from 'blender' to 'stand mixer', preserving the recommendation source and the action of buying." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/revision/revision_01_product_category/681da2ad5e89669c/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "15525e87a755203b01b90fc6fbc3096afd2e023513fb8ed329ff51aa74aa8865", "seconds": 11.64, "spoken_utterances": [ "Buy me a uh... blender as Connor recommended um... in their phone message. ", "Change the product to a stand mixer, not a blender." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blender' but immediately corrects it to 'stand mixer' in the revision clause. The final audible intent is to buy a stand mixer, which matches the original instruction. The revision is clear and applies to the product type, preserving all other conditions.", "heard_transcript": "Buy me a blender as Connor recommended in their phone message. Change the product to a stand mixer, not a blender." }, "candidate_id": "wav_15aecd697058b64f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'uh' and 'um' filler, then delivers a coherent initial request ('Buy me a blender as Connor recommended in their phone message'). The revision ('Change the product to a stand mixer not a blender') is a clear, self-contained correction that flows logically from the initial statement. The prosody is consistent, with natural stress and pacing, and the correction is delivered without awkwardness or repetition. The overall delivery sounds like a single, spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "179101a78dbe2edad7818794a35d892f53fac81fa4a22c1958685382e7f51393", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/cancellation/cancel_019/6e1cd12512b8a2ba/speech.wav", "wer": 0.25, "wav_sha256": "59ccb459db5e0ea10bc1ac50a0c018c7e428abd66a8989f0e7406fb887085827", "seconds": 9.84, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message.", "Hold on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly cancels it with 'Hold on, never mind, don't do that.' This constitutes a final cancellation of the original request. The initial part of the audio is a false start, but the final interjection is a direct and unambiguous cancellation. The control intent matches the expected 'cancel' for this type.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. kind of recommend in their phone message. Hold on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "14eba60e54bfd8673e21cc73d93d50864845fa03ac263847ce8f97c75b95a4ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/ambiguous_halt/halt_021/e90b024643f55008/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "76c657c4558ee8030d29d3d5d22bec0ce9a360906488c0f2a165426f583bd45f", "seconds": 9.32, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial request is not completed, but the speaker does not revoke or modify it; they only signal a need to pause.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. Connor recommended in their phone message. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f2d553e68c719a42b74892b4a99a01e2fca5a4928c9938d123aaed834dfd5a68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/filler/original/0dcee260f221c6bc/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "7529af948439861020bef263f874877fd7aca580c6ad486bd437973c3d6fcb2a", "seconds": 5.32, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh' and 'let me think' but does not change the core request. The final audible intent matches the original instruction. The hesitation and backchanneling are typical of a filler type, and the task is completed as stated.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. Connor recommended in their phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_person_name", "proposed_instruction": "Buy me a backpack as Liam—sorry, I mean Mason recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "f22ce438ac44b0f76cd4dc7e39fcc8a9cc143c4b8668400ddf117a63172d2ee5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misnames the recommender as Liam but immediately corrects to Mason. The action, item, and context (phone message) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/false_start/false_start_02_person_name/b782bdfc13f2e359/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "f68879892c99e8d080d8f832fc7ca80dc824206d7f9f390b6769c0663e0f5e15", "seconds": 8.76, "spoken_utterances": [ "Buy me a backpack as Liam—uh... sorry, I mean Mason recommended um... in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Liam' but immediately self-corrects to 'Mason' with 'sorry, I mean Mason'. This is a clear false start where the final intended target is Mason. The rest of the instruction, including the action 'buy me a backpack' and the context 'recommended in his phone message', remains unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Buy me a backpack as uh Liam uh sorry, I mean Mason recommended um in his phone message." }, "candidate_id": "wav_071fd03514c061f9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural false-start pattern: the speaker begins with 'Liam', self-corrects to 'Mason' with a clear, spontaneous 'sorry, I mean', and continues smoothly. The filler 'uh' and 'um' are placed naturally, and the overall prosody reflects genuine hesitation and correction without robotic pauses or unnatural phrasing." } ] }, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Buy me a backpack as Mason recommended. One more detail: Mason recommended it in his phone message.", "long_pause_offsets": [ 40 ], "source_sha256": "f85bac2ba4ecd89b13956f4aa6e085cf3c8281ba6b3815b154ffc8dbfb9f0e29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail that the recommendation came from a phone message, which was already implied in the original instruction. The core action and subject remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/addition/addition_02/3dd78e2a0cb440a3/speech.wav", "wer": 0.0, "wav_sha256": "0f5fa8a44d2ea28c5a365da0e1ef11eab3187c8068920a8a8cf1774a3aaa8c31", "seconds": 11.48, "spoken_utterances": [ "Buy me a ... backpack as Mason recommended. ", "One more detail: Mason recommended it in his ... phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to buy a backpack as recommended by Mason, then adds the detail that the recommendation came from a phone message. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and complete, matching the original instruction with the added constraint.", "heard_transcript": "Buy me a backpack as Mason recommended. One more detail, Mason recommended it in his phone message." }, "candidate_id": "wav_f2ca66ecde3b76be", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the initial request with natural prosody and a clear, spontaneous-sounding addition. The phrase 'one more detail' acts as a believable conversational cue before the speaker recalls and integrates the missing information ('Mason recommended it in his phone message') smoothly and cohesively. The pacing and intonation suggest genuine recall rather than a rehearsed or robotic insertion." } ] }, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_person_name", "proposed_instruction": "Buy me a backpack as Oliver recommended in his phone message. Sorry, I meant Mason recommended it, not Oliver.", "long_pause_offsets": [ 62 ], "source_sha256": "8e7fbdad633f70ca5223bf2ad2b330cac8236294bede6229762980ba7b4ab1fa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recommender from Oliver to Mason, preserving the action, item, and context of the phone message." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/revision/revision_02_person_name/9701d4869064ce13/speech.wav", "wer": 0.2, "wav_sha256": "57c278f38435d0c4546f8dbef8fec3e3faa8d90e552d6654eeff973d81897c05", "seconds": 10.52, "spoken_utterances": [ "Buy me a uh... backpack as ... Oliver recommended in his phone message. ", "Sorry, I meant Mason recommended it, not Oliver." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misattributes the recommendation to 'Oliver' but immediately corrects themselves to 'Mason' after a pause. The revision is clear and final, changing the recommender from Oliver to Mason. All other conditions (buying a backpack, based on a phone message) remain unchanged. The correction satisfies the revision type criterion.", "heard_transcript": "Buy me a uh backpack as cus Oliver recommended in his phone message. Sorry, I meant Mason recommended it, not Oliver." }, "candidate_id": "wav_27f882a3699de2eb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker starts with a plausible false start ('uh uh backpack as cust...'), self-corrects to 'Oliver recommended', then later corrects to 'Mason recommended'. The correction is natural, with a clear shift in intonation and a believable 'sorry' before the revision. The pacing and filler usage ('uh uh') are consistent with spontaneous speech, and the final sentence flows smoothly after the correction." } ] }, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_016", "hide_variant_original": true, "source_sha256": "05673dfb1439f193ea05b07b082a9c214ea86daeeecf9cd2a2ee89e94b9edfd6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/cancellation/cancel_016/e3e13aa533939081/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ba2a6dd924107d05a1b382f3fa2fc1c860f300a742ad435bb70ded82e79473ca", "seconds": 11.52, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message.", "Oh, wait, you know what? Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Cancel that.' This constitutes a cancellation as per the type-specific criterion. The final audible intent is to cancel, not to execute or pause.", "heard_transcript": "Buy me uh let's see backpack as Mason recommended in his phone message. Oh wait, you know what? Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "e53b68734208eb1d2932837d9d3855c50b07f1be7422bbbe040def3dba1d82e9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/ambiguous_halt/halt_023/05e3b35ce04f315b/speech.wav", "wer": 0.1, "wav_sha256": "d0f5f9f0e8b0b2928ceffc6270e92e53d9302c967ad36b1128da528dbf581697", "seconds": 11.0, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or modified, and the speaker does not resume or correct the instruction. The final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Buy me a uh, let's see, backpack as Mason recommended in his phone message. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "7e1be84_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "426369198e46a044807e7720b324b283b89b7195521a4ced54bd1e81a950a125", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/filler/original/da54d2d968a1ff69/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "61db4c8971ac65c4ec60ba5172e94f78adfd2a5aadb13f14cc09a7d026dcbb06", "seconds": 6.88, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh let's see' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Buy me uh let's see backpack as Mason recommended in his phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_source_collection", "proposed_instruction": "What is the title of the most-liked song in my Spotify library—scratch that, make it playlists.", "long_pause_offsets": [], "source_sha256": "3aa0a5044729ac2c9e7c27cd8f9496697c8aa3aca1d985e1a1ff9f1d2b9a41a3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'library' but immediately corrects it to 'playlists' within the same utterance. The final target is 'playlists', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/false_start/false_start_03_source_collection/1e0629ddedd94852/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "fcf4c1cd1d07e03fc27091a601c04d9545af7fc8885a98f8c8eed088b8d1a342", "seconds": 9.84, "spoken_utterances": [ "What is the title of the ... most-liked song in my Spotify library—uh... scratch that, make it playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify library' but immediately self-corrects to 'playlists' after a pause and the interjection 'scratch that'. According to the 'false_start' type criterion, the final corrected value ('playlists') should be applied, and the original incorrect value ('library') should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "What is the title of the most-liked song in my Spotify library? Uh, scratch that, make it playlists." }, "candidate_id": "wav_1195e4d349c4940b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start: the speaker begins with 'What is the title of the most liked song in my Spotify library' and then immediately self-corrects with 'uh, scratch that, make it playlists' in a spontaneous, conversational tone. The correction is integrated smoothly into the same utterance, with a realistic hesitation ('uh') and a casual, self-directed correction ('scratch that'). The prosody and pacing are consistent with someone thinking aloud and adjusting their request on the fly." } ] }, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "What is the title of the most-liked song in my playlists. I should also mention, this is in my Spotify playlists.", "long_pause_offsets": [ 58 ], "source_sha256": "910f84570691b5b99a887786a613b2dfd69e99a724e88d09283a324188d5dac9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the context of 'playlists' as being on Spotify, which was already implied in the original instruction. The core request for the most-liked song's title remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/addition/addition_02/d4c86e75ccba9d3d/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "f4b711f103c579dba367b166985a147d3adfd32c3d357df7bb3c76a4695c77b5", "seconds": 14.28, "spoken_utterances": [ "What is the title of the uh... most-liked song in my playlists. ", "I should also mention, this is in my ... Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'What is the title of the most liked song in my playlist?' and is followed by an addition clarifying the context as 'this is in my Spotify playlists.' The addition provides a specific target for the initial request, combining the two into a single, coherent instruction. The final intent matches the original instruction with the added specificity, satisfying the addition type criterion.", "heard_transcript": "What is the title of the um most liked song in my playlist? I should also mention, this is in my Spotify playlists." }, "candidate_id": "wav_13c6f849e7126105", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural hesitation ('um') and a believable additive structure, first asking about the most liked song and then adding context about Spotify playlists. The pacing is conversational, with a slight pause before the addition that sounds like spontaneous recall. The intonation and rhythm are consistent with someone formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "What is the title of the most-liked song in my Spotify library. That should be 'playlists', not 'library'.", "long_pause_offsets": [ 64 ], "source_sha256": "9747ad2d2d2a6ff951cd29784fdb5a61cfe022308b207139c724456c463bcb16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'library' to 'playlists', preserving the action and the rest of the query." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/revision/revision_03_source_collection/ffac2f7e7008b3b4/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "1897e39e27a3acb98a02ac560c8375fc22d57b4d4e335f0d0410714c1ec6a734", "seconds": 11.72, "spoken_utterances": [ "What is the title of the let's see... most-liked song in my ... Spotify library. ", "That should be 'playlists', not 'library'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'library' but immediately corrects themselves to 'playlists' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the most-liked song in 'playlists', which matches the original instruction. The correction is clear and unambiguous.", "heard_transcript": "What is the title of the let's see, most liked song in my uh Spotify library. That should be playlists, not library." }, "candidate_id": "wav_6c790149295ec64a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see'), then delivers the initial request fluently. The correction ('That should be playlist not library') is a clear, self-contained revision with a believable shift in intonation and pacing, indicating a spontaneous realization. The flow is continuous and the prosody sounds like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "69f4cfa84c85383581c1fa4f936a54ea44fc9d6c660afcbe00af3d5859fec1e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/cancellation/cancel_001/78c00c00a7cc6610/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "0ac7e1511325c768b7df10fef410962fcec41e034ed3753f51837f0ec9ee2501", "seconds": 11.32, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to ask for the title of the most-liked song in their Spotify playlist but then immediately says 'Oh wait, cancel that.' This constitutes a clear and final cancellation of the initial request. The control intent matches the 'cancellation' type criterion, as the recording ends with an explicit revocation of the task.", "heard_transcript": "What is the title of the most like song in my um Spotify playlist? Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_009", "hide_variant_original": true, "source_sha256": "f595741e77037f82d56016e57db06b55a2faf94fc031d6f04885f8f8d4f9b04f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/ambiguous_halt/halt_009/d2d4d54c09e7a250/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8a571363445944157c4311c40fc11ff5ee2c50cf8081b5c493333424316e72fd", "seconds": 10.6, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists.", "Wait a minute..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait a minute,' which is a clear request to pause or hold. The initial request is not completed, and there is no indication of cancellation or a new instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "What is the title of the most-liked song in my um Spotify playlist? Wait a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "056af54a913006b9a8970143e3a6b5dd8ff54081dd41796878e398896c14e354", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/filler/original/ba18b2f70d034cdb/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "ab9c793532518b696410563966593ad03eceba7e75c7de74558c82be8927b65c", "seconds": 7.56, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most like song' but corrects to 'most-liked song' in the context of the original instruction. The use of 'um' is a filler, and the final intent is clear. The task is to find the title of the most-liked song in the user's Spotify playlist, which matches the original instruction.", "heard_transcript": "What is the title of the most like song in my um Spotify playlist?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most-played—no, I mean least-played song in my Spotify song library.", "long_pause_offsets": [], "source_sha256": "aeb6904eb6b65d52945c6f51aff940c7799681f7851a1a06344fa0977bee9838", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most-played' but immediately corrects to 'least-played'. The final intent matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/false_start/false_start_01_song_selection/c0689bc84c380a9a/speech.wav", "wer": 0.2, "wav_sha256": "dbd6e5a9a6d2d2bcca4ab8e95f3dd7af8ae1640adbaaeaee119ce45c779c244b", "seconds": 8.72, "spoken_utterances": [ "What is the title of the um, let me think... most-played—uh... no, I mean um... least-played song in my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'least played'. According to the 'false_start' type criterion, the immediate self-correction should be applied, and the final intent is to find the least-played song. The correction is clear and unambiguous, so the task is executed as intended.", "heard_transcript": "What is the title of the um, let me think, most played uh, no, I mean um, least played song in my Spotify song library." }, "candidate_id": "wav_0bb7a02c2a69618b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible false-start flow. The initial request 'most played' is clearly a mistake, immediately self-corrected with 'no, I mean least played' in a way that sounds spontaneous and not rehearsed. The filler 'um' and 'let me think' are placed naturally, and the correction is integrated smoothly into the same utterance without awkward pauses or repetition. The prosody and rhythm are consistent with someone thinking aloud." } ] }, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "What is the title of the least-played song in my song library. I left out a detail: the song library is on Spotify.", "long_pause_offsets": [ 63 ], "source_sha256": "f3fbddaa2f22ab60c89954367b0b92be573a52152a9a1555e2cabf5e52643033", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify' to the initial request, which was already about the least-played song in a song library. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/addition/addition_02/0ea61f462abbc714/speech.wav", "wer": 0.03125, "wav_sha256": "0f4835d0d9f0e83e9118baec81c6a623a4b5635e24c85f58aadf46ee2c7379fc", "seconds": 14.2, "spoken_utterances": [ "What is the title of the um, let me think... least-played song in my song library. ", "I left out a detail: the song library is on uh, let me think... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was missing the source 'Spotify'. After a pause and self-correction, the speaker added the missing detail. The final intent combines the original request with the added information, matching the reference instruction. The 'addition' type is satisfied.", "heard_transcript": "What is the title of the um, let me think, least played song in my song library. I left out a detail, the song library is on uh, let me think, Spotify." }, "candidate_id": "wav_3ffb1c8dcb032485", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the most-played song in my Spotify song library. Wait, the selection should be the least-played song, not the most-played.", "long_pause_offsets": [ 70 ], "source_sha256": "20c5cbd942c62eda1e04df37e88de88211622314e2bc72a9f96370b9a474e160", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the selection criterion from 'most-played' to 'least-played', while retaining the app (Spotify), the data source (song library), and the requested output (title)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/revision/revision_01_song_selection/5d2650a0f5ecda0d/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "25e973450004808cb401e0aa1168f02ea4e8e895c17ebb281fc741df65672704", "seconds": 14.0, "spoken_utterances": [ "What is the title of the um... most-played song in my uh... Spotify song library. ", "Wait, the selection should be the ... least-played song, not the most-played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially requested the most played song but immediately corrected themselves to the least played song. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to find the least-played song, which matches the original reference instruction.", "heard_transcript": "What is the title of the most played song in my uh Spotify song library. Wait, the selection should be the least played song, not the most played." }, "candidate_id": "wav_6248ae53d2fa593a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('What is the title of the most played song...') with a natural 'uh' filler and a pause. The revision ('Wait, the selection should be the least played song...') is a believable change-of-mind, with a clear shift in intonation and a smooth reference back to the initial request. The prosody is continuous and the pacing feels like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "1ce513ab97f936f12f35e3a0279ffb6ac5d5e892315378d3812c2bce9569ff0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/cancellation/cancel_040/5ce7cfac50905958/speech.wav", "wer": 0.0, "wav_sha256": "0631fd390025cc1000d6935336493037b490bbd6f92d18cdaeed62fa323af5a3", "seconds": 13.12, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library.", "Hang on, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking for the least-played song in their Spotify library but then immediately follows with 'Hang on, I don't want to proceed with that task. Please cancel it.' This constitutes a clear and final cancellation of the original request. The final audible intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library. Hang on, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "c52f08068ef38ed8944995943e1279239601075bee34932560f9f3e8e91f8921", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/ambiguous_halt/halt_014/91292a2f8e5b924a/speech.wav", "wer": 0.0, "wav_sha256": "33965ef4554690bc3b1ef38e1ad9d9ca1226e9b98bb1fd0c38a0e52b93284fb5", "seconds": 10.92, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original request and then clearly says 'Can you hold on a second?', which is an audible request to pause. There is no cancellation or instruction to resume. The final interjection is a pause request, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bfbf78ec20e7aa55f47c46eae116b9c83d161aecb3b2d93485b42df0ca60238a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/filler/original/ffc786fb027487a9/speech.wav", "wer": 0.0, "wav_sha256": "febf5de92e444ec2d5ae03b50875c016550bb0836aa4f097641e9b9fb91c5d5a", "seconds": 7.52, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'What is the title of the least played song in my uh Spotify song library?' and the 'uh' is a filler word that does not change the task goal. The final intent is to retrieve the title of the least-played song in the Spotify library, which matches the original instruction. The filler word does not alter the meaning or intent of the task.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most-listened—sorry, I meant most-played song in my Spotify album library.", "long_pause_offsets": [], "source_sha256": "ade29adbe2e59e211b9414d415d56170fd10f710a5bece1fd997d000f60f7cf7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most-listened' but immediately corrects to 'most-played'. The final intended action is to identify the most-played song, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/false_start/false_start_01_song_selection/424e1b0f33032153/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "f49e19ba8d2958efe86f25b901509ced0df600029d42da0bd31a4a15c3382c5f", "seconds": 8.0, "spoken_utterances": [ "What is the title of the most-listened—hmmm... sorry, I meant most-played song in my uh... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most listened' but immediately self-corrects to 'most played' with 'sorry, I meant'. This is a false start where the final intended action is the corrected version. The rest of the instruction, including 'song in my Spotify album library', remains unchanged and is preserved. The final audible intent matches the original instruction after the correction.", "heard_transcript": "What is the title of the most listened hmm, sorry, I meant most played song in my uh Spotify album library." }, "candidate_id": "wav_59f16e32fbcaec32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a believable self-correction. The initial 'most listened' is immediately followed by a hesitant 'huh' and a smooth, unforced correction to 'most played song'. The prosody is consistent, and the correction integrates seamlessly into the sentence without awkwardness or repetition. The pacing and intonation suggest genuine spontaneous speech." } ] }, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song in my Spotify album library. Oh, I forgot to mention, it should be the most-played song.", "long_pause_offsets": [ 59 ], "source_sha256": "f97817a63d6dcdc50a0c58e15ebd75be3a5bdd3d8b423ded4e1f047f9aa819d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most-played' to the initial request for a song title, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/addition/addition_01/731dbf3e9b797525/speech.wav", "wer": 0.0, "wav_sha256": "311016679da296e1a2dc3224da0b7b2f598aff13aa4ac02e9db9ce0de6f8a819", "seconds": 11.0, "spoken_utterances": [ "What is the title of the song in my ... Spotify album library. ", "Oh, I forgot to mention, it should be the ... most-played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is to find the title of a song in the user's Spotify album library. After a pause, the user adds the constraint that it should be the 'most played song'. This is a clear addition of a new condition to the original request. The final intent combines both the original scope and the new constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "What is the title of the song in my Spotify album library? Oh, I forgot to mention it should be the most played song." }, "candidate_id": "wav_41fbb0f21dd07ac0", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "What is the title of the most-played song in my Spotify playlist collection. Change the source to album library instead of playlist collection.", "long_pause_offsets": [ 77 ], "source_sha256": "f30b8228515bd35804c895fc38c82d7e9af44344b95f58bcc1aa331621786249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly changes the source from 'playlist collection' to 'album library', preserving the action and the app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/revision/revision_03_source_collection/3a43858820015cce/speech.wav", "wer": 0.0, "wav_sha256": "9e38747e0549aa6e3f7facb34321c63565525b76d5514f3536160b8de6477408", "seconds": 14.52, "spoken_utterances": [ "What is the title of the um... most-played song in my uh, let's see... Spotify playlist collection. ", "Change the source to album library instead of playlist collection." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the most-played song in the 'playlist collection' but then immediately corrects the source to 'album library'. This is a revision where the final instruction overrides the initial incorrect source. The final intent matches the original reference instruction, so the task should be executed as requested.", "heard_transcript": "What is the title of the um most played song in my uh let's see Spotify playlist collection. Change the source to album library instead of playlist collection." }, "candidate_id": "wav_a4fea937a4c98857", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, continuous flow with plausible hesitation ('uh', 'let's see') and a clear, self-contained revision. The correction ('change the source to album library instead of playlist collection') is smoothly integrated into the same utterance, with a believable shift in intonation and pacing that suggests spontaneous thought. The prosody is consistent, and the revision feels like a genuine in-the-moment adjustment rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "d8d5f81443ef2074798897cd523c6ddb181e0352327f9f587a96cb1d638ebabb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/cancellation/cancel_014/ebf539f495b8fdf0/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6ebcb8f597c65ff21fafa398b964d0ffbb22c92ab2481dfeafb951d9cbaa99fb", "seconds": 11.04, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library.", "Wait, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly and unambiguously cancels it with 'cancel that' as the final interjection. This meets the criterion for a cancellation, as the final audible intent is to revoke the task.", "heard_transcript": "What is the title of the uh most played song in my um Spotify album library. Wait, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "3f4d833e353417879440a939fdf0b0612d0bac8965b6237891fdc3b26d853d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/ambiguous_halt/halt_028/5589277f56d66e74/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "8ad29a3422191e0827947a58fea827bfdd4cb241357b26deeaa755aab1b922db", "seconds": 10.8, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request is not completed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "What is the title of the most played song in my um Spotify album library? Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "82e2fac_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9621770f4e1428eb269c4cff2dc4d4cd8be5f0655bda26b0d94de70117119fbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/filler/original/a3768b5e77c49e68/speech.wav", "wer": 0.125, "wav_sha256": "b7ebe4e1f03bede71760c477d86b111c63033b43fad67d34e926affc25ac86f1", "seconds": 6.6, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "What is the title of the uh most played song in my um Spotify album library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks—no, make that recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "92ec7681a07b9f3610cb15e9a54b9564f52c80e7d24079d5a7fa9c49fb7f098f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'top tracks' with 'recommended songs'; all other actions (reset, shuffle, play) and the app (Spotify) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/false_start/false_start_02_recommendation_source/99d3cb94754efce4/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "8d230a134b7cadd42348b59bdf67a179a7c809d68394c7315af7a717ed1ea077", "seconds": 14.32, "spoken_utterances": [ "Reset my Spotify queue with all of its um... top tracks—hmmm... no, make that recommended songs, uh... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately self-corrects to 'recommended songs' after a pause and 'No, make that'. The final instruction matches the original reference. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Reset my Spotify queue with all of its um top tracks. No, make that recommended songs. Uh, shuffle it and play it." }, "candidate_id": "wav_98f796f67cb94fef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural and plausible false start. The initial phrase 'top tracks' is clearly a spontaneous misremembering, immediately corrected with 'no, make that recommended songs' in a way that sounds like genuine self-correction. The prosody is fluid, with natural pauses and filler words ('um') that fit the context of formulating a request. The delivery is continuous and the correction is integrated smoothly into the utterance, making it the most realistic false start." } ] }, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and. One more thing: start playing the shuffled queue.", "long_pause_offsets": [ 75 ], "source_sha256": "e15a4da481b180a9838f5864bd1ffe182cbbec3c368ae99a52639a299f8dc198", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing action to start playing the shuffled queue, completing the original instruction which already included reset, shuffle, and the queue context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/addition/addition_02/162ee9a4f04a7c51/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "bfc76366da4e2f47d043dca69430c41d0c8154241c058974b85665a352b84283", "seconds": 15.64, "spoken_utterances": [ "Reset my Spotify queue with all of its um... recommended songs, uh... shuffle it, and. ", "One more thing: ah... start playing the shuffled queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to reset the Spotify queue with recommended songs and shuffle it. After a pause and an interjection ('one more thing'), the speaker adds the final action to 'start playing the shuffled queue'. This matches the 'addition' type criterion, where the final goal combines the initial instruction with the added information. The final intent is complete and executable, matching the reference instruction.", "heard_transcript": "Reset my Spotify queue with all of its um recommended songs uh shuffle it and one more thing uh start playing the shuffled queue." }, "candidate_id": "wav_f585f07307780ae3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with a believable flow. The initial request is followed by a pause and a filler ('uh'), then a smooth addition of a new instruction ('one more thing') with a second filler ('uh') before the final command. The prosody is consistent, and the intonation reflects a speaker formulating and expanding on a thought in real time, which is highly realistic." } ] }, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Reset my Apple Music queue with all of its recommended songs, shuffle it, and play it. Wait, reset my Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [ 87 ], "source_sha256": "15044bc5b145c1f521f717be79533cb347d407c860be66ab05bae1c4f5ccfd7d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining all other actions and conditions from the initial instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/revision/revision_01_app/fb058a45a61c2e82/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "37dada05ce53e1914b29e0a970e5aa5cc69423d9f5e4abfedc996fa127207c4d", "seconds": 19.76, "spoken_utterances": [ "Reset my Apple Music queue with all of its ... recommended songs, let's see... shuffle it, and play it. ", "Wait, reset my uh... Spotify queue with all of its recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Reset my Apple Music queue with all of its recommended songs, let's see, shuffle it and play it. Wait, reset my uh Spotify queue with all of its recommended songs, shuffle it and play it." }, "candidate_id": "wav_12adfbef453658f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker starts with a clear request, then hesitates with 'let's see' before self-correcting to the correct platform ('Spotify') and smoothly continuing. The correction is integrated into the ongoing utterance without awkwardness, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "f00d9bc204c0d3d1afb6d16f1345c79665e56f5c797ef59ba4eabc3b221ea0f4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/cancellation/cancel_018/41390f2570a7981f/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "65406c959033a8704e855649abbd8404352fd93e94b32a48d59f55089ce9cd3a", "seconds": 15.12, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it.", "Hold on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a complete task request, then immediately follows with 'Hold on, I'd rather not do that after all. Please cancel it.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or modification of the initial request.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it. Hold on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "77813149d291b1b444be9abac9697116c3a0068e2fb04d0c5feb1a3230a841c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/ambiguous_halt/halt_034/92afc6a40039e3af/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "22aba060f930064cf9a39912e9a20d142be9a783ed9f12ea7c4f070772e066bd", "seconds": 12.8, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial task is not canceled, but the speaker indicates they need a moment to reconsider, so the agent should pause.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1ed9d24df49faeadc97e4869828509e3cfbaa894259a475c0e3f6f44c125c723", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/filler/original/d7cad1f892b7d06f/speech.wav", "wer": 0.0625, "wav_sha256": "355bffaaa84a7630a770d37e27a7a2603b6ba2543ef1c5ea8669631852d21b3e", "seconds": 8.56, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Reset my Spotify queue with all of its recommended songs, shuffle it, and play it.' The 'um' is a filler sound that does not alter the task goal. The final intent is to execute the full original instruction, which matches the reference. No correction, cancellation, or pause is requested.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, sort it—sorry, I meant shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "ad13700741022863fe2d967b5f2bdb5938bcdada49fe1d7643e10a128cf3380b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it' but immediately corrects to 'shuffle it' within the same utterance. The final action is shuffle, which matches the original instruction. All other components—resetting the queue, including recommended songs, and playing—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/false_start/false_start_03_queue_ordering_action/d931a092b88b357b/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "315935329b4f5a7492ca983d2b4286f3c571964536bf138cb036678157283032", "seconds": 13.32, "spoken_utterances": [ "Reset my Spotify queue with all of its hmmm... recommended songs, sort it—uh... sorry, I meant um... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it out' but immediately self-corrects to 'shuffle it'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('sort it out') is discarded. The final instruction is to shuffle and play the queue, which is a valid and complete task. The correction is immediate and unambiguous, so the task is executed as intended.", "heard_transcript": "Reset my Spotify queue with all of its hmm recommended songs, sort it out. Sorry, I meant um shuffle it and play it." }, "candidate_id": "wav_6f5cba31c3798d70", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural false start: the speaker begins with a clear, slightly hesitant request, then self-corrects from 'sorted up' to 'shuffle it' with a believable 'sorry, I meant um' transition. The prosody includes a realistic pause and filler ('hmm'), and the correction flows smoothly without awkwardness. The rhythm and intonation suggest genuine spontaneous speech." } ] }, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs,, and play it. Oh, I forgot to mention, shuffle the queue.", "long_pause_offsets": [ 72 ], "source_sha256": "95a58f6ec9fbd63fb42b953e48c18c51e16627d4aa7867b333980d66e1055b55", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing shuffle action, which was not in the initial utterance but is present in the original instruction. The final goal combines all three actions: reset, shuffle, and play." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/addition/addition_01/b12c3c5764fd583d/speech.wav", "wer": 0.0, "wav_sha256": "438190b70b24c95e32fd91d8dd8b478b890b62e41f2ca1c332d7cbb511e0bb7e", "seconds": 13.32, "spoken_utterances": [ "Reset my Spotify queue with all of its ... recommended songs,, and play it. ", "Oh, I forgot to mention, um... shuffle the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to reset the queue with recommended songs and play it. After a pause and an interjection, the speaker added 'shuffle the queue' as an additional instruction. Since the type is 'addition', the final goal is the combination of the initial and added instructions. The final audible intent matches the original reference instruction exactly, including the shuffle action.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs and play it. Oh, I forgot to mention, um, shuffle the queue." }, "candidate_id": "wav_26e6a374a81dcf39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with natural prosody. The addition of 'Oh, I forgot to mention' followed by a brief pause and the insertion of 'shuffle the queue' feels like a spontaneous afterthought. The filler 'um' is placed naturally, and the overall flow is coherent and conversational, with no signs of artificiality or forced correction." } ] }, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, clear it, and play it. I misspoke about the queue action: shuffle it, not clear it.", "long_pause_offsets": [ 81 ], "source_sha256": "827186ebfb0863e415bffcfafa254cbbd3757b2f3c3c5080b4a9116a16953755", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'clear it' to 'shuffle it', preserving the initial reset and play actions, as well as the scope of recommended songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/revision/revision_03_queue_ordering_action/3d71c794258d9ac3/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "fb0a00bf3fe17867f1935c680fbd5dc3aa6c2b99050155e3721d97442757da97", "seconds": 18.28, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, let's see... clear it, and play it. ", "I misspoke about the queue action: hmmm... shuffle it, not clear it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'clear it' but immediately corrects themselves to 'shuffle it, not clear it' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, let's see, clear it and play it. I misspoke about the queue action. Hmm. Shuffle it, not clear it." }, "candidate_id": "wav_bdab6d337cad18cd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a clear revision after a brief pause and filler ('let's see'), correcting 'clear it' to 'shuffle it'. The self-correction is natural and the pacing reflects a moment of reconsideration. However, the initial request is slightly less fluid than audio_03, with a minor stumble on 'recommended songs' and a slightly more abrupt correction." } ] }, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "da49f61f39d09212faef89b5b2e44da1568a6c6ce6fef55b6ed3514fd7a429e4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/cancellation/cancel_032/9cede3f92c46e258/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "e3079af065eef4d73e4c129c71d7a90d8cc3627dfaf66b6a56685dc08786d743", "seconds": 15.2, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it.", "Hold on, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to reset and play the Spotify queue, then immediately follows with 'Hold on, I don't need this anymore. Please cancel the request.' This constitutes a clear and final cancellation of the initial task. The cancellation is explicit and not merely a pause or hesitation.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it and play it. Hold on, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "81bb7371093b47b1f60433777a28a277b21ec8516063400d1202cfe2a175bf16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/ambiguous_halt/halt_010/8685b07a0c13a31f/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "fcdd9dc8af7134b75a1b4cb8e68b20722367fdb8804f05c1f8760f9ac429b235", "seconds": 12.56, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then audibly requests to pause with 'Hold on a second.' This constitutes a final, clear request to pause, which matches the 'ambiguous_halt' type criterion. The initial hesitation ('let me think') is a filler and does not alter the final intent.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it, and play it. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2d4faff18608fad85c5aa7a9f878b220ae0f99852e7440f2a24f5d61d830403d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/filler/original/d51047e6cee1d7f1/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "9417664420b3feab7471568a098edbc666df3cef227f78e41e0b9858987fcb32", "seconds": 9.28, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial hesitation 'let me think' is a filler that does not alter the final task goal. The speaker proceeds to complete the original instruction without correction or change. The final intent matches the reference, and the filler is ignored per the type-specific criterion.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it, and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Reset my Apple Music—I should say Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "4a83d9f82f556153f23704001c47f6dd68ed1c751bb4e8b7165fe71a3c9125f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving all other actions and conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/false_start/false_start_01_app/866c9d2ccceae62e/speech.wav", "wer": 0.0, "wav_sha256": "410951e78b764d5d9e32048e8098749765e6241de9bfdb05c15d428b7785718d", "seconds": 11.4, "spoken_utterances": [ "Reset my Apple Music—um... I should say Spotify queue with all of its uh... recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction. The rest of the task constraints are preserved and correctly executed.", "heard_transcript": "Reset my Apple Music, um, I should say Spotify queue with all of its uh recommended songs, shuffle it and play it." }, "candidate_id": "wav_d02cc4604030f25f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction ('I should say Spotify queue') that flows seamlessly into the rest of the sentence. The prosody is consistent, the pace is conversational, and the correction feels like a spontaneous realization rather than a forced edit. The phrasing is slightly more fluent than others, with minimal hesitation." } ] }, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and. One more detail: start playing the queue.", "long_pause_offsets": [ 75 ], "source_sha256": "d12d23c1eda0a5e1285be9063d4c65f7614b1ada93d861fda5857aa373cfdaa6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing action of starting playback, which was implied but not stated in the original instruction. The original instruction ends with 'and', suggesting an incomplete action list. The follow-up completes the goal by specifying 'start playing the queue'. All other requirements (resetting with recommended songs, shuffling) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/addition/addition_02/659085d35cba2283/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ae0c8b2d293099dc51e8e111bb66de32321c35592dc1f6ef199dae8baf6a233a", "seconds": 12.72, "spoken_utterances": [ "Reset my uh... Spotify queue with all of its recommended songs, ... shuffle it, and. ", "One more detail: ... start playing the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. After a pause and the interjection 'one more detail', the speaker adds 'start playing the queue', which is an addition to the original task. The final goal combines the original action with the new condition, resulting in a complete and coherent instruction that matches the reference when the addition is included.", "heard_transcript": "Reset my uh Spotify queue with all of its recommended songs, shuffle it, and one more detail, start playing the queue." }, "candidate_id": "wav_5d4dc639991c90f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with a clear addition. The initial request is fluent, followed by a brief pause and the filler 'uh' before 'shuffle it'. The second addition, 'one more detail', is a plausible conversational cue, and the final instruction 'start playing the queue' is smoothly integrated with natural intonation and pacing. The flow feels like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that 'recommended songs', not 'top tracks'.", "long_pause_offsets": [ 76 ], "source_sha256": "e92c3389105240d2b844885db8692b4e585c50fa093b8007ff6276c533bcacc0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'top tracks' to 'recommended songs', preserving the app, action sequence, and other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/revision/revision_02_recommendation_source/65c463ecd86e45e0/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "0e976ed994f77ba749aaa2116b9e61ad508252164de3f465884d93025f613561", "seconds": 15.68, "spoken_utterances": [ "Reset my ... Spotify queue with all of its ... top tracks, shuffle it, and play it. ", "Make that 'uh... recommended songs', not 'top tracks'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects to 'recommended songs' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that uh recommended songs, not top tracks." }, "candidate_id": "wav_6b221fc4c9d120ec", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural revision flow. The speaker starts with a clear, fluent request, then hesitates with 'make that uh' before correcting 'top tracks' to 'recommended songs'. The correction is integrated smoothly into the same utterance, with a natural-sounding pause and filler. The prosody is consistent, and the correction feels like a spontaneous realization rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "58c7a66088d339a81d246eb0b037f500018a05c8e2a3ca292c3b94daaf8add47", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/cancellation/cancel_010/edb79120d4be2c7d/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "e527b68a78f155db83af47569a6a9f859cf4a28e25b82f85fb2b20c8ba7bab8a", "seconds": 13.68, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it.", "Hold on, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a complete request to reset, shuffle, and play the Spotify queue. After a brief pause and the interjection 'Hold on', they explicitly state 'cancel the request I just made'. This constitutes a clear cancellation of the entire task. The final control intent is a cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it. Hold on, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "50beb43770e3c58cf1b067a44cc7367e9e12004f2fd846ab3ce19a42512f0f6f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/ambiguous_halt/halt_012/bf8f82e63da6237f/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "143610c7c5ae9e9cae0e5c562a25376ff913fd4aef3b409707610929f21b8226", "seconds": 11.88, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then explicitly requests a pause with 'Give me a second.' This constitutes a valid ambiguous_halt as the final interjection is a clear request to wait, with no cancellation or resumption permission. The initial instruction is fully audible and complete.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "8749218_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "07341e37555d4db43365f7e409a8d48885c73b8375f21d2eebf1a964cb10c6e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/filler/original/404e53bfbc6f4a00/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "527e1240876c45027357d2d0357b30d65d07f0f29c8182255efc2c89dcdb28a1", "seconds": 8.72, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'let me think' but does not change the core task. The final instruction matches the original request in action, target, and sequence. The hesitation is internal and does not alter the goal, so the task is executed as intended.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_destination_city", "proposed_instruction": "I am going on a trip to Shanghai—I meant Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "e2913707d281479114bee04e841093c8a1f772bbf5c3972f54a1fe881826f093", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Shanghai' to 'Beijing' within the same utterance. The correction does not affect the core task of incorporating comments, leaving a specific final comment, and marking the task complete. The original instruction's goal is fully preserved after the correction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_01_destination_city/cc441d4b17c5d615/speech.wav", "wer": 0.06756756756756757, "wav_sha256": "2c68dc8b2eb84b405fcc4927fd0cb1cbbf6d09fcc0d0fc35d5e9228bd7f79b64", "seconds": 27.16, "spoken_utterances": [ "I am going on a trip to Shanghai—uh... I meant ... Beijing with some of my roommates. ... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Shanghai' but immediately self-corrects to 'Beijing'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The rest of the instruction is delivered clearly and matches the original request. The final control intent is to execute the task as corrected.", "heard_transcript": "I am going on a trip to Shanghai, uh, I meant, Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "candidate_id": "wav_13383d1dd3f66b20", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural, spontaneous self-correction ('Shanghai... I meant Beijing') with a clear pause and intonation shift, followed by a fluent and continuous delivery of the rest of the request. The speaker's prosody is consistent and conversational, with no signs of robotic pacing or unnatural pauses. The correction is integrated smoothly, and the overall flow is highly realistic." } ] }, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I am going on a trip to Beijing. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Oh, I forgot to mention, it's just with some of my roommates.", "long_pause_offsets": [ 364 ], "source_sha256": "6cb1bbf09fa1ae219196915364ba41dd06428bd5e79116a96f48e8f6556937ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds a clarification about the group ('just with some of my roommates') but does not change any of the core actions or conditions. The original instruction already specified the task context (Beijing trip, Todoist project, playlist comments), and the follow-up only adds a detail about the group, which is consistent with the original. The final goal remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/addition/addition_01/ac52d0dec5804a62/speech.wav", "wer": 0.0641025641025641, "wav_sha256": "174397d64065d81fc4ed5032fab01e27409aa79ba374ee16b08997d7f1d6dbc8", "seconds": 28.44, "spoken_utterances": [ "I am going on a trip to Beijing. ... We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. uh... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. ", "Oh, I forgot to mention, it's just with some of my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains an addition to the original instruction. The speaker first gives the core task (incorporate suggestions, comment 'Incorporated changes', mark complete) and then adds the detail 'it's just with some of my roommates' after a pause. This addition clarifies the group involved but does not change the core action. The final intent is to execute the original task with the added context, which matches the 'addition' type criterion.", "heard_transcript": "I am going on a trip to Beijing. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. Uh, I've made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes,\" and mark it complete. Oh, I forgot to mention, it's just with some of my roommates." }, "candidate_id": "wav_cd9c76d06ce30cfb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural pacing and a believable addition at the end. The initial statement is fluent, followed by a pause and a self-corrected phrase ('for it') that sounds like a spontaneous recovery. The final addition ('Oh, I forgot to mention...') is a pragmatic, believable afterthought with appropriate intonation and filler, fitting the 'addition' augmentation type well. The overall prosody is consistent and conversational." } ] }, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_destination_city", "proposed_instruction": "I am going on a trip to Tokyo with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Wait, the trip is to Beijing, not Tokyo.", "long_pause_offsets": [ 388 ], "source_sha256": "22fd9c8b4fc31a99221ae019e57bc7e4bef55f7313d95953759c4db18dd5c7a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination city from Tokyo to Beijing, which matches the original instruction. All other conditions, including the app (Todoist), task context, actions (incorporate suggestions, comment, mark complete), and the comment content, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_01_destination_city/0b63cce9d18ea2b5/speech.wav", "wer": 0.075, "wav_sha256": "c802bfa1a0e4fbff6fcc481e61323608c3a20e746a3d037a0af6732cc8ad5fd8", "seconds": 27.16, "spoken_utterances": [ "I am going on a trip to Tokyo with some of my roommates. uh... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. um... Please incorporate them, leave a final comment, ... \"Incorporated changes.\", and mark it complete. ", "Wait, the trip is to Beijing, not Tokyo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the trip is to Tokyo but immediately corrects themselves to Beijing. Since this is a revision of the destination, the final intent is to perform the requested actions (incorporate suggestions, comment, mark complete) for the trip to Beijing. The revision is clear and the final control intent is to execute the task with the corrected information.", "heard_transcript": "I am going on a trip to Tokyo with some of my roommates. Uh, we are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Um, please incorporate them, leave a final comment, \"Incorporated changes,\" and mark it complete. Wait, the trip is to Beijing, not Tokyo." }, "candidate_id": "wav_f2c5d24a0771fb10", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous narrative with a believable revision. The initial mention of 'Tokyo' is followed by a clear, self-corrected realization ('Wait, the trip is to Beijing, not Tokyo') with appropriate intonation and pacing. The filler 'uh' and 'um' are used naturally, and the correction flows seamlessly within the context of the task, making it sound like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "84056d5313a32f379a3b3d2baf20ab46fbe469ce693fc15e2bd343a4ba02be09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/cancellation/cancel_018/43798e97b7e0442f/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "199da543e0ae6a7f0b38b801dff63078d31d3505664485a3da1c66df1cc7b8f8", "seconds": 32.52, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "Wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a complete instruction to incorporate comments, leave a specific comment, and mark a task complete. After a pause and a self-correction ('Wait, I rather not do that after all'), they explicitly state 'Please cancel it.' This constitutes a clear cancellation of the entire prior task request. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Wait, I rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "a60a01bf370c388ecd8f6cbbe7b5a505731a21c9b5b1f53ba2200171902efd75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/ambiguous_halt/halt_015/277ea9dc5088a505/speech.wav", "wer": 0.07792207792207792, "wav_sha256": "c17d81e467a38177c903427ccaefa6d9e10f26850288d407b04a3aa23791ef4d", "seconds": 30.2, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Please wait a moment,' which is a clear request to pause or hold. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or revoked, and the speaker does not provide a new instruction or correction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8b360a6149793c45cf5b24a544ee975e944a8a6f00235f58eb43389c928383e1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/filler/original/d2111f43aba3c11e/speech.wav", "wer": 0.0821917808219178, "wav_sha256": "8e0d7d5778210bbac3184a189b1179f2120b488da8a3880c7c661185214ad44b", "seconds": 27.12, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'Hmm' and a brief pause, but continues with the full original instruction without any correction, addition, or cancellation. The final interjection is the completion of the original task request. The filler does not alter the goal, so the final intent matches the original instruction.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. Hmm. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_destination_city", "proposed_instruction": "I am going on a trip to Paris—sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "72878c2bc1848a083a044845adb7c64a8999506a41485538eaa5f096586fddcd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Paris' to 'Edinburgh' within the same utterance. The correction does not affect the core task, which is to incorporate comments, leave a 'Done!' comment, and mark the task complete. The original instruction's goal is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_01_destination_city/1474dfd57775f408/speech.wav", "wer": 0.0641025641025641, "wav_sha256": "20789c0e0c548e1262dd04ee3f6b1150df0912b38e566277e13c79218a114cae", "seconds": 27.84, "spoken_utterances": [ "I am going on a trip to Paris—uh... sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. let me think... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. hmmm... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Paris' but immediately self-corrects to 'Edinburgh'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The rest of the instruction is delivered without further changes, so the final goal is to incorporate the comments, leave a 'Done!' comment, and mark the task complete. The correction is clear and immediate, satisfying the false_start rule.", "heard_transcript": "I am going on a trip to Paris, uh, sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. Let me think. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "candidate_id": "wav_39cd632516104282", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and varied prosody, with believable hesitations ('Let me think'), self-corrections ('Paris... sorry, I meant Edinburgh'), and filler sounds ('huh') that are well-timed and integrated into the flow. The speaker's intonation reflects genuine thought processes, and the correction is smooth and contextually appropriate, making it sound like a spontaneous, real-time request." } ] }, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist.. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Oh, and one more thing: I've already made the playlist and shared it with others on the project.", "long_pause_offsets": [ 309 ], "source_sha256": "c753d42139a57ed30d53de98c0d1d7e3122b648f5c34cbe421ed0944d4cc0c15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds background context about the playlist already being made and shared, but does not change or contradict the core action: incorporate suggestions, comment \"Done!\", and mark complete. The original instruction already included these steps, and the addition only clarifies the state of the playlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/addition/addition_03/0f4c06d56ed45125/speech.wav", "wer": 0.05333333333333334, "wav_sha256": "a3fb4ad271e18bbe1966fd9dd14a86133d221c385ac61207f58169de2f4d958b", "seconds": 29.32, "spoken_utterances": [ "I am going on a trip to Edinburgh with um... some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist.. ... But they have made some suggestions in comments. Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete. ", "Oh, and one more thing: I've already made the playlist and shared it with others on the project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the original instruction followed by an addition that provides extra context ('I've already made the playlist and shared it with others on the project'). The addition does not change the core task of incorporating suggestions, leaving a 'Done!' comment, and marking the task complete. The final intent is to execute the original task with the added context, which matches the 'addition' type criterion.", "heard_transcript": "I am going on a trip to Edinburgh with um some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. But they have made some suggestions in comments. Please incorporate them, leave a final comment, done, and mark it complete. Oh, and one more thing. I've already made the playlist and shared it with others on the project." }, "candidate_id": "wav_c1d2be6b07cd85b2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural flow. The addition of 'Oh, and one more thing' after the main instruction is a highly plausible conversational device, and the speaker smoothly integrates the previously omitted detail without awkwardness. The pacing, intonation, and use of 'um' are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_task_app", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Let me correct the app: we're using Todoist, not Trello.", "long_pause_offsets": [ 374 ], "source_sha256": "4ad6e8cc9c086baa36e39b31742cd9e7e9dc17e4c0485f1843f6203d2e66d5f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the project management app from Trello to Todoist, while retaining all other requirements: incorporating comments, leaving a 'Done!' comment, and marking the task complete. The initial mention of Trello is replaced by the corrected app name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_03_task_app/1cfe8679bd9342f1/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "dce31afa496c8d2df0315e99b7b4c0b23866891151c6e4fb7d38b2684b618b4a", "seconds": 34.52, "spoken_utterances": [ "I am going on a trip to Edinburgh with hmmm... some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. um... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. ... Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "Let me correct the app: we're using Todoist, not Trello." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Trello but immediately corrects themselves to Todoist. Since this is a revision, the final instruction is to perform the same actions (incorporate comments, leave a comment, mark complete) but on the correct app, Todoist. The core task remains unchanged and is fully specified.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Let me correct the app. We're using Todoist, not Trello." }, "candidate_id": "wav_a37af5f03776689d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant setup, then smoothly corrects the project tool from 'Trello' to 'Todoist' after a brief pause and self-correction. The prosody is consistent, with believable fillers ('um') and a natural, conversational rhythm. The revision is integrated seamlessly into the narrative without awkwardness." } ] }, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "fdc3500768575e63dae3bb1510f619ef2f65dd8ea3e9d92358991cc937fd62b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/cancellation/cancel_001/f73781106dd4ad35/speech.wav", "wer": 0.05405405405405406, "wav_sha256": "388d320ea1b8707c4d435dd817e8ee970cecfd9f991c3f5c65b5679c3a56e29f", "seconds": 30.84, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Oh wait, cancel that,' which is a clear and final directive to revoke the entire preceding task. This matches the 'cancellation' type criterion, where the recording must clearly revoke/cancel the task. The final interjection is not a pause or a revision but a direct cancellation of the prior instruction.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\", and mark it complete. Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "555c716c13e4555d00bcbace4cfa7be92ff9df2b5e615e7564bf7378e06e849b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/ambiguous_halt/halt_005/303c3bb868888a19/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "c306007c3570fe0d668452731f61de554b7077d694655433db03907b3fc4c18f", "seconds": 30.64, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete.", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'One second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or change its goal. The speaker's intent is to temporarily halt the process, not to execute or cancel the original instruction. Therefore, the control intent is 'pause', which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\" and mark it complete. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5e17377cdfcd00c3946b7fa8da03789d3f82cbcf41ce59543f9b25cfe7d25a60", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/filler/original/10ea413568be4891/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "9aa10aa2e0cab47be2e2eb1886e3d4ecbafad6b7d5be0e2078db4c1723d24238", "seconds": 27.2, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'Uh' but immediately proceeds to state the full task request without any correction or change in intent. The final interjection is the complete instruction, which matches the original reference. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_music_app", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—or rather, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "bd5c30f1146d24c07d2e51fd93a5dbbc2e4e591c9b1da43677fa28970f0eca25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' within the same utterance. The final intended app is Spotify, which matches the original instruction. All other requirements—task context, action (incorporate comments), final comment content, and completion status—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_04_music_app/862e0c28e6defae2/speech.wav", "wer": 0.07894736842105263, "wav_sha256": "31278c402f92373a950c664ad1ca1c3d7d31c448257386f3945650abbbeb5bd4", "seconds": 27.08, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—um... or rather, ah... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"... Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlist'. According to the 'false_start' type criterion, the final corrected value ('Spotify playlist') should be executed, and the initial incorrect value ('Apple Music') should be discarded. The rest of the instruction is clear and matches the original. The final control intent is to execute the corrected task.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music um, or rather a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_919931a7be77320c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent delivery. The speaker makes a self-correction from 'Apple Music' to 'Spotify playlist' with a smooth, conversational 'or rather' transition. The prosody is consistent, with a natural hesitation and intonation. The correction is integrated seamlessly into the flow without any awkwardness, and the final command is delivered clearly and confidently." } ] }, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. I left out a detail: the task is about preparing a Spotify playlist.", "long_pause_offsets": [ 348 ], "source_sha256": "2389202e9fe8302b52f0f8591b4c9bb12640a5e46fb3b5f1c94a8478bd5b6eb0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The initial utterance specifies the full action sequence and the final comment. The followup adds the missing detail that the task is about preparing a Spotify playlist. This narrows the scope without contradicting or removing any prior requirement. All actions and conditions from the original instruction are preserved and augmented with the missing context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/addition/addition_03/84d45049616d8498/speech.wav", "wer": 0.0625, "wav_sha256": "f83a8300da14c7de411a36a4e5f6dbd2090aa9d8e77560e5582ef47170c1e25c", "seconds": 30.96, "spoken_utterances": [ "I am going on a trip to uh... Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"um... Thanks! Updated.\", and mark it complete. ", "I left out a detail: the task is about preparing a Spotify playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio starts with the original instruction and then adds a clarifying detail ('the task is about preparing a Spotify playlist') after a longer pause. This matches the 'addition' type, where the final goal combines the original task with the added information. The final intent is clear and executable, and the addition correctly narrows the scope of the task to Spotify playlist preparation, which was only implied in the original. No contradictions or errors remain.", "heard_transcript": "I am going on a trip to uh Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is, I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, um, \"Thanks! Updated.\", and mark it complete. I left out a detail, the task is about preparing a Spotify playlist." }, "candidate_id": "wav_1a4ae5e6d70b6307", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable self-correction. The initial mention of 'Bangkok' is slightly hesitant ('uh'), and the speaker later adds a missing detail ('I left out a detail...') in a way that sounds like a spontaneous realization. The pacing is conversational, with a natural pause and filler ('um') before the correction, and the correction itself is smoothly integrated into the ongoing thought. The prosody and rhythm are consistent with someone recalling and refining a task request in real time." } ] }, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_05_comment_text", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks!\", and mark it complete. Change the comment to \"Thanks! Updated.\", not just \"Thanks!\".", "long_pause_offsets": [ 374 ], "source_sha256": "8629197f8a4d7bf506d51b3eb2e196284823b9046e50cd884998fc5c79613dc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revises the exact text of the comment to include \"Updated.\", replacing the initially requested \"Thanks!\". The rest of the task—incorporating suggestions and marking complete—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_05_comment_text/d171278b03bb9fc0/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "08046421eae804c4314094ef71deacbb4fc51e197e95968ed31ab36e3853ea2c", "seconds": 34.0, "spoken_utterances": [ "I am going on a trip to uh, let me think... Bangkok with some of my friends. We are managing its planning on a Todoist project uh... for it. One of the tasks in it is about preparing a um... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks!\", and mark it complete. ", "Change the comment to \"Thanks! Updated.\", not just \"Thanks!\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'let me think' but correctly states 'Bangkok' and continues the original instruction. The final instruction is a revision of the comment content from 'Thanks!' to 'Thanks! Updated.', which is clearly stated and preserved. The rest of the task (incorporate suggestions, mark complete) remains unchanged. The revision is applied to the earlier instruction, and all other conditions are preserved, satisfying the revision type criterion.", "heard_transcript": "I am going on a trip to uh, let me think, Bangkok with some of my friends. We are managing its planning on a Todoist project uh for it. One of the tasks in it is about preparing a um, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_738370c0c45b85fc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a very natural flow with a realistic 'let me think' pause and a clear self-correction of the comment text. The speaker's intonation and rhythm are consistent with spontaneous speech. The correction is immediate and integrated, and the overall delivery is credible, though slightly less fluid than audio_03 due to a minor stutter ('it's planning') and a slightly longer pause before the correction." } ] }, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "0409946485d5f207a31c6702e890a6ed88a928ad814cd7ec0b2446880c8aaeeb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/cancellation/cancel_012/89bc6fa36114538b/speech.wav", "wer": 0.024691358024691357, "wav_sha256": "c84091e19f7ce2405bb7baa4e3bc4c28a9143480f4a61d70534c6078000f2cfb", "seconds": 32.2, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "Hold on, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation: 'Hold on, actually, I've changed my mind. Don't do that.' This is a direct revocation of the prior instruction to incorporate comments, leave a message, and mark the task complete. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Hold on, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "c28d76b0fdf660b4354d52a42da03c14a5064e327218da9111fd7fe0ecdacea0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/ambiguous_halt/halt_016/872a858e911174e1/speech.wav", "wer": 0.02666666666666667, "wav_sha256": "3480dfdd6db7d18858264a74982e6fe9aad5c6ac69bba34f4de33041347102a6", "seconds": 30.12, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a second please,' which is a clear request to pause or wait. This is a final interjection that does not cancel or modify the task, nor does it execute it. The original instruction remains intact and unaltered, and the speaker is not requesting any change to the task goal. Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a hmm Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Just a second please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "986aa4e_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8dd202ab44ce34669f01b23e6a5c7a729adef536ed667ff52158798457abd8e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/filler/original/de9db71a6c6f2972/speech.wav", "wer": 0.028169014084507043, "wav_sha256": "45a95dea1184ba6049cc2bf78d7522545e69eff31c5caab1d11bce663a58ddb6", "seconds": 26.88, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler word ('hmm') but immediately corrects and continues with the full original instruction. The final interjection is the complete task request, with no cancellation or pause. The filler does not alter the goal, so the execution intent matches the reference.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a hmm Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address—sorry, I meant work address.", "long_pause_offsets": [], "source_sha256": "1eb4a28573ff6c30b91adb8b49162d8df788eaf2322260486fc0d0ced0337fda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects to 'work address' using 'sorry, I meant'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/false_start/false_start_03_destination_address/c8aa5de0a4aa9b2c/speech.wav", "wer": 0.05, "wav_sha256": "ec8cdaa445495de5fa60b77fc972ae6a6210403a1763a73c2089ab87e874bfe3", "seconds": 8.84, "spoken_utterances": [ "Buy everything on my amazon wishlist, and have it delivered to my home address—um... sorry, I meant ... work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task with 'home address' but immediately self-corrects to 'work address' after a brief hesitation. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Buy everything on my Amazon wishlist, and have it delivered to my home address. Um, sorry, I meant work address." }, "candidate_id": "wav_7787890ee206c708", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with a clear request, then after a brief pause and a soft 'um', smoothly corrects 'home address' to 'work address' with a clear 'sorry, I meant' phrase. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels like a genuine moment of recall." } ] }, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my work address.", "long_pause_offsets": [ 38 ], "source_sha256": "bf984e7b97174bc7d9866e13d47d6ed33ab6fc91ec3d1af2286d7c03769bbc20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement without changing any other aspect of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/addition/addition_01/0fd24a071748033b/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "57dc7788eca64e91308235c1831677e3a9e4063a42c1ccf9d2708129fa1d8632", "seconds": 11.0, "spoken_utterances": [ "Buy everything on my um... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to buy everything on the Amazon wishlist is followed by an addition specifying delivery to the work address. The addition completes the original request by adding a missing constraint. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Buy everything on my um Amazon wishlist. Oh, I forgot to mention, have it delivered to my work address." }, "candidate_id": "wav_847e691e85aca228", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then delivers the core request fluently. The follow-up 'Oh, I forgot to mention...' is a highly plausible, spontaneous self-correction with a natural intonation shift and clear reference back to the initial request. The delivery is continuous and the repair feels genuinely integrated into the thought process." } ] }, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Buy everything on my ebay wishlist, and have it delivered to my work address. Sorry, I meant Amazon, not eBay.", "long_pause_offsets": [ 78 ], "source_sha256": "01d921799eecb3bc0c46e496c73f1b7a374832b2e9ee1000321fa90930311379", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the platform from eBay to Amazon, while retaining the action, scope (everything on the wishlist), and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/revision/revision_01_app/6bb5e10a8fa68305/speech.wav", "wer": 0.0, "wav_sha256": "8ea45b86ae2ab3c3ba073a92ddbd801ba8f5734c68ef3f8cf5b1b5b076816eb9", "seconds": 14.56, "spoken_utterances": [ "Buy everything on my um, let me think... ebay wishlist, and have it delivered to uh, let me think... my work address. ", "Sorry, I meant um... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects to 'Amazon not eBay' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with all constraints preserved.", "heard_transcript": "Buy everything on my um let me think eBay wishlist and have it delivered to uh let me think my work address. Sorry, I meant um Amazon not eBay." }, "candidate_id": "wav_3498733fa222a821", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker starts with a clear request, hesitates with 'let me think' before specifying 'ebay wishlist', then pauses again before correcting 'sorry, I meant Amazon not eBay'. The self-correction is immediate, the intonation reflects genuine recollection, and the filler 'let me think' is used at natural points. The prosody is consistent and the delivery sounds like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "d53c2a9ce1099026366de4490886414d0b88a350a041c2b6dee7f1c7a13a3fdd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/cancellation/cancel_004/485f7daa8d515008/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c0e1a6199015856556272632d383c7ca2f620f95d39b55c307a512051f7486e5", "seconds": 11.68, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address.", "Wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete request to buy items from the Amazon wishlist and deliver them to the work address. Immediately after, they say 'Wait, cancel that request,' which is a clear and final directive to revoke the entire task. This matches the 'cancellation' type criterion, where the recording must clearly revoke the task. The final interjection is not ambiguous and explicitly cancels the prior instruction.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um let's see my work address. Wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "183e323ef009a503a6509926bd92dbcaf2476d152e302334a4e57d459786536f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/ambiguous_halt/halt_037/2a74e1633fb98fb3/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "9c426ad7923bd8bb7a98815baac22657869c0a471c63691c7f9e73aba4ea157c", "seconds": 12.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Pause for a moment' repeated twice, which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is not completed, but the final control request is audible and unambiguous.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um, let's see, my work address. Pause for a moment. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "142e3a3ae9c8f7eb2dd34da96027e92b3bc326ef0777a7b0f22c8c83c087790c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/filler/original/d8470961b05d977c/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "25e3683329caccbc31e9f3db45f11814df12b8cdb83ae337b1093d0a25e54a00", "seconds": 7.76, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'let's see') and a pause but does not change the core request. The final intent remains to execute the original instruction. The filler does not alter the goal, so the execution criterion is satisfied.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um, let's see, my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my work address—no, I mean home address.", "long_pause_offsets": [], "source_sha256": "b405bd651cc563e428aab1b3f55e81699e594b5e15ed4ba6aad0548c51b8a943", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately corrects it to 'home address'. The final intended destination is 'home address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/false_start/false_start_03_destination_address/48b6bd6895971994/speech.wav", "wer": 0.0, "wav_sha256": "9a3cd4d5998cfc34cbb5fcd99e56d4bfe52b88411681e6f734634c99ed7762d5", "seconds": 11.08, "spoken_utterances": [ "Buy everything on my let's see... amazon wishlist, and have it delivered to my work address—uh... no, I mean home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately self-corrects to 'home address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Buy everything on my let's see, Amazon wishlist, and have it delivered to my work address. Uh, no, I mean home address." }, "candidate_id": "wav_798f618690404331", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'Let's see' as a filler, then states the initial request, pauses, and clearly corrects 'work address' to 'home address' with a believable 'uh, no, I mean' repair. The intonation and pacing of the correction sound spontaneous and integrated into the flow of thought, making it the most realistic false start." } ] }, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my home address.", "long_pause_offsets": [ 38 ], "source_sha256": "11f5dcd6323bb4cb0194c885ffd7ad961f00f7dddac6281e9c4da534605cea0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery instruction, which combines with the initial request to form the complete goal. The original instruction already included both actions, so the spoken variant completes the intended task." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/addition/addition_01/7109bedf10c61675/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "e8afeb00091d2c70f63d4d02c2c8130bbd9d8a8934e63fed973ea4a90c851d7b", "seconds": 11.96, "spoken_utterances": [ "Buy everything on my uh, let's see... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to uh... my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to buy everything on the Amazon wishlist is followed by an addition specifying delivery to the home address. The addition completes the original instruction by adding a missing delivery detail. The final intent matches the reference instruction, and the addition type criterion is satisfied.", "heard_transcript": "Buy everything on my uh let's see Amazon wishlist. Oh, I forgot to mention, have it delivered to uh my home address." }, "candidate_id": "wav_9b2e64c9adfc5659", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my work address. Change the destination to home address, not work address.", "long_pause_offsets": [ 80 ], "source_sha256": "3e2e131d565686e8565f4022b4cd1ca5a410ecee04e979da1e00c44d7e8f64c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the destination from work address to home address, aligning with the original instruction. The action and item scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/revision/revision_03_destination_address/003957c77512bc12/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "540397c2d9ae4904d751410f55eec1ef3b5ac4957a3dd7e452a10f62e18bf13e", "seconds": 17.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, hmmm... and have it delivered to uh... my work address. ", "Change the destination to um, let's see... home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests delivery to the 'work address' but immediately corrects this by saying 'Change the destination to... home address, not work address.' This constitutes a revision where the final intent is to deliver to the home address. The correction is clear and unambiguous, and the rest of the instruction (buying everything on the wishlist) remains unchanged. Therefore, the final audible intent matches the original reference instruction.", "heard_transcript": "Buy everything on my Amazon wishlist, and have it delivered to uh my work address. Change the destination to uh let's see, home address, not work address." }, "candidate_id": "wav_0e46887a4169c852", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, pauses thoughtfully with a believable 'hmm', then smoothly revises the delivery address from 'work address' to 'home address' with a self-correcting tone and intonation. The revision is integrated into the same utterance, and the prosody reflects genuine reconsideration rather than a mechanical edit." } ] }, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "44d86adb00f6cd0d99476de15faaefc8c5dd290175dc72c8ddcfa22279270d18", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/cancellation/cancel_004/b4add5a28431e5e5/speech.wav", "wer": 0.0, "wav_sha256": "834a18dc2887237257bd8eca666c4e0fdd36d2a2fdd6b369d02bcbaa3dadecae", "seconds": 13.28, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address.", "Oh, wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and self-correction ('let me think'), proceeds to state the full request. Immediately after, they say 'Oh wait, cancel that request,' which is a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. Therefore, the final control intent is cancellation, matching the type-specific criterion.", "heard_transcript": "Buy everything on my um, let me think, Amazon wishlist, let's see, and have it delivered to my home address. Oh wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "df5d92035e66750d2992274a7d5abd682f6e4c953244b0c286f49db1aa117408", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/ambiguous_halt/halt_010/320d1f64fd69fb1a/speech.wav", "wer": 0.0, "wav_sha256": "092e35f8a2452a9f043a582d085878eac98defd9e753d5eecf2f51f04b32f398", "seconds": 12.32, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then says 'Hold on a second' as the final interjection, which is a clear request to pause. The initial hesitation and 'let me think' are fillers within the ongoing task and do not constitute a final control command. The final audible intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Buy everything on my um, let me think. Amazon wishlist, let's see. And have it delivered to my home address. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9eebc7777bf038dc036d893e98b4d465df1c9c900fbdd088281614e08e7d9ab2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/filler/original/ec203ccf3ed6a2c5/speech.wav", "wer": 0.0, "wav_sha256": "d9e678a5072b8a1fa70e9cb45819841d5336d41f9dbb6c823a49406474a1c10c", "seconds": 9.04, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation markers ('um', 'let me think', 'let's see') but does not change the core request. The final intent remains consistent with the original instruction. The filler words do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Buy everything on my um, let me think, Amazon wishlist, let's see, and have it delivered to my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address—sorry, I meant work address.", "long_pause_offsets": [], "source_sha256": "1eb4a28573ff6c30b91adb8b49162d8df788eaf2322260486fc0d0ced0337fda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects to 'work address' using 'sorry, I meant'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/false_start/false_start_03_destination_address/f2b03449280ca906/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "febf19dc50382026282b711cc75520776fd660148a82116938229c8b5fcffe3f", "seconds": 8.16, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to ... my home address—uh... sorry, I meant work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately self-corrects to 'work address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to my home address. Uh, sorry, I meant work address." }, "candidate_id": "wav_f754e953e938ed24", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with a clear request, then after a brief pause and a soft 'uh', corrects 'home address' to 'work address' with a believable 'sorry, I meant' phrase. The prosody includes a slight pitch drop on 'sorry' and a quick, natural recovery, matching how a real person would correct a slip of the tongue mid-sentence. The pacing and intonation are consistent with genuine hesitation and recovery." } ] }, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my work address.", "long_pause_offsets": [ 38 ], "source_sha256": "bf984e7b97174bc7d9866e13d47d6ed33ab6fc91ec3d1af2286d7c03769bbc20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement without changing any other aspect of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/addition/addition_01/736713db9db2bd9b/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "b44eca61c8118c4aad03a2b07bdcada14b2dfbb0509c2058203151ebd8abe595", "seconds": 11.88, "spoken_utterances": [ "Buy everything on my uh, let me think... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Buy everything on my Amazon wishlist' is followed by an addition 'have it delivered to my work address' after a longer pause and self-correction. This matches the 'addition' type criterion, where the final goal combines the original request with the added constraint. The final audible intent fully satisfies the original instruction and adds the delivery address, so the task is complete and correct.", "heard_transcript": "Buy everything on my uh let me think Amazon wishlist. Oh, I forgot to mention, have it delivered to my work address." }, "candidate_id": "wav_29ae69965f317008", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a plausible 'let me think' filler, indicating genuine thought. The addition of the delivery address is smoothly integrated with a believable 'Oh, I forgot to mention' phrase, and the prosody remains consistent and conversational throughout. The pacing and intonation suggest a spontaneous, real-time formulation of the request." } ] }, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address. Wait, deliver it to my home address, not work address.", "long_pause_offsets": [ 80 ], "source_sha256": "7231433dfc4cca5500ae4a88aba7a1cc4a366299dddb15be62a478c716c1d343", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly confirms 'deliver it to my home address, not work address', which corrects the initial delivery location and preserves the original action and target list." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/revision/revision_03_destination_address/dcd9d259b293c1c0/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "a26d551e5a6f3544f7082048a47265d51b315234fe1e9495cd3403629f80af70", "seconds": 11.88, "spoken_utterances": [ "Buy everything on my um... amazon wishlist, uh... and have it delivered to um... my home address. ", "Wait, deliver it to my home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests delivery to the 'work address' but immediately corrects this to 'home address' after a pause and the word 'Wait'. This constitutes a revision where the final intent is to deliver to the home address. The correction is clear and unambiguous, and the rest of the instruction remains consistent. Therefore, the final intent matches the revised instruction and satisfies the revision type criterion.", "heard_transcript": "Buy everything on my um Amazon wishlist uh and have it delivered to um my home address. Wait, deliver it to my home address not work address." }, "candidate_id": "wav_e6b24463a8a5f6de", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous prosody with believable hesitation ('um') and a clear, spontaneous self-correction ('Wait, deliver it to my home address not work address'). The correction is integrated into the flow of the utterance, with a natural shift in intonation and emphasis, making it sound like a genuine moment of reconsideration during a real task request." } ] }, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "69a335cec2a465ac7546723087bb016e2fc4dc7e44feb50018991455f6297246", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/cancellation/cancel_031/7302a100c5e22522/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "a13991068f2eadd46491eeb665fbb7bb04a7c5db80f4aed30da0fc9e82cf9ebd", "seconds": 14.68, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address.", "Oh, wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task request, then immediately follows with 'Oh wait, there's no need to do that anymore. Cancel it.' This constitutes a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. The final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address. Oh wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "70565641b539cc3f7e587a762ec575d62c69b2eef3faa519cf9d7f71f3211cc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/ambiguous_halt/halt_027/aa0b45f1b6ebf651/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "70695cd2f2c7a8ccc1d02bd923218725550fc1e51ec8dc86b48118ccdedfd1b7", "seconds": 13.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly requests a pause with 'Hang on, I need a moment.' This constitutes a final interjection that is a direct request to hold the task, matching the 'ambiguous_halt' type. The original instruction is fully audible and not canceled or revised.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9871968_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4c3629a904f6d58950119594d568a3814bb8c0d7dcb61ccb658dce73fd1de747", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/filler/original/7366d33894d9e32d/speech.wav", "wer": 0.1875, "wav_sha256": "207b97a148b1b2814ecaea78337b4c4e48504740e9a65a0d9ff599cfbbd2ca4a", "seconds": 9.6, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or cancellations. The filler type allows for hesitation and backchannels without altering the task goal.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family—no, I meant friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "d09673a3b2a9a93d503a12a84bfb53d96eb82dc50a75e33ac2d79761f3743f67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance. The final intended group is 'friends', matching the original instruction. All other elements, including the project creation and task completion request, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/false_start/false_start_02_companion_group/bb9e9e351a45fb24/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "468d2a33f3ca93d828b116e55334701438a2512e2776fd3b00d164238c7220e8", "seconds": 13.36, "spoken_utterances": [ "I am going on a camping trip uh... with some of my family—um... no, I meant friends. let me think... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. This is a false start, and the final, corrected intent is to create a Todoist project and mark its tasks as complete. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "I am going on a camping trip uh with some of my family. Um, no, I meant friends. Let me think. I've created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_e47bfc0bac02d184", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'family', hesitates with 'um', self-corrects to 'friends' with a believable 'no, I meant friends' and a brief 'let me think' filler, all while maintaining a consistent, conversational pace and intonation. The correction feels like a genuine moment of recall, and the overall delivery is fluid and unforced." } ] }, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/addition/addition_01/94408e53790596c9/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "099322b636b0af79f52a13ceb224672dba0a827d6bf83b83cb4baf80be88b24f", "seconds": 20.44, "spoken_utterances": [ "I am going on a camping trip. uh, let me think... I have created a new Todoist project for managing its tasks. um... Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with hmmm... some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to mark all tasks as complete. The addition of 'it's with some of my friends' is a clarifying detail about the trip, not a change to the core action. The final intent remains to mark all tasks in the project as complete, which matches the original instruction. The addition does not alter the primary goal.", "heard_transcript": "I am going on a camping trip. Uh, let me think. I have created a new Todoist project for managing its tasks. Um, please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_cbe4f216f38d7798", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and realistic speech patterns. The speaker uses a believable 'uh' and 'um' as fillers, a clear 'let me think' hesitation, and a spontaneous-sounding 'oh, I forgot to mention' that introduces a new detail. The intonation and rhythm are consistent with someone formulating a request in real time, and the flow from initial statement to addition feels organic and unscripted." } ] }, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_activity_description", "proposed_instruction": "I am going on a hiking trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. I misspoke about the activity: it should be a camping trip, not a hiking trip.", "long_pause_offsets": [ 165 ], "source_sha256": "0aa30377907f693e09e7dc936dc5b57cac9d5f93e3d309b0c3acb34d91b7bb40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the activity from 'hiking trip' to 'camping trip' but retains all other requirements: the group ('some of my friends'), the app ('Todoist'), the action ('do the tasks'), and the completion status ('mark them all complete')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/revision/revision_01_activity_description/d902c602284359ad/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "cc947d10db34abdcdf5ba466fdf42dd4544bfd8fa08a646537c011da1f3a8bc8", "seconds": 21.12, "spoken_utterances": [ "I am going on a uh, let me think... hiking trip with some of my friends. uh... I have created a new Todoist project for managing its tasks. ... Please do them for me, and mark them all complete. ", "I misspoke about the activity: it should be a camping trip, not a hiking trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects it to 'camping trip' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark tasks complete for a camping trip, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints are changed.", "heard_transcript": "I am going on a uh let me think hiking trip with some of my friends. uh I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. I misspoke about the activity. It should be a camping trip, not a hiking trip." }, "candidate_id": "wav_59ca78b814880712", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear false start ('hiking trip'), self-corrects with a believable 'I misspoke about the activity' followed by a smooth, contextually appropriate correction to 'camping trip'. The prosody includes natural hesitation and a clear change-of-mind intonation. The initial pause and filler 'uh' are well-placed, and the overall rhythm mimics genuine spontaneous speech." } ] }, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "a13c155d13b53b6297c7b5d048a6c5ff21a03542a439ae76590b30429b733dfc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/cancellation/cancel_023/6d1357d1ce30be50/speech.wav", "wer": 0.020833333333333332, "wav_sha256": "3b926c36d3485b77288fb0e3b30c7245cdb3cddc98c8064d30eaaaf4e1425d81", "seconds": 18.44, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete.", "Oh, wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a complete instruction to mark all tasks in a Todoist project as complete. After a pause and a self-correction ('Oh wait, never mind about that'), they explicitly state 'Don't carry out the request,' which constitutes a clear cancellation of the original task. The final intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Oh wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "e3ca08d6575ea37c19309b3df34c2b081b3cf656e4a646a19b45ad649e0bcd32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/ambiguous_halt/halt_033/2c999e08141e7412/speech.wav", "wer": 0.024390243902439025, "wav_sha256": "ce64a1792fbd63df0a1ef45b22c2079e6fa1bc6dc55e0fbda76d8b54bc5077fc", "seconds": 17.0, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, there's something,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request to mark tasks complete is not executed due to the halt, which is correct for this type.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Hold on, there's something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bc51b34e902458e83d403a3dde4286de9f8a1a7f222aef1015fe07c6f118b1dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/filler/original/faad954a72ff2f35/speech.wav", "wer": 0.02702702702702703, "wav_sha256": "cbc452e4229854e96e6050f0612c2abc9eb0a7a1bbf3a03e9ed045921456be70", "seconds": 13.08, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and a pause but does not change the core request. The final instruction is identical to the original, with no correction or modification. The fillers are internal hesitations and do not constitute a final directive to pause or cancel. Therefore, the final intent matches the original instruction, and the control intent is to execute.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family—no, I mean friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "b2fb8d9d47aefefda25e8b7bfe18eacc84a9598fcf99b083eea9d25e8644de98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'family' to 'friends' within the same utterance, replacing the group reference. The rest of the instruction—creating a Todoist project and marking all tasks complete—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/false_start/false_start_02_companion_group/88a13d27b4e333f0/speech.wav", "wer": 0.0, "wav_sha256": "7d9a570736c1bc417f83a98b0071e1da17f0f0c407ca0bef2b5d0c26db66b2fa", "seconds": 17.72, "spoken_utterances": [ "I am going on a camping trip with some of my um... family—uh... no, I mean friends. I have created a new Todoist project for managing its tasks. ... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('family') is discarded. The rest of the instruction, including the creation of a Todoist project and the request to mark all tasks complete, remains unchanged and is fully audible. Therefore, the final intent matches the original instruction after the correction.", "heard_transcript": "I am going on a camping trip with some of my um family uh no, I mean friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_9fc265efae9f851e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a very natural false start with 'family' followed by a clear self-correction to 'friends'. The intonation and pacing are consistent with spontaneous speech, and the correction is immediate and fluent. The overall prosody is smooth and the task request is completed without further hesitation." } ] }, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/addition/addition_01/ff39845b8636e8d3/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "6a05ddb7da9c04958b591d0d6d74456d25d78b48e2424c7d717057415e4af8e6", "seconds": 19.96, "spoken_utterances": [ "I am going on a camping trip. uh, let's see... I have created a new Todoist project um, let me think... for managing its tasks. Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition 'it's with some of my friends' is a clarifying detail that is incorporated into the final goal. The addition does not contradict or remove any part of the original request; instead, it adds a relevant context. The final intent matches the original instruction with the added detail, satisfying the 'addition' type criterion.", "heard_transcript": "I am going on a camping trip. Uh, let's see. I've created a new Todoist project, um, let me think, for managing its tasks. Please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_6cc4e30208525c5b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow of spontaneous speech. It begins with a clear initial statement, includes a believable 'let me think' filler, and smoothly integrates a later addition ('Oh, I forgot to mention...') that feels like a genuine recollection. The prosody is varied and conversational, with pauses and fillers placed in a way that mimics real-time thought. The speaker's voice remains consistent, and the overall delivery is cohesive and unforced." } ] }, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them as done. Change 'mark them all complete' to 'mark them as done'.", "long_pause_offsets": [ 161 ], "source_sha256": "dd9c6daf4d991dfa0f7d3bb1e95352be4b29b569f28b54deecfdbd2d78bd5934", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the phrase 'mark them all complete' with 'mark them as done', which is a synonymous instruction for completing the tasks. The action, target, and intent remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/revision/revision_04_completion_action/cb51f73c358f9f34/speech.wav", "wer": 0.0, "wav_sha256": "1fe3014b75050f6fbcb2707953fc55d096100410e168320dd408a013d594722f", "seconds": 21.92, "spoken_utterances": [ "I am going on a camping trip um... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them as done. ", "Change 'mark them all complete' to 'mark them as done'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark them all complete' but then immediately corrects this to 'mark them as done' in a revision. The final audible instruction is 'mark them as done', which is the intended goal. The revision type criterion is satisfied as the correction is applied to the earlier instruction, and the final control intent is to execute the revised task.", "heard_transcript": "I am going on a camping trip um with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them as done. Change mark them all complete to mark them as done." }, "candidate_id": "wav_9e3b77cd883400af", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody. The speaker's intonation and rhythm are consistent with spontaneous speech, including a believable pause and filler ('um') at the start. The revision is a smooth, self-contained correction of the phrase 'mark them all complete' to 'mark them as done,' delivered with a natural change in intonation and pacing that suggests a genuine moment of reconsideration. The overall flow is coherent and unforced." } ] }, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "a56ddc41e4e94fdcd7399ec8f15a92df5eadf7a4e1bb4e72f020af07e197ea53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/cancellation/cancel_027/9cf9d0de95188eab/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "e4278ac5930781a7f775e6001e5f9dd07674528cf0a46f0f6b662b3ea76e3387", "seconds": 18.12, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete.", "Hold on, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a clear request to mark all tasks in a Todoist project as complete. Immediately after, they say 'Hold on, please disregard my last request,' which is a direct and unambiguous cancellation of the preceding instruction. The final control intent is clearly to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Hold on, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "82b87ea78b159df7e562331a18344f1c5ad08ae32601fad1debe9415d0ee186b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/ambiguous_halt/halt_036/76e6f782b6779de6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "498d85b098122d0cdc320018e28fa6f2f00c8bc6d542519f4855f543d566db75", "seconds": 17.0, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second, before you go ahead,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request to 'do them for me and mark them all complete' is not executed due to the final pause command.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2947b09dfd9e2cd63821f16f4fd4a4403ef374cf05560e5eda708904341420c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/filler/original/099f8e576763f865/speech.wav", "wer": 0.11428571428571428, "wav_sha256": "4f4691f4ba766aadbc1e932772f2be99c3a62e7261db1c2a0085c1e2d03b37da", "seconds": 12.8, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'let's see,' which is a filler, and then proceeds to state the original instruction without any correction or change to the goal. The final interjection is the completion of the original request, so the control intent is to execute the task as stated.", "heard_transcript": "I am going on a camping trip, let's see, with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and check them all off—sorry, I mean mark them all complete.", "long_pause_offsets": [], "source_sha256": "b2d926aaa2b6e59465dd213dfe7780453239f3d00ac22f64672a486d990ac272", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately corrects to 'mark them all complete'. These phrases are functionally equivalent in the context of task management, both meaning to finalize all tasks in the project. The app (Todoist), the action (completing tasks), and the scope (all tasks in the project) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/false_start/false_start_04_completion_action/01ed5d2cebb557a9/speech.wav", "wer": 0.020833333333333332, "wav_sha256": "1ce93b8c484feee4e9b9a006d18a30e20570a79f9f2f5556a3f3a8b5a06f39ab", "seconds": 19.96, "spoken_utterances": [ "I am going on a camping trip with some of my friends. I have created a new Todoist project um, let me think... for managing its tasks. Please do them for me, and um, let's see... check them all off—ah... sorry, I mean mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately self-corrects to 'mark them all complete'. According to the 'false_start' type criterion, the final corrected instruction should be applied. The final intent is to mark all tasks as complete, which matches the original instruction's goal. The correction is immediate and unambiguous, so the task should be executed as corrected.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project, um, let me think, for managing its tasks. Please do them for me, and um, let's see, check them all off. Uh, sorry, I mean mark them all complete." }, "candidate_id": "wav_feadcce02dedbc58", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant statement, uses a believable 'let me think' filler, and then smoothly self-corrects 'check them all off' to 'mark them all complete' with a natural, corrective intonation. The pauses and fillers are well-placed, and the overall prosody sounds like a person genuinely formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/addition/addition_01/fe7dab962476809f/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "64c17f658f3c0c5e280a5f81fe4372273e678cc69b87df1e7bcec6c455b9448e", "seconds": 21.92, "spoken_utterances": [ "I am going on a um, let me think... camping trip. ... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with ... some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates but then clearly states the original task. After a pause, they add the detail 'it's with some of my friends' as an additional constraint. The final intent combines the original instruction with this new information, which is a valid addition. The core action of marking all tasks in the project as complete remains unchanged and is fully audible.", "heard_transcript": "I am going on a um, let me think, camping trip. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_3c5bc10fcb69f7cf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a realistic 'um' and 'let me think' filler, indicating genuine hesitation. The addition of 'Oh, I forgot to mention' is a believable, pragmatic follow-up that integrates smoothly with the prior request. The prosody is varied and conversational, with appropriate pauses and intonation that suggest a real-time thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_activity_description", "proposed_instruction": "I am going on a hiking trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. I misspoke about the activity: it should be a camping trip, not a hiking trip.", "long_pause_offsets": [ 165 ], "source_sha256": "0aa30377907f693e09e7dc936dc5b57cac9d5f93e3d309b0c3acb34d91b7bb40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the activity from 'hiking trip' to 'camping trip' but retains all other requirements: the group ('some of my friends'), the app ('Todoist'), the action ('do the tasks'), and the completion status ('mark them all complete')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/revision/revision_01_activity_description/c84e28ad698c792e/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "03489cece9430e291fe1ff91808664c475be05b183a305667407b83c043ceac7", "seconds": 21.76, "spoken_utterances": [ "I am going on a um... hiking trip uh... with some of my friends. um... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. ", "I misspoke about the activity: it should be a camping trip, not a hiking trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects it to 'camping trip' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark tasks complete for a camping trip, which matches the original instruction's goal except for the activity type. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "I am going on a um hiking trip uh with some of my friends. um I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. I misspoke about the activity, it should be a camping trip, not a hiking trip." }, "candidate_id": "wav_295114179c0ad68a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, coherent initial request with natural hesitation ('um') and a believable, smoothly integrated revision. The correction ('I misspoke about the activity...') is a complete, self-contained sentence that logically follows the initial statement, with a natural change-of-mind intonation. The prosody is consistent, and the pacing feels like a genuine moment of recollection." } ] }, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "c236b39b5dff725219f1a8791fa498216c67f1293cac991360c6a14eead4c587", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/cancellation/cancel_006/6854c92c7c880190/speech.wav", "wer": 0.0, "wav_sha256": "d7baf9e5c72d4efea0ffee035a210079ea4956ad2f59f2e592fc8bcdee411413", "seconds": 23.0, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete.", "Oh, wait, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to mark all tasks as complete, but immediately corrects themselves with 'Oh wait, please cancel the entire task.' This constitutes a clear cancellation of the previously stated task. The final interjection is a direct command to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete. Oh wait, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "a787f124a621e87bb4fb402160c561fa362320bd1a2f95465b5fd5070ba73c8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/ambiguous_halt/halt_040/138227d19ec4fc91/speech.wav", "wer": 0.0, "wav_sha256": "cfd7ec1e9dd0e956e389d3b0c208356dff2566273dda0caed978ddc36fbd4a25", "seconds": 21.8, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Hang on, just a minute,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or request a rollback, but rather asks for a temporary halt. The initial request to 'do them for me and mark them all complete' is not executed due to this pause command. The audio ends with this pause request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "988af8e_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c533cb96a82bc2e3e650b2d4a7fca6712dd8d69406b9f6ce56005c3d4268244e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/filler/original/301cb740fa1bc864/speech.wav", "wer": 0.0, "wav_sha256": "6cd1ae007f472518fa95ff6a21fd2004e27a99b5440ca10f4f3bb4c32c9a15d9", "seconds": 18.08, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let me think') and a pause ('Let's see') but does not change the core request. The final instruction 'Please do them for me and mark them all complete' clearly retains the original intent to mark all tasks in the project as complete. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_recipient_reference", "proposed_instruction": "Email the driving license found in my file system to my colleague—no, I mean partner.", "long_pause_offsets": [], "source_sha256": "bf3499dffd5a3995b57b3ebcb375efe3430bd81b0c7728a0184a792c32f0ff1d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'colleague' with 'partner'; the app, source, and content remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/false_start/false_start_03_recipient_reference/dedd3fcfda4825e2/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "59b13f15af749caaa5c43db1c8d98ce928d3f1be47a8b46177854b02e5f3c7da", "seconds": 9.8, "spoken_utterances": [ "Email the uh... driving license found in my file system to my ah... colleague—um... no, I mean partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'colleague' but immediately self-corrects to 'partner' with 'no, I mean partner.' This is a clear false start where the final intended target is 'partner'. The rest of the instruction matches the original, so the final intent is fully aligned with the reference.", "heard_transcript": "Email the uh driving license found in my file system to my uh colleague um no, I mean partner." }, "candidate_id": "wav_827fd6c90107bf41", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false-start pattern. The speaker begins with 'email the uh driving license found in my file system to my uh colleague um no I mean partner.' The initial hesitation ('uh') and filler ('um') are placed naturally, and the self-correction ('no I mean partner') is a smooth, spontaneous-sounding revision of the recipient. The prosody includes a slight pitch drop on 'colleague' and a rise on 'partner,' indicating a genuine change of mind. The pacing and rhythm feel like a real person formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Email the driving license to my partner. Oh, I forgot to mention, it's the one found in my file system.", "long_pause_offsets": [ 41 ], "source_sha256": "f26224cb6fb635691f9733d395076adda87e4271b6e41918af1b7955f30813da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint 'found in my file system' to the initial email request, preserving the recipient and item. The final goal combines both utterances." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/addition/addition_01/76b12dbd4b6c1c42/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "705542864cc1da53e1afecb475840034ead5db05529be4a5fc1825028a46f592", "seconds": 12.84, "spoken_utterances": [ "Email the uh... driving license ah... to my partner. ", "Oh, I forgot to mention, it's the one um... found in my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially begins the task and then adds a clarifying detail ('found in my file system') after a pause. According to the 'addition' type criterion, the final goal is the combination of the original instruction and the added information. The final audible intent matches the original instruction with the added constraint, so the task is correctly executed.", "heard_transcript": "Email the uh driving license uh to my partner. Oh, I forgot to mention, it's the one um found in my file system." }, "candidate_id": "wav_aa3615dafb55acfe", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-repair ('uh' after 'the'), which is followed by a plausible, additive clarification ('Oh, I forgot to mention...') that smoothly integrates the missing detail. The prosody is consistent, and the pause before the addition feels like genuine recollection rather than a scripted insertion. The flow from initial request to addition is very natural." } ] }, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_attachment_source", "proposed_instruction": "Email the driving license found in my email inbox to my partner. Make that the driving license from my file system, not my email inbox.", "long_pause_offsets": [ 65 ], "source_sha256": "10c08b31933b40c29a40785930eaf34c41a7b540784d317cb310c387035d600d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the source of the driving license from 'email inbox' to 'file system', while retaining the recipient and the item to be sent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/revision/revision_02_attachment_source/5ee65ec888a5aa14/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "26421618520fcb15a34ac6489f68a5da95b73626aced8d95e43b7135865224e3", "seconds": 11.56, "spoken_utterances": [ "Email the driving license found in ah... my email inbox uh... to my partner. ", "Make that the driving license from my file system, not my email inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'email inbox' but immediately corrects it to 'file system' with a clear revision. The final intent matches the original instruction exactly, so the revision is correctly applied and the task is valid.", "heard_transcript": "Email the driving license found in my email inbox to my partner. Make that the driving license from my file system, not my email inbox." }, "candidate_id": "wav_cb3da53df96bfa43", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous prosody with a believable self-correction. The speaker starts with 'email the driving license found in my email inbox' and then smoothly revises to 'Make that the driving license from my file system, not my email inbox.' The correction is integrated into the same utterance with a natural shift in intonation and pacing, suggesting spontaneous thought. The pauses are brief and contextually appropriate, and the overall delivery sounds like a single, coherent thought process." } ] }, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "d959eb36102dd4f7b3c9c1a89622b5773749b65b1347190c78db9f4e067ceed3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/cancellation/cancel_011/05ada78dc0c0853b/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "fd1ad12949261bf5378f23d150692e06bf761e7a95dcba3c07744cf97ae64319", "seconds": 13.88, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner.", "Hang on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the original instruction but then clearly revokes it with 'Hang on, I've changed my mind. Cancel that.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is unambiguous and complete, leaving no room for execution of the initial request.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner. Hang on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "4c9a1a3e65dd73eb1ea27618de62b8aec30a12cf7a57e541840fcab716ed393e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/ambiguous_halt/halt_037/fac1b4b49c68acb6/speech.wav", "wer": 0.1, "wav_sha256": "07f9de549c46d91a1c5a8f6a607788de5c42a2c8310fd2ce91bcca026f6bbef1", "seconds": 12.36, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This aligns with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request to email the document is not canceled, but the final audible intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "32846bf8d67158d5c3df6d128d0b36ae3202d4b501c68b803a734cd70e043318", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/filler/original/f0032f033f17d681/speech.wav", "wer": 0.125, "wav_sha256": "3ce2a58a0cd46651cc8d6c64cbdb0daf206ff5c875dfc4f11398a34bd0aeddbc", "seconds": 9.08, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final intent remains to email the driving license to the partner, matching the original instruction. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_attachment_description", "proposed_instruction": "Email the resume—sorry, I meant headshot found in my file system to my manager.", "long_pause_offsets": [], "source_sha256": "d3b331d3056d8a9a2ff2986913a3607a3216695858e045d68aaa318797d6e2c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'resume' to 'headshot' within the same utterance, preserving the action, source, and recipient." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/false_start/false_start_01_attachment_description/6ab153fdbcefb98b/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "5e0711fb367d1187d43674b08aac25a9f1798c8e0566e81e16a98ffce1a512f4", "seconds": 9.04, "spoken_utterances": [ "Email the resume—um... sorry, I meant headshot found in my file system to let's see... my manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'resume' but immediately self-corrects to 'headshot'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the 'false_start' type criterion.", "heard_transcript": "Email the resume. Sorry, I meant headshot found in my file system to let's see, my manager." }, "candidate_id": "wav_e8cc112d3b213ad3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false-start pattern. The speaker begins with 'resume', self-corrects to 'headshot' with a clear, audible hesitation ('um'), and then continues with a believable, slightly hesitant formulation of the recipient ('to let's see, my manager'). The pacing, filler usage, and repair are all highly realistic and mimic genuine human speech." } ] }, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Email the headshot to my manager. Oh, I forgot to mention, it's the headshot found in my file system.", "long_pause_offsets": [ 34 ], "source_sha256": "2733ce2549e96ae1516ad9dd565b5d098514d26160081560eedce7350dfab26b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source restriction 'found in my file system' to the initial email request, preserving the recipient and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/addition/addition_01/078c365e6a93e4d1/speech.wav", "wer": 0.0, "wav_sha256": "f7a8115a24af3aece19b409271400eecfab148633eed19e87818a71b010e5628", "seconds": 13.36, "spoken_utterances": [ "Email the um, let me think... headshot to my manager. ", "Oh, I forgot to mention, it's the headshot found in um... my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email the headshot to my manager' and then adds the clarifying detail 'it's the headshot found in my file system' after a longer pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is fully aligned with the original instruction, and the addition correctly specifies the source of the headshot.", "heard_transcript": "email the um, let me think. headshot to my manager. oh, I forgot to mention it's the headshot found in um, my file system." }, "candidate_id": "wav_b591f84560cb92d5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_recipient_reference", "proposed_instruction": "Email the headshot found in my file system to my my assistant. Wait, the email should go to 'manager', not 'my assistant'.", "long_pause_offsets": [ 63 ], "source_sha256": "5c47cbd5e95fa00f89cb9247271d5fd8cfd481c00be38eb3f27cf4c1b1f1a2ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recipient from 'my assistant' to 'manager', aligning with the original instruction. All other elements, including the subject (headshot), source (file system), and action (email), remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/revision/revision_03_recipient_reference/97f7ff3c482a264b/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "f069d93994cdfa86c5449a1d035b0dcb0b5f6af5158f0fd0450b2defb2563f11", "seconds": 17.2, "spoken_utterances": [ "Email the um, let me think... headshot found in my file system to hmmm... my my assistant. ", "Wait, the email should go to '... manager', not 'my assistant'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to my assistant' but immediately corrects themselves with 'Wait, the email should go to manager, not my assistant.' This constitutes a revision where the final intent is to send the email to the manager. The correction is clear and unambiguous, and the rest of the instruction remains consistent with the original. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Email the um, let me think, headshot found in my file system to my my assistant. Wait, the email should go to manager, not my assistant." }, "candidate_id": "wav_06ab41022843467c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear task, hesitates with 'let me think' and 'um', then self-corrects from 'my assistant' to 'manager' with a believable 'wait' and intonation shift. The prosody is continuous, the pauses are appropriately placed, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "c2361e37719efa765fbe18c14783779e295807740e333c817ee086e3b40a40b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/cancellation/cancel_002/920ad87051c51c2b/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "4de8be8887d35cbc63bdc980beb253f43427f3fb627b7ddfdb2c712eeb861aa8", "seconds": 14.48, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager.", "Oh, wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original email task but then clearly revokes it by saying 'cancel the whole task' as the final interjection. This meets the criterion for a cancellation, as the final audible intent is to cancel the task, not to proceed with it.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager. Oh wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "6385f84085e1a2151ed37a7978dbb3696a8de36dc52235d3793798ac239df53e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/ambiguous_halt/halt_004/41daca11dece2256/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "8ce7f641a5717c005d9b98baee7403ac0c7cdd067dd8c66fd04f089db0a20724", "seconds": 13.6, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager.", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second,' which is a clear request to pause or wait. The initial part of the audio is a hesitant but complete instruction to email a headshot, but the final spoken control intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager. Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5812330f99e9596a8ef44f1e9807deab7852352296d68571ad558bf63c03d5d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/filler/original/a9cf26c9520973cb/speech.wav", "wer": 0.0625, "wav_sha256": "ce5f55c1405a2531b481f328df28bb2e8dfccd37c6bfffafa8e48b2f97c2f8ae", "seconds": 10.08, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core task. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_attachment_source", "proposed_instruction": "Email the birth certificate found in my email inbox—I should say my file system to my husband.", "long_pause_offsets": [], "source_sha256": "959b7d31c56f65961df8358ad36b25190900cd6f6928b55d53a4da873ede8f63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email inbox' to 'file system' within the same utterance, preserving the recipient and the document identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/false_start/false_start_02_attachment_source/a606bd12cf30640f/speech.wav", "wer": 0.1, "wav_sha256": "cb382ae04dd3c7f4f579fc8c6264c48d91b55a0eb88fc78a8b8f9ef31433a788", "seconds": 10.36, "spoken_utterances": [ "Email the birth certificate found in um... my email inbox—hmmm... I should say my file system to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email inbox' but immediately self-corrects to 'file system'. This is a false start, and the final, corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "Email the birth certificate found in um my email inbox. Hmm. I should say my file system to my husband." }, "candidate_id": "wav_de5d0968f286f585", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('in um... my email inbox'), followed by a natural self-correction ('I should say my file system') and a smooth continuation. The intonation and pacing reflect genuine hesitation and correction, with a believable pause and filler ('hmm') before the correction. The flow is continuous and the repair is integrated into the utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Email the birth certificate found to my husband. Oh, I forgot to mention, the birth certificate is in my file system.", "long_pause_offsets": [ 49 ], "source_sha256": "80c8faceaa07904248bbdb698b6478d676ca535df2dae4cf374dd270fa2f86be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint 'in my file system' to the initial email request. The recipient and object remain unchanged, and the added detail clarifies the source without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/addition/addition_01/fedd3ed60a77778b/speech.wav", "wer": 0.12, "wav_sha256": "91d75615fa40ae9dfa0b4f50b9686a4f26d679d3d08036bfefa12a122fcba580", "seconds": 14.76, "spoken_utterances": [ "Email the uh... birth certificate found ... to my husband. ", "Oh, I forgot to mention, the birth certificate is uh, let's see... in my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Email the birth certificate found to my husband' and then adds the location 'in my file system' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and complete, matching the original instruction with the added constraint.", "heard_transcript": "Email the uh birth certificate found to my husband. Oh, I forgot to mention, the birth certificate is uh, let's see, in my file system." }, "candidate_id": "wav_b3f53cd8d8017975", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, then pauses and adds a relevant detail ('the birth certificate is... let's see... in my file system') with a believable 'let's see' filler and a slight hesitation before completing the thought. The prosody is consistent, and the addition feels like a spontaneous memory rather than a rehearsed or awkward insertion." } ] }, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_attachment_source", "proposed_instruction": "Email the birth certificate found in my email inbox to my husband. Let me correct the source: find the file in 'my file system', not 'my email inbox'.", "long_pause_offsets": [ 67 ], "source_sha256": "99b13b1ad3eb47365b167d4321747ba8c27c6d66cbcdf14e3920b87bc95b0e4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'email inbox' to 'my file system', while retaining the recipient and the document identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/revision/revision_02_attachment_source/a1c1ee6f7ad22e13/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "15be6f487487089c46f356f2a7385b54eadba35a9d11190cecbb6969d5e89d8c", "seconds": 16.88, "spoken_utterances": [ "Email the uh... birth certificate found in my email inbox ... to my husband. ", "Let me correct the source: find the file in 'um... my file system', not 'my email inbox'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'email inbox' but immediately corrects it to 'file system' after a pause. The revision is clear and final, and the rest of the instruction (action, target, and object) remains consistent. The final intent matches the original instruction, so the revision criterion is satisfied.", "heard_transcript": "Email the uh birth certificate found in my email inbox to my husband. Let me correct the source. Find the file in um my file system, not my email inbox." }, "candidate_id": "wav_e2658ea703f8c412", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural false start ('uh') and a self-corrected revision ('find the file in um my file system not my email inbox'). The correction is smoothly integrated into the same utterance, with a believable hesitation and a clear, audible change of mind. The prosody and pacing are consistent with spontaneous speech, and the repair is linguistically and pragmatically plausible." } ] }, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "b664a77b6f2ef53817e2f7feb74a3863261723f091210b574afccea13ae2eef2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/cancellation/cancel_026/400ca2ba6f6d36d8/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "86868e0e0f42b02869fb058dc0f76d944cb40646b89dd102e176d934d1880a7a", "seconds": 13.2, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband.", "Hang on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by attempting to issue the email task but then explicitly states 'Hang on, forget what I just asked, cancel it.' This constitutes a clear and final cancellation of the original request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Email the birth certificate found, let me think, in my file system, uh, let's see, to my husband. Hang on, forget what I just asked, cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "37e03bb8a121edac9c454576ae16b317153b2acf502aa0ce1e2e313d5b310096", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/ambiguous_halt/halt_024/fa21b5cdc3f61bd6/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "f8b2d0c1a965040e77739c00ccae09fd53ce5500f7cbd1a9c48508a107fddbc5", "seconds": 12.36, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but then repeatedly uses hesitation fillers ('let me think', 'uh, let's see', 'give me a moment to think') and does not complete the task. The final interjection is 'Give me a moment to think,' which is a clear request to pause. There is no cancellation or correction, and the speaker does not resume the task. This matches the 'ambiguous_halt' type, where the final intent is to pause.", "heard_transcript": "Email the birth certificate found, let me think. In my file system, uh, let's see, to my husband. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "98d2608_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2f8faad742c79c14584f8b43edf3b8b63a08216d7d664801d83c08f5bd36c235", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/filler/original/7f49e1bad71506d9/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "2345e377afa7ade0d3b93287c5db31e1bef9c0cf5846143da7a6a1b676305ec8", "seconds": 8.4, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation fillers ('let me think', 'uh', 'let's see') but does not change the core request. The final intent remains to email the birth certificate to the husband, matching the original instruction. The filler type allows ignoring such hesitations, so the task is executed as intended.", "heard_transcript": "Email the birth certificate found, let me think, in my file system, uh, let's see, to my husband." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all blenders—no, I meant food processors from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "1354b5fa6bd6a4d3cfac981a728a31f4be62ea54f4f683d0428794cc5e9aa40d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'blenders' to 'food processors' within the same utterance, preserving the action, source, destination, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/false_start/false_start_01_product_category/a2342ad636f0a7a8/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "e4eeb554589a50b4b1cde5008bf7f5473c87985887b1c7ad214334cc70ee5491", "seconds": 11.4, "spoken_utterances": [ "Move all uh... blenders—um... no, I meant hmmm... food processors from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blenders' but immediately self-corrects to 'food processors'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as stated.", "heard_transcript": "Move all uh blenders um. No, I meant um food processors from my Amazon cart to wish list." }, "candidate_id": "wav_fd4ecaf733aabe64", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible false start. The speaker initially says 'blenders', then clearly hesitates with 'um' and 'no', followed by a self-correction to 'food processors'. The correction is smooth, with a natural pause and intonation shift, and the final sentence flows without awkwardness. The prosody and pacing are consistent with someone thinking aloud and correcting a mistake in real time." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Move food processors from my amazon cart to wish list. Oh, I forgot to say, move all food processors.", "long_pause_offsets": [ 55 ], "source_sha256": "d33402703f09acaf74256ba6aece5ce7feaf5397af7cd898490340ae0539577b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly adds the 'all' modifier to the initial request, which was missing in the first utterance. The final goal now includes the scope 'all' while preserving the app (Amazon), item type (food processors), and action (move from cart to wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/addition/addition_01/1e339a22d7de75c3/speech.wav", "wer": 0.12, "wav_sha256": "c7ec7bc3ee24e70dbcdb924159163417b77a54d61993e0ae2bfe38f33883664e", "seconds": 16.72, "spoken_utterances": [ "Move ... food processors from my amazon cart let's see... to wish list. ", "Oh, I forgot to say, move um, let me think... all food processors." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Move food processors from my Amazon cart to wish list' and then adds 'all food processors' after a pause and self-correction. According to the 'addition' type criterion, the final goal is the combination of the initial instruction and the added information. The final audible intent matches the original instruction exactly, with the addition of 'all' to clarify the scope. Therefore, the task is correctly executed.", "heard_transcript": "Move food processors from my Amazon cart, let's see, to wish list. Oh, I forgot to say, move, um, let me think, all food processors." }, "candidate_id": "wav_a899f5a82c59c02f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' and 'oh I forgot to say,' followed by a self-correction ('move um... let me think') before completing the full, coherent instruction. The pauses and fillers are well-placed and sound spontaneous, and the correction flows naturally within the same utterance. The prosody is consistent and the delivery is fluent, making it the most realistic." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all food processors from my amazon inbox to wish list. Make that from 'cart', not 'inbox'.", "long_pause_offsets": [ 60 ], "source_sha256": "0c44f3601f8f3b2458e703d45d3ef078d55fa50532884c45d5f962b579a27bf4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'inbox' to 'cart', preserving the app (Amazon), item type (food processors), and destination (wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/revision/revision_03_source_collection/fb955ca9000bc648/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "7704250f7e8312f0f5f8081d2ae96cee1bab2d89b28ddcf65e09a78836d99da4", "seconds": 15.08, "spoken_utterances": [ "Move all food processors ... from my amazon inbox uh, let's see... to wish list. ", "Make that from 'uh... cart', not 'inbox'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'cart' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "Move all food processors from my Amazon inbox, uh, let's see, to wish list. Make that from uh, cart, not inbox." }, "candidate_id": "wav_c4e3d9b65d9c0be9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and plausible revision flow. The speaker begins with a clear request, then hesitates with 'uh let's see'—a believable filler indicating momentary recall—before correcting 'inbox' to 'cart'. The correction is smoothly integrated into the same utterance with a natural back-reference ('Make that from...'), and the prosody maintains a consistent, conversational rhythm without robotic pauses or unnatural stress." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "0abaf1978c9bd8e3c86c1a111056d4797f44e468e75fde9eb96e4f8961c24cbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/cancellation/cancel_029/231d4b1ae47b1802/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "63d2a5386215f3e1ca9d9f7896d1900bf8dad234aaf5da3e8488339362a0e111", "seconds": 9.76, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task, then clearly states 'Hang on, I'm withdrawing that request,' which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "f2b7301e749e625f4ae686bdd85a42538c4c29ff5c0bc7ab65027768511ff086", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/ambiguous_halt/halt_031/e4aabbe4b0e3edc7/speech.wav", "wer": 0.1875, "wav_sha256": "dff00bb604dce449bf246ff9f513b2389fbe828e2bf0dfe549c1423b4eeb592d", "seconds": 9.44, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or hold. There is no cancellation of the task, nor is there a new instruction. The speaker is halting the process mid-sentence, indicating a temporary stop. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c3bf83013b3e12a431839ecb5f759eff63209ce363e5e96a0a6ddf447c079de3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/filler/original/1067a65260b4b848/speech.wav", "wer": 0.25, "wav_sha256": "69b227ab94606f5ae9175dda4d418c6f22e26c54bef3cafa5ee20e6ac4537b19", "seconds": 5.4, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'uh' but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all shoes—sorry, I mean watches from my amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "ea6c86358b7f0dfebece744cbe440f440065ad9bed3b6531ae75473660a528e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shoes' to 'watches' within the same utterance, preserving the app (Amazon), source (wish list), destination (cart), and action (move all)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/false_start/false_start_01_product_category/dbac95b97f453eb1/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "8061ee1afffad11df493a925ed7ae8b495dd25e513ebc315c841a4f8788e7655", "seconds": 8.48, "spoken_utterances": [ "Move all hmmm... shoes—um... sorry, I mean watches from uh... my amazon wish list to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shoes' but immediately self-corrects to 'watches'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, with only a minor grammatical repair ('I mean watches') replacing the mistaken 'shoes'. All other constraints (from wishlist to cart, on Amazon) are preserved.", "heard_transcript": "Move all um, shoes um, sorry, I mean watches from uh my Amazon wish list to cart." }, "candidate_id": "wav_f4f863e2aaf78c1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a believable 'uh' and 'um' hesitation before correcting 'shoes' to 'watches'. The correction is smooth, with a natural pause and intonation shift. The final request is fluent and complete, with no awkward phrasing or unnatural pauses. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Move watches from my amazon wish list to cart. Oh, I forgot to mention, move all watches.", "long_pause_offsets": [ 47 ], "source_sha256": "29fc221f0ce3c0c711cc8c7157ad6ac184c0139a714f82d2d02f6c07813ac02c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the 'all' modifier to the initial request, which was missing in the first utterance. The app (Amazon), source (wish list), destination (cart), and item type (watches) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/addition/addition_01/8a8acaf958d47776/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "61542bc82dc5a553fba13eb3fefe767368de3a0ba098cb8795bca26ef43473db", "seconds": 10.52, "spoken_utterances": [ "Move watches from um... my amazon wish list to cart. ", "Oh, I forgot to mention, move ... all watches." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Move watches from my Amazon wish list to cart.' The speaker then added 'move all watches' after a pause, which clarifies the scope of the watches to be moved. The final intent combines the original action with the added constraint, resulting in 'Move all watches from my Amazon wish list to cart.' This matches the original instruction, which already specified 'all watches'. The addition does not introduce any conflicting or incorrect information.", "heard_transcript": "Move watches from um my Amazon wish list to cart. Oh, I forgot to mention, move all watches." }, "candidate_id": "wav_5adbb084c7b83aef", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all watches from my amazon gift registry to cart. I misspoke about the source: move all watches from my amazon wish list to cart, not gift registry.", "long_pause_offsets": [ 55 ], "source_sha256": "1cba491add5de1c26bd921773990dd2e5447b2b6f12e0f91114467c699a05d78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'gift registry' to 'wish list', aligning with the original instruction. The action, item, and destination remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/revision/revision_03_source_collection/5e9ffb289d35f918/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "e97ae6fb68552c51ecfbc5efa1348a5b411706c1140ebb82817587053bab097b", "seconds": 14.32, "spoken_utterances": [ "Move all watches from ... my amazon gift registry uh... to cart. ", "I misspoke about the source: move all watches from my amazon ... wish list to cart, not gift registry." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'gift registry' but immediately corrects themselves with 'I misspoke about the source' and clearly states the intended action as 'move all watches from my Amazon wish list to cart'. This is a revision where the final, corrected instruction matches the original reference. The correction is clear and unambiguous, so the task should be executed as revised.", "heard_transcript": "Move all watches from my Amazon gift registry uh to cart. I misspoke about the source, move all watches from my Amazon wish list to cart, not gift registry." }, "candidate_id": "wav_04af30934b368148", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects with a natural-sounding 'I misspoke about the source' followed by a smooth, grammatically correct revision. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The pacing and intonation are very natural, with a believable pause before the correction." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "de00bf7949c53c1a1d0eb55e67c75fdf9405b4eae1bf69a3ca6c470710c06efb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/cancellation/cancel_008/dd847dc2ffa21fa0/speech.wav", "wer": 0.08, "wav_sha256": "bcaaeecc09c5a665c28f247db12be994789fd3b649264a7811806781f42b3908", "seconds": 13.08, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart.", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hang on, I want to cancel that task.' This constitutes a final cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart. Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "d02e3ad52d9ed45446825fd3bda909f9303e59acdd05ae896cec6056bdf655a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/ambiguous_halt/halt_023/e92d80531911bdfc/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "bbc884424a95fdcc12482437c5dc6ba20fad6b7ee488c040011e388f5f102702", "seconds": 12.6, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to move watches from the Amazon wish list to the cart is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1821d3b69d7bbf622d99ade23b2dcf80ec90c492f1c352d604d7f114fbf56cb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/filler/original/345c162fb50a3e90/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "744ef41c7bc153abee4b372434d389d3bc6de1be7a75976dc128bdf9b5265bc7", "seconds": 8.44, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler phrases ('um', 'let me think', 'uh', 'let's see') but does not change the core request. The final audible intent is to move all watches from the Amazon wish list to the cart, matching the original instruction. The filler does not alter the goal, so the task should be executed.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all chairs—I meant sofas from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "28f7f757c56b1a0d3d730eb683e96d166d4305677463c4c484df04f28f2a8b43", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'chairs' to 'sofas' within the same utterance, preserving the app (Amazon), action (move), source (cart), destination (wish list), and scope (all)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/false_start/false_start_01_product_category/94e8ba26b95768b9/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "88a9e8b1ca80fbf23b915f48967e6470cfb4ae81430e32137164c6829aea928e", "seconds": 8.16, "spoken_utterances": [ "Move all chairs—ah... I meant sofas um... from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'chairs' but immediately self-corrects to 'sofas' with 'I meant sofas'. This is a false start where the final corrected value replaces the initial mistaken one. The rest of the instruction matches the original, so the final intent is clear and executable.", "heard_transcript": "Move all chairs out. I meant sofas, um, from my Amazon cart to wish list." }, "candidate_id": "wav_e68ecd07af6db05f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('chairs'), immediately self-corrects with 'I meant sofas' in a natural, conversational tone, and then delivers the full, correct request smoothly. The correction is integrated into the flow without awkwardness, and the prosody sounds spontaneous and unscripted." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Move all sofas to wish list. Oh, and one more thing: move them from my Amazon cart.", "long_pause_offsets": [ 29 ], "source_sha256": "8062e2841858964da7cd64e9f9fdfab508799f4abc69683d6d2a65bf98b6a151", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source ('from my Amazon cart') to the initial instruction, completing the full context. The action, target, and object remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/addition/addition_02/7d86f303c945f4e5/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "c738ed546d5510bcb1638a669f72a1097d54497da5b3480c9280793c0fc19a42", "seconds": 12.84, "spoken_utterances": [ "Move all um, let me think... sofas to wish list. ", "Oh, and one more thing: move them ... from my Amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Move all um, let me think, sofas to wish list' is a filler-interrupted version of the original task. The speaker then adds the missing constraint 'move them from my Amazon cart' after a pause, which completes the original instruction. The addition clarifies the source of the sofas, matching the original intent. The final audible goal is fully aligned with the original instruction, so the addition type criterion is satisfied.", "heard_transcript": "Move all um, let me think, sofas to wish list. Oh, and one more thing, move them from my Amazon cart." }, "candidate_id": "wav_e9154ae626e25fad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a plausible self-initiated pause ('let me think'), followed by a clear initial request. After a brief pause, they add a clarifying detail ('Oh, and one more thing') with a conversational tone, smoothly integrating the additional instruction ('move them from my Amazon cart'). The prosody is varied, the pacing is human-like, and the addition feels spontaneous rather than forced." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all sofas from my amazon cart to shopping list. Change 'shopping list' to 'wish list'.", "long_pause_offsets": [ 53 ], "source_sha256": "ecc686430fba1d1fb1f36fb40e44664e457bb9251f253e2c67c48724382130f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination from 'shopping list' to 'wish list', preserving the action, item, and source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/revision/revision_04_destination_collection/a6411487987031dd/speech.wav", "wer": 0.125, "wav_sha256": "357643a0bd82d7fe49dfc1cf5dd0ac1842c9d4720a4139d33832246ed2c9d828", "seconds": 18.92, "spoken_utterances": [ "Move all uh, let's see... sofas uh, let me think... from my amazon cart hmmm... to shopping list. ", "Change 'shopping list' to 'wish list'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to shopping list' but then corrects it by saying 'Change shopping list to wish list.' This is a revision where the final intent is to move the sofas to the wish list, matching the original instruction. The correction is clear and final, so the task should be executed as revised.", "heard_transcript": "Move all uh let's see sofas uh let me think from my Amazon cart to shopping list. Change shopping list to wish list." }, "candidate_id": "wav_eef4189165cb18ed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and complex revision flow. It begins with multiple fillers ('uh', 'let's see', 'let me think') and a drawn-out pause, suggesting genuine hesitation. The speaker then makes a clear initial request, followed by a pause and a believable correction of 'shopping list' to 'wish list'. The prosody is varied, with natural intonation shifts and a conversational rhythm, including a self-correcting 'change shopping list to wish list' that sounds like a spontaneous thought process rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "4d4ce8afb9b99ae0c5ad81aaa492ae90f80721247638e76e8a8436dd3d830230", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/cancellation/cancel_037/0f2defce77972544/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "13ba680e86e2b7d2f318c9cf54f5cf732b68e3e1e41bcf72377b0efffeb964d4", "seconds": 10.8, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list.", "Wait, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but immediately corrects themselves with 'Wait, stop there' and explicitly states 'Cancel the whole task.' This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list. Wait, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "e16d4960cee5fee81601d6f605093d4236b5e739e848b71ff451a981a53ddc1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/ambiguous_halt/halt_010/dba5f39bd81de734/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "0e2d3eb52d3e59f76fe1d501c1d9912ce5a9e28733689f5a2c57de7cbfef011b", "seconds": 9.2, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then audibly requests to pause with 'Hold on a second.' This constitutes a final, clear request to halt, matching the 'ambiguous_halt' type criterion. No cancellation or execution is requested.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "9bf2c8a_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "131a656060e4c52159c2ba0afd5cf39d674dda7a1f8d813b16e63cfed9632e90", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/filler/original/8fda3afedef963c2/speech.wav", "wer": 0.36363636363636365, "wav_sha256": "0ba9005e15a4e643c969bc90c6cfb1da156f088c29610dd5185f0bd2c2821309", "seconds": 5.48, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'uh' but does not change the core task. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all artists of all jazz-genre—no, I meant classical-genre songs in any of my playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "9ac5e38c400d64dd1efe03423fa2e24e205c0ec2a1dea285462ea129f25194ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'jazz-genre' to 'classical-genre' within the same utterance. The rest of the instruction, including the app (Spotify), scope ('all artists'), and context ('in any of my playlists'), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/false_start/false_start_01_genre/252939907f9103b9/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "05488d130a0921791d35e2af859bcdde8a4f42187d8c7b8f2979176b53c793e9", "seconds": 10.84, "spoken_utterances": [ "Follow all artists of all jazz-genre—ah... no, I meant um... classical-genre songs in any of my playlists on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz genre' but immediately self-corrects to 'classical genre songs'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference instruction in both action ('Follow all artists') and target ('classical-genre songs in any of my playlists on Spotify').", "heard_transcript": "Follow all artists of all jazz genre. Ah, no, I meant um, classical genre songs in any of my playlists on Spotify." }, "candidate_id": "wav_ff091fd1c3796ad4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start and self-correction. The initial 'jazz genre' is clearly a slip, followed by a believable 'no, I meant um' with a filler that sounds like spontaneous hesitation. The correction to 'classical genre' is delivered with a slight shift in intonation and pacing, and the rest of the sentence flows smoothly. The prosody and rhythm are consistent with someone thinking aloud and correcting themselves in real time." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Follow all artists of all classical-genre songs in on Spotify. I left out a detail: this should apply only to my playlists.", "long_pause_offsets": [ 63 ], "source_sha256": "a4cf7bc8193cca50da0607cac6750041d7efa4c1a305e248818872ebac1aaeb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only to my playlists', which narrows the scope of the initial instruction without contradicting it. The app, genre, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/addition/addition_02/952fe4ab44a77a5a/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "07cae4b374c11d8f55e2dc37959f1c431b84d11783f9c04725f787b88f0cca23", "seconds": 16.88, "spoken_utterances": [ "Follow all artists of all uh... classical-genre songs in on um... Spotify. ", "I left out a detail: this should apply only to um, let's see... my playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a broad instruction to follow all classical artists across any playlists. After a pause and self-correction, they narrow the scope to 'my playlists' using 'This should apply only to...'. Since the type is 'addition', the final goal combines the original instruction with the added constraint. The final audible intent is a narrower version of the original, which matches the expected behavior for an addition. No other issues are present.", "heard_transcript": "Follow all artists of all uh classical genre songs in on um Spotify. I left out a detail. This should apply only to um let's see, my playlists." }, "candidate_id": "wav_745ffc675101b3b5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Follow all artists of all classical-genre songs in any of my playlists on Apple Music. Let me correct the app: follow the songs on Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "eea727e4b0b824862dd83db2ce8655ab25496422c4f60abbd97c6f101ae8734d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the genre, scope, and action. The final goal matches the original instruction with the correct app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/revision/revision_03_app/efba3c545ecaa6d3/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "e2266e0bca1ad40d3bbbbca6d9de903917f9287cee218ee457f1c774922262f7", "seconds": 15.92, "spoken_utterances": [ "Follow all artists of all uh... classical-genre songs in ... any of my playlists on um... Apple Music. ", "Let me correct the app: follow the songs on Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with only a minor paraphrase ('songs' vs 'artists of songs') that does not change the core task. The correction is clear and unambiguous.", "heard_transcript": "Follow all artists of all uh classical genre songs in any of my playlists on um Apple Music. Let me correct the app, follow the songs on Spotify not Apple Music." }, "candidate_id": "wav_72bd52d66a94b6db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant request, uses a believable 'let me think' filler, and then smoothly self-corrects the platform from Apple Music to Spotify. The prosody is consistent, the pauses are appropriately placed, and the correction is integrated into the same utterance without awkwardness. The speaker sounds like a real person formulating and refining a request in real time." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "80bd387e1b16a216a8db02575e5eacd6a1d873d935af757adceae0433d747171", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/cancellation/cancel_039/afcd8d7e81b3e92c/speech.wav", "wer": 0.08, "wav_sha256": "3aa7e120810d2059bbaebba4275c581ba6ebf02e45f83c11c83e95edd08c70a4", "seconds": 15.76, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify.", "Wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the task to follow artists of classical songs in playlists on Spotify. After a brief hesitation, they explicitly say 'Wait, don't go ahead with that. Cancel it.' This constitutes a clear cancellation of the entire task. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Follow all artists of all classical genre songs in hmm any of my playlists on uh Spotify. Wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "432c351d0eaa67b3faa012f1e18475a286feb6cb5125995a632f9e979069a478", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/ambiguous_halt/halt_016/f574d0194ccd4ed6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3c667e57a128d7ab07d4154a9d2c6e4b0e5c96f886f7f1cdcb19e84e9c21a11b", "seconds": 14.96, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify.", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause request without cancellation or resumption. The initial instruction is not revoked or corrected, but the speaker explicitly halts further action.", "heard_transcript": "Follow all artists of all classical genre songs in any of my playlists on uh Spotify. Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e2fbe70145d03d745e38061acd69f23cd17b8e5f2eb216e3575e993cdfd55bc7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/filler/original/066fef5a5fbc1d78/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "9c88ee9211757627c5f454eb9d53e075b2b596d2d15025a2346506ed59d9eed1", "seconds": 10.96, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('hmm', 'uh') but does not change the core task. The final intent remains to follow all artists of all classical-genre songs in any of the user's playlists on Spotify, matching the original instruction. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "Follow all artists of all classical genre songs in hmm any of my playlists on uh Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Follow all artists of all reggae-genre songs in any of my albums—make that playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "e0af0ccdd8a864e119d4d98afe2f319ae98b2d17ab903cfa8293d515403b94a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately corrects it to 'playlists' within the same utterance. The final target is 'playlists', which matches the original instruction. All other elements—genre, action, scope, and app—remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/false_start/false_start_02_source_collection/5e96b2409c4004e6/speech.wav", "wer": 0.1, "wav_sha256": "abc2f3b7b5838002a448c57ffcd3dcaf118b1b1cea7cdd8620a85d0650447359", "seconds": 10.84, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in uh... any of my albums—hmmm... make that ... playlists on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with the phrase 'Make that playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only the target object corrected from 'albums' to 'playlists'. All other conditions, including 'reggae-genre songs', 'all artists', 'any of my', and 'on Spotify', are preserved. Therefore, the task is executed as intended.", "heard_transcript": "Follow all artists of all reggae-genre songs in uh any of my albums. Hmm. Make that playlists on Spotify." }, "candidate_id": "wav_889c39346633ce5f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'albums', hesitates with 'hmm', then smoothly revises to 'playlists' with a clear, conversational 'make that' transition. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pacing and intonation suggest genuine thought processing." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Follow all artists of all reggae-genre songs on Spotify. I left out a detail: this should only apply to songs in my playlists.", "long_pause_offsets": [ 57 ], "source_sha256": "8bf4ef2e79c9cf82b4431c9d9a22426afb7ce89286f9e3e9e9161186e09af64a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint that the songs must be in the user's playlists, which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/addition/addition_02/b90d4fe68317d67a/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "ef83a4bcc98c3faa538185571e5d7c5578a36bee1b7d5d82184d76c35bf6fe92", "seconds": 14.8, "spoken_utterances": [ "Follow all artists of all ... reggae-genre songs on uh... Spotify. ", "I left out a detail: this should only apply to songs um... in my playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to follow all artists of all reggae-genre songs on Spotify. After a pause, they add a clarifying constraint: 'This should only apply to songs um in my playlists.' Since the addition is made after a longer pause and modifies the scope of the original request, the final intent is the combination of the original instruction with the added constraint. The final goal matches the reference instruction with the addition of a specific scope ('in my playlists'), which is a valid addition. Therefore, the task is executed with the combined intent.", "heard_transcript": "Follow all artists of all reggae-genre songs on uh Spotify. I left out a detail. This should only apply to songs um in my playlists." }, "candidate_id": "wav_f8313b4d112478b5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then naturally pauses and adds a crucial detail ('this should only apply to songs...') with a believable hesitation ('um') and a smooth, additive flow. The correction feels spontaneous and contextually grounded, fitting the 'addition' augmentation type well. The prosody is consistent, and the speaker's voice remains stable throughout, suggesting a single, continuous thought process." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all artists of all rock songs in any of my playlists on Spotify. I misspoke about the genre; use reggae-genre instead of rock.", "long_pause_offsets": [ 72 ], "source_sha256": "8719a872a59d6d12560d6e9a29ad815071f115247817fd53c98751c75b2e9854", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from 'rock' to 'reggae-genre', while retaining the app, scope ('all artists'), and context ('in any of my playlists')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/revision/revision_01_genre/9fcbd5dd9a474444/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "d5196a1b8f7d9d93b2940b876b25cd826ccc776dd1caf6a01f61308e8b572954", "seconds": 16.76, "spoken_utterances": [ "Follow all artists of all uh... rock songs in any of my playlists on uh, let me think... Spotify. ", "I misspoke about the genre; use ... reggae-genre instead of rock." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock songs' but immediately corrects themselves with 'I misspoke about the genre. Use reggae genre instead of rock.' This constitutes a revision where the final intent is to follow artists of reggae songs, matching the original instruction. The correction is clear and applied to the earlier instruction, with no other conflicting conditions.", "heard_transcript": "Follow all artists of all uh rock songs in any of my playlists on uh let me think. Spotify. I misspoke about the genre. Use reggae genre instead of rock." }, "candidate_id": "wav_43e586437682c2d5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker begins with a clear request, inserts a realistic 'let me think' pause, then self-corrects the genre from 'rock' to 'reggae' with a believable 'I misspoke about the genre' phrase. The correction is smoothly integrated into the same utterance, maintaining a natural rhythm and prosody. The speaker's hesitation and self-correction feel spontaneous and not rehearsed." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "fbb4d99727d9a6ad1ce394d12df160bcac1248c5ccc97ff74bab569e5aa0af3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/cancellation/cancel_012/6c4dede73afabcbb/speech.wav", "wer": 0.0, "wav_sha256": "3131b577f8a720b32a3a9abfa2bf0b901293fd38c8153dd5850d36ffc9095e34", "seconds": 13.04, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify.", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then after a pause and the phrase 'Wait, actually, I've changed my mind,' explicitly says 'Don't do that.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute, and the cancellation applies to the entire task as originally stated.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify. Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "9b9f40bbf7eb04423068fd36cd231eb99eae6220fd42899d26fd8d595fe66949", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/ambiguous_halt/halt_011/96383dbaefd9a8cd/speech.wav", "wer": 0.0, "wav_sha256": "f0464c6466c27948dfcde66291bfd5cc02fae18ccff35806a80b5ab81fc236a0", "seconds": 11.04, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on a second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task but asks for a temporary halt. The initial instruction is fully audible and complete, and the final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "eafd29ea67b1c644f7b34391419ed6675928bf854f4efefa22e97f534b32ba66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/filler/original/2e70c3c6849b50a9/speech.wav", "wer": 0.0, "wav_sha256": "9184d9ec1374be12f78ae93286d62cebbd6a8b120f7a566c5703f48df32384ed", "seconds": 7.76, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not change the goal. Since no such hesitation is present, the task is executed as stated.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Follow all artists of all indie-genre songs in any of my albums—no, make that playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "9f71b987c873e3a63065f989ce6d63e6a72b6e608b19b82a14de1ad87fd9579a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'albums' to 'playlists' within the same utterance. The final target is 'playlists', which matches the original instruction. All other conditions—genre, app, and scope—remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/false_start/false_start_02_source_collection/8d6cd57b5082367c/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "ae3d39d19d36217c372e6e905b9cae2a6ae82c210afade2de3f7212e83501ef4", "seconds": 10.28, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my ... albums—uh... no, make that playlists on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with 'no make that playlists'. This is a false start where the final corrected value is 'playlists'. The rest of the instruction matches the original. The final intent is clear and executable.", "heard_transcript": "Follow all artists of all indie genre songs in um any of my albums uh no make that playlists on Spotify." }, "candidate_id": "wav_d904a3fcfeafeff9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' and 'uh' before self-correcting 'albums' to 'playlists' with a smooth, plausible intonation shift. The correction is immediate and integrated into the flow, sounding like a genuine moment of recall. The prosody is consistent, and the delivery feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "b7a9ee9_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "f81b3b35d440637ab27a0bf272f2641dc6c07fcc48d9b081d8753b6f4e6b6324", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/cancellation/cancel_032/e2962d0aeae4ee42/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "f3628cc9de57a95cd522a951e31f0a3a5344ca4820eae32c202e70252249d436", "seconds": 12.72, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify.", "Hold on, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue a request but immediately self-cancels with the explicit phrase 'Hold on, I don't need this anymore. Please cancel the request.' This constitutes a clear and final cancellation of the task. The initial part of the audio is a false start, but the final control intent is unambiguous.", "heard_transcript": "Follow all artists of all indie-genre songs in um any of my playlists on let's see. Hold on, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "604b6723e0843020634b15cc25c4d9d2794f4216923366d892bd35ba06396cfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/ambiguous_halt/halt_006/18f2f83b5381590e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3fe24a5180a8473d8f1ed9fdcdc152c8fdf36eaf562d44de329fffe8bf891fba", "seconds": 10.72, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a moment', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is incomplete and contains hesitation ('um'), but the final interjection is a pause command, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Follow all artists of all indie genre songs in um any of my playlists on let's see just a moment" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "b7a9ee9_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a2eeff774663533b3d94281c449044330bf29dcfa973ee92175671ae04c9de27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/filler/original/8171b8e4177a5c5d/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5c32a9a047976c59245c1de3a69619fd3e4da4f411cb48369bfab1cd3f8b617d", "seconds": 7.48, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let's see') but does not change the core request. The final intent is to follow all artists of all indie-genre songs in any of the user's playlists on Spotify, which matches the original instruction. The hesitation and fillers are ignored as per the 'filler' type criterion.", "heard_transcript": "Follow all artists of all indie-genre songs in um any of my playlists on let's see" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I should say Venmo anyone I have sent or received money from this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "5562916415081a91ae13903221c683f143dafff5a9626e4d0ac56361fa1fbd6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, the condition about money history, and the unfriend directive." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/false_start/false_start_01_app/51145399e847ba40/speech.wav", "wer": 0.36, "wav_sha256": "7023f87177fddb3884b145413c771ad9bf18c79fe43613b7d5e28ed85c0a5f48", "seconds": 13.6, "spoken_utterances": [ "Befriend on CashApp—uh... I should say um... Venmo anyone I have sent or received money from let's see... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'I should say, um, Venmo'. This is a false start where the initial incorrect value is corrected. The final instruction matches the original reference, so the task should be executed as corrected.", "heard_transcript": "Befriend on Cash App. I should say, um, Venmo anyone I have sent or received money from, let's see, this month and unfriend everyone else." }, "candidate_id": "wav_430f051af49dbd50", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural false start with 'Befriend on Cash App' followed by a self-correction to 'Venmo'. The speaker uses a realistic filler 'um' and a brief pause before continuing. The correction is immediate and the flow remains smooth, with a plausible hesitation ('let's see') before completing the request. The prosody and pacing are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month and unfriend everyone else. One more thing: only befriend people I've sent or received money from.", "long_pause_offsets": [ 71 ], "source_sha256": "c10b2b09885ae63b650044d537e5cb1d7c2246b9be731b106e93ea57a93c9942", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction that only people with whom money was sent or received should be befriended. The initial instruction is broad ('anyone I have this month'), but the addition clarifies the condition. The action (befriend/unfriend) and scope (this month) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/addition/addition_02/17b324ecc0df98fe/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "34d871c6c286604318e297f9fb1ca77abdf28dc28f33bfe83626a87fd5974cdc", "seconds": 17.08, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have this month and ... unfriend everyone else. ", "One more thing: only befriend people I've let me think... sent or received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'anyone I have this month' which is incomplete, but then adds the clarifying condition 'only befriend people I've sent or received money from' after a pause. This addition correctly specifies the intended constraint, combining with the original instruction to form the complete and correct goal. The final intent matches the original instruction exactly.", "heard_transcript": "Befriend on uh Venmo anyone I have this month and unfriend everyone else. One more thing, only befriend people I've let me think sent or received money from." }, "candidate_id": "wav_784526c1d243963d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have sent or received money from this month and unfriend everyone else. Wait, the app should be Venmo, not PayPal.", "long_pause_offsets": [ 100 ], "source_sha256": "199ae49208ffca74caf758b9a75715659e8f9a7c3f8eae6b3e24c77c5cb37d1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the action, condition, and scope of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/revision/revision_01_app/87ca115e080c25af/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "acdb4e50142d818e8c85a519893eb4db68ff79da0d4de73dfa740acc4d60ee8a", "seconds": 16.96, "spoken_utterances": [ "Befriend on uh... PayPal anyone I have ... sent or received money from this month and unfriend everyone else. ", "Wait, the app should be ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions PayPal but immediately corrects themselves to Venmo after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to perform the action on Venmo, which matches the original reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Befriend on uh PayPal anyone I have sent or received money from this month and unfriend everyone else. Wait, the app should be Venmo, not PayPal." }, "candidate_id": "wav_abfcc8e95ed2e20a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker starts with a clear request, then pauses with a believable 'wait' before self-correcting the app name from 'PayPal' to 'Venmo'. The correction is integrated smoothly into the ongoing utterance with a natural tone of realization, and the prosody remains consistent throughout." } ] }, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "1e2d2e3c163cf3cc4cfc62f241509ad2a43bdf51bb47905fb339c7216cc9d775", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/cancellation/cancel_037/b84eb75983e737fe/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "4298c69a36b9f5aae272b3571a8454d60cd95039b2019aaf9ffb261503aac78c", "seconds": 17.04, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else.", "Hang on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request and then clearly and unambiguously cancels the entire task with 'Cancel the whole task.' This meets the cancellation type criterion, as the final audible intent is a complete revocation of the original instruction.", "heard_transcript": "Befriend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else. Hang on, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "f2d40145b64fd575378b4aeb02a061c0e6d521f1deefeea57e44a6fce2f735c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/ambiguous_halt/halt_022/e00932b55938b33d/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "eaada0c86528e6884177653bbf05f93ad4c3aa5ba41844c9480ae4b34856c001", "seconds": 16.08, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or a new instruction. The initial request is not executed, but the intent to pause is audible and final.", "heard_transcript": "befriend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f8d716f1d38b308596bdbf3e6763d4ae4feaf73eafd433ba6f7773fa9e21b036", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/filler/original/29463bc346ed6791/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d638f0dd5e199db06b0bbf30a62f33f5b9ca1249d5442c414d66f4614d1f193b", "seconds": 12.04, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('uh', 'let's see', 'um') but completes the full original instruction without any correction or change of intent. The final interjection is the completion of the task request, not a pause or cancellation. The hesitation and fillers are ignored as per the 'filler' type criterion, and the final audible task matches the original instruction.", "heard_transcript": "be friend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have received money from—no, I meant sent money to this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "55edea40df4150e92824d71c99c28bb0eb654a7daece89c710f193953b739a49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'received money from' to 'sent money to' within the same utterance. The final action and condition match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/false_start/false_start_02_transaction_relation/501928e09e6f9194/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "a1d046efa9bb70b01a7d63b89ef4ea374dc5a90d987b623ab1cd753bd0da7890", "seconds": 10.2, "spoken_utterances": [ "Befriend uh... on Venmo anyone I have received money from—hmmm... no, I meant sent money to uh... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received money from' but immediately self-corrects to 'sent money to'. According to the 'false_start' type criterion, the final corrected intent should be executed. The correction is immediate and unambiguous, and the rest of the instruction remains unchanged. Therefore, the final intent matches the original instruction.", "heard_transcript": "Befriend uh on Venmo anyone I had received money from hmm no I meant sent money to uh this month and unfriend everyone else." }, "candidate_id": "wav_d9aa09731bb83a33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'befriend uh on Venmo anyone I had received money from' and immediately self-corrects to 'no I meant sent money to' with a believable hesitation ('hmm') and a smooth, conversational tone. The correction is integrated into the same utterance without awkward pauses or robotic phrasing, and the prosody reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and. One more detail: unfriend everyone else.", "long_pause_offsets": [ 62 ], "source_sha256": "65b9c8d670ecb8063230ca1546bdd357085beb7d8c37e30a07ad1d993fa55f80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'unfriend everyone else' condition, completing the original instruction. The initial utterance already specified the befriending action and its condition, and the addition supplies the complementary unfriending action without altering the first part." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/addition/addition_02/5fd093d8acf8eca2/speech.wav", "wer": 0.12, "wav_sha256": "78463b39d6aee15e274ddf596f5fea0f84487da95904d14b1685fe3d6c536bda", "seconds": 14.08, "spoken_utterances": [ "Befriend um, let's see... on Venmo anyone I have sent money to uh, let me think... this month and. ", "One more detail: ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task and then adds the missing constraint 'unfriend everyone else' after a pause and self-interruption. This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is clear and complete, matching the reference instruction exactly.", "heard_transcript": "Befriend, um, let's see, on Venmo, anyone I have sent money to, uh, let me think, this month, and one more detail, unfriend everyone else." }, "candidate_id": "wav_701055d82f2ed8c8", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last month and unfriend everyone else. Make that this month, not last month.", "long_pause_offsets": [ 85 ], "source_sha256": "05458ded87c854bfdc1f4048d30bc05ccde4dfa3b45cd6cc0ef3fb799bbba1c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this month', preserving the action and scope of befriending and unfriending." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/revision/revision_03_time_reference/26ec4dfaaa93c859/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "70b53ec8626b55539e4832d0ce6bd1334fe7fb4e390065e6fe016d62f9db24dc", "seconds": 13.68, "spoken_utterances": [ "Befriend uh... on Venmo anyone I have sent money to let me think... last month and unfriend everyone else. ", "Make that ... this month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Befriend uh on Venmo anyone I have sent money to let me think last month and unfriend everyone else. Make that this month, not last month." }, "candidate_id": "wav_191893e05252a462", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('uh') and a believable self-correction of the time frame from 'last month' to 'this month'. The correction is smoothly integrated with a natural-sounding 'make that' phrase, and the overall prosody is consistent and conversational. The pacing and intonation suggest spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "4e2dde7ef0ced5758dccd42a396e8c95f27584770a370063f6d0ebf78a5f89bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/cancellation/cancel_037/66df4b2774f01d14/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "5c740fd4508b2bc2b60662c5dc2dd1de971d9a350cf7c91bc69274c577691e05", "seconds": 15.52, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else.", "Hold on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial instruction and then clearly states 'Cancel the whole task' as a final interjection, which constitutes a direct cancellation of the entire task. The cancellation is unambiguous and not followed by any resumption or modification of the original goal.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. Hold on, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_035", "hide_variant_original": true, "source_sha256": "119064280d27122ea1776de3f006ea06a098c47878d288f7005445bfd4ceb899", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/ambiguous_halt/halt_035/e357967abafa9321/speech.wav", "wer": 0.125, "wav_sha256": "d669579a2117b6d8465d2bc23028ac24577ffd8048e958efabb9acafc1a819f7", "seconds": 14.8, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else.", "Hold on before you do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, before you do that,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or provide a new instruction, but rather halts execution temporarily. The initial instruction is not revoked, and no new goal is stated. Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. Hold on, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b8704b517eb7e010364f63832ed1a65e5905feda395fb557f0de65af3d34e59f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/filler/original/c595b0f27364d459/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "ee9512726d09f6dfdbb6a559b149bc69e9a7562fe09e568b4df6383cbfc9ef0e", "seconds": 10.6, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Befriend let's see' which is a hesitation or filler, but the core instruction remains unchanged. The final intent is to befriend those sent money to this month and unfriend others. The filler does not alter the goal, so the task should be executed as per the original instruction.", "heard_transcript": "Befriend let's see on Venmo anyone I have sent money to this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have received money from last month—no, that should be this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "cc232526b9f358d390edbf71f5b7a1a1d4b8bf82439985ff922c1a88d7c3b3d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month' within the same utterance, preserving the action, app, and the unfriend-all-else condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/false_start/false_start_03_time_reference/ec25e9a617f3bb92/speech.wav", "wer": 0.12, "wav_sha256": "0a4d8f4c2a1f0e2ebe1c34b82f74133aaf5faf7fe0e907af3ba3a5c7b009c88f", "seconds": 15.48, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have received money from last month—hmmm... no, that should be ... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'No, that should be this month'. According to the 'false_start' type criterion, the final corrected value ('this month') overrides the initial mistaken value ('last month'). The rest of the instruction, including 'befriend on Venmo anyone I have received money from' and 'unfriend everyone else', remains unchanged and is preserved. The final intent matches the original instruction with the corrected time frame.", "heard_transcript": "Befriend on Venmo anyone I have received money from last month. No, that should be this month, and unfriend everyone else." }, "candidate_id": "wav_54dc053714c7a20d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic speech patterns. The speaker uses a genuine 'let me think' filler, followed by a spontaneous self-correction from 'last month' to 'this month' with a believable hesitation and intonation shift. The pauses are appropriately placed, and the overall flow mimics a real person formulating a complex request in real time." } ] }, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month and. I left out a detail: unfriend everyone else after befriending those you've received money from.", "long_pause_offsets": [ 68 ], "source_sha256": "7e55af03458e05f05fa88198ddada273b8ff82cea1693fd24cfedb9298f2ed32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing unfriend instruction, which completes the original goal. The final action is to befriend the specified group and then unfriend all others, matching the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/addition/addition_02/6f91a2f2b5bb23ec/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "4a867687f07a307f8ac345283748b5944eee8ebbed5a74096d56ae75b620fa13", "seconds": 14.12, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have received money from ... this month and. ", "I left out a detail: ... unfriend everyone else after befriending those you've received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the initial instruction, then adds a clarifying detail ('Unfriend everyone else after befriending those you've received money from') after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final audible intent is complete and matches the reference instruction.", "heard_transcript": "Befriend on uh Venmo anyone I have received money from this month and I left out a detail. Unfriend everyone else after befriending those you've received money from." }, "candidate_id": "wav_a932fda4521336a5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. I misspoke about the transaction relation: it should be 'received money from', not 'sent money to'.", "long_pause_offsets": [ 85 ], "source_sha256": "0fc41415c4e157d0592c53211c06d2083ff75ca5b01dd03702067c2a94f3c2c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the transaction relation from 'sent money to' to 'received money from', while retaining the app, time frame, and unfriend-all action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/revision/revision_02_transaction_relation/d32c27c6ee59e49c/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "41968adb7661901dc88fbe65dcdd6cd3d425ac1919b2105f5882639754cf18e8", "seconds": 17.8, "spoken_utterances": [ "Befriend on ah... Venmo anyone I have sent money to ... this month and unfriend everyone else. ", "I misspoke about the transaction relation: it should be 'um... received money from', not 'sent money to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the incorrect condition 'sent money to' but immediately self-corrects after a pause, revising it to 'received money from'. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. I misspoke about the transaction relation. It should be um received money from, not sent money to." }, "candidate_id": "wav_ac0d4a2b36a00182", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then naturally pauses and self-corrects the transaction direction ('sent money to' to 'received money from') with a plausible 'I misspoke' opener. The correction is smoothly integrated into the same utterance, maintaining a natural conversational flow and realistic hesitation." } ] }, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "181e9b7587e5c9e12777d158cb3dafe2250a5b208aff8caf7ae100fa947821c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/cancellation/cancel_025/be3b33abd9b320d0/speech.wav", "wer": 0.16, "wav_sha256": "8e1bd13b298fe469b0e43e374797b3158d39b3b645c09eeca901df304eaddd84", "seconds": 14.8, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else.", "Hang on, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then clearly revokes it with 'Hang on, forget that request. Don't do it.' This constitutes a final, unambiguous cancellation of the entire task. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else. Hang on, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "8c47bdfd38e2b6999b588211ccaf34e7531817518ee500e8880bbc0b7174a11a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/ambiguous_halt/halt_038/03e5432d867e15a2/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "264464defb21e853f4508322fda03b7701d59ad160cd8696d61396be7753eeb2", "seconds": 13.92, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full original instruction and then makes a final interjection requesting to pause. The phrase 'Could you pause there for a second?' is a clear, audible request to hold the current operation, with no indication of cancellation or resumption. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c77c005_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7997c1c839172c662973b73a7bbd739a5cefb4bb96d7db9e2f7f7a571b20458c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/filler/original/126fc4eea33ef8ae/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "9cfa7eeb527a6cc230f9cce3fe88660b876d3ad73be501d0c462ed47e1cdb58a", "seconds": 10.12, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core instruction. The final intent matches the original instruction exactly, with no correction or modification. The 'filler' type allows for such hesitations without altering the task goal.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_collection_name", "proposed_instruction": "Play the least listened to song on Spotify from the The Velvet Echo—scratch that, make it Echo Chamber Chronicles album.", "long_pause_offsets": [], "source_sha256": "a9529b45a9f2f6fa79aed7944ba3fe0dae6e8d235585a03e38a26ac11233b8ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'The Velvet Echo' but immediately corrects it to 'Echo Chamber Chronicles' using 'scratch that, make it'. The final target album is correctly specified, and all other conditions (action, app, and song selection criteria) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/false_start/false_start_03_collection_name/d8b7bc10bd9dadf0/speech.wav", "wer": 0.0, "wav_sha256": "394290555adec032afc436ec9d0ae9e631dce4f272466ad61678536467f83abb", "seconds": 13.08, "spoken_utterances": [ "Play the least listened to song on ... Spotify from the uh... The Velvet Echo—um... scratch that, make it Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Velvet Echo' but immediately self-corrects with 'scratch that' and specifies 'Echo Chamber Chronicles album'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Play the least listened to song on Spotify from the uh the Velvet Echo um scratch that, make it Echo Chamber Chronicles album." }, "candidate_id": "wav_1abfd8d9648854ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a spontaneous-sounding 'scratch that' correction. The initial error ('The Velvet Echo') is plausible, and the correction is immediate and integrated into the flow. The prosody is consistent, and the self-correction feels like a genuine moment of recall, making it the most realistic false start." } ] }, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Play the song on Spotify from the Echo Chamber Chronicles album. Oh, I forgot to mention, play the least listened to song.", "long_pause_offsets": [ 65 ], "source_sha256": "3a4c255c81fcff071b02b462fc045f85e90445c023365945105462ec7e938a34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least listened to' to the initial request, refining the selection from any song to the specific one with the lowest listen count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/addition/addition_01/8e1c626d58d22d44/speech.wav", "wer": 0.04, "wav_sha256": "652dbbdf7d99dd9dbc3d355cd473569b67286c03034e8b74956aba40aef7034d", "seconds": 13.44, "spoken_utterances": [ "Play the song on hmmm... Spotify from the um... Echo Chamber Chronicles album. ", "Oh, I forgot to mention, play the uh... least listened to song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests playing 'the song' from the specified album on Spotify, then adds the constraint 'least listened to song' after a pause. This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "Play the song on um Spotify from the um Echo Chamber Chronicles album. Oh, I forgot to mention, play the uh least listened to song." }, "candidate_id": "wav_8f943520c21fbd18", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "Play the least listened to song on Apple Music from the Echo Chamber Chronicles album. Let me correct the app: play the song on Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "a27b5fc655ba5979db5a7db50f2355b9992052b62b5729992202807ee98a0fae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the album and the 'least listened to song' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/revision/revision_02_app/ba7c48b4941bc577/speech.wav", "wer": 0.027777777777777776, "wav_sha256": "15e8f4070229850d0f7a573e0df5d9af5ecf3f4958073b03435c2511e41ae9aa", "seconds": 19.08, "spoken_utterances": [ "Play the um, let me think... least listened to song on Apple Music from the um... Echo Chamber Chronicles album. ", "Let me correct the app: play the song on let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the task on Apple Music but immediately corrects it to Spotify after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, with the correct app (Spotify) and album ('Echo Chamber Chronicles').", "heard_transcript": "Play the um, let me think, least listened to song on Apple Music from the um, Echo Chamber Chronicles album. Let me correct the app, play the song on, let me think, Spotify not Apple Music." }, "candidate_id": "wav_9517a11cd794125f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural 'let me think' filler, followed by a believable correction of the platform from Apple Music to Spotify. The self-repair is smooth, with a pause and a rephrased, corrected version that flows logically. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a scripted edit." } ] }, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_020", "hide_variant_original": true, "source_sha256": "80804cf19dc2812a38f8ac09f4bba4845e4ccbf25cc997022117eafdec3a8ad3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/cancellation/cancel_020/bb1687a73898ac21/speech.wav", "wer": 0.25, "wav_sha256": "bc0149fc7323f4cfe632457c9b32ee4ad1d86d3d39303b01d275478a0e0eedba", "seconds": 13.08, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album.", "Hang on, never mind, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hang on, never mind, cancel that request,' which constitutes a direct and unambiguous cancellation of the previously stated task. The final control intent is to cancel, matching the expected behavior for the 'cancellation' type. No execution of the original task is implied.", "heard_transcript": "Play the um least listened to song on uh Spotify from the um Echo Chamber Chronicles album. Hang on, never mind, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "4866f4f47f8cba78e9f2e6c40e6bab18776afb112780dfb97281c1e73b46041b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/ambiguous_halt/halt_037/0f6a5d1c666b4283/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "c050cc9ebc0728c9fef87b58bdd93c0a30531168b1bd1aff765537b13fb20f58", "seconds": 11.72, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a clear final interjection 'Pause for a moment' after the completion of the initial request. This constitutes an audible request to pause, with no indication of cancellation or permission to resume. The initial request to play a song is not revoked, but the final control command is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Play the least listened to song on uh Spotify from the um Echo Chamber Chronicles album. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1db01108f1c3ec79329e1822a6fd3e07459703dd3b36558ecf38046892ee16ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/filler/original/4fb7e632479bf551/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "8b7b7d21c38188fe5c0cec9c147499a621d808536f7f8ef72d48b6b6ec6791f3", "seconds": 8.28, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') and pauses but does not change the core request. The final intent matches the original instruction, with only hesitation markers present. The filler type allows ignoring such hesitations, so the task is executed as intended.", "heard_transcript": "Play the um least listened to song on uh Spotify from the um Echo Chamber Chronicles album." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_selection", "proposed_instruction": "Play the most played—sorry, I meant most listened to song on Spotify from my Woodstock Reimagined: Festival Vibes playlist.", "long_pause_offsets": [], "source_sha256": "94a10c6b2064c4b8b7c803d9a0c83cf4feae3deacf746301f693f3539b477a21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'most played' to 'most listened to' within the same utterance, preserving the app, playlist, and action. The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/false_start/false_start_01_song_selection/28ad8a10201fd4f7/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "167ce66c2f5908790e573991750d891d382aa307bd63e5dc7d98de36f99904f6", "seconds": 13.28, "spoken_utterances": [ "Play the most played—hmmm... sorry, I meant ah... most listened to song on hmmm... Spotify from my Woodstock Reimagined: Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'most listened to'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction, including the platform (Spotify) and playlist name, remains consistent and clear. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Play the most played [inaudible] sorry, I meant uh most listened to song on [inaudible] Spotify from my Woodstock Reimagined: Festival Vibes playlist." }, "candidate_id": "wav_6b85307ab11abe27", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'most played' corrected to 'most listened to song' and 'Spotify' corrected to 'Spotify from my Woodstock Reimagined Festival Vibes playlist'. The self-correction is immediate and fluid, with a believable hesitation and intonation shift. The speaker's prosody is consistent and the pacing reflects genuine thought processing." } ] }, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Play the most listened to song on from my Woodstock Reimagined: Festival Vibes playlist. I should also mention, play it on Spotify.", "long_pause_offsets": [ 89 ], "source_sha256": "334a2439887ff5ac32e84bb986a8d7a4d070eca948baf58ab20e61a37ea68633", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify', which was implied in the original instruction but not explicitly stated. The core action, playlist, and ranking criterion remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/addition/addition_02/bf24f218e37b0540/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "1fcb1a1450ba50b84e07b238859a3ccdeb36cde3b54c9fb5359bc86be1ee7bb4", "seconds": 15.4, "spoken_utterances": [ "Play the uh... most listened to song on ... from my Woodstock Reimagined: Festival Vibes playlist. ", "I should also mention, play it on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is partially garbled ('Play the uh most listened to song on from my...'), but the final addition 'I should also mention, play it on Spotify' clarifies the platform. The addition specifies the context for the song playback, combining with the original request to form a complete and coherent task. The final intent matches the reference instruction with the addition of the platform, which is a valid refinement under the 'addition' type.", "heard_transcript": "Play the uh most listened to song on from my Woodstock Reimagined: Festival Vibes playlist. I should also mention, play it on Spotify." }, "candidate_id": "wav_918dd892785a9aee", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_song_selection", "proposed_instruction": "Play the most played song on Spotify from my Woodstock Reimagined: Festival Vibes playlist. Wait, that should be the most listened to song, not the most played.", "long_pause_offsets": [ 92 ], "source_sha256": "ec3b4488f319a3f66e9ca8121020aefddc014d5b6a8d1b03eced42add6fd38cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the phrase 'most played' to 'most listened to', which is semantically equivalent in this context. The app, playlist, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/revision/revision_01_song_selection/9933c7206863a222/speech.wav", "wer": 0.1, "wav_sha256": "f51ccf7676eace0974a10f7757e4deafe722a87e28697eef3454515138a9e8f8", "seconds": 17.12, "spoken_utterances": [ "Play the let's see... most played song on Spotify uh... from my Woodstock Reimagined: Festival Vibes playlist. ", "Wait, that should be the ... most listened to song, not the most played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played song' but immediately corrects themselves to 'most listened to song' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to play the 'most listened to song', which matches the original instruction. All other conditions, including the app (Spotify) and playlist, remain unchanged.", "heard_transcript": "Play the uh let's see most played song on Spotify uh from my Woodstock Reimagined: Festival Vibes playlist. Wait, that should be the most listened to song, not the most played." }, "candidate_id": "wav_03a3c95347891491", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous revision. The speaker begins with a slightly hesitant 'uh' and 'let's see,' then makes a clear mistake ('most played') which is immediately self-corrected with 'Wait, that should be the most listened to song.' The correction is fluent and integrated into the same utterance, with a natural pause and intonation shift. The prosody is consistent, and the flow feels like a genuine moment of recall and correction." } ] }, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "6147b2d0288de8b36992d526cd2402617612c995bf5f928298e226f6320f4c53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/cancellation/cancel_038/bfbaef0b133a05f2/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "45bfd4be35bccac4ac8604211c4fb8d368c13f5f8ca4c57c1da560b75cbb2443", "seconds": 12.52, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist.", "Hang on, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially issues a request to play a song from a playlist, then immediately follows with 'Hang on, please stop and cancel that request,' which is a clear and unambiguous cancellation of the prior task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist. Hang on, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "705fb2d8927888440ee5c39b97ce8c19bed0109d4bc1e2dcac2f9da658368d63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/ambiguous_halt/halt_023/1b7b621af27aaed6/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "27e4cce5c7598879715142c9fa048fba28780302a8ba6bab4ab69b1ec2665a61", "seconds": 11.52, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Hold on, I need a second' is a clear request to pause or wait, with no indication of cancellation or resumption. The initial request is fully audible and complete, and the final control intent is a pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "44523e640e617f18394f25016dcbf2c4e9d5048e2295cfa71c25cae1673be7d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/filler/original/ce6b094e0af9b6b5/speech.wav", "wer": 0.125, "wav_sha256": "aa2be070c12f4f02c4bae665013b81d60d0615670b50d358dc16978ea7dee7ea", "seconds": 7.6, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core task. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_collection_name", "proposed_instruction": "Play the most listened to song on Spotify from the The Velvet Underground—no, I meant Velvet Underground album.", "long_pause_offsets": [], "source_sha256": "a02d8294a564bd9e96e633c82cab67f78f0e8d3c6e3556ec5c7f088ae27af1bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the album name as 'The Velvet Underground' but immediately corrects it to 'Velvet Underground'. The final intended album name matches the original instruction, and all other requirements (action, app, and song selection criteria) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/false_start/false_start_03_collection_name/1d3a0c57deb99047/speech.wav", "wer": 0.3181818181818182, "wav_sha256": "18bff83e8e622e3c39ba65b9e4462c6d5f1b9c11ac6ae1887d485fa3cfc61c33", "seconds": 13.08, "spoken_utterances": [ "Play the ah... most listened to song on uh... Spotify from the The Velvet Underground—um... no, I meant Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces 'Velvet Underground' as 'Develvet Underground' but immediately self-corrects with 'No, I meant Velvet Underground album.' This is a classic false_start where the final, corrected version of the target is clearly stated. The control intent is to execute the corrected request, which matches the original instruction.", "heard_transcript": "Play the uh most listened to song on uh Spotify from the Develvet Underground um. No, I meant Velvet Underground album." }, "candidate_id": "wav_f9b6691f75c0f270", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker initially says 'Velvet Undergroundm' with a clear mispronunciation, then immediately self-corrects with 'No, I meant Velvet Underground album' in a way that sounds like a genuine, spontaneous recovery. The prosody is consistent, the pause before the correction is natural, and the intonation reflects a real-time realization of the error." } ] }, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Play the most listened to song on from the Velvet Underground album. I should also mention, do this on Spotify.", "long_pause_offsets": [ 69 ], "source_sha256": "62628513d25e1328709ca0069eecffc220c266570d1ebd52be64876018f84803", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the platform 'Spotify', which was implied but not explicitly stated in the original instruction. The core action, target album, and ranking criterion remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/addition/addition_02/d786761b3db0c735/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "cab87056e75c80f45e2bd14b888162b46a62fb9d8d946e71fd8acc89fab5d24f", "seconds": 11.28, "spoken_utterances": [ "Play the ah... most listened to song on from the uh... Velvet Underground album. ", "I should also mention, do this on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is slightly hesitant with fillers ('uh') but clearly requests playing the most listened to song from the Velvet Underground album. The addition 'do this on Spotify' is a clear specification of the platform, which is a valid addition to the original instruction. The final intent combines both the original request and the added constraint, matching the expected behavior for an 'addition' type task.", "heard_transcript": "Play the uh most listened to song on from the uh Velvet Underground album. I should also mention do this on Spotify." }, "candidate_id": "wav_06048f1bd4ce7cef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural prosody and a believable addition. The initial request is slightly hesitant with 'uh' fillers, which is realistic. The addition 'I should also mention do this on Spotify' is smoothly integrated with a natural pause and intonation shift, sounding like a spontaneous afterthought rather than a rehearsed or robotic insertion." } ] }, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_collection_kind", "proposed_instruction": "Play the most listened to song on Spotify from the Velvet Underground playlist. Let me correct: it should be the Velvet Underground album, not playlist.", "long_pause_offsets": [ 80 ], "source_sha256": "59ddeab867f14550ae77d6e58a1b8d318ee99627c5ff53e69b64aa9dd784ba23", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'playlist' to 'album', preserving the artist, app, and action. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/revision/revision_04_collection_kind/0d7e28cea339635c/speech.wav", "wer": 0.0, "wav_sha256": "8f5ef3d4b20dea3888f437ad931501a7e149cd9e84f7232865556be82e83393f", "seconds": 16.8, "spoken_utterances": [ "Play the uh... most listened to song on ... Spotify from the let's see... Velvet Underground playlist. ", "Let me correct: it should be the Velvet Underground album, not playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially refers to the 'Velvet Underground playlist' but immediately corrects themselves to 'the Velvet Underground album, not playlist'. This constitutes a revision where the final, corrected target is the album. The rest of the instruction, including the action (play), the metric (most listened to), and the platform (Spotify), remains unchanged. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Play the uh most listened to song on Spotify from the let's see, Velvet Underground playlist. Let me correct, it should be the Velvet Underground album, not playlist." }, "candidate_id": "wav_ba332f356f60914b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a plausible initial request, then hesitates with 'let's see' before correcting 'playlist' to 'album' with a clear, self-aware, and smoothly integrated self-correction. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "ad948726c13fba957cba4ce55c4a955211111da4d8d69fffbac1efae22cd295e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/cancellation/cancel_008/6078ba593fc204bf/speech.wav", "wer": 0.04, "wav_sha256": "b7436a7700d3465877bec63e3021c0a574328e3bebdb1186766dd67db9c02928", "seconds": 12.96, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album.", "Hold on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a request to play a song from the Velvet Underground album on Spotify, followed by a clear and final statement, 'Hold on, I want to cancel that task,' which constitutes a cancellation of the initial request. The final intent is to cancel, not to execute the original task. The cancellation is explicit and leaves no ambiguity about the speaker's intent.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album. Hold on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "37e668a22ffd2364c842e2e3acba268c2f8cd887c26eb294dc5a000f8599d190", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/ambiguous_halt/halt_034/e6e26c15057512bc/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "67f03cad47dab5bda6503867d6d9cdc1e8d899ad37c8ffc81415f9235c2db002", "seconds": 12.6, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Wait, I was just thinking,' which is a clear request to pause and reconsider. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption. The initial task is not executed, but the intent to pause is audible and final.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "c901732_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5567648ff51b7508c6feb99614da9c97bf5083fc1e2e3b5316e1b5384fef771d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/filler/original/34d6354a63da80f4/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "83a79d25c2e3325edaf90f24b5b115308696e2f3899b5c1cfb3a7d8420e3df97", "seconds": 8.08, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let's see') and pauses but does not change the core request. The final intent is to play the most listened to song on Spotify from the Velvet Underground album, which matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_billing_period", "proposed_instruction": "Make me an amazon account with yearly—no, that should be monthly prime membership.", "long_pause_offsets": [], "source_sha256": "d00fb345af0d08306c1fb593705fe582cc69ad5b2c768ebeedccc543f5195595", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly' but immediately corrects it to 'monthly' within the same utterance. The final intended membership type is 'monthly', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/false_start/false_start_02_billing_period/ad85fe3c9c0dadfb/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "d0f9e15b67e559c8fe041cb407b0da7dd643d2dd8bb980898e10c17cd841abce", "seconds": 8.48, "spoken_utterances": [ "Make me an amazon account with uh... yearly—um... no, that should be hmmm... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly' but immediately self-corrects to 'monthly' with 'no, that should be monthly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as requested.", "heard_transcript": "Make me an Amazon account with a yearly um, no, that should be monthly Prime membership." }, "candidate_id": "wav_35f9922ec3db2141", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with 'uh', then self-corrects 'yearly' to 'monthly' with a natural-sounding 'no, that should be' and a brief pause. The correction is integrated smoothly into the flow, and the final phrase is delivered with consistent prosody and a believable change-of-mind intonation. The filler and repair are well-placed and sound spontaneous." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Make me an amazon account with prime membership. Oh, right, make it a monthly prime membership.", "long_pause_offsets": [ 49 ], "source_sha256": "3e5e827ab34110b171daa96c9d0c7b3b4bd8dc9029f3878660e36e03a67977c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the specific term 'monthly' to clarify the Prime membership type, which was unspecified in the initial utterance. The core action and target remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/addition/addition_01/f65ad621b4f7db5c/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "6f714879ca6b5af3fac7a9c7c3c5b04c30fb9d2294773f8c308f4df3ff15d0a7", "seconds": 10.24, "spoken_utterances": [ "Make me an ah... amazon account with prime membership. ", "Oh, right, make it a ... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to create an Amazon account with Prime membership. The speaker then added 'monthly' as a specific detail to the Prime membership type. The final intent combines the original request with the added constraint, resulting in a complete and coherent task that matches the reference instruction. The addition is clear and does not contradict the original request.", "heard_transcript": "Make me an Amazon account with Prime membership. Oh right, make it a monthly Prime membership." }, "candidate_id": "wav_828bd5db7550a0c6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the initial request fluently and with natural prosody. After a brief pause, they self-correct with 'Oh right' and revise the membership type to 'monthly prime membership'. The revision is pragmatic and plausible, with a believable change-of-mind intonation and clear reference back to the earlier request. The flow feels spontaneous and conversational." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Make me an amazon account with monthly basic membership. That should be prime membership, not basic membership.", "long_pause_offsets": [ 57 ], "source_sha256": "7c49a5e2ce3cfeaf942f3f830bf66542991f758ae00d8ebb72745defc2e0a2c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the membership type from 'basic' to 'prime', while retaining the account creation and monthly duration. The original instruction's intent is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/revision/revision_03_subscription_type/824cdfe7b67b01de/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "28f8573f3f86f567fe46de0483416608726825b918ce91a8108f420a5fcb7a4a", "seconds": 11.24, "spoken_utterances": [ "Make me an uh... amazon account with monthly basic membership. ", "That should be um... prime membership, not basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately corrects it to 'Prime membership' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to create an Amazon account with monthly Prime membership, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "Make me an ah Amazon account with monthly basic membership. That should be um Prime membership, not basic membership." }, "candidate_id": "wav_267c5f678500cd1b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with 'uh' and 'um', then self-corrects the membership type from 'basic' to 'Prime' in a natural, spontaneous manner. The correction is integrated smoothly into the same utterance, with a believable shift in intonation and pacing. The flow feels like a genuine moment of recollection, and the prosody remains consistent throughout." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "020a4cdb13ee4bf3acef433259c7416a54e28cb92d8c1ea1b58035687b7403d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/cancellation/cancel_004/95880e3a9c39a70a/speech.wav", "wer": 0.2, "wav_sha256": "b63cb3b8ffc49526f3eefd4e47cc62007ae35ac22c88e1b9e6b677f9bf123e0d", "seconds": 9.04, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership.", "Wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a request to create an Amazon account with monthly Prime membership, then immediately follows with 'Wait, cancel that request.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and applies to the entire task as requested.", "heard_transcript": "Make me an Amazon account with monthly Prime membership. Wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "aac045e55352edc799321403fafe675eb1ea8b086ecec4a4f59dd0a5c7f04286", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/ambiguous_halt/halt_018/2b686123d145d9d2/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "1a6b2b9c66f9857b144a5e74e5d3394a64effcd031094e2201fd852084ef24ba", "seconds": 8.76, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then ends with a clear request to pause. The final interjection 'Could you pause for a second?' constitutes an unambiguous halt command. The type is 'ambiguous_halt', which requires only a pause request, not execution or cancellation. The final intent is to pause, which matches the type's expected control intent.", "heard_transcript": "Make me an Amazon account with monthly Prime membership. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "96bebf8f156c171bda68a052f21ec6172c028d62ec563e87fcaeab32f68b57ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/filler/original/aaf15e943f747e15/speech.wav", "wer": 0.2727272727272727, "wav_sha256": "dd6bcf6bb265a8b7af5b47b610c9cf1516a01e780293ee65ade8ebbc0ea83651", "seconds": 5.12, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon account with monthly Prime membership' with a slight hesitation and correction ('monthly Prime membership' instead of 'monthly Prime membership'), but the final, corrected utterance matches the original instruction. The hesitation and self-correction are typical filler behavior and do not alter the final goal. The task is executed as intended.", "heard_transcript": "Make me an Amazon account with monthly Prime membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Make me a Apple Music—I should say spotify account with yearly premium membership.", "long_pause_offsets": [], "source_sha256": "e927da6c0de48bd305559e3865b019484c8c34c44bb20a6c9716c69d55f2aff6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, app, and membership type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/false_start/false_start_01_app/1b9147c71c8f9c87/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "57271b68891d6b6dd7676be931ee69b5ef21e67887c3e9456cd47bff90a136a0", "seconds": 7.76, "spoken_utterances": [ "Make me a Apple Music—uh... I should say spotify account with ... yearly premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including 'yearly premium membership', remains unchanged and is preserved. Therefore, the final audible intent matches the original instruction, and the task should be executed.", "heard_transcript": "Make me an Apple Music uh, I should say Spotify account with yearly premium membership." }, "candidate_id": "wav_508097a933604589", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a natural-sounding self-correction from 'Apple Music' to 'Spotify account' with a believable 'I should say' phrase. The prosody is consistent, the correction is smoothly integrated into the sentence, and the overall delivery sounds like a spontaneous, real-time formulation of a request without robotic pauses or unnatural phrasing." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "cbed9b68b427bdc58fa8bf8341130d7c8b709585fd45524fe09656ce6fc6fefa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/cancellation/cancel_001/573e889c748e845f/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "ac801395574937df5e67d5a3c2be152ab7461c618b5049aaac5d908cf7311a5d", "seconds": 10.12, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'cancel that' after initially stating the task. This constitutes a direct cancellation of the entire task as requested in the reference. The final control intent is a cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Make me a Spotify account with a yearly premium membership. Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "df5e94ad0a78348c896c46b1e050917a00b5827c0da1fc0b33b31db40214b633", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/ambiguous_halt/halt_028/67fe27c2daf86acf/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "dd1433c3e65d2329e74a3fbe4f731b28e4fdef8fa8e91c1a9bd3de23f2e1cdcf", "seconds": 10.76, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request is not revoked, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Make me a Spotify account with a yearly premium membership. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "84d613d692b1b5c873dbd157e56a7b83ae44ad93d11f0f926b1fec23b804e53e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/filler/original/a51efbb454e2fd27/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "aafbf613c9dae99605b6160fef44177536eb1f93d8b5772de4f6885246b01b67", "seconds": 6.32, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Make me a', which is a hesitant start, but immediately proceeds to state the full, correct instruction without any correction or change in intent. The hesitation and filler word 'a' do not alter the final goal, which matches the original instruction. Therefore, the task is executed as intended.", "heard_transcript": "Make me a Spotify account with a yearly premium membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Make me an amazon account with yearly basic membership—scratch that, make it prime membership.", "long_pause_offsets": [], "source_sha256": "bacce68ffa0608dbeb4d7b8a7816729d9f8a40d153f4c714744bd73aedd993c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly basic membership' but immediately corrects it to 'prime membership'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/false_start/false_start_03_subscription_type/9bd0fd7de1a32505/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d56e68d239d17ed951a20286eb310404ae7dad771b8b5a96260153440d8d8a7e", "seconds": 10.8, "spoken_utterances": [ "Make me an amazon account with let me think... yearly basic membership—um... scratch that, make it prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly basic membership' but immediately self-corrects with 'scratch that' and specifies 'Prime membership'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent is clear and executable.", "heard_transcript": "Make me an Amazon account with, let me think, yearly basic membership. Um, scratch that. Make it Prime membership." }, "candidate_id": "wav_cf968bd4a1850811", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and layered self-correction. It begins with a clear false start ('Make me an Amazon account with'), followed by a believable 'let me think' filler, then a specific but incorrect detail ('yearly basic membership'). The speaker then self-corrects with 'scratch that' and a smooth, natural revision to 'make it Prime membership'. The prosody includes appropriate hesitation, a change in intonation during the correction, and a plausible final phrasing. The flow mimics a real person reconsidering a detail mid-sentence." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Make me an amazon account with yearly. I left out a detail: include prime membership.", "long_pause_offsets": [ 39 ], "source_sha256": "5b746600bf35c515d665a5a2ebb4af88794a5899fd6e4e358b00dd2b18356855", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'prime membership' requirement, completing the initial request for a yearly Amazon account. The final goal combines both the account creation and the membership type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/addition/addition_02/4390f97d80f02592/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "9e9d4da7b4daacef624eb0850a2fffec0bb7fdd818b802dcadf6d1ae5dd95861", "seconds": 12.52, "spoken_utterances": [ "Make me an um, let me think... amazon account with ... yearly. ", "I left out a detail: include prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make me an um, let me think, Amazon account with yearly' and then adds 'include Prime membership' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added detail. The final audible intent is to create an Amazon account with yearly Prime membership, which matches the original instruction. The hesitation and self-correction do not change the final goal.", "heard_transcript": "Make me an um, let me think, Amazon account with yearly. I left out a detail, include Prime membership." }, "candidate_id": "wav_94bad1a94d15d74c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a plausible self-pause ('let me think'), followed by a clear initial request. The speaker then remembers a missing detail and adds it smoothly with a self-corrective tone ('I left out a detail'), which is a realistic and common pattern in spontaneous speech. The prosody is consistent, and the addition flows naturally without awkwardness or robotic delivery." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_billing_period", "proposed_instruction": "Make me an amazon account with monthly prime membership. Make that yearly prime membership, not monthly.", "long_pause_offsets": [ 57 ], "source_sha256": "b436cecd0a43c131af82a925a6b912fe675b078e08396ce5b034e96e1a8a6a69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the membership type from monthly to yearly, preserving the account creation and Amazon app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/revision/revision_02_billing_period/76b14fb71ddfda69/speech.wav", "wer": 0.15, "wav_sha256": "4f17b1f0fa90c4ccfbc961ea13608a7a004e73bbeb2158601af743a6ddd6b59b", "seconds": 13.36, "spoken_utterances": [ "Make me an let's see... amazon account with um... monthly prime membership. ", "Make that hmmm... yearly prime membership, not monthly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly Prime membership' but immediately corrects it to 'yearly Prime membership, not monthly'. This is a clear revision where the final intent is to request a yearly membership. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Make me an let's see Amazon account with monthly Prime membership. Make that yearly Prime membership, not monthly." }, "candidate_id": "wav_c3baae213be67cd0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The initial hesitation ('Let's see') is followed by a clear, self-corrected revision of the membership duration. The correction ('Make that... yearly prime membership not monthly') is smoothly integrated into the utterance with a natural pitch drop and emphasis on 'yearly', showing a believable internal thought process. The prosody is consistent and the pacing feels like a person formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "7573baf7f47fa312dd5812b0184aec2464c9e24be507c48f168fb9ed48d4a8b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/cancellation/cancel_015/f6897404283869ff/speech.wav", "wer": 0.0, "wav_sha256": "fb98ce16271d8dd88d5e1b24ae9480896ceb08d38f6afbf12effd2e4741f532c", "seconds": 11.28, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership.", "Wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to create an Amazon account with a yearly Prime membership, but immediately corrects themselves by saying 'Wait, actually, cancel the whole thing.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction.", "heard_transcript": "Make me an Amazon account with yearly Prime membership. Wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "c653b23a42c9965931bda70a4327a827c61deba6dd0bcc4a01f5a495d9bc83b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/ambiguous_halt/halt_038/34516a7966b65997/speech.wav", "wer": 0.0, "wav_sha256": "38bc1932aacc54b4045784475b950c8cd6d826d8c5c066a854d24948164fc59a", "seconds": 10.88, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Could you pause there for a second?', which is a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The original instruction is not canceled or revised, but the speaker explicitly asks to pause, so the control intent is to pause.", "heard_transcript": "Make me an Amazon account with yearly Prime membership. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cb4ee7a_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7a5653ebcf66151db4928aa746241504404acaccbd1155d524c4f46f21dd5705", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/filler/original/e778bb53e5a56c41/speech.wav", "wer": 0.0, "wav_sha256": "7dab79974a0a05970f9eb1d18943f17b762eab303d174868d6cfda35d5bf79dc", "seconds": 6.4, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Make me an', which is a hesitant start, but immediately proceeds to state the full, correct instruction without any correction or change in intent. The hesitation is a filler and does not alter the final goal. The final audible intent matches the original instruction exactly.", "heard_transcript": "Make me an Amazon account with yearly Prime membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_artist_selection", "proposed_instruction": "Who is the least followed—no, that should be most followed artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "cb81def7a8436650d25427e9885ea1bc373a7bfe3235949cf118fbd3462d3a8f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least followed' but immediately corrects it to 'most followed'. The rest of the query, including the app (Spotify) and the scope ('artist I follow'), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/false_start/false_start_01_artist_selection/0b1e6190ea64bcf4/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "daf95b134fadf0c06db3b7d83b7e99804b7b988a0778aee04345ad9abe87491a", "seconds": 9.76, "spoken_utterances": [ "Who is the um... least followed—uh... no, that should be most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least followed' but immediately self-corrects to 'most followed' with 'no, that should be most followed'. This is a clear false start where the final intended value is 'most followed'. The rest of the instruction matches the original, so the final intent is executable and correct.", "heard_transcript": "Who is the um least followed uh no, that should be most followed artist I follow on Spotify?" }, "candidate_id": "wav_be3e8a97b1a47f46", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible self-correction. The initial 'least followed' is clearly a slip, and the immediate, smooth correction to 'most followed' with a self-aware 'no, that should be' is highly realistic. The filler 'um' and 'uh' are well-placed, and the overall prosody reflects genuine hesitation and recollection rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Who is the most followed on Spotify? Oh, I forgot to mention, I mean the artist I follow.", "long_pause_offsets": [ 37 ], "source_sha256": "10ac0732ac67790538f37ad4f875f2436374b8fa747ffed08d840ac6a5d5e8b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'artist I follow' to the initial broad question, narrowing the scope to only artists the user follows. The core action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/addition/addition_01/c2404edefcdcdd68/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "ac79264c01a13b13683f9eaaa1ff0780384c6316b6ae5155245b98061818574e", "seconds": 12.4, "spoken_utterances": [ "Who is the uh... most followed on Spotify? ", "Oh, I forgot to mention, I mean the let me think... artist I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request 'Who is the most followed on Spotify?' is incomplete. The speaker then adds the missing constraint 'artist I follow' after a pause and self-correction, completing the original instruction. The addition clarifies the scope of 'most followed' to be limited to the user's own followed artists, which is the intended final goal. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the uh most followed on Spotify? Oh, I forgot to mention, I mean the, let me think, artist I follow." }, "candidate_id": "wav_033a926ee528f501", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music? Let me correct the app: the question is about Spotify, not Apple Music.", "long_pause_offsets": [ 57 ], "source_sha256": "df06d67adaa44ef8fbf143fdc61b3afb0905dcbdb8d59a8ab1f7196627fb5e1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (most followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/revision/revision_03_app/43c1566a9ae427c4/speech.wav", "wer": 0.0, "wav_sha256": "ee33e0cf8cdf24c2cc9bd2c32186f44ce5c714a86471c08d607d2fba83cd141c", "seconds": 12.24, "spoken_utterances": [ "Who is the most followed artist I follow on uh... Apple Music? ", "Let me correct the app: the question is about let's see... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects the app to Spotify. The revision is clear and unambiguous, and the final intent matches the original instruction. The type is 'revision', and the correction is applied as required.", "heard_transcript": "Who is the most followed artist I follow on uh Apple Music? Let me correct the app. The question is about, let's see, Spotify, not Apple Music." }, "candidate_id": "wav_0c136f1c702e44c7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects the platform from 'Apple Music' to 'Spotify' with a believable hesitation ('let me correct the app') and a brief, plausible thought process ('the question is about... let's see'). The correction is smoothly integrated into the utterance, and the prosody reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "d135287deeccdeeef8e93ca3b1ee887ee3de8a61bdfff4357bb0916b7a62fa6e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/cancellation/cancel_012/642e81ded75b6248/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "7684ef3ce5b1f08e904cbb38b74c2ca15852216d70c9527df68f0817eea5911d", "seconds": 11.04, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task but then clearly revokes it with 'Wait, actually, I've changed my mind. Don't do that.' This constitutes a direct cancellation of the original request. The final intent is to cancel, not to execute or pause.", "heard_transcript": "Who is the um, let me think, most followed artist I follow on Spotify? Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "b2533de5616b97bfec6b4969efa50bd1d63824492179d874768b30efe9206ae8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/ambiguous_halt/halt_003/0996e10c5e039c9f/speech.wav", "wer": 0.25, "wav_sha256": "20472b8d6e6ccac2dedc75de0f23f119bde8738806aad09f50fd4c294d4340f6", "seconds": 8.64, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on,' which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed. This matches the 'ambiguous_halt' type, where the final interjection is a pause command.", "heard_transcript": "Who is the um, let me think. Most followed artist I follow on Spotify. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8f37a2fbcdc4be4a5615c56bde3e2787b960037deab9aa7166402a83d15e0e9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/filler/original/8289fa2671261680/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "cf11951443ac5249810f5bd97be7802c3d0bf757e0ac8935e855db606109b6d9", "seconds": 5.8, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler phrases 'um' and 'let me think' but does not alter the core request. The final intent remains unchanged and matches the original instruction. The type is 'filler', which permits such hesitations and backchannels as long as the task goal is preserved.", "heard_transcript": "Who is the um, let me think, most followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_artist_selection", "proposed_instruction": "Who is the most followed—no, I mean least followed artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "cf616eee8c6a973eaeddc3abcd01f76721d0cc3e8678807fca65e63f12a9f380", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most followed' but immediately corrects to 'least followed' within the same utterance. The final intended comparison is 'least followed', and all other elements (scope: 'artist I follow', app: 'Spotify') match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/false_start/false_start_01_artist_selection/343a484b44bcbaed/speech.wav", "wer": 0.0, "wav_sha256": "ee628fc7529147a52044538d9811692e2440e1fcbee92c5aa30713cc04e9753c", "seconds": 10.04, "spoken_utterances": [ "Who is the most followed—um... no, I mean ... least followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most followed' but immediately self-corrects to 'least followed'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the most followed um, no, I mean, least followed artist I follow on Spotify?" }, "candidate_id": "wav_46dbad165adfd86f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural and plausible false start. The initial request 'most followed' is clearly spoken, followed by a genuine-sounding hesitation ('um') and a self-correction to 'least followed'. The intonation and pacing of the correction ('No, I mean...') are highly realistic, and the overall flow is continuous and unforced. The speaker's voice and prosody remain consistent throughout, with no signs of artificiality or disjointedness." } ] }, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Who is the artist I follow on Spotify? Oh, I forgot to mention, I want the least followed artist.", "long_pause_offsets": [ 39 ], "source_sha256": "eecb0d2eec1c7e86810c2f9ee273c1032e70635c6175eba6125824f798f29617", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least followed' to the initial question about artists I follow, refining the scope to the specific artist with the minimum follower count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/addition/addition_01/0ef27f9ac602ae7e/speech.wav", "wer": 0.0, "wav_sha256": "688a3ab7d1594064f954b80fb04929a35d6ac327d20c9fe25c2fbf3ed0946230", "seconds": 11.76, "spoken_utterances": [ "Who is the artist I follow on ... Spotify? ", "Oh, I forgot to mention, I want the ... least followed artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Who is the artist I follow on Spotify?' is a broad request. The speaker then adds the constraint 'I want the least followed artist' after a pause, which narrows the scope. The final intent is to identify the least followed artist from the user's followed artists, matching the original reference instruction. The addition refines the initial request without contradicting it, satisfying the addition type criterion.", "heard_transcript": "Who is the artist I follow on Spotify? Oh, I forgot to mention, I want the least followed artist." }, "candidate_id": "wav_936d365d48c88e1c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, coherent request with natural prosody and a believable, spontaneous-sounding addition. The initial request is grammatically complete, followed by a pause and a self-correcting addition ('Oh, I forgot to mention...') that is smoothly integrated with appropriate intonation and pacing. The flow feels like a real-time thought process, making it highly realistic." } ] }, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Who is the least followed artist I follow on Apple Music? Let me correct the app: the least followed artist I follow on Spotify, not Apple Music.", "long_pause_offsets": [ 58 ], "source_sha256": "efacde022b2c80d5e16e123d5e1b787e250786417ee9af1773f9c116a4b28ce1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (least followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/revision/revision_03_app/0629ad33dec72288/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "56a870bebb1dff142707b074ccf26c82c7c9d06b51ff6d5060cda9d355411f08", "seconds": 18.96, "spoken_utterances": [ "Who is the uh, let's see... least followed artist I follow on um, let me think... Apple Music? ", "Let me correct the app: the least followed artist I follow on um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects themselves to Spotify, which is the final and intended target platform. The revision is clear and unambiguous, and the rest of the instruction remains consistent with the original. Therefore, the final intent matches the original instruction and the revision criterion is satisfied.", "heard_transcript": "Who is the uh, let's see, least followed artist I follow on um, let me think, Apple Music? Let me correct the app. The least followed artist I follow on um, Spotify, not Apple Music." }, "candidate_id": "wav_0df70748f7cd4cef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage thought process with plausible fillers ('uh', 'let me see', 'let me think') and a clear, self-corrected revision of the platform from 'Apple Music' to 'Spotify'. The correction is integrated smoothly into the ongoing utterance with a natural shift in intonation and pacing, suggesting genuine recall and adjustment rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "19cedf95620d1a38d5ba389523ddd44a0a9c25bff4064f9bb65475d560637b58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/cancellation/cancel_019/5ea554f839f051cb/speech.wav", "wer": 0.0, "wav_sha256": "6ae621e77e3ed1145bc7f3818bf4d8d85120a4c98e21329415c3c45d19b0b5e6", "seconds": 11.6, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, never mind, don't do that,' which is a clear and unambiguous cancellation of the initial request. The final interjection explicitly revokes the task, satisfying the cancellation type criterion.", "heard_transcript": "Who is the least followed artist I follow on Spotify? Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "4b1693b5b6095ef0b48f85ffca0855fb19b31417787671ec9c055033d01727b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/ambiguous_halt/halt_002/09d4532b88f8089d/speech.wav", "wer": 0.0, "wav_sha256": "7fcda7be0f36b21d3e993cc4170a6583ab841d1bd70de6a59a232ca8d01e8cf2", "seconds": 9.88, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or modified, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "Who is the least followed artist I follow on Spotify? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a01889086f5f6207363dfc7bfb23a565c9e9308811c2654f9d8c1977ea54492a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/filler/original/074aba93571633e5/speech.wav", "wer": 0.0, "wav_sha256": "43f1586ff27b777f3ca27727f7cfdb0bd30ed718fc1bcdd2611ac1bcb7624e4d", "seconds": 6.6, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or pause that would indicate a pause or cancellation. The speaker delivers the full request clearly and without correction. As this is a 'filler' type, any initial hesitation or pause within the instruction is ignored, and the final task meaning matches the original instruction.", "heard_transcript": "Who is the least followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music—make that Spotify?", "long_pause_offsets": [], "source_sha256": "3e482f0ec7197ae22fc15bd4f1e06507b814071d953b1b91524489648cd2ce66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says Apple Music but immediately corrects to Spotify. The action (finding the most followed artist) and the scope (artists I follow) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/false_start/false_start_03_app/70b4768f7953a357/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "7ec4c13d7c5f8e05989f90b2657322b316aaac1d68d0a19e84cde6d7d969aeca", "seconds": 7.88, "spoken_utterances": [ "Who is the ... most followed artist I follow on Apple Music—um... make that Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the most followed artist I follow on Apple Music? Um, make that Spotify." }, "candidate_id": "wav_b783420938b428db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and fluent self-correction. The initial request is slightly hesitant ('who is the...'), followed by a clear, spontaneous correction ('make that Spotify') that flows directly from the prior statement. The prosody is consistent, and the correction feels like a genuine moment of realization rather than a rehearsed edit. The pause before the correction is brief and plausible, fitting a moment of recall." } ] }, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Who is the most followed on Spotify? Oh, I forgot to say, I mean the artist I follow.", "long_pause_offsets": [ 37 ], "source_sha256": "52c5f3994bf3915823179b210e3f065f8c034da3fa38aaba21bd334452d57449", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'artist I follow' to the initial broad question, narrowing the scope to only artists the user follows. The core action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/addition/addition_01/7d2ea8e5bad2d253/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "e534d9179ab2e49413d17d3bbc48c144b1e9e079a535d995e98968af5608cb4e", "seconds": 14.12, "spoken_utterances": [ "Who is the um... most followed on uh, let me think... Spotify? ", "Oh, I forgot to say, I mean the ... artist I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the most followed artist on Spotify, then adds a clarifying phrase 'I mean the artist I follow' after a pause. This addition narrows the scope of the original request, and the final intent is a combined version of both. The addition is clear and directly modifies the initial request to specify the user's personal context, matching the expected behavior for an 'addition' type. The final goal is coherent and executable.", "heard_transcript": "Who is the um most followed on uh let me think Spotify? Oh, I forgot to say, I mean the artist I follow." }, "candidate_id": "wav_82b0e8303309459e", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music? Let me correct the app: it should be Spotify, not Apple Music.", "long_pause_offsets": [ 57 ], "source_sha256": "8d8c1d626351067f15b5ca0e833ca742997616e1cd8ab022a5cec7431967a18f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (most followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/revision/revision_03_app/45e0625e6b4a86e0/speech.wav", "wer": 0.0, "wav_sha256": "5ad8ceaeb3899658d72ac27a775f4d5cc2b5fb64ece40dad80283872b912f92c", "seconds": 11.84, "spoken_utterances": [ "Who is the ... most followed artist I follow on let me think... Apple Music? ", "Let me correct the app: it should be Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects it to 'Spotify' after a pause, which aligns with the revision type criterion. The final audible intent is to perform the task on Spotify, matching the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Who is the most followed artist I follow on let me think, Apple Music? Let me correct the app, it should be Spotify, not Apple Music." }, "candidate_id": "wav_2b8ffe671f07fa0d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "cd451c406d278775060e054438161f9f70bb8b132455983030bd5278105a7ec6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/cancellation/cancel_023/33f4672d2aa6938a/speech.wav", "wer": 0.0, "wav_sha256": "d1d65541beaa99da6d22b88baebdeca657198ef95f2dc901315b94200619d7d5", "seconds": 12.08, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?", "Oh, wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Oh wait, never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly instructs not to execute the prior request, satisfying the cancellation type criterion.", "heard_transcript": "Who is the most followed artist I follow on Spotify? Oh wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "6159b7f8af247a5fbf74392dcb46fa9b0e4404fe248362bd2ce164393e4ef56c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/ambiguous_halt/halt_006/f23c31d396dbad19/speech.wav", "wer": 0.0, "wav_sha256": "88c3eade080e95b3712894f0a78152e3f288c7b34425319ed7366cfc3d954677", "seconds": 9.84, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment,' which is a clear request to wait or pause. The initial instruction is not canceled or revised, and the final interjection is a temporary halt, not a cancellation. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Who is the most followed artist I follow on Spotify? Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cef9191_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9e69d5cb832bb4b3b3b62eaf38ab8534113c4c60276e141a25d3ef28efe20f30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/filler/original/75f6140c430ff3bc/speech.wav", "wer": 0.0, "wav_sha256": "59f903cce77eedd3ee4d4bee49bfa87f2db652696a110fe4df92d8a86eca06c7", "seconds": 6.56, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or pause that changes the goal. The speaker delivers the full request clearly and without correction. As per the 'filler' type criterion, any hesitation or pause within the instruction is ignored, and the final task meaning must match the original instruction. The final intent is fully aligned with the original instruction.", "heard_transcript": "Who is the most followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_note_description", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Daily Tasks—no, I mean Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "c14fa01b45cedd68564f24dacf94d22fede2182384908f1333d417f938c2de68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'Daily Tasks' to 'Bucket List Simple Note' within the same utterance, preserving the action, note description, and completion status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/false_start/false_start_02_note_description/64a6c064b83729af/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "8b488d9fe3a749778fce09aef41cc9a1a2e2390d4a93478cd5c5b947de1f0d29", "seconds": 13.64, "spoken_utterances": [ "Mark um, let's see... \"Learning to cook a signature dish from scratch\" in my Daily Tasks—um... no, I mean ... Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the location as 'daily tasks' but immediately self-corrects to 'bucket list simple note'. According to the 'false_start' type criterion, the final corrected target should be applied. The action ('mark as done') and the task description ('learning to cook a signature dish from scratch') are preserved. The correction is immediate and unambiguous, satisfying the type-specific requirement.", "heard_transcript": "Mark, um, let's see, learning to cook a signature dish from scratch in my daily tasks, um, no, I mean, bucket list simple note as done." }, "candidate_id": "wav_d30d668982ebebd9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. The speaker begins with a filler ('um'), hesitates with 'let's see', then makes a clear false start ('daily tasks') before a smooth, self-correcting revision ('no, I mean bucket list'). The prosody is consistent, and the correction is integrated into the ongoing thought process without awkwardness. The pacing and intonation suggest genuine spontaneous formulation." } ] }, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note. One more detail: mark it as done.", "long_pause_offsets": [ 85 ], "source_sha256": "3bfd81f40127f49c67281ccab8c96c9436a865cd0358475a8e40284f3cb03df0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the 'mark it as done' instruction, which is consistent with the original goal of marking the note as done. The initial utterance already specifies the note and its content, and the addition completes the intended action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/addition/addition_02/3cdc69f7ad8e36ac/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "fa96e7a45b87d05e16fd4a3b2ddf6d097f6d7875873daa8ba0144a12c9046d9a", "seconds": 13.88, "spoken_utterances": [ "Mark ... \"Learning to cook a signature dish from scratch\" in my um... Bucket List uh... Simple Note. ", "One more detail: mark it as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to mark a specific task in a note as done. The speaker then adds the detail 'One more detail, mark it as done,' which reinforces and clarifies the original intent. Since the addition is consistent with the original instruction and does not introduce conflicting or contradictory information, the final intent is to execute the task as originally stated. The addition serves to emphasize the completion status, not to alter the target or action.", "heard_transcript": "Mark learning to cook a signature dish from scratch in my um bucket list uh simple note. One more detail, mark it as done." }, "candidate_id": "wav_22508d41319b3404", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to mark a task as done. After a brief pause and filler ('um'), they add a contextual detail ('in my bucket list simple note') in a way that sounds like a spontaneous memory. The prosody is consistent, and the addition flows naturally without awkwardness or robotic phrasing. The speaker's hesitation and correction feel authentic and integrated into the flow." } ] }, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_item_text", "proposed_instruction": "Mark \"Learning to cook a signature dish from home\" in my Bucket List Simple Note as done. Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\".", "long_pause_offsets": [ 90 ], "source_sha256": "4df65ee2b5f6dbbd65415246f7c927cbe6cc5b584e3fb2982dd0a85a7781b856", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the item description from 'from home' to 'from scratch', preserving the app, action, and status. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/revision/revision_01_item_text/63af8cfb45bec2cc/speech.wav", "wer": 0.0, "wav_sha256": "cceba2ff2ff0bfe815062fe1b49a293383f8bd9ac811d2ff04ead3c8be549d7e", "seconds": 18.16, "spoken_utterances": [ "Mark ... \"Learning to cook a signature dish from home\" in my ... Bucket List Simple Note as done. ", "Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the item as 'from home' but immediately corrects it to 'from scratch' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Mark \"Learning to cook a signature dish from home\" in my Bucket List Simple Note as done. Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\"." }, "candidate_id": "wav_0ce658818ccb3858", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a clear initial request, then pauses with 'Wait' before smoothly correcting the specific phrase 'from home' to 'from scratch'. The correction is integrated into the same utterance with a natural change of mind intonation, and the flow is continuous and believable without awkwardness or robotic delivery." } ] }, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "fff6cac785542edebe8927b1fa341a4c7d0b80ea1fbacdaa4f1a7c8c80f8d696", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/cancellation/cancel_024/b4678578ec3a325b/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "cf5f7447e4b30f6298805f642ceae97c1a59720482c524bb9bdc94cee27e3873", "seconds": 16.72, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done.", "Wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task but immediately corrects themselves with 'Wait, oh never mind, cancel the whole thing,' which is a clear and final cancellation of the task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done. Wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "8a3c1921fb3648dff8889f71af49249a9911534fda636b7132e9c86418213b7a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/ambiguous_halt/halt_006/b18990a40cbdcc83/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "54f1c8cfb753471ec4b86382d9646e3fa302a23fd2cecff417f04d27b78cc8fd", "seconds": 14.92, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial instruction is not revoked, and the speaker does not resume or correct it, so the final intent is to pause.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done. Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "50234f82310b915614798d96be76412f39368fd59592cb97345b30a452bed248", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/filler/original/6d834e580b4f3476/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "c51898a97323ec1df73ece730cb7e63a0cb4389458f8c4fa18d25ecb7e9893ae", "seconds": 11.88, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a brief hesitation ('uh') but does not alter the core instruction. The final intent matches the original request exactly, with no correction or change in goal. The hesitation is a filler and does not constitute a change in control intent.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_item_text", "proposed_instruction": "Mark \"Going to the beach\"—I meant \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done.", "long_pause_offsets": [], "source_sha256": "6d85864e1a82050840793e48f6709357b0aee7226fece9b4ad6b6e5f7be393e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the note title from \"Going to the beach\" to \"Taking a solo backpacking trip\" within the same utterance. The app, action, and status ('not done') remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/false_start/false_start_01_item_text/f421e4653da53f89/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "1d76a1fe1ed3dcfd75cbe8fd1f143c723af8d3749c0d2e89a97c1272f47622a6", "seconds": 12.72, "spoken_utterances": [ "Mark um, let me think... \"Going to the beach\"—uh... I meant \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Going to the beach' as a false start, then immediately self-corrects with 'I meant taking a solo backpacking trip'. The correction is clear and immediate, and the rest of the instruction matches the original. The final intent is to mark the correct task as not done, so the false_start type is satisfied.", "heard_transcript": "Mark, um, let me think. Going to the beach. Ah, I meant taking a solo backpacking trip in my bucket list simple note as not done." }, "candidate_id": "wav_926354e956d7e542", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible false start. The speaker begins with a clear hesitation ('um, let me think'), then makes a distinct, unrelated initial attempt ('going to the beach') before self-correcting with 'I meant...' to deliver the correct request. The correction is smooth, the intonation reflects genuine recollection, and the pacing feels spontaneous rather than rehearsed. The filler words and pauses are well-placed and contribute to the realism of a person formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note. Oh, I forgot to add, mark it as not done.", "long_pause_offsets": [ 69 ], "source_sha256": "69961315a95abb9b21c2b3c264864d2a4318e980d4b1bb354438c623153309ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'not done' status after the initial marking action, completing the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/addition/addition_01/2280a904a8619603/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "bd747e527702c5e53b00b1151d02c18e42a58c8ec1967acc3bac157cef3d4fed", "seconds": 16.48, "spoken_utterances": [ "Mark hmmm... \"Taking a solo backpacking trip\" in um... my Bucket List Simple Note. ", "Oh, I forgot to add, mark it as let's see... not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to mark the item as 'not done' but then adds the explicit phrase 'mark it as, let's see, not done' after a pause, which clarifies and completes the original instruction. The final audible intent matches the original instruction exactly, with no contradiction or ambiguity.", "heard_transcript": "Mark, hmm, taking a solo backpacking trip in um, my bucket list simple note. Oh, I forgot to add, mark it as, let's see, not done." }, "candidate_id": "wav_7fede9813e526ac8", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'oh I forgot to add') and a believable self-correction ('mark it as let's see not done') that flows smoothly within the utterance. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition of the missing detail feels like a genuine memory, not a forced insertion." } ] }, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_note_description", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Travel Goals Simple Note as not done. Let me correct the note description: use \"Bucket List\", not \"Travel Goals\".", "long_pause_offsets": [ 82 ], "source_sha256": "d1f3509c99ec11faf0c75d96e4f34244ad0d3da4404d9b5b521b345dd6b67196", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the note name from 'Travel Goals' to 'Bucket List', preserving the action, item, and status. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/revision/revision_02_note_description/9fba94e12798c81a/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "02089912c392f400e12dfb471d81639a7e5bf609f27a6407211a169c51f2aeae", "seconds": 19.6, "spoken_utterances": [ "Mark uh, let's see... \"Taking a solo backpacking trip\" in uh... my Travel Goals Simple Note as not done. ", "Let me correct the note description: use \"... Bucket List\", not \"Travel Goals\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'travel goals' but immediately corrects it to 'bucket list' after a pause, which is the revision type. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Mark uh let's see taking a solo backpacking trip in uh my travel goals simple note as not done. Let me correct the note description, use bucket list not travel goals." }, "candidate_id": "wav_1e66aa9b3d77bfad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody with believable hesitation and self-correction. The speaker uses 'uh' and 'let me see' as genuine fillers, pauses naturally, and smoothly revises 'travel goals' to 'bucket list' with a plausible justification ('let me correct the note description'). The intonation and rhythm mimic a real person thinking aloud and adjusting their request, making it the most realistic." } ] }, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "d842df6653a527e511d21050b4b102b42b6c2524bbc7d56f204f03954a51046a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/cancellation/cancel_003/7f47edfed656f2cf/speech.wav", "wer": 0.0, "wav_sha256": "71dd74b1c2e4d093da888a18177460900b74a83c885f64744a412070c23fe7a7", "seconds": 12.84, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done.", "Hang on, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, followed by a clear and final self-correction: 'Hang on, please cancel that.' This constitutes a direct cancellation of the entire task. The final audible intent is to cancel, not to execute the original request. The cancellation is unambiguous and complete, satisfying the 'cancellation' type criterion.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done. Hang on, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "921619842b7c2ac3eefdfb345c251b1caaf8590d21d32255d2f3fce750d913b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/ambiguous_halt/halt_021/f6387170c44213e5/speech.wav", "wer": 0.0, "wav_sha256": "dc75d715a969760e6c71af425032da57a96a7af563223810c89f36c14f51622d", "seconds": 12.72, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial task is not revoked, and the speaker does not provide a new goal, so the correct control intent is 'pause'.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "36555ac41d1b7b80164c287a816090d996d131af8f09cfab60baf9e9a752bb2a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/filler/original/da42fda639b01ebb/speech.wav", "wer": 0.0, "wav_sha256": "5d7b69f8e361c3f7c44e3609e89e40344ef323a8a5fd34a33c32ef7532e9a42e", "seconds": 8.52, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not alter the goal. Since the final utterance is a complete and correct execution of the task, the verdict is pass.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_item_text", "proposed_instruction": "Mark \"Hiking Mount Everest\"—I meant \"Witnessing a total solar eclipse\" in my Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "9fb6f2786dfb49f1ea968d8c6fce64a0dbf4d7f61db101e009537d0085e2fe66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the item name from \"Hiking Mount Everest\" to \"Witnessing a total solar eclipse\" within the same utterance. The app, action, and status ('as done') remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/false_start/false_start_01_item_text/c2e785b824ed0c36/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "be5bc0da0ad6b0fd8eddbc1cb1d2b1e8b473dba693b45e352d672c35dfcd72f4", "seconds": 11.52, "spoken_utterances": [ "Mark um... \"Hiking Mount Everest\"—hmmm... I meant \"Witnessing a total solar eclipse\" in uh... my Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking Mount Everest' but immediately self-corrects to 'witnessing a total solar eclipse'. According to the 'false_start' type criterion, the final intent is the corrected version, and the initial mistake is discarded. The final instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "Mark um, hiking Mount Everest. Hmm. I meant witnessing a total solar eclipse in uh my Bucket List Simple Note as done." }, "candidate_id": "wav_bee2b5bde89b90bc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a false start ('hiking Mount Everest'), followed by a believable self-correction ('I meant...') with a realistic 'hmm' hesitation. The correction is delivered with a natural shift in intonation and pacing, and the final request is clear and complete. The prosody and rhythm are consistent with someone thinking aloud and adjusting their request in real time." } ] }, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Simple Note. Oh, I forgot to mention, mark it as done.", "long_pause_offsets": [ 71 ], "source_sha256": "d68a3490f9188391473dc89e4e09b274554f9cea1c6f46a3607ed9fb0184d8e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'mark it as done' instruction, which completes the original task. The note title and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/addition/addition_01/5eddfa1a41e3e581/speech.wav", "wer": 0.16, "wav_sha256": "a783726ae0f6bd49a65b6353462d0fce5a5a050db4336b30ed5c2edf000329db", "seconds": 13.88, "spoken_utterances": [ "Mark \"Witnessing a total solar eclipse\" in ... my Bucket List Simple Note. ", "Oh, I forgot to mention, mark it um, let me think... as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to mark an item in a note. The speaker then adds 'as done' after a brief hesitation, which is a clear addition to the original task. The final intent is to mark the specified item as done, which matches the expected behavior for an 'addition' type. The correction is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "Mark witnessing a total solar eclipse in my bucket list simple note. Oh, I forgot to mention, mark it, um, let me think, as done." }, "candidate_id": "wav_68634ef5dffb4703", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then pauses with a believable 'oh I forgot to mention' and 'let me think' before adding the correction 'mark it um as done'. The prosody is consistent, the fillers are placed naturally, and the self-correction flows smoothly within the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_item_text", "proposed_instruction": "Mark \"Hiking Mount Everest\" in my Bucket List Simple Note as done. Wait, the item should be \"Witnessing a total solar eclipse\", not \"Hiking Mount Everest\".", "long_pause_offsets": [ 67 ], "source_sha256": "954738b984e7be680f0812aa22f903494b7efe9497e712812f0b0c959e100937", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly identifies the intended item to mark as done, replacing the initially mentioned 'Hiking Mount Everest' with 'Witnessing a total solar eclipse'. The app ('Bucket List Simple Note') and action ('mark as done') remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/revision/revision_01_item_text/70ce748b63748f16/speech.wav", "wer": 0.0, "wav_sha256": "34bf9c5e8fcf6fe152dee0fdd4043e021288b7bb692165c5fdcd9fef3cd3d380", "seconds": 16.2, "spoken_utterances": [ "Mark \"Hiking Mount Everest\" in um... my Bucket List Simple Note as done. ", "Wait, the item should be \"uh... Witnessing a total solar eclipse\", not \"Hiking Mount Everest\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the item as 'hiking Mount Everest' but immediately corrects it to 'witnessing a total solar eclipse' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Mark hiking Mount Everest in um my Bucket List Simple Note as done. Wait, the item should be uh witnessing a total solar eclipse, not hiking Mount Everest." }, "candidate_id": "wav_1ead70089dd9c80e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('um') and a believable self-correction. The revision ('the item should be...') is smoothly integrated into the same utterance, with a pause and filler ('uh') before the correction, matching a spontaneous thought process. The prosody is consistent, and the correction flows logically without awkwardness." } ] }, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "c4e5af0eea02ab89abf32c6e3ffb5d515f6fe7e91697662cebf204e871fcabc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/cancellation/cancel_029/0dfe1800804c820a/speech.wav", "wer": 0.04, "wav_sha256": "e351f98e7250ace8dbdf88b7927b93669dffb7a787b39cd763a6f6fb528cfabf", "seconds": 15.44, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly states 'Hang on, I'm withdrawing that request,' which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, not to execute or pause. The cancellation is unambiguous and applies to the whole task.", "heard_transcript": "Mark um witnessing a total solar eclipse in uh let me think my bucket list simple note as done. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "d100f9d1347074a3eba07112545583f5dfe593faa49ce4d6c8fa655e25ed2b8f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/ambiguous_halt/halt_017/651603dbdd92a286/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "9c1aed03b7d4c098b966bd4309c5b032f37d114a03eaeda232a4c864163a7ad4", "seconds": 14.36, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a clear request to pause with 'One moment please' following a hesitation and self-interruption ('let me think'). This constitutes an unambiguous halt request. The original instruction is not canceled or revised, and the speaker does not complete the task. The final interjection is a pause directive, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Mark um, witnessing a total solar eclipse in uh, let me think, my bucket list simple note as done. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "cf6abd2_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2702e1d27a2c3278a1efeba3182daca7513ae35b251ee4ecf32f63ea59155728", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/filler/original/e6e0255326aff7c4/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "fb03f58878bed37e413027d8de4aa4c65be32b734dcb47cddadaa0cd10c9fc6e", "seconds": 10.92, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and pauses but does not change the core task. The final audible intent is to mark the specified task as done in the specified note. The hesitation and self-interruption do not alter the goal, so the filler type criterion is satisfied.", "heard_transcript": "Mark um, witnessing a total solar eclipse in uh, let me think, my bucket list simple note as done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my friends—no, I mean roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "8701a82923531144f4703f2395a746d8982729c5f1da2e345dbaa29133fa3285", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'friends' to 'roommates' within the same utterance, preserving the app (Venmo), action (approve), scope (all), and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/false_start/false_start_02_requester_group/650465dc45bd3fe2/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "c7ddb996c7561b8df2619dd77cef43f7638021b990ae28a973dbaf9c21a60e5d", "seconds": 13.2, "spoken_utterances": [ "Approve all venmo payment requests from my hmmm... friends—um... no, I mean ... roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'roommates' with 'no, I mean'. According to the 'false_start' type criterion, the final corrected value ('roommates') should be applied, and the original incorrect value ('friends') should be discarded. The rest of the instruction matches the original. Therefore, the final intent is clear and executable.", "heard_transcript": "Approve all Venmo payment requests from my um friends um no, I mean roommates from this calendar month." }, "candidate_id": "wav_eae585138e7ca5a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'my friends' and then clearly self-corrects to 'roommates' with a believable hesitation and intonation shift. The pause and filler 'um' before the correction sound spontaneous, and the overall flow is continuous and human-like. The correction is integrated smoothly into the sentence without awkwardness." } ] }, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Approve all venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month.", "long_pause_offsets": [ 54 ], "source_sha256": "973aa8196080eda7df24eb7a4ce01f67543e0cdfa0306aeadb97c8271b837d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial request; the action, app, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/addition/addition_01/9a3bd5ef2b9bc8df/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "521e5f44f879bdef3c131fba703196eea1eed0a312ab1841a84cc6f55633f8d5", "seconds": 12.92, "spoken_utterances": [ "Approve all ah... venmo payment requests from um... my roommates. ", "Oh, I forgot to mention, only approve requests from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests approval of all Venmo payment requests from their roommates. After a pause and an interjection, they add the constraint 'only approve requests from this calendar month.' This addition narrows the scope of the original request, and the final intent is to approve only those requests that fall within the current calendar month. The addition is clear and modifies the original instruction without introducing conflicting or ambiguous conditions. Therefore, the final intent matches the expected behavior for an 'addition' type task.", "heard_transcript": "Approve all uh Venmo payment requests from um my roommates. Oh, I forgot to mention, only approve requests from this calendar month." }, "candidate_id": "wav_430db484a7516350", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding instruction ('Approve all Venmo payment requests from...') and then adds a restriction ('Oh, I forgot to mention...') with a believable hesitation ('um') and a smooth, self-contained follow-up. The prosody is consistent, and the addition feels like a spontaneous afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last month. Make that this calendar month, not last month.", "long_pause_offsets": [ 70 ], "source_sha256": "79c3fc67bcc745a71b39426144e8af1b7caf5da2d2a1c5f17649ffd226ce4b24", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this calendar month', preserving the app, action, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/revision/revision_03_time_reference/a62acd36f0e3cbe3/speech.wav", "wer": 0.125, "wav_sha256": "d89e6bba4c108f35a0a8c617cd32c81064a1508c6871f7a64d1f7754b75d242b", "seconds": 14.12, "spoken_utterances": [ "Approve all let me think... venmo payment requests from uh... my roommates from last month. ", "Make that ah... this calendar month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but then corrects it to 'this calendar month, not last month' after a pause. This is a revision where the final intent is the corrected version. The final goal matches the original instruction, so the revision is correctly applied.", "heard_transcript": "Approve all, let me think, Venmo payment requests from uh my roommates from last month. Make that uh this calendar month, not last month." }, "candidate_id": "wav_8071a21f1cf14e58", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), then delivers a clear initial request. The revision ('Make that... this calendar month, not last month') is a smooth, self-corrected continuation with a believable change-of-mind intonation and clear reference back to the original request. The prosody is consistent, and the correction flows naturally within the same utterance." } ] }, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "dbb22e8962efa7355a34cbaeb828e0c8e9e17e0c8ee5155ae661762f4c16599f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/cancellation/cancel_011/31fabe3a166f4794/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c01de2ca818583d1c6f784eb78a416f2fac39784ebdb47f8b64a616fa01d7c31", "seconds": 13.12, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month.", "Oh, wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task to approve Venmo requests, then immediately follows with 'Oh wait, I've changed my mind. Cancel that.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Oh wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "39d7f7d93cdcef825b96b376c0f38a7bdc3a02aefd553a19b258d69dfa9708eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/ambiguous_halt/halt_040/8fa22d602f3758f9/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "b4c58aae35ef79bda21209d34cb723bed938ce70400d2699f0cdbb0b581da84f", "seconds": 11.36, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, just a minute,' which is a clear request to pause or wait. There is no cancellation of the task, and the speaker does not resume or modify the instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a74caea109a60677728002ffff93a94324ba2c05c54f3349604fec4f973085ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/filler/original/4ac76bffbbd40d96/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "20102f3ae4a3cd0e90e51570ffe9e8438ff732594e549154671df7d4df2545f5", "seconds": 7.96, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a brief hesitation ('um') but the final, complete instruction matches the original reference exactly. The hesitation is a filler and does not alter the task goal. The instruction is fully audible and unambiguous, so the final intent is to execute the approval task as stated.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family—sorry, I meant roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "b8e9c9e5fdb923e129bea27dfca9cba625588555e24557da16dde040aaf8d7db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'roommates' within the same utterance, preserving the app (Venmo), action (approve), scope (all payment requests), and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/false_start/false_start_02_requester_group/56c67bf7411eef08/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "57998f00a157245e8ecd4946e56ef9e94fd0cb6ee6388cb4fc279b8af23d39ce", "seconds": 12.2, "spoken_utterances": [ "Approve all um... venmo payment requests from ... my family—hmmm... sorry, I meant roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'roommates' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value ('roommates') should be applied, and the original incorrect value ('family') should be discarded. The rest of the instruction, including 'Venmo payment requests' and 'this calendar month', remains unchanged and correct. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Approve all um Venmo payment requests from my family. Sorry, I meant roommates from this calendar month." }, "candidate_id": "wav_631e4c80d93f41a5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'approve all um Venmo payment requests from my family,' then self-corrects with 'sorry, I meant roommates from this calendar month.' The correction is smooth, with a believable shift in intonation and a clear reference back to the initial request. The pacing and prosody are consistent with spontaneous speech, and the repair feels integrated rather than forced." } ] }, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Approve all venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month.", "long_pause_offsets": [ 54 ], "source_sha256": "973aa8196080eda7df24eb7a4ce01f67543e0cdfa0306aeadb97c8271b837d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial request; the action, app, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/addition/addition_01/5a29339d637d50bf/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "d38ee2782a7b27d9ea5919a7985c805676b36ba6f700c61ced999f22bd6ee98d", "seconds": 14.12, "spoken_utterances": [ "Approve all let me think... venmo payment requests from my roommates. ", "Oh, I forgot to mention, only approve requests from hmmm... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Approve all Venmo payment requests from my roommates' and then adds a constraint 'only approve requests from this calendar month' after a longer pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is a narrowed version of the original instruction, which is the expected behavior for an addition. The original instruction is not fully met without the addition, but the addition is clearly spoken and final, so the combined goal is complete and matches the reference.", "heard_transcript": "Approve all let me think Venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month." }, "candidate_id": "wav_ac792937889f785e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Approve all Venmo payment requests from my roommates') and then adds a restriction ('only approve requests from this calendar month') after a believable hesitation ('Oh, I forgot to mention...'). The addition is smoothly integrated with a natural pause and filler ('um'), and the correction feels spontaneous rather than forced. The prosody is consistent, and the speaker's intonation reflects genuine recollection." } ] }, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last month. That should be from this calendar month, not last month.", "long_pause_offsets": [ 70 ], "source_sha256": "ff9da03f1317465fffb22f8287c4b15c9458092d9303d3f4c4becaf0d7685357", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this calendar month', while preserving the action, app, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/revision/revision_03_time_reference/97e38a18f9d68183/speech.wav", "wer": 0.0, "wav_sha256": "56297150db0d03724b625f0c20dc54e50251012be7dc54991d7c9d312d13e59d", "seconds": 14.92, "spoken_utterances": [ "Approve all venmo payment requests from uh, let's see... my roommates um... from last month. ", "That should be from this calendar month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'from last month' but immediately corrects themselves to 'from this calendar month, not last month'. This is a clear revision where the final intent is the corrected version. The original instruction matches the final audible intent, so the task is executed as revised.", "heard_transcript": "Approve all Venmo payment requests from uh let's see my roommates um from last month. That should be from this calendar month, not last month." }, "candidate_id": "wav_963f5fa10ad05e4e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous flow with a believable self-correction. The initial request is clear, followed by a plausible hesitation ('let's see'), a pause, and a smooth revision from 'last month' to 'this calendar month'. The prosody is consistent, and the correction is integrated naturally into the sentence without awkwardness." } ] }, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "f1a9d3d575bec66d315bbced99ce046cbfacd468e30e11f96a2e9ed626bfd2d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/cancellation/cancel_018/773bae80039bc85c/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "6a734f32bf54c8a8a09f53ba690ba7b15d042e27db648110c3b8d046151aeeb3", "seconds": 13.88, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month.", "Hold on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to approve Venmo requests, but then clearly revokes it by saying 'Hold on, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hold on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "6feb41706b92c78962fc94432bdfcedd39238da16c3935f909123d5ba0d649f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/ambiguous_halt/halt_010/0ddb3aa9134ad847/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "310543f113173c096cbd9f924a90ae9e9e6673e953dc124b14d02299c077bf75", "seconds": 11.12, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Hold on a second,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there any indication that the task should be resumed. The final audible intent is to halt execution temporarily, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b7a83e650c661eda7759521e7eae794efc16827723bafbbedc5cd5129465dcab", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/filler/original/dc47f5f5df514aa1/speech.wav", "wer": 0.42857142857142855, "wav_sha256": "163e21f52de684897b607ac98f42bfa9b6626e563d6bf520f1dfbf4ef8f751a2", "seconds": 7.64, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or modifications. The filler type allows ignoring hesitations and backchannels, so the task is executed as stated.", "heard_transcript": "Approve all um Venmo payment requests from my roommates uh from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Approve all PayPal—I should say venmo payment requests from my roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "b7ea7c76f02aef8468424142862edd5e173b232aa9d5f158c8717ff17ab3f5fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the recipient group, time frame, and action. The final intended app is Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/false_start/false_start_01_app/c35fadff821d3748/speech.wav", "wer": 0.0, "wav_sha256": "a161250ff704b0bda23f1887f8dd89a966bd17b60a6c9e5a0358ff11e3509393", "seconds": 8.24, "spoken_utterances": [ "Approve all PayPal—uh... I should say ... venmo payment requests from ... my roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say'. This is a false start where the final, corrected value is 'Venmo'. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final intent is to approve Venmo requests, not PayPal.", "heard_transcript": "Approve all PayPal uh, I should say Venmo payment requests from my roommates from this calendar month." }, "candidate_id": "wav_d254d3ee714f62c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'PayPal' followed by a smooth self-correction to 'Venmo'. The intonation and pacing are consistent with spontaneous speech, and the correction flows seamlessly into the rest of the sentence without awkwardness or hesitation. The prosody is continuous and the delivery sounds like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Approve all venmo payment requests from from this calendar month. One more detail: only approve requests from my roommates.", "long_pause_offsets": [ 66 ], "source_sha256": "95a75229e1c63353aa1c00d1d9a14abd003c7ebeee8514220a4f6ddbf0d30452", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the scope to only requests from roommates, which is consistent with the original instruction's intent to approve such requests. The original instruction already specified 'from this calendar month' and 'all venmo payment requests', so the addition refines the recipient condition without contradicting or removing any prior requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/addition/addition_02/9bfb1710136caeef/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "dfe09f7de9629231b84dcae965b6d5c8f3311cdc5f0e11599c515300e0c42f54", "seconds": 12.6, "spoken_utterances": [ "Approve all ... venmo payment requests from ... from this calendar month. ", "One more detail: only approve requests from uh, let me think... my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to approve all Venmo payment requests from this calendar month. Then, after a pause and a self-correction ('let me think'), they add the constraint 'Only approve requests from my roommates.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original instruction with the added constraint, matching the reference instruction exactly. Therefore, the task passes.", "heard_transcript": "Approve all Venmo payment requests from from this calendar month. One more detail. Only approve requests from uh let me think. My roommates." }, "candidate_id": "wav_25c3099e72d7d500", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Approve all PayPal payment requests from my roommates from this calendar month. Wait, use Venmo instead of PayPal for the payment requests.", "long_pause_offsets": [ 80 ], "source_sha256": "ab7995084f7778725fc3f22891aab230188a1419437a44d82665e60c0100b0a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the recipient group (roommates) and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/revision/revision_01_app/f04ce5992db6a1f2/speech.wav", "wer": 0.04, "wav_sha256": "7178dd9917f43fc21c4e8db98ec8a1611f6b7aea88c3d345b5394e8d7cfcfd85", "seconds": 13.88, "spoken_utterances": [ "Approve all PayPal payment requests from let me think... my roommates from this calendar month. ", "Wait, use ... Venmo instead of PayPal for the payment requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction ('Wait, use Venmo instead of PayPal'). This is a revision where the final intent is to use Venmo. The rest of the instruction, including the target ('my roommates') and time frame ('this calendar month'), remains unchanged. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Approve all PayPal payment requests from let me think my roommates from this calendar month. Wait, use Venmo instead of PayPal for the payment requests." }, "candidate_id": "wav_c43b1bbbfccd3372", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') before stating the initial request. The correction ('wait, use Venmo instead of PayPal') is delivered with a clear shift in intonation and a self-repair that sounds spontaneous. The flow from initial request to correction is smooth and credible, with no unnatural pauses or robotic delivery. The speaker's voice remains consistent and the prosody reflects genuine thought process." } ] }, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "407b6770174e37a3177bf27be16c4f96f0394caa2fee7227e2e8bd3185770f6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/cancellation/cancel_038/2bd0f382858c2c4e/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "d14aa0bca1ad336036e39f8d970701068b9ed34e0bc2e261115ee6027de07d29", "seconds": 15.32, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month.", "Oh, wait, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'Oh wait, please stop and cancel that request.' This constitutes a final, unambiguous cancellation of the task. The original instruction is irrelevant after this cancellation.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month. Oh wait, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "3cc44651638e916bca3d59f70891d99323a68b3e9f977b40db61b78b1333fdfc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/ambiguous_halt/halt_026/7e9569d2f3abbe52/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "d5b7781a759996ca23e33479e3a7bc4dbf1f1b9da8b1ad98e3a719653b81a096", "seconds": 14.36, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but then repeatedly hesitates ('uh', 'let me think') and ends with 'Just a moment, let me think about that,' which is a clear request to pause. The final interjection indicates a temporary halt, not a cancellation or a change in the task goal. The instruction is not completed, but the speaker's intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d18139b_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "fbb1d5e5aac5380aff925eab5efb03d79d0d005db1f2bbb9aee745e0ca216657", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/filler/original/b45a591f5f3e3524/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5e7c93de69355e2919d7041a2e4bc7463d2796c2e8f7178aa44328be663af7fa", "seconds": 10.28, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let me think') and pauses but does not change the core request. The final intent matches the original instruction, with only hesitation and backchanneling. The 'filler' type allows for such hesitations as long as the final goal remains unchanged.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists—I should say songs and albums I have ever liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8759de467cb3ca2bcb9ab277eaffbc0c85b65411dd68822377ac1186fd9abeef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but immediately corrects to 'songs and albums'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/false_start/false_start_01_media_scope/7661328389d659f0/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "7939369532d83d1d7a9d241323c99cd4c68938c0f02771c9ea8575f81a2c590f", "seconds": 9.52, "spoken_utterances": [ "Follow artists of all the uh... tracks and playlists—um... I should say songs and albums I have ever liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlist them', which is a false start. They immediately self-correct to 'songs and albums I have ever liked on Spotify'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistake should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Follow artists of all the uh tracks and playlist them. I should say, songs and albums I have ever liked on Spotify." }, "candidate_id": "wav_533d1405a50a4a59", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and fluent self-correction. The initial false start 'uh tracks and playlists' is immediately and smoothly repaired to 'songs and albums' with a believable 'I should say' transition. The prosody is consistent, the pace is conversational, and the correction feels spontaneous rather than rehearsed. The speaker also correctly specifies 'on Spotify' without further revision, maintaining task clarity." } ] }, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Follow artists of I have ever liked on Spotify. Oh, I forgot to mention, this should apply only to all the songs and albums.", "long_pause_offsets": [ 48 ], "source_sha256": "846cee9956c9197a779aad28d711499de5e6e79c355fcebea05387de65d3f6ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the action should apply only to 'all the songs and albums', which refines the initial ambiguous phrase 'artists of I have ever liked'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/addition/addition_01/9c6b86f174b379d0/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "3108cd8f41699cb453edbe0b9b31d0540c4027381e5123601a76ac16cca96b06", "seconds": 14.04, "spoken_utterances": [ "Follow artists of um... I have ever liked on uh... Spotify. ", "Oh, I forgot to mention, this should apply only to um, let's see... all the songs and albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('Follow artists of um, I have ever liked on uh Spotify'), but the speaker adds the missing constraint 'all the songs and albums' after a pause. This addition completes the original instruction, making the final goal match the reference. The type is 'addition', and the final combined task matches the original instruction exactly.", "heard_transcript": "Follow artists of um, I have ever liked on uh Spotify. Oh, I forgot to mention, this should apply only to um, let's see, all the songs and albums." }, "candidate_id": "wav_75da393e2c53e3d3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable addition. The initial request is stated with slight hesitation ('um'), followed by a pause and a plausible, smoothly integrated restriction ('Oh, I forgot to mention...'). The prosody is consistent, and the correction feels spontaneous rather than forced. The pacing and intonation suggest genuine thought processing." } ] }, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists I have ever liked on Spotify. Change the media scope to 'songs and albums' instead of 'tracks and playlists'.", "long_pause_offsets": [ 77 ], "source_sha256": "de075b75bc2ce0dddb63a34ceead7edb74302fae3a2ddd597f7e23ec392e9573", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the media scope from 'tracks and playlists' to 'songs and albums', aligning with the original instruction. The action and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/revision/revision_01_media_scope/5bb2d20f3667646a/speech.wav", "wer": 0.12121212121212122, "wav_sha256": "afde0a8540fbf0eb10f0e50c5bc3a1027c5547dcedb40d5a5d7e1993d7fd6bd0", "seconds": 17.0, "spoken_utterances": [ "Follow artists of ah... all the tracks and playlists uh, let me think... I have ever liked on Spotify. ", "Change the media scope to 'um... songs and albums' instead of 'tracks and playlists'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but then corrects to 'songs and albums' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction, so the task passes.", "heard_transcript": "Follow artists of uh all the tracks and playlists uh let me think. I have ever liked on Spotify. Change the media scope to um songs and albums instead of tracks and playlists." }, "candidate_id": "wav_9706b1aa911bc8e3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses a realistic 'uh' and 'let me think' filler, followed by a clear, self-corrected revision ('change the media scope to...') that sounds like a spontaneous thought process. The prosody is consistent, and the correction is smoothly integrated into the ongoing utterance without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e166825667fc670d3b0c125f4b26a4cb54e277fd2e400304735f8154b403c324", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/cancellation/cancel_009/9815681b3ce210d6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "8af6acdd2f76ca3296ac665843534a33c3b3234d879ac6fbf68ff1cb7c5a2ea2", "seconds": 12.56, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately follows with 'Hang on, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "560f71b252045b27376c661480bbf2466658d12ea1b247d2f15e87fca9dd8171", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/ambiguous_halt/halt_030/e70ca37c4b97c33c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "730a17513c69eeb929fe959156b7ab6e58048166849d9180673860e13d649443", "seconds": 11.52, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify.", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause and reconsider the task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or execution. The initial request is not completed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4477e664b62b203b430e648c733322275d027a513b9a99ba971bc1c0bb9bcdff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/filler/original/2161fbf520479464/speech.wav", "wer": 0.125, "wav_sha256": "0f5c8ea7d92fa9c8862f424c69a23edfa32d984e3a8d586f6266c65662f9760b", "seconds": 8.28, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'uh' as a filler, then proceeds to deliver the full original instruction without any correction, cancellation, or pause. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved—sorry, I meant liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8a6078a6c0167c65d7136f1b45233af212bf6f365e0f462344ee7b8074cdd287", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked'. The final intended action is to follow artists of songs and albums the user has liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/false_start/false_start_02_interaction_type/4fa6b6022e00d842/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5b4ecdd1f2036351ed49e27da76f68b6841de5d81e51a0b5b5ab58eae1d3b3bc", "seconds": 10.88, "spoken_utterances": [ "Follow artists of ... all the songs and albums I have ever saved—hmmm... sorry, I meant ... liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked' after a brief pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('saved') is discarded. The final instruction matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Follow artists of all the songs and albums I have ever saved. Hmm. Sorry, I meant liked on Spotify." }, "candidate_id": "wav_4d4aa5057661a8d4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'saved' to 'liked' with a plausible 'sorry I meant' phrase. The correction is smoothly integrated, and the prosody remains consistent throughout. The pause before the correction is brief and realistic, suggesting genuine hesitation rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Follow artists of all the songs and albums on Spotify. I didn't mention this, but only follow artists of songs and albums I have ever liked.", "long_pause_offsets": [ 55 ], "source_sha256": "99af53a8d4796ca49a46d8f0ffb91343131252a38bac9cecb456f6c5dfd4b3d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'I have ever liked', which narrows the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/addition/addition_02/405e8c8bc9c04db3/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "22efba7153651ff273886d1a993cd6b1f21c6a94daf86ce4f943b788c50b9d4c", "seconds": 12.76, "spoken_utterances": [ "Follow artists of ah... all the songs and albums on ... Spotify. ", "I didn't mention this, but only follow artists of songs and albums ... I have ever liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow artists of all the songs and albums on Spotify.' This is followed by a longer pause and an addition: 'I didn't mention this, but only follow artists of songs and albums I have ever liked.' This addition narrows the scope of the original request by specifying that only artists of songs and albums the user has ever liked should be followed. The final intent is a combination of the original instruction and the added constraint, which matches the reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Follow artists of uh all the songs and albums on Spotify. I didn't mention this, but only follow artists of songs and albums I have ever liked." }, "candidate_id": "wav_968c16b2a58f708b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved on Spotify. Wait, I meant 'liked' rather than 'saved'.", "long_pause_offsets": [ 73 ], "source_sha256": "8632071240fa4ba1cd7330bc6c75e1932ce3961a8b09c8df7d6db7d7c107005a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'saved' to 'liked', preserving the app, scope (all songs and albums), and action (follow artists)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/revision/revision_02_interaction_type/df0940570eee8af1/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "37d9af1581194a3f0198620d5b01bc139d5aa61ad6e0315636453690bd4d0dc9", "seconds": 11.4, "spoken_utterances": [ "Follow artists of ... all the songs and albums uh... I have ever saved on Spotify. ", "Wait, I meant '... liked' rather than 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as intended.", "heard_transcript": "Follow artists of all the songs and albums uh I have ever saved on Spotify. Wait, I meant liked rather than saved." }, "candidate_id": "wav_5fb57da87060aeb6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural hesitation ('uh') and a self-repair ('Wait, I meant...') that sounds spontaneous and integrated into the flow. The correction is immediate and the intonation reflects a genuine moment of recollection. The pacing is conversational, and the overall delivery is fluent and continuous without robotic pauses or distortions." } ] }, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_017", "hide_variant_original": true, "source_sha256": "bf262404db2655c530c6d83a703afdda5a457b7078782b7d956bd8dc29b94d38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/cancellation/cancel_017/7490e09f62877755/speech.wav", "wer": 0.0, "wav_sha256": "7bb4e8929221b040c4d1fffd1d1aac73476b27387d630603bfba117bb759a8fc", "seconds": 10.32, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify.", "Wait, I don't want that done anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request, then clearly revokes it by stating 'Wait, I don't want that done anymore. Cancel it.' This constitutes a direct cancellation of the task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify. Wait, I don't want that done anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "1c111b8f68666f8433ab2b4db6f49922acbede966415ca9dece2ceca0a56b7f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/ambiguous_halt/halt_024/41dc5bf6c17b550a/speech.wav", "wer": 0.0, "wav_sha256": "6ee230b1bb179b92f9a2050e354fbdc2f16cfafa4865f3f73fb0885d2c17be93", "seconds": 9.4, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e59887305786f4274ca9bde3d3898c173efda0744eaa633ff1e85870014e7ad3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/filler/original/abfbc219f8463c5e/speech.wav", "wer": 0.0, "wav_sha256": "c47f3c64b11951efcbca60f55a15486d832c4296e3e3d35ae8d861092aaa70da", "seconds": 5.68, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the filler type allows for minor hesitations or pauses that do not change the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved—I meant liked on Spotify.", "long_pause_offsets": [], "source_sha256": "c943377f239c95be0a4ca7cc2a5f82c61291cb2c65a7b7cc5ba675ebd35417bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended action is to follow artists of liked content, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/false_start/false_start_02_interaction_type/8974bb6b5ac391a9/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "98e97ab4ac626df71364a59cbd8f818f462b6ff4e2c25f94149023b4b244526f", "seconds": 13.12, "spoken_utterances": [ "Follow artists of uh... all the songs and albums uh, let's see... I have ever saved—uh... I meant liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked'. According to the 'false_start' type criterion, the final corrected value ('liked') should be applied, and the original incorrect value ('saved') should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Follow artists of uh all the songs and albums uh let's see, I have ever saved uh, I meant liked on Spotify." }, "candidate_id": "wav_e584d1d80bbade13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, uses a plausible filler ('uh') to pause, then self-corrects 'saved' to 'liked' with a smooth, spontaneous-sounding repair. The intonation and rhythm are consistent with someone formulating a thought aloud, and the correction feels integrated into the sentence rather than tacked on. The pacing and prosody are very realistic." } ] }, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Follow artists of all the songs and albums on Spotify. I didn't mention this, but it's only for the ones I have ever liked.", "long_pause_offsets": [ 55 ], "source_sha256": "1b7756f0719335ac203c257aeefb890e1c084d63125a96b9128ac2cdb4ca2e8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'I have ever liked', which narrows the scope of the initial broad request to only those songs and albums that were liked. The action, app, and target remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/addition/addition_02/5be4112ecb5282c1/speech.wav", "wer": 0.0625, "wav_sha256": "6fe2cdc1ada3678ebe976308befd73fff8485065578915bbbf71b6e3c1b6fd4b", "seconds": 17.52, "spoken_utterances": [ "Follow artists of um, let's see... all the songs and albums on ah... Spotify. ", "I didn't mention this, but it's only for the ones uh, let me think... I have ever liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to follow artists of all songs and albums on Spotify. After a pause and a self-correction, they add the constraint 'but it's only for the ones I have ever liked'. This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Follow artists of um, let's see, all the songs and albums on uh, Spotify. I didn't mention this, but it's only for the ones uh, let me think, I have ever liked." }, "candidate_id": "wav_c1e3816d5729540e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and varied prosody, with believable pauses and fillers ('um', 'let me think') that align with spontaneous speech. The addition of 'I didn't mention this, but it's only for the ones...' is a clear, pragmatic follow-up that sounds like a remembered detail being added mid-thought. The intonation and rhythm feel organic, not rehearsed or robotic." } ] }, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "4318ed46b578d6e45566504ba8a7a1abe49a3bee04ce61cb297190d47a4bacbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the rest of the instruction including the scope of 'all the songs and albums I have ever liked'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/revision/revision_03_app/9add0b9e221a07c3/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "efe9474c6310c17edbdd311f36fbe676dec56b6401faaed9061ee6d9103f5cc4", "seconds": 17.08, "spoken_utterances": [ "Follow artists of all the songs and albums let me think... I have ever liked on uh... Apple Music. ", "Wait, the app should be hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to follow artists on Spotify, which matches the original instruction. The correction is clear and unambiguous, so the task should be executed as revised.", "heard_transcript": "Follow artists of all the songs and albums that me think I have ever liked on uh Apple Music. Wait, the app should be hmm Spotify, not Apple Music." }, "candidate_id": "wav_3aa2feefb4cb52ca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. The speaker begins with a clear request, hesitates with 'let me think', then corrects the platform from 'Apple Music' to 'Spotify' with a believable 'wait' and a self-repair that sounds like a genuine moment of recollection. The prosody is consistent, and the correction is integrated smoothly into the utterance without awkwardness." } ] }, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "9b7df41f219a3471f9618ba995ce645edc28a6227e39259d6d2373a6da559ee5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/cancellation/cancel_028/2569f13bfd06f5cc/speech.wav", "wer": 0.0, "wav_sha256": "70b026f642bff60c719e9fdea166e9b867310f08f320a3db492bc9adae6ecbe0", "seconds": 13.4, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify.", "Oh, wait, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original instruction, then immediately follows with 'Oh wait, disregard that instruction. Don't carry it out,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify. Oh wait, disregard that instruction. Don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "64f8df398973341b0801dcdbf9ce193f227104f02fb15045bc13efcc8a1baf94", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/ambiguous_halt/halt_008/942762b7cc4bf8ac/speech.wav", "wer": 0.0, "wav_sha256": "ebe3c6c35b582899a7e88aca3251153b5245bc714515b896f9c1971df38cc69b", "seconds": 11.08, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second,' which constitutes a clear request to pause or hold. The initial instruction is not canceled, but the final control intent is to pause, matching the 'ambiguous_halt' type criterion. No execution occurs, and the task is not resumed.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "d4e9306_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3eeb1f1012a0b330bffc37a078a54be3aebe5d70aa6839443f1fd3789d1afb33", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/filler/original/16bd6c31aba54c36/speech.wav", "wer": 0.0, "wav_sha256": "32af3d19a77e2046c4b50c72c3de6c6ffad3e0e30ef63573d6d603f9815734d5", "seconds": 7.76, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um', a common filler, but continues with the full original instruction without any correction or change in intent. The final audible content matches the original instruction, with only a minor hesitation that does not alter the task goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of February—sorry, I meant January this year?", "long_pause_offsets": [], "source_sha256": "b0ada947be65758c7b3ee2c765edc7fa4c3900de3eea685490e85fea32f38a10", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'February' but immediately corrects to 'January' using 'sorry, I meant'. The final intended month is January, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/false_start/false_start_02_month_name/92619fd8c0689f67/speech.wav", "wer": 0.1, "wav_sha256": "e6d42569dd1b79a8c8df1e6a35edf18455f00b206eefaf9942a294b1fd59166b", "seconds": 9.48, "spoken_utterances": [ "How many ... venmo friends have I made uh... since the start of February—hmmm... sorry, I meant January this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'February' but immediately self-corrects to 'January this year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many Venmo friends have I made uh since the start of February? Hmm. Sorry, I meant January this year." }, "candidate_id": "wav_b8e68e9a32f1d722", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then self-corrects from 'February' to 'January this year' with a smooth, spontaneous-sounding repair. The intonation and pacing are consistent with a person recalling information, and the correction is integrated into the same utterance without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "dac78d9_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How many friends have I made since the start of January this year? One more thing: I'm asking about Venmo friends.", "long_pause_offsets": [ 67 ], "source_sha256": "ec37d1eb0e90fb289eccba5d371f4981f67a6f69105a52f7f6bccc5fc409e169", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Venmo' to the initial question, which was otherwise identical in scope and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/addition/addition_02/6f2b1fecd33ca0dd/speech.wav", "wer": 0.08, "wav_sha256": "c2e73073bca9a05d18338303193d4fa7cfbf4a31bfc44ab404c27cfc67a65cf5", "seconds": 12.12, "spoken_utterances": [ "How many friends have I made uh, let me think... since the start of January this year? ", "One more thing: I'm asking about ... Venmo friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many friends have I made since the start of January this year?' which is a broad request. The speaker then adds 'I'm asking about Venmo friends' as a clarifying addition. The final intent combines the original time frame with the specific target 'Venmo friends'. The addition narrows the scope of the initial request, and the final goal matches the reference instruction. Therefore, the addition type is satisfied.", "heard_transcript": "How many friends have I made, uh, let me think, since the start of January this year. One more thing, I'm asking about Venmo friends." }, "candidate_id": "wav_722e5f5aa87a8a3c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant thought process ('uh, let me think') and a plausible self-correction from 'friends' to 'Venmo friends'. The addition of the time constraint ('since the start of January this year') is smoothly integrated with a conversational tone and appropriate prosody. The pacing and intonation suggest genuine spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "dac78d9_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "How many PayPal friends have I made since the start of January this year? Wait, I meant Venmo, not PayPal.", "long_pause_offsets": [ 74 ], "source_sha256": "69d8944c3131fc3aad1d1acb4f3d17f3ea155270201a194f4c9b517297a82eb2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the rest of the query including the time frame and the action of counting friends." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/revision/revision_01_app/43ad48ff00556372/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6100bc26004a22825b8a6bd80d35f171fc4261fb54c39c6cb6dbf82bbc971028", "seconds": 12.64, "spoken_utterances": [ "How many um... PayPal friends have I made ah... since the start of January this year? ", "Wait, I meant Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to query Venmo friends, which matches the original instruction. All other conditions, including the time frame 'since the start of January this year', are preserved.", "heard_transcript": "How many um, PayPal friends have I made uh, since the start of January this year? Wait, I meant Venmo, not PayPal." }, "candidate_id": "wav_74a0cc5367df3bfb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, plausible request, then self-corrects from 'PayPal' to 'Venmo' with a natural hesitation ('wait, I meant...') and a smooth, conversational tone. The correction is integrated into the flow without awkwardness, and the prosody reflects genuine recall and correction." } ] }, "is_default": true } }, { "task_id": "dac78d9_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "How many PayPal—I should say venmo friends have I made since the start of October last year?", "long_pause_offsets": [], "source_sha256": "142d48e52cb4cb0d5bc54dcddddadf97877cfe4d6e7bd1a88819f0098d2fb1c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the query including the app, the action, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/false_start/false_start_01_app/bd5093a0c798b001/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "c2d5bcf8217f6d93ed9de1647fa4e72f1375f532770709bf500516c3a2376e7f", "seconds": 11.48, "spoken_utterances": [ "How many PayPal—hmmm... I should say uh... venmo friends have I made um, let's see... since the start of October last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say uh Venmo friends'. This is a clear false start where the initial incorrect value is replaced by the correct one. The final intent matches the original instruction exactly, and the control intent is to execute the corrected task.", "heard_transcript": "How many PayPal, I should say uh Venmo friends have I made um let's see, since the start of October last year?" }, "candidate_id": "wav_abd8f09415de9abc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The false start 'PayPal' is immediately and smoothly corrected to 'Venmo' with a believable self-repair. The subsequent 'let's see' filler is a very plausible hesitation before continuing the request, and the overall prosody, pacing, and intonation sound like a genuine person formulating a question in real time." } ] }, "is_default": true } }, { "task_id": "dac78d9_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of September last year? Make that since the start of September last year, not October.", "long_pause_offsets": [ 75 ], "source_sha256": "90c09785dcaef1bee2c512b44a66e8cc398997043b504defa8061ddeeb967cdc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the month from October to September, while retaining the app (Venmo), the action (counting friends), and the time frame (start of the month last year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/revision/revision_02_month_name/a638667565b21763/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "66bbb8f857a1c41f97604b07a24fcb110eba83c077abe25ac0ba6b7030fc7e6c", "seconds": 12.92, "spoken_utterances": [ "How many venmo friends have I made ... since the start of September last year? ", "Make that since the start of um... September last year, not ah... October." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about October, then corrects to September after a pause and self-repair. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final audible intent is to query September, which matches the revised goal. All other constraints remain unchanged.", "heard_transcript": "How many Venmo friends have I made since the start of September last year? Make that since the start of um uh September last year, not uh October." }, "candidate_id": "wav_3e3d87982ce49603", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question. After a brief pause and filler ('um'), they self-correct 'October' to 'September' with a plausible hesitation and a smooth, continuous flow. The correction is integrated into the same utterance without awkwardness, and the prosody reflects genuine recall and adjustment." } ] }, "is_default": true } }, { "task_id": "dac78d9_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "3ae5d8197bcee1f280b27768978e24208a009b554936e62d03bc992756d8fb16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/cancellation/cancel_027/4ab7b424e7deee8a/speech.wav", "wer": 0.0, "wav_sha256": "64573fe3bdd70190f7f17526ff8fc03f3fcac184e0422f53ad30ae5e7ec27576", "seconds": 11.36, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?", "Hang on, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request, then immediately follows with 'Hang on, please disregard my last request,' which is a clear and unambiguous cancellation of the prior task. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How many Venmo friends have I made since the start of October last year? Hang on, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "6647f5842ae3ecaa765e2a6d232f6bce6e779cedde564412e67e50f9b5a62e39", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/ambiguous_halt/halt_011/5d66b0a3592b5888/speech.wav", "wer": 0.0, "wav_sha256": "3414d69cddaa8217bbd656c276147a307dbce179a97c6af0509070888ef819b8", "seconds": 10.2, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled, and the speaker does not resume or modify it. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "How many Venmo friends have I made since the start of October last year? Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ef1db8970b7906047548b633ced4b771863ec68f643e3220f711fcc9f554df42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/filler/original/0bbe7ecc60af8433/speech.wav", "wer": 0.0, "wav_sha256": "86b96e22a6507d5a6e718a87297a4518dbfe3af7ce2b95951ebcb46217fc3b7d", "seconds": 6.8, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full original instruction clearly and without modification. The final intent matches the original instruction exactly, and the 'filler' type allows for such minor hesitations or pauses that do not change the goal. Therefore, the task is executed as intended.", "heard_transcript": "How many Venmo friends have I made since the start of October last year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "How many PayPal—sorry, I meant venmo friends have I made since the start of March this year?", "long_pause_offsets": [], "source_sha256": "ab0e1b9c05cee1517dae909e63bac1a16e28ee5894afc233df767be6f52b60fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the query including the app, the action, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/false_start/false_start_01_app/f3b098483c1a0bf7/speech.wav", "wer": 0.0, "wav_sha256": "e5f2bcba82d4ecf7e4042b1b7608d57baf841c0b0f2760b40302944a8877b640", "seconds": 9.76, "spoken_utterances": [ "How many PayPal—uh... sorry, I meant ... venmo friends have I made ... since the start of March this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'sorry, I meant'. The final audible intent is to ask about Venmo friends, which matches the original instruction. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "How many PayPal uh, sorry, I meant Venmo friends have I made since the start of March this year?" }, "candidate_id": "wav_25949ea26ab2c57b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a different service (PayPal), immediately self-corrects with 'sorry, I meant Venmo' in a conversational tone, and then continues with the full, correct request. The false start and correction are seamlessly integrated, and the prosody remains consistent and human-like throughout." } ] }, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many venmo friends have I made? Oh, I forgot to mention, I want to know about the period since the start of March this year.", "long_pause_offsets": [ 36 ], "source_sha256": "0ba9b36ec2af38baacdb267f1f26d7e2a95cb5f1779d4216bfa6b1c7c51f38e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint to the initial question, combining both parts into a single, complete query that matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/addition/addition_01/ec62921b16fbce14/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "25ea048b27e4836fe08eb783efbd49c9d4335a314ebe36e9b757942a21ff3065", "seconds": 15.04, "spoken_utterances": [ "How many ah... venmo friends have I made? ", "Oh, I forgot to mention, I want to know about the period uh, let's see... since the start of March this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, asking only 'How many Venmo friends have I made?'. After a pause and self-correction, the speaker adds the missing time constraint 'since the start of March this year'. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, so the task should be executed.", "heard_transcript": "How many uh Venmo friends have I made? Oh, I forgot to mention, I want to know about the period uh let's see, since the start of March this year." }, "candidate_id": "wav_b489bdac48058377", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding 'uh' before stating the core question. The phrasing 'Oh, I forgot to mention' is a very common and realistic way to add a detail after an initial request. The subsequent addition of the time period is delivered with a believable 'let's see' pause and a smooth, conversational tone, maintaining a consistent rhythm and prosody throughout. The overall flow feels like a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of April this year? I misspoke about the month; it should be March, not April.", "long_pause_offsets": [ 71 ], "source_sha256": "733087b94fabf6629a00fdc12a66bb9d3c852ae38ded0840fb81e26a5d129ccf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the month from April to March, preserving the app (Venmo), the action (counting friends), and the time frame (since the start of the month this year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/revision/revision_02_month_name/3459b46812a41592/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "bdc82806c71156c9e9001e0315cbc2e00e3810bbc5b5f4cc0aa6bf6cd5b257da", "seconds": 13.88, "spoken_utterances": [ "How many uh... venmo friends have I made ... since the start of April this year? ", "I misspoke about the month; it should be um... March, not April." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'April' but immediately corrects themselves to 'March' after a pause, which aligns with the revision type criterion. The final audible intent is to query Venmo friends since the start of March, matching the original instruction. The correction is clear and applied to the earlier instruction.", "heard_transcript": "How many uh Venmo friends have I made since the start of April this year? I misspoke about the month, it should be um March, not April." }, "candidate_id": "wav_d748ae9259f87790", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural hesitation ('uh') and a plausible self-correction of the month. The revision ('it should be um March not April') is spoken with a believable change-of-mind intonation, and the flow from the initial request to the correction is smooth and continuous, without any jarring edits or unnatural pauses. The prosody and rhythm are consistent with spontaneous speech." } ] }, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "4b1dc3258357742a82f413f9c86e73e766b3d153866dd9235e7ab55c9baf602b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/cancellation/cancel_010/c313c10d0a0d3524/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "846e45f46bf09779dca9096e9a0c81dfe797f3ceb69430c988365b3fb534a7e9", "seconds": 12.48, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a cancellation directive ('cancel the request I just made') spoken immediately after the initial request. This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final intent is to cancel, not to execute the original request.", "heard_transcript": "How many Venmo friends have I made since the start of March this year? Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "94c7c0d6221ae029f5cd8dd2489d9cb486fb51a1ac173ee6906b3d24598ac877", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/ambiguous_halt/halt_010/6784e29036479009/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "0ac257b74e49849319f1b38d6ccb2e3aa6863936eec0d371d5b094ae4649f46b", "seconds": 11.48, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on a second,' which is a clear request to pause or wait. This matches the expected behavior for an 'ambiguous_halt' type, where the final interjection is a pause command and not a cancellation or execution. The initial request is not revoked, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "How many Venmo friends have I made since the start of March this year? Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "dac78d9_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4c6334c1909ce3c142f3211e5600425870553dc79d69ab2c215bd8da12527288", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/filler/original/aefbb4e6c00e2f4f/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "3aa3a64d6e63fb84e415fba24c482d479d22a7f59050cf507495b41d55011926", "seconds": 7.88, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker delivers the full original instruction without hesitation, backchanneling, or self-correction. The audio contains no pauses or fillers that alter the task's goal. The final intent matches the original instruction exactly, satisfying the 'filler' type criterion where hesitations are ignored and the remaining task meaning must match the original.", "heard_transcript": "How many Venmo friends have I made since the start of March this year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family—sorry, I meant roommates.", "long_pause_offsets": [], "source_sha256": "0b1cbdde7cd5f686b149fd4e54cd9442af0dae703a580d2a07ab27450b9b835e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'roommates' within the same utterance, preserving the app, time scope, and action. The final intended target is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/false_start/false_start_04_counterparty_group/0f69821f7ce58e61/speech.wav", "wer": 0.05, "wav_sha256": "3793333547910bc8ec61d312b92628227c5893fccfd1ef1acf44c4ca70b2c851", "seconds": 7.8, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing year to and from my family—um... sorry, I meant ... roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'roommates' with 'sorry, I meant roommates.' This is a false start, and the final audible intent is the corrected version. The correction is immediate and unambiguous, so the task should be executed with 'roommates' as the target.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing year to and from my family um, sorry, I meant roommates." }, "candidate_id": "wav_92792fe12b214a07", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a clear false start ('family') followed by a natural self-correction ('sorry, I meant roommates') with a believable tone of mild embarrassment. The initial hesitation and correction are smoothly integrated into the flow, and the speaker's voice remains consistent and conversational throughout. The pacing and intonation suggest spontaneous speech rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Like all the venmo transactions of the ongoing year. Oh, and one more thing: only include transactions to and from my roommates.", "long_pause_offsets": [ 53 ], "source_sha256": "39b8b0ed0baaf0b74b8bf15fe3f0c836d1e7dc36e0be7e90a59ddf00ac4e26cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial broad request, narrowing the scope to only transactions involving roommates. The action, app, and time frame remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/addition/addition_03/dd9469c1667a7422/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "275f1fac746fb4804aab757abbb711921a562b5c96478a8131ce5804bacf70a1", "seconds": 14.6, "spoken_utterances": [ "Like all the uh, let's see... venmo transactions of the ... ongoing year. ", "Oh, and one more thing: only include transactions um... to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Like all the Venmo transactions of the ongoing year'. After a pause and an interjection ('Oh, and one more thing'), the speaker adds a constraint: 'only include transactions to and from my roommates'. This is a clear addition that narrows the scope of the original request. The final intent combines both the original and the added constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Like all the uh, let's see. Venmo transactions of the ongoing year. Oh, and one more thing, only include transactions um, to and from my roommates." }, "candidate_id": "wav_99ce8b540bc3e702", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses a realistic 'uh' filler, a plausible 'let's see' hesitation, and a smooth addition of a new constraint ('only include transactions...') after a brief pause. The prosody is consistent, and the addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to my roommates. Wait, it should be to and from, not only to.", "long_pause_offsets": [ 74 ], "source_sha256": "d49d4b9a3ee52276dc3fea33bcdaeafb25ddc10bb1bf50241b71738b180aeebd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the direction from 'only to' to 'to and from', which matches the original instruction. The app, time scope, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/revision/revision_03_transaction_direction/1b5121e7a26b512e/speech.wav", "wer": 0.0, "wav_sha256": "40fac7461572612a76a4fb407dc05b6eeda0bfe1a2d4084af9e27745c18ce81a", "seconds": 12.4, "spoken_utterances": [ "Like all the venmo transactions of the ongoing year um... only to my roommates. ", "Wait, it should be let's see... to and from, not only to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to my roommates' but then corrects themselves with 'to and from, not only to' after a pause. This is a revision where the final intent is the corrected version. The original instruction matches the final intent, so the task is executed as revised.", "heard_transcript": "Like all the Venmo transactions of the ongoing year, um, only to my roommates. Wait, it should be, let's see, to and from, not only to." }, "candidate_id": "wav_2749405c73e6a624", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, multi-stage revision. The initial request is clear, followed by a believable self-correction ('wait it should be...') with a slight pause and intonation shift. The correction is smoothly integrated into the same utterance, maintaining a consistent rhythm and prosody. The flow feels like a genuine moment of recollection and adjustment." } ] }, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "13efead27c3e722ce3ee4fbae51d25bef351e9fe82c25c766d358ca116d9438b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/cancellation/cancel_023/fa3345bff617e376/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "90958a68ea42b5a7e0b677077bfcb90fc2027356dfde292ade26f0252d45bad4", "seconds": 12.88, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates.", "Hold on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original request, then immediately follows with 'Hold on. Never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly instructs not to execute the request, satisfying the cancellation type criterion.", "heard_transcript": "Like all the um Venmo transactions of the ongoing year uh to and from my roommates. Hold on. Never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "e55edf826bc990612fa96812cf2f14ef2f1304df4a213a3c9bee7d12fc1cb094", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/ambiguous_halt/halt_032/58b73fea5dfef3ec/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "0a2954168ca6b65cfdad60eab7884d4f53bee4e27a8ea5742cab7d516d524312", "seconds": 11.24, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, before you do that,' which is a clear request to hold or pause the task. This is a final interjection that does not cancel the task but asks for a temporary halt. The initial request is fully audible and complete, and the final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Like all the um Venmo transactions of the ongoing year uh to and from my roommates. Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1c7eb58a1f9a9fcf29b23269eacd1ac11cb47284dcf24a71d0f4bb9578062df8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/filler/original/ea5f5165903f4376/speech.wav", "wer": 0.1875, "wav_sha256": "ee917735df8366f2b2309968f24ac6f3b248fb9e3263c8a612db373cdc93a1e2", "seconds": 7.24, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses filler words ('um', 'uh') and pauses but does not change the core request. The final utterance clearly retains the original intent to retrieve Venmo transactions with roommates for the ongoing year. The filler does not alter the goal, so the task should be executed as originally intended.", "heard_transcript": "Like all the um, Venmo transactions of the ongoing year uh, to and from my roommates." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing month to and from my family—I meant friends.", "long_pause_offsets": [], "source_sha256": "f139f4b90f1cb21bcf5ca2fff393261f16000ff41b8bbd9c9f51fd988dee5bff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance, replacing the group of people. The app, time frame, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/false_start/false_start_04_counterparty_group/646b8d3ad00a95c3/speech.wav", "wer": 0.15, "wav_sha256": "090e7dad07aeaaf936c5e81e64a0099d66715b3357884502e4d2a38d68c990bd", "seconds": 10.68, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing month um... to and from my family—hmmm... I meant friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects with 'I meant friends.' This is a clear false start where the final intended target is 'friends.' The correction is immediate and unambiguous, so the final intent matches the original instruction. The control intent is to execute the corrected task.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing month um to and from my family. I meant friends." }, "candidate_id": "wav_0514e22fc6e29b18", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent false-start pattern. The speaker begins with 'like all the uh Venmo transactions of the ongoing month,' then hesitates with 'um,' and after a brief pause, self-corrects with 'I meant friends' in a way that sounds like a genuine, spontaneous recollection. The prosody is consistent, the correction is immediate and integrated into the flow, and the intonation clearly marks the shift from the initial mistake to the intended meaning." } ] }, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Like all the venmo transactions to and from my friends. Oh, I forgot to mention, this should only apply to the ongoing month.", "long_pause_offsets": [ 56 ], "source_sha256": "552a291866bb1c8bf7d89b6935aaf76dc650e952dc026cb2b9e9f601f866ec2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction that narrows the scope of the initial request. The final goal combines the original action and target with the new temporal constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/addition/addition_01/e262f2270ff6b935/speech.wav", "wer": 0.2, "wav_sha256": "fdd66b82f897478f801141bf31208143cf5454f36e4642e993d5251f96ed5d91", "seconds": 13.12, "spoken_utterances": [ "Like all the ah... venmo transactions to and from my friends. ", "Oh, I forgot to mention, this should only apply to the um... ongoing month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. The speaker then adds a clarifying constraint ('only apply to the ongoing months') after a pause, which is a valid addition. The final intent combines the original request with the new condition, satisfying the addition type criterion.", "heard_transcript": "Like all the Venmo transactions to and from my friends. Oh, I forgot to mention, this should only apply to the um ongoing months." }, "candidate_id": "wav_33061a8d20b04e68", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent initial request with natural prosody and a believable, smoothly integrated addition. The filler 'uh' and the phrase 'Oh, I forgot to mention' are placed at a realistic point in the utterance, and the correction is seamlessly woven into the flow without awkwardness. The pacing and intonation suggest spontaneous thought rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "Like all the PayPal transactions of the ongoing month to and from my friends. Wait, the transactions should be on Venmo, not PayPal.", "long_pause_offsets": [ 78 ], "source_sha256": "581ec569270ee8d19c0704433dc51834ff415b458ff456304776cb1fda9bce29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the rest of the instruction including the transaction scope, time frame, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/revision/revision_01_app/7987f0197a344937/speech.wav", "wer": 0.0, "wav_sha256": "51c84033e7e42851568c834754a5768c7866af99b002a9a70f5b8bb94a64536a", "seconds": 16.96, "spoken_utterances": [ "Like all the uh, let me think... PayPal transactions of the ongoing month ... to and from my friends. ", "Wait, the transactions should be on ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal transactions' but immediately corrects themselves to 'Venmo not PayPal' after a pause. This is a revision where the final intent is to use Venmo, which matches the original instruction. The correction is clear and the final goal aligns with the reference.", "heard_transcript": "Like all the uh let me think. PayPal transactions of the ongoing month to and from my friends. Wait, the transactions should be on Venmo not PayPal." }, "candidate_id": "wav_fd0e5543d75b7020", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker starts with a clear initial request, uses a believable filler ('uh let me think'), then makes a clear mistake ('PayPal'), immediately self-corrects with a plausible reason ('Wait, the transactions should be on Venmo not PayPal'), and maintains consistent prosody and pacing throughout. The correction is integrated smoothly into the utterance, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "a0a5eb55f2599b984a82cffa7e600b75b9828f2b606f32410ff7ef27f4173009", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/cancellation/cancel_007/bf95c09fcd64bc36/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "d123b87627a4160b68c68a9e24c3e621701e392b4966c5884b8859370e49095a", "seconds": 14.2, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends.", "Oh, wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues the original task and then immediately follows with a clear cancellation command: 'Oh wait, cancel my last request.' This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final interjection is not ambiguous and explicitly cancels the prior instruction.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends. Oh wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "8483fb3e209d3f859370b562238a25ec072d6df71fdc13089a72144dc8faf993", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/ambiguous_halt/halt_027/3deaf7443081a92b/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "525b2902b34ce71c963ef3ee65080ff9a4c4ac44d81473747b22eaa06cb7fd8a", "seconds": 13.68, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, I need a moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "27476db2a76a9a9f8027f3dfbd27bdcee09856fd2915c992bf9c408a6b1cdcca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/filler/original/91e07ba076101837/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "36f90ecada7096246bd6f5220da9456922e89f3e2223c72cc7b5ddb3b6da6af3", "seconds": 9.56, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction, and the filler type allows for minor hesitations or pauses that do not change the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family—I should say coworkers.", "long_pause_offsets": [], "source_sha256": "32b9cdfc91e8655a5312ede11cbe97225e833615c5e2af8ca9c9d5d9870eb4dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'coworkers' within the same utterance, preserving the app, time scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/false_start/false_start_04_counterparty_group/3e3053a8be8a8438/speech.wav", "wer": 0.2, "wav_sha256": "3598b1396345321be46694bdc299f204bc06c3be5329f9132b23daea954c34c9", "seconds": 11.28, "spoken_utterances": [ "Like all the venmo transactions of the uh... ongoing year to and from my ... family—ah... I should say coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'familia' but immediately self-corrects to 'coworkers'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is discarded. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my famlia. I should say coworkers." }, "candidate_id": "wav_e37e21e8fc7363b4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects 'family' to 'coworkers' with a smooth, spontaneous-sounding false start. The prosody is consistent, and the correction is integrated into the flow without awkwardness or hesitation." } ] }, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_03", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from. I left out a detail: only like transactions with my coworkers.", "long_pause_offsets": [ 65 ], "source_sha256": "99090b6a24eeff27af4b0e5e574bea513dfd48a562ba143a4fc8e5d0dfa4d526", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'with my coworkers', which narrows the initial broad request to the exact scope specified in the original instruction. The action, app, time frame, and direction are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/addition/addition_03/72b7845420453d70/speech.wav", "wer": 0.12, "wav_sha256": "128da53e13b651f2c115b306652dfd176783c54ca38891c16fbacdec63a1b4d5", "seconds": 11.48, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing year ah... to and from. ", "I left out a detail: only like transactions with my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Like all the Venmo transactions of the ongoing year to and from my coworkers.' The speaker then adds a clarifying detail: 'Only like transactions with my coworkers.' This is a clear addition that narrows the scope of the original request. The final intent combines the original instruction with the added constraint, resulting in a more specific but coherent task. The addition is spoken after a pause and is not a revision or cancellation. Therefore, the task is executed with the combined intent.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing year uh to and from. I left out a detail. Only like transactions with my coworkers." }, "candidate_id": "wav_83dfeea6bbdd2fbc", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to my coworkers. Wait, the transactions should be to and from, not only to.", "long_pause_offsets": [ 74 ], "source_sha256": "a5ea011d16d6beaf2c3fed61433f44ac34ac2cc66f650a3957f19fc680c5be13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the restriction from 'only to' to 'to and from', restoring the original bidirectional scope. The app, time frame, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/revision/revision_03_transaction_direction/c8bd037d3e525679/speech.wav", "wer": 0.04, "wav_sha256": "8892f5eee83e7ff7a0411b3593a1f00bdc74148e0bf1fee638e4352668071060", "seconds": 14.16, "spoken_utterances": [ "Like all the venmo transactions of the ongoing year uh... only to ... my coworkers. ", "Wait, the transactions should be ... to and from, not only to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to my coworkers' but then corrects themselves with 'Wait, the transactions should be to and from not only to.' This is a clear revision where the final intent is to include both 'to and from' as specified in the original instruction. The revision applies to the earlier instruction, and the final goal matches the original instruction exactly.", "heard_transcript": "Like all the Venmo transactions of the ongoing year, uh, only to my coworkers. Wait, the transactions should be to and from not only to." }, "candidate_id": "wav_9ab1bfcc747812de", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'uh' and 'wait,' followed by a self-correction that smoothly revises the scope from 'only to' to 'to and from.' The intonation and pacing reflect genuine reconsideration, and the correction is integrated into the same utterance without awkwardness. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "db6abbef4a8ae6a7dabae9063d80f97bb7a8d15b62636cd1ddcdf0c4361ce6c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/cancellation/cancel_002/a8237ee8edf147a2/speech.wav", "wer": 0.0, "wav_sha256": "9a0cdf3dd6a5c2cfd0fb0dd7a6f1ccba8bbc813bae0e40933b6c1d15e2d06eb8", "seconds": 10.52, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers.", "Hold on, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hold on, cancel the whole task,' which is a direct and unambiguous cancellation of the entire preceding instruction. The initial request is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers. Hold on, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "a6b49f84292dc9e11bc9dac7aa0a1d8a9aafe262d24fd13018ffe55e33e979dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/ambiguous_halt/halt_034/1193accd17f14299/speech.wav", "wer": 0.0, "wav_sha256": "ac637a3aa981ed12cb2e1ed3069695ddfbf7472f4a9306582c40eb230ea0d45f", "seconds": 10.0, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking...', which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or execution. The initial request is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers. Wait, I was just thinking..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "df61dc5_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ceaa11393dd0b6a3e77c4f89d6faf65780de303749f981dffbabfa34c4a46695", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/filler/original/451df7f59f841b02/speech.wav", "wer": 0.0, "wav_sha256": "d85ac310211583fd4f0c8c1def91134c0a094ee54171ee954cb5c4520f145ede", "seconds": 6.04, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker completes the original instruction without any change in meaning or intent. The 'filler' type allows for such natural speech patterns, and the final audible task matches the original instruction.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Add all spotify-recommended classical songs released in last year—no, that should be this year to a new \"Spotify Recommended Songs\" playlist.", "long_pause_offsets": [], "source_sha256": "62aad9cd8e00e90a806396f89a6bfaabe61a091fe53fef691d9610c6725ed8bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'last year' with 'this year'; all other conditions including genre, recommendation source, playlist creation, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/false_start/false_start_03_time_reference/abb0e2da3cf1a23e/speech.wav", "wer": 0.12, "wav_sha256": "e3b6ec6d5bd43f1b31dc6d804e79c826bbf70d9062f8f17c3efbeaf80180b12d", "seconds": 12.8, "spoken_utterances": [ "Add all hmmm... spotify-recommended classical songs released in last year—um... no, that should be ... this year to a new \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference instruction in action, target, and constraint.", "heard_transcript": "Add all um, Spotify recommended classical songs released in last year. Um, no, that should be this year to a new Spotify recommended songs playlist." }, "candidate_id": "wav_d07d9c31f192cb88", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Spotify recommended classical songs released in last year', then self-corrects with 'No, that should be this year'—a believable and smoothly integrated revision. The intonation and pacing of the correction sound spontaneous, and the pause before the correction is appropriately placed. The overall prosody is consistent and conversational, with no signs of robotic delivery or unnatural pauses." } ] }, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Add all spotify-recommended songs released in this year to a new \"Spotify Recommended Songs\" playlist. I left out a detail: only add classical songs.", "long_pause_offsets": [ 103 ], "source_sha256": "bc272aa82e0322b97fb53fd834ea2339b5a9f38242d05c0d5d9dba34b45c7576", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing genre restriction 'classical songs' to the initial broad request for 'spotify-recommended songs', preserving the app, release year, and playlist creation intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/addition/addition_02/7a0ea74067348125/speech.wav", "wer": 0.03125, "wav_sha256": "64cf5b20239935ac4d2b3dc6d16d1b3f62420bdf30555334f827a41554863a60", "seconds": 16.08, "spoken_utterances": [ "Add all uh, let's see... spotify-recommended songs released in this year to a new um... \"Spotify Recommended Songs\" playlist. ", "I left out a detail: only add let me think... classical songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is broad, requesting all Spotify-recommended songs from this year. After a pause, the speaker adds a clarifying constraint: 'only add classical songs'. This addition narrows the scope of the original request, and the final intent is to execute the original task with the added constraint. The addition is clear and directly modifies the initial instruction without introducing new conflicting goals.", "heard_transcript": "Add all uh let's see, Spotify recommended songs released in this year to a new um, Spotify recommended songs playlist. I left out a detail, only add, let me think, classical songs." }, "candidate_id": "wav_d2b1254f07694c75", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('uh', 'let's see', 'um') and a believable hesitation ('I left out a detail') before adding a specific restriction ('only add classical songs'). The self-correction is integrated smoothly into the flow, with a pause and 'let me think' that sounds like genuine recollection. The prosody is conversational and the pacing includes realistic pauses, making the addition feel spontaneous." } ] }, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_time_reference", "proposed_instruction": "Add all spotify-recommended classical songs released in last year to a new \"Spotify Recommended Songs\" playlist. Make that songs released this year, not last year.", "long_pause_offsets": [ 113 ], "source_sha256": "f904d2d8d30fbc4d60e0804cf53adc53fc6ee3ddf0097fd1dd353748d732be5e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the release year from 'last year' to 'this year', preserving the app, genre, action, and playlist name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/revision/revision_03_time_reference/b8c666c1545ace5c/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "fc62c3322a70b4d418e873f181889b7749004acdc456383179984a825609a533", "seconds": 14.72, "spoken_utterances": [ "Add all spotify-recommended classical songs released in ... last year to a new ... \"Spotify Recommended Songs\" playlist. ", "Make that songs released uh... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects to 'this year' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the playlist name and genre. Therefore, the task should be executed as revised.", "heard_transcript": "Add all Spotify recommended classical songs released in last year to a new Spotify recommended songs playlist. Make that songs released uh this year, not last year." }, "candidate_id": "wav_e0c939efc81909c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant request, then corrects the year from 'last year' to 'this year' with a smooth, self-repairing intonation. The pause before the correction is brief and plausible, and the correction itself is integrated naturally into the sentence without awkwardness. The prosody is consistent, and the speaker sounds like they are formulating and adjusting a request in real time." } ] }, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "3316b2d1803626121d417827294f29a695244c6f2c30abb01da93665f5727799", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/cancellation/cancel_027/bcfb2afde762ac40/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "c5443db00fc06cd2901ea48aa2ce4651a5ba2c0c100eb89ab04e5cee2444c1ae", "seconds": 17.04, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist.", "Oh, wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to add songs to a playlist, then immediately follows with 'Oh wait, please disregard my last request,' which is a clear and unambiguous cancellation of the prior instruction. The final interjection is a direct cancellation command, satisfying the criterion for the 'cancellation' type. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist. Oh wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "f4acd0b227fdb208b38d6b62a78cc71898581978a0eef0a8b782130fe39dcce8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/ambiguous_halt/halt_019/f0a3e881b8d5688b/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "2ae1a841617b9033ddf74288faf6af0c192c0e8300138b11838e2ea690ff0f53", "seconds": 16.0, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the initial request with 'Can you give me a moment?', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not execute or modify the task. The final intent is to halt execution temporarily, not to cancel or proceed.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "65f40a852549fb7fbe8bcc5e95cd4cfa12e57e5b8d38976e06e641fab48e2235", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/filler/original/d184f4512e14b67d/speech.wav", "wer": 0.05, "wav_sha256": "f654b804a2fe68d48107a246f582fffa4cefa60a612136b8c711fd24d8df6b48", "seconds": 12.4, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh') but does not change the core request. The final intent matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "Add all Apple Music—I should say spotify-recommended R&B songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist.", "long_pause_offsets": [], "source_sha256": "5622f340535a039de3ba65fdb3969ee2c7afcb64b4d3b8462a32b4d549bf898f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'spotify-recommended' within the same utterance, preserving the genre, release timeframe, playlist name, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/false_start/false_start_01_app/378fe263f5833ac0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "296fae5c8735280b87b292bd1e9c6e34415d146a84cbc3b9858f886ed1a70d4b", "seconds": 14.2, "spoken_utterances": [ "Add all ... Apple Music—um... I should say spotify-recommended R&B songs released in ... this or last year to a new \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('Apple Music') is discarded. The rest of the instruction, including the genre (R&B), release timeframe ('this or last year'), and playlist name ('Spotify R&B Recommendations'), is preserved and matches the original instruction. Therefore, the final audible intent is correct and complete.", "heard_transcript": "Add all Apple Music um, I should say Spotify recommended R&B songs released in this or last year to a new Spotify R&B Recommendations playlist." }, "candidate_id": "wav_7e4fcc6b9829ce60", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural false start with 'Apple Music' followed by a self-correction to 'Spotify recommended R&B songs'. The correction is smooth and the prosody remains consistent. The speaker's voice is clear, and the pacing feels like a genuine moment of recall. The flow from initial error to correction to completion is the most seamless and realistic among the four." } ] }, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "Add all spotify-recommended R&B songs to a new \"Spotify R&B Recommendations\" playlist. I left out a detail: only include songs released in this or last year.", "long_pause_offsets": [ 87 ], "source_sha256": "a8e523b2683e86f71f2b95f4de12bed05e7cd0a450edb973418c6027380affec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing release-year constraint, which complements the initial instruction without contradicting any other requirement. The app, genre, playlist name, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/addition/addition_02/b94a627fa9f31c80/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "fdd9fe1db9bc72ffa9b5696e89ebeccb6421643272ae955b25fb94e2b043f6df", "seconds": 14.32, "spoken_utterances": [ "Add all spotify-recommended ... R&B songs to a new uh... \"Spotify R&B Recommendations\" playlist. ", "I left out a detail: only include songs released in this or last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Add all Spotify recommended R&B songs to a new Spotify R&B recommendations playlist.' The speaker then adds a constraint: 'only include songs released in this or last year.' This is a clear addition of a new condition to the original task. The final intent combines the original action with the added constraint, matching the reference instruction. The type is 'addition', and the final goal is fully specified and executable.", "heard_transcript": "Add all Spotify recommended R&B songs to a new uh Spotify R&B recommendations playlist. I left out a detail, only include songs released in this or last year." }, "candidate_id": "wav_fab53e4643ac8464", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this or last year to a new \"R&B Hits 2023\" playlist. Use \"Spotify R&B Recommendations\" instead of \"R&B Hits 2023\".", "long_pause_offsets": [ 103 ], "source_sha256": "77e32f51d5e748ec140d9401856970cb566cdb93dd4c8f438813247354696798", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the playlist name from \"R&B Hits two thousand twenty-three\" to \"Spotify R&B Recommendations\" while preserving all other conditions: genre (R&B), source (Spotify-recommended), release timeframe (this or last year), and the creation of a new playlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/revision/revision_04_playlist_name/3f4f401d2269d936/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "9a7e309d74755099d2669ec1e2d0dfb1b4ee11db0db9b4875c20bb09c2c8e13e", "seconds": 20.96, "spoken_utterances": [ "Add all spotify-recommended um... R&B songs released in this or last year to a new um, let's see... \"R&B Hits two thousand twenty-three\" playlist. ", "Use \"um, let me think... Spotify R&B Recommendations\" instead of \"R&B Hits two thousand twenty-three\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B hits 2023' as the playlist name but then corrects it to 'Spotify R&B Recommendations' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction with the corrected playlist name.", "heard_transcript": "Add all Spotify recommended um R&B songs released in this or last year to a new um let's see, R&B hits 2023 playlist. Use um let me think, Spotify R&B Recommendations instead of R&B hits 2023." }, "candidate_id": "wav_675ccf432f756ab4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses multiple fillers ('um', 'let me think') at plausible points, with a clear self-correction from 'R&B hits 2023' to 'Spotify R&B recommendations'. The intonation and rhythm are consistent with someone formulating a request in real time, and the correction is smoothly integrated into the sentence without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "fc76128583693cc0d5a0327648ab349f24637601cd844c22c683a9997bc17af2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/cancellation/cancel_003/3df2e8f97093862c/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "ae4f9fad478395bc7cbfc8f5db8058435d9c4308e9f269846097e6eefc9b028d", "seconds": 15.88, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist.", "Oh, wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The recording ends with the explicit cancellation request 'Oh wait, please cancel that.' This clearly revokes the previously stated task. The cancellation is not ambiguous and applies to the entire task as originally described. The final intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist. Oh wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "62e0bd8112f3a1944f4b534e60b3e35f3f69fdc2036c306318b82ffd547582d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/ambiguous_halt/halt_012/99e2eb93a153269e/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "37975659926a24570c1d1a9424bac1fec20c9ca90e906ff4c8f83ff41603f4c8", "seconds": 15.0, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or change its goal. The initial request is fully audible and complete, and the speaker does not attempt to revise, add to, or cancel it. Therefore, the final intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "06a859e365dca15a5d1bed550ba0c72734e39546586b571a8c458c6dd1e19b16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/filler/original/7220793e4131b5b9/speech.wav", "wer": 0.08, "wav_sha256": "a37c53db4e83ff3969753caa0e30f26cfc1d1eb2e64e0dad8fa0ef9dfb580486", "seconds": 11.96, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see', 'um') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or modifications. The fillers are typical hesitation markers and do not alter the task goal.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_genre", "proposed_instruction": "Add all spotify-recommended hip-hop—or rather, R&B songs released in this year to a new \"R&B Recommendation\" playlist.", "long_pause_offsets": [], "source_sha256": "2263d6dab590e6760cd2fe73c7cae5e0db4554a9e0cff9aca35ae73bab31e120", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'hip-hop' but immediately corrects to 'R&B' within the same utterance. The final genre is R&B, matching the original instruction. All other elements—source (Spotify-recommended), release year condition, and playlist creation—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/false_start/false_start_02_genre/677e7a3bcdd2f227/speech.wav", "wer": 0.12, "wav_sha256": "bf0be965fc3a28e73c2d5097ece0c36420e0f7ceb2458d3472a161e1e48bee9a", "seconds": 14.48, "spoken_utterances": [ "Add all spotify-recommended hip-hop—um... or rather, R&B songs released in this year hmmm... to a new uh... \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hip hop' but immediately self-corrects to 'R&B' with 'or rather'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion. All other constraints (Spotify-recommended, released in this year, new playlist named 'R&B Recommendation') are preserved.", "heard_transcript": "Add all Spotify recommended hip hop um, or rather R&B songs released in this year to a new uh R&B recommendation playlist." }, "candidate_id": "wav_f84c66affc58eb13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'hip-hop', hesitates with 'um', then self-corrects to 'or rather, R&B' in a way that sounds spontaneous and conversational. The prosody is consistent, with a natural rise and fall in intonation, and the correction flows seamlessly into the rest of the sentence without awkward pauses or robotic phrasing. The filler 'um' is placed at a typical point of hesitation, and the overall rhythm feels like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Add all R&B songs released in this year to a new \"R&B Recommendation\" playlist. Oh, I forgot to mention, only add songs that are Spotify-recommended.", "long_pause_offsets": [ 80 ], "source_sha256": "a3def8469d1bb5518801bedebd63710f2fd2310d174a48f292b32e035c9a38ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the songs must be Spotify-recommended, which is consistent with the original instruction and narrows the scope without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/addition/addition_01/aa537a28fabb596d/speech.wav", "wer": 0.03125, "wav_sha256": "c365f45fda70e4fb5dd1291462292cda91ea361977c1abd351f9155ee0da529e", "seconds": 18.08, "spoken_utterances": [ "Add all um... R&B songs released in this year ... to a new \"R&B Recommendation\" playlist. ", "Oh, I forgot to mention, only add songs that are let me think... Spotify-recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Add all R&B songs released in this year to a new R&B recommendation playlist.' The speaker then adds a constraint: 'only add songs that are Spotify recommended.' This is a clear addition of a new condition to the original task. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Add all um R&B songs released in this year to a new R&B recommendation playlist. Oh, I forgot to mention, only add songs that are let me think, Spotify recommended." }, "candidate_id": "wav_d6a85ee6536df991", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural hesitation and self-correction. The initial request is clear, followed by a believable 'oh I forgot to mention' revision that adds a constraint. The filler 'let me think' is placed naturally before the final detail, and the prosody maintains a conversational rhythm without robotic pauses or unnatural word boundaries." } ] }, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this year to a new \"R&B Hits\" playlist. Change the playlist name to \"R&B Recommendation\" instead of \"R&B Hits\".", "long_pause_offsets": [ 90 ], "source_sha256": "0b64d6ec8ac035318819d97d2225b39e54f340097a4c6291f00668d90a7625ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the playlist name from \"R&B Hits\" to \"R&B Recommendation\"; all other requirements (genre, source, release year, and creation of a new playlist) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/revision/revision_04_playlist_name/5a300fcd9d22f94f/speech.wav", "wer": 0.1875, "wav_sha256": "a785d4f2d2f4878acc83407261c92e3b03fa41de4e2351e1945620931d563a25", "seconds": 17.48, "spoken_utterances": [ "Add all spotify-recommended ... R&B songs released in this year ... to a new um... \"R&B Hits\" playlist. ", "Change the playlist name to \"R&B Recommendation\" instead of \"R&B Hits\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B hits playlist' but immediately revises it to 'R&B recommendation' in the same sentence. The revision is clear and final, and the rest of the instruction matches the original. The final intent is to create a playlist with the corrected name, so the revision criterion is satisfied.", "heard_transcript": "Add all Spotify recommended R&B songs released in this year to a new um R&B hits playlist. Change the playlist name to R&B recommendation instead of R&B hits." }, "candidate_id": "wav_25255deba81ebdb6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then adds a follow-up revision ('Change the playlist name...') that is smoothly integrated into the same utterance. The prosody is consistent, with a believable 'um' filler and a natural-sounding correction. The revision is not abrupt but feels like a spontaneous addition to the original plan, making it sound like a single, coherent thought process." } ] }, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "cfd684fd9c8d276274b5a0c7a66e31452604dfca5e3facd7b3d0f9138cf5932f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/cancellation/cancel_003/8a4e831e7dc006b6/speech.wav", "wer": 0.12, "wav_sha256": "5fd931289bff105538a34a96fad720aaedc3c45a63b5531460f2d297ba5cf2ff", "seconds": 14.28, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist.", "Oh, wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a complete instruction to add songs to a playlist, then immediately corrects themselves by saying 'Oh wait, please cancel that.' This constitutes a clear cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Add all Spotify recommended R&B songs released in um this year uh to a new R&B recommendation playlist. Oh wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "c5ba41fe20eff523630636d242075aecb733f420f7266f8626239e86f24e45db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/ambiguous_halt/halt_007/34b38c9f0033bc9e/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "0710aac2792c53cd068bacc3456e71d87e1306100abbca3ab78ad768664372cb", "seconds": 13.6, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the final control intent is to halt, which satisfies the type-specific criterion.", "heard_transcript": "Add all Spotify-recommended R&B songs released in um this year uh to a new R&B Recommendation playlist. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e3d6c94_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6cc2832dd9873bebd3ff83fb6b51dc768db4630b324eea1ce33a285cd6e49a0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/filler/original/1b68a2ce9e7dcdca/speech.wav", "wer": 0.15, "wav_sha256": "e9671dde50e921e3e0848c03e526cf44fae9f6582470c00fae8692672bfe6188", "seconds": 10.6, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core task. The final intent matches the original instruction. The filler words do not alter the goal, so the task should be executed as intended.", "heard_transcript": "Add all Spotify recommended R&B songs released in um this year uh to a new R&B recommendation playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_thread_priority", "proposed_instruction": "How many priority-2—I meant priority-1 unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "9e56e36bdc4f44db1aa8353a176f06a3232598dd9307d7d37c5e4d8471305806", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'priority-two' to 'priority-one' within the same utterance, preserving the app (Gmail), the object (unread email threads), and the action (counting)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/false_start/false_start_01_thread_priority/9d0a9d893323cd52/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "4bba2a67546648f5f0f89bc56446ce59dfd75600c8794ecd44b6e8603d30cb1f", "seconds": 7.48, "spoken_utterances": [ "How many priority-two—um... I meant ah... priority-one unread email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority two' but immediately self-corrects to 'priority one'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is executed with the correction. The rest of the instruction matches the original.", "heard_transcript": "How many priority two um, I meant uh priority one unread email threads are in my Gmail inbox?" }, "candidate_id": "wav_a19fe8f615a7f8f4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker begins with a clear false start ('priority two'), immediately self-corrects to 'priority one' with a natural-sounding 'I meant', and smoothly continues the full request. The correction is concise, the prosody is consistent, and the flow feels spontaneous and unforced." } ] }, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Oh, I forgot to mention, it should be priority-1 threads.", "long_pause_offsets": [ 53 ], "source_sha256": "91e8d809b7f4e8d4bdbe9252ffb6190af569f5664b52252cef0eb9f872a29dae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'priority-one' restriction to the initial question about unread threads, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/addition/addition_01/25b333fd30f0a46c/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "9fd2ee46ae985d0ac1f2c59128bc7d8dbbd64eaa5717b323f647c79292b275da", "seconds": 12.32, "spoken_utterances": [ "How many unread email threads are ... in my Gmail inbox? ", "Oh, I forgot to mention, it should be um... priority-one threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How many unread email threads are in my Gmail inbox?' and was followed by an addition specifying 'priority one threads'. The addition clarifies the type of unread threads to count, combining with the original request to form the complete instruction. The final intent matches the reference instruction, and the addition type is correctly applied.", "heard_transcript": "How many unread email threads are in my Gmail inbox? Oh, I forgot to mention it should be um priority one threads." }, "candidate_id": "wav_7ac6e79ff48b8ec5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. After a brief pause and filler ('oh'), they add a clarifying detail ('priority one threads') in a way that sounds like a spontaneous afterthought. The prosody is consistent, and the addition flows smoothly without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_thread_priority", "proposed_instruction": "How many priority-2 unread email threads are in my Gmail inbox? Make that priority-1 unread email threads, not priority-2.", "long_pause_offsets": [ 64 ], "source_sha256": "d7019d29e3a2d20bc636841f7005d814b75ed0fb17e4d837f572d7d7d11604e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the priority level from two to one, while retaining the other conditions: unread status, email threads, and Gmail inbox context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/revision/revision_01_thread_priority/c5b9bd4f6a5a2807/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "54e4a3aa6bfc5e130d1d50d2650ed5a6d3145c0bf9f5e2fc1c4854326008e91a", "seconds": 16.16, "spoken_utterances": [ "How many um, let me think... priority-two unread email threads are ... in my Gmail inbox? ", "Make that hmmm... priority-one unread email threads, not priority-two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority two' but immediately corrects to 'priority one' after a pause, which matches the original instruction. The revision type criterion requires applying the correction spoken after a longer pause to the earlier instruction, which is satisfied here. The final audible intent is to count priority-1 unread threads, matching the original instruction.", "heard_transcript": "How many um, let me think, priority two unread email threads are in my Gmail inbox? Make that priority one unread email threads, not priority two." }, "candidate_id": "wav_db3e5fa4156b87dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a realistic 'um' and 'let me think' filler, then delivers the initial request with a natural, slightly hesitant prosody. The correction ('make that...') is smoothly integrated into the same utterance, with a clear self-repair of the priority level. The intonation and pacing are consistent with spontaneous speech, and the correction feels like a genuine moment of recollection rather than a mechanical edit." } ] }, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "1c197057a0a2f8ddb5459f4b2b6623c85dacac569a9c08375d57f52f3ab34abb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/cancellation/cancel_021/2082ac9c7b9c26b1/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "e57c7fcc5748ef3df0a1c0c12f3b717ff523d2aac04a7fc75b82d2be35d1b428", "seconds": 11.32, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question, then clearly revokes the entire task with 'Hold on, never mind, I don't need that done anymore.' This constitutes a complete cancellation of the original request. The final intent is to cancel, not to execute or pause, and the cancellation is unambiguous and applies to the whole task.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox? Hold on, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "0506275714eca4d9d46cc278750db646a5dbfc710b6547887dd8b581b8d51d8b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/ambiguous_halt/halt_037/a855ff1a1b5cc330/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "56000963a728f408d7a92e13f1a1f327a7b1d8b855032a8b96dff7ec1261a6e9", "seconds": 9.28, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause command. The initial request is not canceled or revised, but the final control intent is to pause, not execute.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox? Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0e583ea672591dd803ec163d7837c19393ed289cdd82f82f2e027c68271193d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/filler/original/b7ccd815ddeb9571/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "fd2a7db52b2aa7b43a5e3805470fcc458ad1e13ce58e669b693f198b2b67b960", "seconds": 6.08, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'um' but continues with the complete and correct original instruction. The final intent matches the reference exactly, with no changes or omissions. The filler does not alter the task goal, so the execution criterion is satisfied.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_thread_priority", "proposed_instruction": "How many priority-1—no, I meant priority-2 read email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "db78a20d3f6959745f5279d8fc3aff4624f0986e66da46c949982b7527fe1eb6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'priority-one' to 'priority-two' within the same utterance, preserving the rest of the query including the app (Gmail), status (read), and thread count request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/false_start/false_start_01_thread_priority/cb99fe5cd14e35ab/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "44ba109269c9e4b86115ccb272bd428ad44779342f00d0c34e8a9d8d98904f0a", "seconds": 10.56, "spoken_utterances": [ "How many ... priority-one—um... no, I meant priority-two uh... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority one' but immediately self-corrects to 'priority two'. The final audible intent matches the original instruction exactly, with the correction applied. The 'false_start' type requires execution of the corrected version, which is fully present and clear.", "heard_transcript": "How many priority one um no I mean priority two uh read email threads are in my Gmail inbox?" }, "candidate_id": "wav_2dfa7f3d85e89974", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and spontaneous-sounding false start. The speaker begins with 'priority one', immediately self-corrects to 'priority two' with a clear, audible hesitation ('uh'), and continues smoothly. The prosody is consistent, and the correction feels like a genuine moment of recall. The pacing and intonation are very human, with no signs of artificiality or forced delivery." } ] }, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How many priority-2 email threads are in my Gmail inbox? I left out a detail: only count read email threads.", "long_pause_offsets": [ 57 ], "source_sha256": "5de72d55971910941e8ffd462d39a81aeef3fba2177d2a0f176dfa33f59e66d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count read email threads', which refines the initial broad request. The priority level and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/addition/addition_02/5b72987e28a17505/speech.wav", "wer": 0.12, "wav_sha256": "d977dd9af9fd39d74b9c76461bdc45a945fa89cb043da3547bcdb668224812ce", "seconds": 15.84, "spoken_utterances": [ "How many um... priority-two email threads are in my ... Gmail inbox? ", "I left out a detail: only count um, let's see... read email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many priority-2 read email threads are in my Gmail inbox?'. The speaker then adds a clarifying condition 'Only count... read email threads' after a pause, which refines the original request by adding a constraint. This matches the 'addition' type, where the final goal is the combination of the initial instruction and the added information. The final audible intent is to count priority-2 threads that are read, which is a valid and complete task.", "heard_transcript": "How many um priority two email threads are in my Gmail inbox? I left out a detail. Only count um let's see read email threads." }, "candidate_id": "wav_a58e32add5395e94", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a natural-sounding addition, but the flow is less smooth. The speaker starts with a hesitant 'How many um' and then makes a self-correction ('I left out a detail') before adding the restriction. However, the addition ('only count um let's see read email threads') is less coherent and the phrase 'read email threads' is awkwardly phrased, possibly due to a slip of the tongue or unclear thought. The prosody is slightly less fluid, and the final phrase sounds like a mumbled afterthought rather than a fully formed addition, reducing its realism compared to audio_02." } ] }, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_thread_priority", "proposed_instruction": "How many priority-1 read email threads are in my Gmail inbox? Wait, the priority should be priority-2, not priority-1.", "long_pause_offsets": [ 62 ], "source_sha256": "07f4217469d36d3d22dad7226a9dc3ee8569ecfecb0322f2b949331204399e53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the priority level from 'priority-one' to 'priority-two', while retaining the other conditions: 'read email threads' and 'in my Gmail inbox'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/revision/revision_01_thread_priority/3b922129fec1e7de/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "ff9f6d426e7222ff8e53ccbc0dcf3592fb6cbf52577c1dfcc1cf6d874e468596", "seconds": 15.16, "spoken_utterances": [ "How many uh, let's see... priority-one read email threads are in my Gmail inbox? ", "Wait, the priority should be uh... priority-two, not priority-one." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority one' but immediately corrects themselves to 'priority two, not priority one' after a pause. This is a clear revision where the final audible intent is to request priority-2 threads. The original instruction matches this final intent, so the revision is correctly applied.", "heard_transcript": "How many uh let's see, priority one read email threads are in my Gmail inbox? Wait, the priority should be uh priority two, not priority one." }, "candidate_id": "wav_f77d006ba2c14377", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a plausible hesitation ('uh'), formulates the initial request, then self-corrects the priority level with a believable change-of-mind intonation ('Wait, the priority should be...'). The correction is smoothly integrated into the same utterance, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "de9930a3bb9b897f2aa2745668cf8ee8bf7bbbfcc31a2db3eb7408534e390f22", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/cancellation/cancel_014/473af8b288f9ad50/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "f74eaaf4a59d77acddcaf178f7affc0147c2e6aae66dedd09494e628c9a5aae0", "seconds": 13.12, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?", "Oh, wait, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to ask a question but then explicitly says 'cancel that' after a brief pause and self-correction. This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute the original task.", "heard_transcript": "How many uh let's see priority two read email threads are in my Gmail inbox? Oh wait, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "02ec18a2e714016d63ba8aaf6f2c4b5e6ca0d5afb8fc0af9795866b0eeca2d62", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/ambiguous_halt/halt_024/0ef88e27673472dc/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "cb00f20782ac76fa9c4f2365cb791f0fdc634b07ca164f9a496bd78eda539229", "seconds": 11.56, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Give me a moment to think,' which is a clear request to pause. The audio ends with this interjection, and there is no cancellation or further instruction. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "How many uh, let's see, priority two, read email threads are in my Gmail inbox? Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e3d755dfb8ce8a7626032d924a2c9707b373d9e919f320892eb0f60070a131e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/filler/original/ba1ed4c7dec7d9a7/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "e073e60c5819bd864a707553cab0a611db5fbe26ca3a5ae1bb89fcd4be709b0d", "seconds": 7.92, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('uh', 'let's see') but completes the original instruction without changing the action, target, or value. The final intent matches the reference, and the filler words do not alter the task goal.", "heard_transcript": "How many uh let's see priority two read email threads are in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_03_app", "proposed_instruction": "How many priority-3 unread email threads are in my Outlook—wait, I meant Gmail inbox?", "long_pause_offsets": [], "source_sha256": "f002198d6c5862ded7974e426bc06b20a442c7c98e903c96f911bbd60a6c2429", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail inbox'. The rest of the query, including the priority level and unread status, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/false_start/false_start_03_app/69c5d63c41cc9b83/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "b8739db5a80aeeda18598f5c25d51508e79251fe0d2b85a6fd71061ff117d90d", "seconds": 8.16, "spoken_utterances": [ "How many priority-three ... unread email threads are in my Outlook—ah... wait, I meant Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook app' but immediately self-corrects to 'Gmail inbox'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many priority three unread email threads are in my Outlook app? Wait, I meant Gmail inbox?" }, "candidate_id": "wav_abebe5d3ae0e3c7c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker starts with 'Outlook', then immediately self-corrects with 'wait, I meant Gmail inbox?' in a way that sounds like a genuine, real-time thought process. The prosody is fluid, with a believable shift in intonation and pacing during the correction. The pause before 'wait' is short and natural, fitting a moment of realization. The correction is integrated smoothly into the utterance without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "e775c78_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How many priority-3 email threads are in my Gmail inbox? One more detail: only count unread email threads.", "long_pause_offsets": [ 57 ], "source_sha256": "c473dbc940b3bfd66fdee059dd1bb155a9fb9b62a10f09e141adddd4b20cf41c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'unread' constraint, which is a refinement of the initial request. The app (Gmail), priority level (priority-3), and action (count) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/addition/addition_02/dba0b02c2274e073/speech.wav", "wer": 0.0, "wav_sha256": "729d8b85e41c1e39a0ce5ffd140e932a54522f7288c6e6dfff5758b92af0144a", "seconds": 11.76, "spoken_utterances": [ "How many ... priority-three email threads are ... in my Gmail inbox? ", "One more detail: only count unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How many priority-3 unread email threads are in my Gmail inbox?'. The speaker first asks about 'priority three email threads' and then adds the constraint 'only count unread email threads'. This matches the 'addition' type, where the final goal combines the initial request with the added detail. The final intent is fully specified and executable, matching the original instruction's intent.", "heard_transcript": "How many priority three email threads are in my Gmail inbox? One more detail, only count unread email threads." }, "candidate_id": "wav_1f80c0fd89a8cbae", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "e775c78_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "794f3d928dd27ff07c0bdb79e7e4397399fa04c5946640a6cb919a372b82e46f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/cancellation/cancel_008/128bcf6b3e481e1e/speech.wav", "wer": 0.0, "wav_sha256": "c945a41cb8c4936657c0b3ceddc1a9423aea873c9be71900f6d16a5bba14c58b", "seconds": 11.64, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?", "Wait, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to state the original task but immediately corrects themselves with 'Wait, I want to cancel that task,' which is a clear and final cancellation of the initial request. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox. Wait, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "cb80ffafbb4a51edba2546b879c0ab9b88dd6b8af4ee43ae0f97bfdad259ed5d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/ambiguous_halt/halt_015/1d851698578ceadd/speech.wav", "wer": 0.0, "wav_sha256": "ec06d1d073b44789a6e453f34e807ecaba034f8b8270faf9f1f461872f93d474", "seconds": 10.4, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Please wait a moment,' which is a clear instruction to pause. The initial part of the audio is a request for information, but the final interjection is a pause command. Since the type is 'ambiguous_halt,' the final intent to pause is what matters, and it is clearly audible and unambiguous.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e775c78_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4b5a7d51bd87095a3ed30d5d553d174b39d84c37aab26740715b07c1951921da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/filler/original/66ae3b296ee9465f/speech.wav", "wer": 0.0, "wav_sha256": "d2cff44138988eeaf66960d21ad2e19432e3ebd637f3379cf6837a3fb944c2ef", "seconds": 6.88, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'How many let's see' as a filler, then proceeds to state the full original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my shortest—I meant longest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "5fe05e48808eac6e7806cfae31d41cc72991330d5a70cc217da7b4aca658c4bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shortest' to 'longest' within the same utterance, preserving all other conditions including the app (Spotify), metric (minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/false_start/false_start_01_playlist_selection/3057ec531b689866/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d7c703f6cc81b3d1ad3aa360080cc1bbc001ac23136f8365ac26cf0c2d790d9d", "seconds": 7.84, "spoken_utterances": [ "How long is my shortest—um... I meant uh... longest Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shortest' but immediately self-corrects to 'longest' with 'I meant uh longest'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('shortest') is discarded. The rest of the constraints ('Spotify playlist', 'in minutes', 'rounded to the nearest number') are preserved and match the reference. Therefore, the final audible intent matches the original instruction.", "heard_transcript": "How long is my shortest um, I meant uh longest Spotify playlist, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_3b06b48da417bf57", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with 'shortest' and immediately self-corrects to 'longest' with a natural-sounding 'I meant' repair. The correction is fluent and integrated into the sentence, and the rest of the utterance is delivered at a steady, conversational pace with a single, brief filler ('um') at the start. The prosody is consistent and the phrasing sounds like a spontaneous correction during a real task." } ] }, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How long is my Spotify playlist, in minutes, rounded to the nearest number? Oh, and one more thing: find the longest playlist.", "long_pause_offsets": [ 76 ], "source_sha256": "126e927ea7926ee6428d12197b7a7a62c94c1ec8b6c44d13d8fb60fff60c3109", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the explicit condition to find the longest playlist, which clarifies the ambiguous 'my Spotify playlist' in the initial utterance. The units, rounding, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/addition/addition_01/d4556b77122394e0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "f05288b44702bec1b5f20a1945f16a5282d82c6c09be34a3ab760daf895398ab", "seconds": 15.2, "spoken_utterances": [ "How long is my Spotify playlist, hmmm... in minutes, um... rounded to the nearest number? ", "Oh, and one more thing: find the um, let's see... longest playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How long is my Spotify playlist?' with constraints 'in minutes' and 'rounded to the nearest number'. After a pause, the user adds 'find the... longest playlist'. This addition modifies the initial request by specifying which playlist to measure, changing the scope from 'my playlist' to 'my longest playlist'. The final goal combines the original constraints with the new target, matching the reference instruction. The addition is clear and unambiguous, satisfying the addition type criterion.", "heard_transcript": "How long is my Spotify playlist? Hmm. In minutes. Um. Round it to the nearest number. Oh, and one more thing, find the um, let's see, longest playlist." }, "candidate_id": "wav_4dd5c87149565572", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('How long is my Spotify playlist?') followed by a believable hesitation ('Hmm...') and a pause. The addition of 'Oh, and one more thing' is a pragmatic, conversational way to introduce a new request. The second part ('Find the... longest playlist') is delivered with a natural, slightly searching intonation ('let's see'), suggesting spontaneous thought. The prosody is continuous and the pacing realistic, with pauses that feel like genuine thinking rather than robotic breaks." } ] }, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_playlist_selection", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes, rounded to the nearest number? Wait, I meant the longest Spotify playlist, not the shortest.", "long_pause_offsets": [ 85 ], "source_sha256": "ab4febbc98728ae4abcd40d2b01281b950c150148adf2e54b2ee5d5fba3c5d88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial request from 'shortest' to 'longest', preserving the app (Spotify), metric (duration in minutes), and rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/revision/revision_01_playlist_selection/979de3d511f76e5c/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "9924fcff1a4df1733cdee395a73a9771c44399a86b9d754686e272aae031c4c7", "seconds": 13.28, "spoken_utterances": [ "How long is my uh... shortest Spotify playlist, in minutes, rounded to the nearest number? ", "Wait, I meant the um... longest Spotify playlist, not the shortest." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the shortest playlist but immediately corrects themselves to request the longest playlist. This is a revision where the final intent is the corrected version. The final audible request matches the original instruction in meaning, with only the target (longest vs. shortest) being revised. The rest of the constraints (Spotify playlist, minutes, rounded) are preserved.", "heard_transcript": "How long is my uh shortest Spotify playlist in minutes, rounded to the nearest number? Wait, I meant the um longest Spotify playlist, not the shortest." }, "candidate_id": "wav_1c447440039101ee", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural revision flow. The speaker starts with a clear request, hesitates with 'uh' and 'wait', then self-corrects 'shortest' to 'longest' with a believable change-of-mind intonation and a smooth, continuous correction. The prosody is conversational, with appropriate pauses and filler usage that reflect genuine thought processing." } ] }, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "bebcc38a6d6e809b47da2ba323f5a9973e1216861dbbed4540d981bf108c5491", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/cancellation/cancel_021/7b6b25b942a0d622/speech.wav", "wer": 0.08, "wav_sha256": "3b831792cf8c48a73b21c26245bdd62feea1b4efb1f5842e9fa48e6728b8d162", "seconds": 12.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?", "Wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the length of their longest Spotify playlist, but immediately follows up with 'Wait, never mind, I don't need that done anymore,' which is a clear and final cancellation of the task. The audio ends with this cancellation, so the final control intent is to cancel, matching the 'cancellation' type criterion.", "heard_transcript": "How long is my hmm longest Spotify playlist in minutes, um rounded to the nearest number. Wait, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "e91f23a0a7f8c8972d7e399c446bc6fc60d3857e539d94f619f9cd9e1b9dfab2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/ambiguous_halt/halt_022/854dd8c0bf2a3378/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "bc366daa19bf877a61b434f17c5480e52c0a7bd1ae82d8c69b71fcc24f3846e6", "seconds": 11.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the user does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How long is my longest Spotify playlist in minutes, um, round it to the nearest number. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3467c579f5e2259aefae8776a0516f63a67f9fb1809b7e9dd22df556c875637f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/filler/original/96845a344b805d03/speech.wav", "wer": 0.125, "wav_sha256": "8c49cf52e2b0b5f3c8d7cbbdaac48aa4021c9e1fdf55aaceb96091d5b52fbfea", "seconds": 7.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um') but does not change the core request. The final intent matches the original instruction, and the filler does not alter the task goal. The instruction is executed as intended.", "heard_transcript": "How long is my longest Spotify playlist in minutes, um, rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my longest—I meant shortest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "317305f2eacaeda6eba583f213c4ab70b8e2be21fc866f8c81e1b8e73e1eb9da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'longest' to 'shortest' within the same utterance, preserving all other conditions including the app (Spotify), metric (duration in minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/false_start/false_start_01_playlist_selection/6cc777ed5e25843f/speech.wav", "wer": 0.05, "wav_sha256": "8db40a99269c394024ae7508cf51253447f669c83c90f7c116b36b7298aaa3ce", "seconds": 11.0, "spoken_utterances": [ "How long is my let's see... longest—ah... I meant shortest Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'longest' but immediately self-corrects to 'shortest'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final audible intent matches the original instruction exactly, with only a minor grammatical repair ('I meant' instead of 'I meant to'). All constraints (Spotify playlist, duration in minutes, rounded to nearest number) are preserved.", "heard_transcript": "How long is my, let's see, longest uh, I meant shortest Spotify playlist, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_be91bdc0c9f6a2c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent self-correction. The speaker begins with 'How long is my... let's see...', which sounds like genuine hesitation. The false start 'longest' is immediately and smoothly corrected to 'shortest' with a believable 'I meant' phrase. The rest of the sentence flows naturally without further disruption, and the prosody is consistent and conversational." } ] }, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes,? Oh, I forgot to mention, round the result to the nearest number.", "long_pause_offsets": [ 55 ], "source_sha256": "abe2d8d096d10314f626dcb37c87f831219fdb60abcc5bd3211de6fc707d12cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rounding instruction, which was not in the initial utterance but is now included in the final goal. The original instruction already specified the playlist and unit, and the addition completes the requirement without changing any other aspect." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/addition/addition_01/e3762ebae063370b/speech.wav", "wer": 0.0, "wav_sha256": "879200178d082e6483af1c3e85f399372025fbabeaac80bc45ea76ae3ca61263", "seconds": 14.96, "spoken_utterances": [ "How long is my let me think... shortest Spotify playlist, ... in minutes,? ", "Oh, I forgot to mention, round the result um... to the nearest number." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is a slightly hesitant version of the original instruction. The speaker then adds the constraint 'round the result to the nearest number' after a pause, which is an addition to the original request. The final intent combines the original task with the added rounding requirement, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "How long is my, let me think, shortest Spotify playlist, in minutes? Oh, I forgot to mention, round the result, um, to the nearest number." }, "candidate_id": "wav_1e4e70ccc435cab6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') and then adds a missing detail ('round the result') after a brief pause, which is followed by a second, more specific addition ('to the nearest number'). The flow is continuous and the speaker's intonation reflects genuine recollection and correction, making the addition feel spontaneous and realistic." } ] }, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_playlist_selection", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, rounded to the nearest number? Wait, I meant the shortest playlist, not the longest.", "long_pause_offsets": [ 84 ], "source_sha256": "01e0807656c22d2494456cb37b63dbc88e2ac6c398ab67db112f78b06a4c3048", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects 'longest' to 'shortest', preserving the app (Spotify), unit (minutes), and rounding instruction. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/revision/revision_01_playlist_selection/fc8bce102fcc81b7/speech.wav", "wer": 0.08, "wav_sha256": "3aa3cc2bbeeca6e3772e6830d8bc4065f998eb0f31d102f34537a53b78e84489", "seconds": 13.48, "spoken_utterances": [ "How long is my let's see... longest Spotify playlist, ... in minutes, rounded to the nearest number? ", "Wait, I meant the shortest playlist, not the longest." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'longest' playlist but immediately corrects themselves to 'shortest' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the initial instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "How long is my let's see, longest Spotify playlist in minutes, rounded to the nearest number? Wait, I meant the shortest playlist, not the longest." }, "candidate_id": "wav_829c682bec6dfbd6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects with a plausible 'Wait, I meant...' followed by a smooth revision. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause before the correction is brief and realistic, fitting a moment of realization." } ] }, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "8d5367d3ce28babfa5dc203cc0d1ca850c83b1b12f3c15dc2115ef69ddf51008", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/cancellation/cancel_008/95a8806dcae486d9/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "d786c2dff34119bcbbb2792668a1127491666afbab6e7159f7ba008a6f2c68a1", "seconds": 12.96, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, I want to cancel that task.' This is a clear and direct cancellation of the entire task. The final interjection is a cancellation command, not a pause or a revision. Therefore, the control intent is correctly identified as 'cancel'.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes rounded to the nearest number? Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "b3a9d638d33bb07a4e6dcaf9221d1dcbd35ef02aa042099a0e882de0a79f095d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/ambiguous_halt/halt_033/5bbf39d0b8b4bafa/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "5e75c414c361598d651d5b0a1b43a4a9cbf2d073f05313852b7d0a0b9cadc525", "seconds": 12.16, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then audibly pauses with 'Hold on, there's something', which is a clear request to hold/wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker intends to resume after a brief interruption.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes, rounded to the nearest number? Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "594bd3005e1345aba27cb4f9bb8eeac28e54fc240cea02fe6d792e4b65e2e66e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/filler/original/3e2e7e3a4cf9f1a0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "b70ba9b48abc18d2418ca3fd9331a2549d3dcdbd28eae26d059b87eaf04b169f", "seconds": 8.28, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see', 'uh') but does not change the core request. The final intent is clear and matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my shortest—no, I mean longest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "0588c694f1fee35923ac5d7b49363571e30c274811b8559009d0db91e58a3783", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shortest' to 'longest' within the same utterance, preserving all other conditions including the app (Spotify), metric (duration in minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/false_start/false_start_01_playlist_selection/6f7b78e672bd2e33/speech.wav", "wer": 0.0, "wav_sha256": "4b3f082a2ce7434258487b79c36ebd12a4616aa6392feda2fef821347a40d939", "seconds": 7.4, "spoken_utterances": [ "How long is my shortest—um... no, I mean longest Spotify playlist, in minutes, ... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shortest' but immediately self-corrects to 'longest'. According to the 'false_start' type criterion, the final corrected value ('longest') should be applied, and the original incorrect value ('shortest') should be discarded. The rest of the instruction, including 'Spotify playlist', 'in minutes', and 'rounded to the nearest number', remains unchanged and is preserved. Therefore, the final intent matches the original instruction.", "heard_transcript": "How long is my shortest um, no, I mean longest Spotify playlist in minutes, rounded to the nearest number?" }, "candidate_id": "wav_babf050632e2b64d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with 'How long is my shortest' and immediately self-corrects to 'longest Spotify playlist'. The correction is smooth, with a natural pause and intonation shift. The phrasing 'rounded to the nearest number' is delivered without hesitation, suggesting a spontaneous, realistic thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How long is my longest Spotify playlist,, rounded to the nearest number? Oh, and one more thing: return the duration in minutes.", "long_pause_offsets": [ 73 ], "source_sha256": "2e72a95757dad187a9abdd09e53180d9bfa2ba4b7376a57625282a9c3d5e4704", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing unit 'in minutes' to the initial request, which was otherwise complete. The original instruction already specified rounding to the nearest number, so the final goal matches exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/addition/addition_02/ad8cc0501faff1e5/speech.wav", "wer": 0.0, "wav_sha256": "d4a05d882f8f090a05ce1f6e883af047f77c6e0b5effd8b68336ef36a2bef25c", "seconds": 14.8, "spoken_utterances": [ "How long is my let me think... longest Spotify playlist,, um... rounded to the nearest number? ", "Oh, and one more thing: return the duration ... in minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance contains hesitation ('let me think') and filler ('um'), but the core request 'How long is my longest Spotify playlist' is clear. The addition 'return the duration in minutes' after a pause refines the output format. The final intent combines the original request with the added constraint, matching the reference instruction. The 'rounded to the nearest number' is preserved as part of the original request.", "heard_transcript": "How long is my let me think longest Spotify playlist um rounded to the nearest number. Oh, and one more thing, return the duration in minutes." }, "candidate_id": "wav_fe2ffd227d572d03", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers like 'let me think' and 'um' with plausible timing and intonation. The addition of 'Oh, and one more thing' is a realistic, conversational way to append a detail. The prosody is continuous and the pacing reflects genuine thought, making the delivery highly spontaneous and believable." } ] }, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How long is my longest Apple Music playlist, in minutes, rounded to the nearest number? Make that Spotify, not Apple Music.", "long_pause_offsets": [ 88 ], "source_sha256": "54787e8d5f87dfd45247e45a9835db5d3698bfa33e9e5670aecbf7ff2b96ff9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the rest of the query including the metric (duration in minutes), the scope (longest playlist), and the rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/revision/revision_02_app/629a26a5328b80d9/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "8d2ff85d864d0163809f83f48e8cee9d18e48b314b4babbfa0a0a373bb4e8c71", "seconds": 14.88, "spoken_utterances": [ "How long is my uh... longest Apple Music playlist, ah... in minutes, rounded to the nearest number? ", "Make that uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects to Spotify. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent matches the original reference instruction exactly, with only a minor grammatical difference ('Make that' vs. 'How long is') that does not alter the meaning.", "heard_transcript": "How long is my uh longest Apple Music playlist uh in minutes, rounded to the nearest number. Make that uh Spotify, not Apple Music." }, "candidate_id": "wav_0cd8f7ea3fe22564", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('How long is my...') and then self-corrects the platform from 'Apple Music' to 'Spotify' with a natural-sounding 'Make that's uh...'. The correction is integrated smoothly into the same utterance, maintaining a consistent rhythm and prosody. The pause before the correction is brief and plausible, fitting a moment of realization. The overall delivery sounds like a spontaneous, real-time correction without any robotic or disjointed qualities." } ] }, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "c13f69783c13ef2d2b02fd0dc4cc491e7ccf03f886c1dac5832afab4b6d43a11", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/cancellation/cancel_021/a3db7f6f7ea14a26/speech.wav", "wer": 0.0, "wav_sha256": "8cab0c9fbe731962fd50bcce52d8ba44fca622e1f73aa9214924de586528cc71", "seconds": 13.08, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?", "Wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Wait, never mind, I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final intent is to cancel, not to execute.", "heard_transcript": "How long is my uh, let me think. longest Spotify playlist in minutes, rounded to the nearest number. Wait, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "c73522420e129c08c49acbc20cb6d5cff2834712a5b27f921f23edae84e79a68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/ambiguous_halt/halt_004/3599c3c22bcf6f79/speech.wav", "wer": 0.0, "wav_sha256": "982f5077b01afe0bdb401b48a95ca417f0b41eec0c4c392f9dcb67224a3119ca", "seconds": 11.48, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler phrase 'let me think' followed by a clear request to pause, 'Just a second.' This constitutes an unambiguous halt request. The original instruction is not canceled or revised, and the user's intent is to delay execution, not to execute or cancel the task. Therefore, the final control intent is 'pause', which matches the type-specific criterion.", "heard_transcript": "How long is my uh, let me think. longest Spotify playlist in minutes, rounded to the nearest number. Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e7a10f8_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d5e86970559f550ec887d5ef5a7dee439369aabdecf7d6213c9f00784afc63a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/filler/original/38998782e0c0acc0/speech.wav", "wer": 0.0, "wav_sha256": "eb6498a4ccee2fd7ac0b1813a37d0b30ac8005b8203a63f25c71b95075da7334", "seconds": 7.96, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh' and 'let me think' but does not change the core request. The final, complete instruction matches the original reference in both action and constraints. The 'filler' type allows for such hesitations and backchannels without altering the task goal.", "heard_transcript": "How long is my uh, let me think, longest Spotify playlist in minutes, rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the least played—no, I mean most played song by Velvet Echo on Spotify.", "long_pause_offsets": [], "source_sha256": "785912f5d65d7b64d343eec054393633cd5cd5ebb9f2b636016b1ab75d4d0a0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately corrects to 'most played' within the same utterance. The final intended action is to identify the most played song, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/false_start/false_start_01_song_selection/0d6aa09a4b0b927f/speech.wav", "wer": 0.125, "wav_sha256": "382742177e7122f05a27daabbf400bae5acf5403cd7f788b9584dfa6f0285db3", "seconds": 11.56, "spoken_utterances": [ "What is the title of the least played—uh... no, I mean hmmm... most played song by um, let's see... Velvet Echo on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately self-corrects to 'most played' with 'no, I mean'. This is a false start where the final intent is the corrected version. The rest of the instruction, including the artist 'Velvet Echo' and platform 'Spotify', remains unchanged and clear. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "What is the title of the least played uh no, I mean most played song by um let's see Velvet Echo on Spotify." }, "candidate_id": "wav_a6e5ab623649f821", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The false start ('least played') is immediately and smoothly corrected with 'no, I mean, most played', followed by a brief, believable hesitation ('hmm') and a natural recovery ('let's see') before completing the request. The prosody is consistent, and the self-correction feels spontaneous rather than rehearsed." } ] }, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Velvet Echo on Spotify. Oh, I forgot to mention, it should be the most played song.", "long_pause_offsets": [ 57 ], "source_sha256": "fae52e994e107531e02d3d2469ee67b012d0997501cbf45bf557819f7e457957", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most played' to the initial request about a song by Velvet Echo, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/addition/addition_01/d2f34a2530d3d6a0/speech.wav", "wer": 0.04, "wav_sha256": "a802935440d8f661ad100fd54b86f84db56536bc1385dcf751115842589b0499", "seconds": 13.04, "spoken_utterances": [ "What is the title of the song by ... Velvet Echo on um... Spotify. ", "Oh, I forgot to mention, it should be the ... most played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, asking for 'the title of the song' without specifying 'most played'. After a pause, the speaker adds the constraint 'it should be the most played song'. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, so the task passes.", "heard_transcript": "What is the title of the song by Velvet Echo on um Spotify. Oh, I forgot to mention, it should be the most played song." }, "candidate_id": "wav_20af90e86e049d91", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. After a brief pause and filler ('um'), they add a crucial detail ('most played song') in a way that sounds like a spontaneous memory. The intonation and pacing of the addition feel organic, not forced or rehearsed. The overall prosody is consistent and conversational." } ] }, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the top rated song by Velvet Echo on Spotify. Wait, the song selection should be 'most played', not 'top rated'.", "long_pause_offsets": [ 67 ], "source_sha256": "2759bc425a5abcb51224213c588870d40bb09aed6f8a130410ac95e9e3250b0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the selection criterion from 'top rated' to 'most played', while preserving the artist, app, and query type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/revision/revision_01_song_selection/c4b3ebe0d9f51ae6/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "eecfb31988e10ff39c024dbf1204b64dc9d7746dd4a83c9b005c67ce09adb0bd", "seconds": 16.72, "spoken_utterances": [ "What is the title of the top rated song by ... Velvet Echo on hmmm... Spotify. ", "Wait, the song selection should be 'uh, let me think... most played', not 'top rated'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'top rated' song but immediately corrects themselves to 'most played' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction, so the task should be executed as requested.", "heard_transcript": "What is the title of the top rated song by Velvet Echo on Spotify? Wait, the song selection should be, uh, let me think, most played, not top rated." }, "candidate_id": "wav_889c80957fff1c05", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible revision process. The initial request is slightly hesitant ('top-rated'), followed by a believable self-correction ('wait... let me think... most played') with a realistic filler and a pause. The prosody is varied, with a genuine-sounding shift in intonation when correcting the metric, and the overall rhythm mimics a person recalling and adjusting their request in real time." } ] }, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "9bd75fbd248f054cd7000ffebbfae70e528f75a1e5e452e89e6290f1c8062a0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/cancellation/cancel_014/60035d3410603b93/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "934e70afdf0935952b1dbac49cb4839127da7dda7a4bfa44cc7185333aa9dd6a", "seconds": 12.08, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify.", "Hold on, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Hold on, on second thought, cancel that,' which is a clear and unambiguous cancellation of the task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify? Hold on, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "431abf1b574623604c221e5533b78aa2a5e74630e5eb361d6b6f84e1e6584679", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/ambiguous_halt/halt_014/013eb87e1d4506f4/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "01499f8cfe23176442268cdbc2f11f0fc199933beee041dff9696b10d62684df", "seconds": 10.96, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Can you hold on a second?', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, but the speaker is asking for a momentary halt.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_1", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ae084364de2983b6668598e8c2327d7b17bb03bed4ba6fc8215aeac6e22f9490", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/filler/original/81d8ba65ea175606/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "dd58735fa4ab3963b0d2dfb2bea284d65567891d968e173dd5bb3a7d3cf58a53", "seconds": 7.28, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent is to retrieve the title of the most played song by Velvet Echo on Spotify, which matches the original instruction. The filler does not alter the task goal.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_artist_name", "proposed_instruction": "What is the title of the least played song by Zoey Jones—sorry, I meant Zoey James on Spotify.", "long_pause_offsets": [], "source_sha256": "d7935fda17b89ae88455f21e2ec241e3a51a9c92ecf8cbd9bacbd91d431afbc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Zoey Jones' to 'Zoey James' within the same utterance. The final artist name matches the original instruction, and all other conditions (least played, song title, Spotify) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/false_start/false_start_02_artist_name/e7d0dd7a358ea9b2/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "fff4378aa11dfcdef2e33e78002f5328b8955564d9483f29efa87c604b2e294f", "seconds": 10.92, "spoken_utterances": [ "What is the title of the let me think... least played song by Zoey Jones—um... sorry, I meant Zoey James on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Zoey Jones' but immediately self-corrects to 'Zoey James' after a pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "What is the title of the least played song by Zoey Jones, um, sorry, I meant Zoey James on Spotify." }, "candidate_id": "wav_91c25eb7d3ebdd71", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'What is the title of the' and then hesitates with 'let me think'. After a brief pause, they incorrectly say 'least played song by Zoe Jones' and immediately self-correct to 'Zoe James' with a clear, spontaneous-sounding 'sorry, I meant'. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and recollection. The pacing and intonation are consistent with a person formulating a request in real time." } ] }, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Zoey James on Spotify. Oh, I forgot to mention, it should be the least played song.", "long_pause_offsets": [ 56 ], "source_sha256": "2087918e6149c92fd745a1c982713e2f332bcbb1e0ed20fed02fcf15147b96c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least played' to the initial request about Zoey James's songs, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/addition/addition_01/bafda86b631304c6/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "d0025b70644f18ce063c789678cf0c9b2be1573bda96c0d566f7d5536738b66d", "seconds": 16.44, "spoken_utterances": [ "What is the title of the song by Zoey James on um, let me think... Spotify. ", "Oh, I forgot to mention, it should be the hmmm... least played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, missing the 'least played' constraint. After a pause and self-correction, the speaker adds the missing constraint. The final intent combines the original request with the added condition, matching the reference instruction. The addition is clear and directly modifies the initial request without introducing new errors.", "heard_transcript": "What is the title of the song by Zoey James on um, let me think, Spotify. Oh, I forgot to mention it should be the least played song." }, "candidate_id": "wav_9b6ab0b0dcbe5886", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_app", "proposed_instruction": "What is the title of the least played song by Zoey James on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 73 ], "source_sha256": "dfc22e3a58989c5a1bb60875964a09a92c710f52922e1ccd93d3e72faa2d4a70", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, preserving the artist, the query type (least played song), and the intent to retrieve the title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/revision/revision_03_app/5cfef3b25f43049c/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "a72f534d5d1df12ce4a406e3749f4181ef44fa95068b7358150cca7181892823", "seconds": 13.64, "spoken_utterances": [ "What is the title of the least played song by Zoey James on ... Apple Music. ", "Make that let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the least played song on Apple Music, then immediately corrects the platform to Spotify. This is a clear revision where the final intent is to query Spotify, not Apple Music. The rest of the instruction (artist, metric, action) remains unchanged. The revision is spoken with a natural pause and self-correction, satisfying the revision type criterion.", "heard_transcript": "What is the title of the least played song by Zoey James on Apple Music. Make that, let me think, Spotify, not Apple Music." }, "candidate_id": "wav_c0b83910f83e90d9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and filler ('let me think') before a smooth, self-corrected revision. The correction ('Spotify not Apple Music') is integrated fluidly into the same utterance, maintaining a consistent rhythm and intonation. The prosody is continuous and the correction feels spontaneous, not forced." } ] }, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "ad51cfb912059a15e9f595f4513dbac46996e3c3b03b7304e92072e7f1789f13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/cancellation/cancel_039/06e6649211f4d747/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "605386ac1d0b83252d78699c5083c64a7a61db27a90bf2bf15c17850a97c4136", "seconds": 14.24, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify.", "Hold on, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly and unambiguously cancels it with 'Hold on, don't go ahead with that. Cancel it.' This constitutes a final cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify. Hold on, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "5d1de2c34df55c60637f8543c99851fb2328fdf0d872b8d4fc7cddf6838b9d25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/ambiguous_halt/halt_005/1009dfdd2f22431e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "1a5a46837c1e960b87651ad2087f1026ae92e474ac5ea9fcd298af936d041bce", "seconds": 13.2, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify.", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request with 'What is the title of the least played song by Zoey James on Spotify' and then immediately says 'One second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is fully audible and unaltered, and the final control intent is to pause, not execute or cancel.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_2", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b782130edda27a3e3537b1551d0912bacf296366717e1c61b607f5860c3d3877", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/filler/original/2e0d4004558f299c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a040c1635bb69c0249c857f8b0406fc16225b5e494e8dea957a9a51397f6670a", "seconds": 9.56, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let me think') but does not change the core request. The final audible intent matches the original instruction. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the least played—no, I meant most played song by Jasper Skye on Spotify.", "long_pause_offsets": [], "source_sha256": "40e5b72a29ff88ff3a02eefd1405b3bde8c6134d9859e05017e0b5b7d9c984b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'least played' to 'most played' within the same utterance, preserving the artist, app, and the intent to retrieve the song title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/false_start/false_start_01_song_selection/b74bda724bc41df6/speech.wav", "wer": 0.05, "wav_sha256": "d9c4d714e73e494e0f3ed2e732a5dfc9ac04a550ddc6ad2337fe16eb64faa4c8", "seconds": 10.88, "spoken_utterances": [ "What is the title of the least played—um... no, I meant most played song by ... Jasper Skye on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately self-corrects to 'most played' with 'no, I meant'. This is a clear false start where the final intended value is 'most played'. The rest of the instruction, including the artist 'Jasper Skye' and the platform 'Spotify', remains unchanged and correct. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "What is the title of the least played um, no, I meant most played song by Jasper Skye on Spotify." }, "candidate_id": "wav_a4a0ba001d4a70d1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a very natural-sounding false start. The initial request 'least played' is clearly spoken, followed by a believable self-correction to 'most played' with a smooth intonation shift and a pause that sounds like genuine reconsideration. The phrasing 'no, I meant' is a common, spontaneous way to correct oneself, and the overall pace and prosody are consistent with someone thinking aloud." } ] }, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Jasper Skye on Spotify. Oh, I forgot to mention, it should be the most played song.", "long_pause_offsets": [ 57 ], "source_sha256": "466e958a76fbf028d599571f9e2bc51ae261c87df10d3fc247c2b4fdbb0d065f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most played' to the initial request about a song by Jasper Skye, refining the scope to match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/addition/addition_01/d9bcd3104cd1876f/speech.wav", "wer": 0.08, "wav_sha256": "420c73d63ddb1f7b268453f80f19a1193eea7f371fceba159d74486075051f1c", "seconds": 11.4, "spoken_utterances": [ "What is the title of the song by um... Jasper Skye on ... Spotify. ", "Oh, I forgot to mention, it should be the most played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was to find the title of a song by Jasper Skye on Spotify. The speaker then added the constraint 'it should be the most played song' after a pause, which is a clear addition to the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "What is the title of the song by um Jasper Skye on Spotify? Oh, I forgot to mention, it should be the most played song." }, "candidate_id": "wav_d563c96f73c98756", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody and a believable, additive follow-up. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to add a detail, and the intonation and pacing of the entire utterance sound spontaneous and conversational. The addition is smoothly integrated without awkwardness or hesitation." } ] }, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_artist_name", "proposed_instruction": "What is the title of the most played song by Luna Moon on Spotify. Sorry, I meant Jasper Skye, not Luna Moon.", "long_pause_offsets": [ 67 ], "source_sha256": "2b63c526cc16822886e6c42bcbc8fbf00983b6f37fe40c2bf273cabe5b7bbdc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the artist name from Luna Moon to Jasper Skye, preserving the rest of the query including the action, app, and the 'most played song' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/revision/revision_02_artist_name/7748499ac1d69228/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "9a01da0739d076f138fc7aa915fbf6fa695205d7435fc91817b2f6f413a42eca", "seconds": 15.0, "spoken_utterances": [ "What is the title of the most played song by ... Luna Moon on uh... Spotify. ", "Sorry, I meant ... Jasper Skye, not Luna Moon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Luna Moon' but immediately corrects to 'Jasper Skye' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the most played song by Jasper Skye on Spotify, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "What is the title of the most played song by Luna Moon on uh Spotify. Sorry, I meant Jasper Skye, not Luna Moon." }, "candidate_id": "wav_735a5af5f1cef5eb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'Luna Moon' to 'Jasper Skye' with a believable hesitation and intonation shift. The correction is smoothly integrated into the same utterance, and the overall prosody is consistent and conversational, suggesting a spontaneous thought process." } ] }, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_036", "hide_variant_original": true, "source_sha256": "7aa6f2c27450d68a3e1c5c892c1ebb53075aa8e60982022a6ae75d7e51b5ad5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/cancellation/cancel_036/6115fefc3eb3cbc1/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "37d5ae8607ac1ae110c4b12d1ea604dc73c8971a21481cf738be8ca408c79cb1", "seconds": 13.16, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify.", "Oh, wait, don't continue with that request. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a final directive to cancel the previously stated task. The speaker first formulates the request, then immediately interrupts with 'Oh wait, don't continue with that request. Cancel it.', which constitutes a clear cancellation of the entire task. The final control intent is a cancellation, matching the expected behavior for this type.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify. Oh wait, don't continue with that request. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "a34dfe1f284ec07c3f3ad43339168b541c33e06aba6f7d2d9229a163eeeae40b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/ambiguous_halt/halt_021/efc92fa2a3eace61/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "f80c092a080565005c479fb903265903e23ef61a77ec84f204082e493317adb9", "seconds": 11.76, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the speaker does not provide a corrected or final instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "e85d92a_3", "split": "train", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1fb5e8b3f63feb1cd0cc37696c8dc589dea08c4b58b8e34e8e2aec21a05b9d35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/filler/original/7f5f5a842e0f72d9/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "397c076cec6365e36701a6830a21170bbcfc89e53eb17398f3f533bd9724e748", "seconds": 7.8, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'let's see') and pauses but does not change the core request. The final intent is clear and matches the original instruction. The filler type allows for such hesitations without altering the task goal.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in this calendar year?", "long_pause_offsets": [], "source_sha256": "7b8af3f643590b6bef72d83eb0fcfe79eb8f3bbc813f91e667c80749bdd12be2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the app and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/false_start/false_start_01_app/f45de312f5a4855a/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "91a13f5083d6fcb218d1cd4d3aedd301ea7dd74a049c426feaa1e57052c3bf8c", "seconds": 7.84, "spoken_utterances": [ "How much did I spend on ebay—um... sorry, I meant amazon in uh... this calendar year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. This is a clear false start where the final intended target is Amazon. The time frame 'this calendar year' remains unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "How much did I spend on eBay um sorry, I meant Amazon in uh this calendar year?" }, "candidate_id": "wav_c2a43e89df78a8ef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('eBay') followed by a smooth self-correction ('I meant Amazon') with a believable hesitation ('uh') and a pause. The intonation and rhythm are consistent with spontaneous speech, and the correction flows naturally within the same utterance without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How much did I spend on amazon? I left out a detail: it's for this calendar year.", "long_pause_offsets": [ 32 ], "source_sha256": "8493fc56f91ff7c2f74dacd3734366fea6a626bbab77c32e917b03d5145742f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'this calendar year' to the initial question about Amazon spending, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/addition/addition_02/bf9c67e0f8472785/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "837bf8fbe1df9d1b1089d8342e211ca23ecddda44bf1ba9e627da676024c1640", "seconds": 11.16, "spoken_utterances": [ "How much did I spend on um... amazon? ", "I left out a detail: it's for uh... this calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('How much did I spend on um Amazon?'), but the speaker immediately adds the missing constraint ('it's for uh this calendar year'). The final intent combines the original request with the added detail, matching the reference instruction exactly. The addition type criterion is satisfied.", "heard_transcript": "How much did I spend on um Amazon? I left out a detail, it's for uh this calendar year." }, "candidate_id": "wav_b5455d953936dee7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant start ('How much did I spend on um...'), followed by a plausible self-correction ('I left out a detail') and a smooth, additive continuation ('It's for this calendar year'). The intonation and pacing reflect genuine thought process, and the addition feels like a remembered clarification rather than a forced insertion." } ] }, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_time_reference", "proposed_instruction": "How much did I spend on amazon in last calendar year? Make that this calendar year, not last calendar year.", "long_pause_offsets": [ 54 ], "source_sha256": "449a5aa7a5f8154429d4d6166240a092ca926515e5e88a18b8c75e3deeb09b8b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last calendar year' to 'this calendar year', aligning with the original instruction. The app (amazon) and action (query spending) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/revision/revision_02_time_reference/10043b9e91a82e51/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "e34f1411caad16cef333f261c305fbe5db4acbaeb9243430b760b65409354d4a", "seconds": 13.44, "spoken_utterances": [ "How much did I spend on ah... amazon in um... last calendar year? ", "Make that ah... this calendar year, not last calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'last calendar year' but immediately corrects it to 'this calendar year' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to inquire about spending on Amazon in 'this calendar year', which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "How much did I spend on uh Amazon in um last calendar year? Make that uh this calendar year, not last calendar year." }, "candidate_id": "wav_a8fabce4a4dec45d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much did I spend on...') with typical fillers ('uh', 'um') and a smooth, self-corrected revision ('Make that uh this calendar year not last calendar year'). The correction is integrated into the same utterance with a believable shift in intonation and pacing, showing a spontaneous change of mind without awkwardness or repetition. The prosody is consistent and the flow is continuous, making it sound like a real-time thought process." } ] }, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "a44c560059e32350370e750e51179abe112d97928161869ad6cce43a37fa5cd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/cancellation/cancel_019/4b700486d345f2e3/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "bc05fd8412851297a149f40d6f14375bf1ba2f3c52a3035535e9c4abab34c4c7", "seconds": 11.48, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?", "Wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but immediately cancels it with 'Wait, never mind, don't do that.' This constitutes a clear cancellation of the task. The final audible intent is to cancel, not to execute the original query.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year? Wait, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "45046fbfcc8b860f05d886b6dcf71de9af1f299faa6411011387039eac77f5e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/ambiguous_halt/halt_036/6e0d49a08743f4f5/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "bea447a01b2dc224c3950fd0a992734c304920133755f0b8e8ae5c4132f69491", "seconds": 11.12, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Wait a second before you go ahead,' which is a clear instruction to pause. The initial query is not canceled, and the final interjection is a directive to hold, not to execute or cancel. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year? Wait a second before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b48011f66c2d9e28e92b6b81166ebaf96107effb881d4703bdfb4bcad4d967bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/filler/original/a48de506a2e1b1ad/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "c15b5ed629cf30b761d54fa396e8af419d0e314cc393a44673e9b7b0a357e8fa", "seconds": 6.84, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in the last calendar month?", "long_pause_offsets": [], "source_sha256": "b043561fce559bd065f666cd563baf679857a1074d2a96566202341e26d7e780", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the time frame and intent to retrieve spending information." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/false_start/false_start_01_app/88ae93a4f89bda10/speech.wav", "wer": 0.0, "wav_sha256": "2d9b09354655b442b4c0ae4ebd5de06cc427726c22721a03bac01cb78833d0ab", "seconds": 7.52, "spoken_utterances": [ "How much did I spend on ebay—um... sorry, I meant ... amazon in the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much did I spend on eBay, um, sorry, I meant Amazon in the last calendar month?" }, "candidate_id": "wav_48a2473ff4cd0a9e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('eBay') and immediately self-corrects to 'Amazon' with a natural-sounding 'sorry, I meant'. The prosody is smooth, the correction is integrated into the flow without awkwardness, and the intonation reflects genuine recall. The phrase 'in the last calendar month' is delivered with appropriate emphasis and clarity, suggesting a spontaneous, realistic formulation." } ] }, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_02", "proposed_instruction": "How much did I spend on amazon? I left out a detail: this should apply only to the last calendar month.", "long_pause_offsets": [ 32 ], "source_sha256": "bd80c61503f78c8508259ea9d789d1f8abe1bd9eb3a8422315abf9085c4e5f37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint, refining the initial broad question to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/addition/addition_02/e34e55636bcd2778/speech.wav", "wer": 0.0, "wav_sha256": "98c2a0597b167f010cf206a14d4670249800b1ce7df5fc989e6b3361180f7272", "seconds": 13.84, "spoken_utterances": [ "How much did I spend on let me think... amazon? ", "I left out a detail: this should apply only to um, let's see... the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about spending on Amazon, then adds the constraint 'the last calendar month' after a pause and self-correction. This matches the 'addition' type criterion, where the final goal combines the original request with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "How much did I spend on, let me think, Amazon? I left out a detail. This should apply only to, um, let's see, the last calendar month." }, "candidate_id": "wav_ff912c90ac9071bd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), formulates the initial request, then adds a clarifying detail ('this should apply only to...') with a plausible follow-up pause ('let's see') before stating the time frame. The prosody is continuous and the flow reflects a genuine attempt to recall and refine a request, making it sound like a spontaneous, real-time thought process." } ] }, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_app", "proposed_instruction": "How much did I spend on ebay in the last calendar month? Wait, I misspoke about the app: use Amazon, not eBay.", "long_pause_offsets": [ 57 ], "source_sha256": "a42469761ef535af69c4cdf43ca62e38553bae24c4b183dce72f773df5786c6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from eBay to Amazon, while retaining the time frame and the action of querying spending. The original instruction's app and the final spoken instruction's app are the only differing elements, and the correction resolves the mismatch." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/revision/revision_01_app/ea8f72b958eb77a8/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "81a1674bb1b94201d6463062b040873a21a277289b2693b457c30a4c116bb3ef", "seconds": 14.2, "spoken_utterances": [ "How much did I spend on ... ebay ... in the last calendar month? ", "Wait, I misspoke about the app: use uh... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested information for eBay but immediately corrected themselves to Amazon after a pause and self-correction. The revision clearly replaces the original app name with the correct one, and the rest of the instruction remains unchanged. The final intent matches the original reference instruction, so the task should be executed as revised.", "heard_transcript": "How much did I spend on eBay in the last calendar month? Wait, I misspoke about the app. Use uh Amazon, not eBay." }, "candidate_id": "wav_8b24888a8fd3727d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "7080236019fe5cadc45e9c9d24429f0e992ad898494e9f1c81422379500444ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/cancellation/cancel_030/827a476ee9ac3835/speech.wav", "wer": 0.04, "wav_sha256": "d93282c3978e321e5425a2b6a2506cb3d01501fd3bac55361b9d20cdfdb41d7d", "seconds": 12.4, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?", "Oh, wait, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Oh wait, you can drop that task. I no longer need it.' This constitutes a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or modification of the original task. Therefore, the final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month? Oh wait, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "d07a35530b12d78d9e07678cd1126cfbfd2a9f1a47e6a889324bb6cd969fc3af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/ambiguous_halt/halt_004/e94252f66c75f83e/speech.wav", "wer": 0.0625, "wav_sha256": "c839c8b171ad6194b5c9ca4fe39feff242786e0d69030934cadd800c4d27423b", "seconds": 10.52, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then says 'Just a second,' which is a clear audible request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or revised, and the speaker does not provide a new instruction. The final intent is to halt execution temporarily.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month? Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a58076a6d629c67866efe7d02e22e800321f9137a1d6f04956f4b37df708f258", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/filler/original/2415a5576c36d8d7/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "442fea92f36b40b0b23b37bbd2d53cc08f7bb059122962d1d77b4b9a2bf800dc", "seconds": 7.2, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in this or the last calendar month?", "long_pause_offsets": [], "source_sha256": "58da42790d0dea596b793d0061a8eb04d361c69f64135e13258b7d0d2df81504", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the time frame and intent to retrieve spending information." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/false_start/false_start_01_app/4f1091d827561915/speech.wav", "wer": 0.1, "wav_sha256": "c445ca605c336339f173529e617e76b6daabf52263e7dc0f06761934252ca8a7", "seconds": 10.0, "spoken_utterances": [ "How much did I spend on ebay—uh... sorry, I meant amazon in um... this or the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. This is a clear false start where the final intent is the corrected version. The time frame 'this or the last calendar month' is preserved. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "How much did I spend on eBay up, sorry, I meant Amazon in um, this or the last calendar month?" }, "candidate_id": "wav_bc8b72f0680db47b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('How much did I spend on eBay...'), immediately self-corrects with a believable 'sorry, I meant Amazon' that matches the intended target, and continues with a plausible, slightly hesitant phrasing ('in um...') before the clip ends. The flow from error to correction to continuation is smooth and mimics genuine spontaneous speech." } ] }, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How much did I spend in this or the last calendar month? Oh, I forgot to mention, I want to know about Amazon.", "long_pause_offsets": [ 57 ], "source_sha256": "53937311c88be638b840e83213a83aa367a87c987a5c6d202d8165a35300b9f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context about Amazon, which was omitted in the initial utterance. The time frame and the action of querying spending remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/addition/addition_01/7d1e86012a835be2/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "1832fbe4be3b1fc058b2435558b04a2c4c7bf2b5a6981960f120d601cab3a9bf", "seconds": 14.68, "spoken_utterances": [ "How much did I spend in uh... this or the last calendar month? ", "Oh, I forgot to mention, I want to know about uh, let's see... Amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about spending in 'this or the last calendar month' without specifying a merchant. After a pause and an interjection, they add 'Amazon' as a new constraint. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent correctly combines the time frame with the merchant, matching the original instruction's intent.", "heard_transcript": "How much did I spend in the this or the last calendar month? Oh, I forgot to mention, I want to know about uh, let's see, Amazon." }, "candidate_id": "wav_4ab416ebf0b05b38", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then pauses and adds a specific detail ('Amazon') with a believable 'let's see' filler. The intonation and rhythm are consistent with spontaneous speech, and the addition flows smoothly without awkwardness or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_time_reference", "proposed_instruction": "How much did I spend on amazon in this calendar month? Make that 'this or the last calendar month', not just 'this calendar month'.", "long_pause_offsets": [ 55 ], "source_sha256": "0b0889d78a33422bfb9d8b884fd80a37e3d158e9fd8d3f4cfa996f4efa49a17b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time range from 'this calendar month' to 'this or the last calendar month', which matches the original instruction. The app (amazon) and action (query spending) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/revision/revision_02_time_reference/29a60fd3dcb78a15/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "a9bf4c798100ebd5ee8f0ff7ce28da5da361a5b45928213b98f36fdd3b306040", "seconds": 14.04, "spoken_utterances": [ "How much did I spend on um... amazon in uh, let me think... this calendar month? ", "Make that 'this or the last calendar month', not just 'this calendar month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for spending in 'this calendar month', then immediately revises the request to 'this or the last calendar month, not just this calendar month'. This is a clear revision of the time constraint. The final intent matches the original instruction, so the task should be executed as revised.", "heard_transcript": "How much did I spend on um Amazon in uh let me think this calendar month. Make that this or the last calendar month, not just this calendar month." }, "candidate_id": "wav_73320c2d77ac3466", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant formulation ('How much did I spend on...') and includes a plausible 'let me think' pause. The revision ('Make that this or the last calendar month') is a smooth, self-corrected rephrasing of the time frame, delivered with a believable change-of-mind intonation and a clear reference back to the original request. The prosody is continuous and the pacing realistic, with no signs of artificiality or disjointedness." } ] }, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "1d609c7dff15e4b1eea3dd5f31ecf3f472c8c8351b20867b1a60d641a2c43289", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/cancellation/cancel_039/c5905ad28ea6cc31/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "e7a8920551d25a314656c63c07943b6460809368a8b58ae78f7978b1c803ff67", "seconds": 11.8, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?", "Hold on, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, don't go ahead with that. Cancel it.' This is a clear and final cancellation of the initial request. The control intent matches the 'cancellation' type criterion, as the task is revoked and not executed.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month? Hold on, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "2ce9579b3970a2f4b393b0541c05b0d41e4ed8bd51f172a9f3cb3aca7aca0d0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/ambiguous_halt/halt_022/1eff909463e91470/speech.wav", "wer": 0.05, "wav_sha256": "5be2147c83524352fed19fa6f1b1cb056da867b9c12d3a52d7c07351398f3443", "seconds": 11.36, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. No cancellation or execution is requested, and the initial query is not revoked. The final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month? Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "eb5ad85_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f29f41f585ae800b7a48c3bb184ca22030257ef25e5f232b75d40af19e0b2c56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/filler/original/11f1aae34a827ba4/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "44601832083d0003c644ba9757a2042fce280e04f26bd2ef10dcca2da4ace655", "seconds": 6.88, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker repeats 'in in' as a filler, which is a common hesitation pattern. The core request remains unchanged and fully intact. The repetition does not alter the goal, so the final intent matches the original instruction. The type is 'filler', and the final control intent is to execute the task.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "How many read email threads are in my Outlook—I should say Gmail inbox?", "long_pause_offsets": [], "source_sha256": "73bee7e6a6d722b0cae3d90dcd7b7c0faa472ebaa0c6aae76facaa8c8858c534", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail inbox' within the same utterance, preserving the rest of the query including the metric (read email threads), the scope (inbox), and the intent (count)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/false_start/false_start_02_app/9fc8e4b1f5f0fa7c/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "587b6b06b557f9fd7c2aebf32a3b6338a82b3a157bb81e40c7eae100b509e421", "seconds": 9.12, "spoken_utterances": [ "How many read email threads are in my um... Outlook—uh... I should say ah... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many read email threads are in my um Outlook uh of I should say uh Gmail inbox?" }, "candidate_id": "wav_f93c9b3a4fd2efaf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction from 'Outlook' to 'Gmail'. The false start is brief and the correction is smoothly integrated into the sentence. The prosody is consistent, and the speaker sounds like they are formulating the request in real time without robotic pauses or unnatural word boundaries." } ] }, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail inbox? Oh, I forgot to mention, only count the read email threads.", "long_pause_offsets": [ 46 ], "source_sha256": "efebe0baecd57180d3eb6d68785bc4afa75869bedab7e84500e3d705146e0e1c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count the read email threads', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/addition/addition_01/3623ed455790b455/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "aa0cfede64baee23f28726cdae18b588bd7b613812064fa1f35fac64756b4c83", "seconds": 13.36, "spoken_utterances": [ "How many email threads are um... in my Gmail inbox? ", "Oh, I forgot to mention, only count the let's see... read email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is a general count of email threads. The speaker then adds a constraint to only count 'read email threads'. This addition refines the original request, and the final audible intent is to count only the read threads. The addition is clear and unambiguous, satisfying the 'addition' type criterion.", "heard_transcript": "How many email threads are um in my Gmail inbox? Oh, I forgot to mention, only count the, let's see, read email threads." }, "candidate_id": "wav_07d9fa3960e78000", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question about email threads. The addition of 'Oh, I forgot to mention' followed by a pause and a specific clarification ('only count the... let's see... read email threads') is a highly plausible spontaneous memory recall. The filler 'let's see' and the slight hesitation before specifying the condition sound authentic, and the overall prosody is smooth and conversational." } ] }, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_email_folder", "proposed_instruction": "How many read email threads are in my Gmail sent? Change the folder to inbox instead of sent.", "long_pause_offsets": [ 50 ], "source_sha256": "be6caf8340cb3dcca5d4888557219fc7af18cc94736734380af358ca867393a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from 'sent' to 'inbox', preserving the action, the entity (read email threads), and the app (Gmail)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/revision/revision_03_email_folder/073ff67cd7325d7d/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "9645b7a2784d724999525848fd2f6717d69498a232b1cae523fd0990f4ceea53", "seconds": 13.44, "spoken_utterances": [ "How many uh... read email threads are ... in my Gmail sent? ", "Change the folder to let's see... inbox instead of sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'Gmail sent' but then corrects to 'inbox' after a pause, which is the final intent. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final goal matches the original instruction.", "heard_transcript": "How many uh read email threads are in my Gmail sent? Change the folder to let's see inbox instead of sent." }, "candidate_id": "wav_8561a32fa6e7fe13", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a hesitant 'how many uh' and a filler 'uh', then asks about read email threads in 'sent'. After a pause, they revise the folder to 'inbox'. While the revision is present, the second part of the utterance sounds more like a follow-up thought than a natural self-correction. The intonation and pacing of the revision are slightly abrupt, and the phrase 'change the folder to let's see inbox' feels more like a task instruction than a spontaneous correction, reducing its realism compared to audio_01." } ] }, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "10ac01bda22e844ca3aac72f20a988afb1acf2265b725917e5ffcc602422c27b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/cancellation/cancel_012/4d0f30fc362a458f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "b5016d7acf976b3abed9af905938918183f2689d195522300cbde78a4a90ac88", "seconds": 11.6, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the task with 'Wait, actually, I've changed my mind. Don't do that.' This constitutes a direct cancellation of the original request. The final intent is to cancel, not to execute the counting task.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox? Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "95b8ee26be82e6b67f5e4300f66a72fca38e8af6dd8d9f145da8174045cae19a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/ambiguous_halt/halt_032/4f4d983280dd4a38/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "0b5ad73e4638718aed89cf90e81e77499bdcaa57e0bbbc744d5f622ef3dcbb3a", "seconds": 9.72, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then immediately says 'Wait, before you do that,' which is a clear request to pause the current action. The final interjection is a pause command, not a cancellation or a new instruction. The audio ends with this request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox? Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f5e7b573e2889c590e349806155632c9efa15d21ec2b96d49adcdce99e271140", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/filler/original/ce4b9d3d35e01740/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "03f2169ac672bbf93b9a39b7ea0b0832d1e53c863b4d7161b8af836cb4536b36", "seconds": 5.6, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent is to count read email threads in the Gmail inbox, matching the original instruction. The fillers do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_email_state", "proposed_instruction": "How many read—no, I mean unread email threads are in my Gmail outbox?", "long_pause_offsets": [], "source_sha256": "8ebdc4424bd3a7c227fbf08e49eab0eb74f9cabd524f869cb08dcaf3af6fc899", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'read' to 'unread' within the same utterance, preserving the app (Gmail), the object (email threads), and the location (outbox)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/false_start/false_start_01_email_state/4148469929c78f11/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "a1a0d22b31c3441541e0eea0ede0926a827cc82924a9724bb1ebb58ad5948bcb", "seconds": 5.32, "spoken_utterances": [ "How many read—uh... no, I mean unread email threads are in my Gmail ... outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'read' but immediately self-corrects to 'unread' with 'no, I mean unread'. This is a false start where the final, corrected value is the intended goal. The rest of the instruction matches the original, so the final intent is clear and executable.", "heard_transcript": "How many read uh, no, I mean unread email threads are in my Gmail outbox?" }, "candidate_id": "wav_e735cb0d2112226e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural false start with 'read' followed by an immediate self-correction to 'unread'. The correction is smooth and spoken with a realistic, slightly hesitant tone. The rest of the sentence flows without further issues, and the speaker maintains a consistent, believable pace and prosody throughout." } ] }, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail outbox? Oh, I forgot to mention, only count the unread email threads.", "long_pause_offsets": [ 47 ], "source_sha256": "bc811b322114d20b7a55c42347db9d6c0836b116f0617fd612e11d9df3a43200", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'unread' to the initial question, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/addition/addition_01/2ce6dcbad4e92582/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "015a5d554100ee7dfb7492538089671a4cc34a7e8e434729ebf4500d860391d6", "seconds": 12.0, "spoken_utterances": [ "How many email threads are in my Gmail um... outbox? ", "Oh, I forgot to mention, only count the ... unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was to count email threads in the Gmail outbox, which is a broad request. The speaker then adds a constraint to only count unread email threads. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final audible intent is to count unread email threads, which is a narrower version of the original request and fully satisfies the addition type criterion.", "heard_transcript": "How many email threads are in my Gmail um, outbox? Oh, I forgot to mention, only count the unread email threads." }, "candidate_id": "wav_779e150fcc901d1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The pause and 'um' are brief and plausible, followed by a self-correcting addition ('Oh, I forgot to mention...') that smoothly integrates the new constraint ('only count the unread email threads'). The intonation and pacing reflect genuine thought processing, and the correction is grammatically and semantically coherent within the context." } ] }, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_email_folder", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Change the folder to outbox, not inbox.", "long_pause_offsets": [ 53 ], "source_sha256": "21b7ab4cd522b67232122920cfb9054d4696547be305d0a35a102625505dfabc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from inbox to outbox, preserving the action, metric, and app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/revision/revision_03_email_folder/182b1569fe59aa93/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "0baabc2e9bb61c83c38509309e5eabac53131d137124bb46520218724425285b", "seconds": 11.84, "spoken_utterances": [ "How many um, let's see... unread email threads are in my uh... Gmail inbox? ", "Change the folder to outbox, not inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'inbox' but immediately corrects the folder to 'outbox' with the phrase 'Change the folder to outbox, not inbox.' This constitutes a revision where the final intent is to query the outbox, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "How many um, let's see, unread email threads are in my uh Gmail inbox? Change the folder to outbox, not inbox." }, "candidate_id": "wav_5d813af545ed71cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a self-pacing phrase ('let's see'), followed by a clear initial request. The correction ('change the folder to outbox, not inbox') is delivered as a smooth, immediate follow-up with a natural tone shift, indicating a spontaneous realization. The prosody is consistent, and the correction is integrated into the flow without awkwardness or repetition." } ] }, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "c576ddc884f349fa266fc058911395191812450d2aecad641ebaaa5ad8af7248", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/cancellation/cancel_001/0e788788c57e0f83/speech.wav", "wer": 0.0, "wav_sha256": "d5e75ec667b7a2d320b7c278287a86481e66be4e66aa96006c4bbf3fe391b3f7", "seconds": 9.88, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?", "Wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Wait, cancel that,' which is a clear and direct cancellation of the task. The final interjection 'Um' does not introduce a new task or modify the cancellation. The audio ends with a cancellation intent, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "How many unread email threads are in my Gmail outbox? Wait, cancel that. Um" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "007ad2cc2b2b1bb4ec73f36817492983eb751816001eb035da61351af403a005", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/ambiguous_halt/halt_016/1f708f1fe403b408/speech.wav", "wer": 0.0, "wav_sha256": "dcf1a6d5add9de98b9880d60cf5cdacce7d24cdc17f5e79552a9fea711427d38", "seconds": 9.32, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the user does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "How many unread email threads are in my Gmail outbox? Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9eb3bb845076bf27aa79a24bdee7f5187cec03777ef6e5e5540e499708a0ab96", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/filler/original/939e53f0303b4858/speech.wav", "wer": 0.0, "wav_sha256": "1962ff2efecbed21d13e3c9382629facdb2131a89e9b7fb2f624a88091a9580a", "seconds": 5.84, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchannel, or repair. The speaker delivers the full request clearly and directly, matching the reference instruction. As this is a 'filler' type task, any initial hesitation or pause within the instruction is ignored, and the final goal is to execute the complete request as heard.", "heard_transcript": "How many unread email threads are in my Gmail outbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_email_state", "proposed_instruction": "How many read—no, I meant unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "97567802033a201be991ad656e9df2ac1e23c7821a472e20ddfe20b3c2c77932", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'read' to 'unread' within the same utterance, preserving the rest of the query including the app (Gmail), the object (email threads), and the context (inbox)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/false_start/false_start_01_email_state/a70a024834fcd4df/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "aa61bcb834f93f5b9ced42b0fac837fbb17b5fe6008b8a0650bc8e35cc154214", "seconds": 11.24, "spoken_utterances": [ "How many um... read—uh... no, I meant hmmm... unread email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'red' but immediately self-corrects to 'unread' after a pause and a laugh. The final, corrected value 'unread' is the intended target. The rest of the instruction matches the original. This is a clear false start where the correction is applied to the initial request.", "heard_transcript": "How many um red uh no, I meant unread email threads are in my Gmail inbox?" }, "candidate_id": "wav_195fad98d7e37b32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start with 'red' followed by a self-correction to 'unread' and a brief, genuine laugh. The prosody is fluid, the pause before the correction is realistic, and the intonation suggests spontaneous thought. The correction is immediate and integrated smoothly into the sentence, making it sound like a real-time mental correction." } ] }, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail inbox? Oh, I forgot to mention, only count the unread email threads.", "long_pause_offsets": [ 46 ], "source_sha256": "fd954453fa265184322c30ebc134385fc9ff5bae3932b05e0b7f5f6402440ac8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count the unread email threads', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/addition/addition_01/f7e97292b60e68d9/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "22887b74377a28ab7aa77cf8d1fcf75bd88b637a256e8cd661f9c6e3d99f26ac", "seconds": 12.84, "spoken_utterances": [ "How many email threads are in my let me think... Gmail inbox? ", "Oh, I forgot to mention, only count the um... unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How many email threads are in my Gmail inbox?' and was followed by an addition specifying 'only count the unread email threads.' This addition narrows the scope of the original request to only unread threads, which matches the final intent of the reference instruction. The addition is clear and directly modifies the original task without introducing conflicting or incorrect constraints.", "heard_transcript": "How many email threads are in my let me think Gmail inbox? Oh, I forgot to mention, only count the um unread email threads." }, "candidate_id": "wav_d92125f77741f2ce", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly natural and plausible sequence of self-correction and addition. The speaker begins with a clear request, hesitates with a believable 'let me think' pause, then adds a specific restriction ('only count the unread email threads') after a second pause. The prosody includes natural intonation shifts, filler words ('um'), and a smooth, continuous flow that mimics genuine spontaneous speech. The addition feels like a remembered detail rather than a rehearsed command." } ] }, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_02_app", "proposed_instruction": "How many unread email threads are in my Outlook inbox? Let me correct the app: check Gmail, not Outlook.", "long_pause_offsets": [ 55 ], "source_sha256": "5aea73f5b031c59cd73b14b83712398707ed3d62b6eaabb05038386013034f4e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Outlook to Gmail, preserving the original query about unread email threads and the inbox context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/revision/revision_02_app/441633135a2edf38/speech.wav", "wer": 0.05, "wav_sha256": "31a08928ede55cb955a29e0dd823671d94188a8955d8c4ca6327ad10ecb8cd2c", "seconds": 12.16, "spoken_utterances": [ "How many um... unread email threads are in my Outlook inbox? ", "Let me correct the app: check ... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Outlook but immediately corrects the target application to Gmail. The revision is clear and final, and the rest of the instruction (counting unread email threads in the inbox) remains unchanged. The final intent matches the original reference instruction.", "heard_transcript": "How many um unread email threads are in my Outlook inbox? Let me correct that. Check Gmail, not Outlook." }, "candidate_id": "wav_acf4e4c6a10cacaa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear but slightly hesitant request ('How many um unread email threads are in my Outlook inbox'), then self-corrects with a believable 'Let me correct the app' followed by a smooth, contextually appropriate correction to 'Gmail not Outlook'. The prosody is consistent, the pause before the correction feels like genuine recollection, and the intonation reflects a spontaneous change of mind rather than a rehearsed or robotic edit." } ] }, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "d76ca8f07e5b622a7fd028cacd1642d7793e1eec0b816fd36e5b37b58cf79696", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/cancellation/cancel_023/fc4c85a32f68f87c/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "e82c599c031f202f6a3f34b30e11402bad16fd914b9764ffb6a553b4e0ca3e21", "seconds": 12.76, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly instructs not to execute the request, satisfying the cancellation type criterion.", "heard_transcript": "How many unread email threads are in my um Gmail inbox? Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "4ede491384041ee352659aaeed5b56489e1c7a27a94f83e0fdd560ef0c28407e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/ambiguous_halt/halt_002/0eeb718341e14428/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "9f26e7add47cf5c0141bc2c0a8914989f052f4336a55c7c68fbf72272fc22d98", "seconds": 9.72, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled, but the speaker explicitly halts further action.", "heard_transcript": "How many unread email threads are in my um Gmail inbox? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "f691597_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d6b556f2851b8e5d9f1b91403e652e733d6ece66f46024cc67a0895d69b67ddc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/filler/original/abc49c8643fa99bb/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "229018add82654cdafc7694cce3c8258c95a24079bcaeda767da0189d7f8c3e3", "seconds": 6.64, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent is to count unread email threads in the Gmail inbox, matching the original instruction. The filler does not alter the task goal.", "heard_transcript": "How many unread email threads are in my um Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_01_contact_group", "proposed_instruction": "Add all my family—I meant friends and roommates as friends on venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "2453711e96b061a415a5e75894e1b3679a22f3835f1df1caf0b311d906c8d437", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends and roommates' within the same utterance, preserving the app, action, and conditional scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/false_start/false_start_01_contact_group/a59836a908e5f589/speech.wav", "wer": 0.1, "wav_sha256": "cdf454030d9baa29af4b8e93a373a1c44190579ab532190d747a3d09571d4545", "seconds": 10.0, "spoken_utterances": [ "Add all my uh... family—hmmm... I meant friends and roommates as friends on venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends and roommates'. According to the 'false_start' type criterion, the final corrected intent overrides the initial mistake. The corrected instruction matches the original reference exactly, including all conditions and targets. Therefore, the task should be executed as stated in the correction.", "heard_transcript": "Add all my uh family, I meant friends and roommates as friends on Venmo, if they are not already." }, "candidate_id": "wav_0c8c5d4c08ecd103", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'family' and immediately self-corrects to 'friends and roommates' with a smooth, conversational tone and a believable 'I meant' phrase. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pacing and intonation suggest genuine thought-in-progress, making it the most realistic." } ] }, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Add all my friends and roommates as friends on venmo,. And make sure to add them only if they are not already friends.", "long_pause_offsets": [ 55 ], "source_sha256": "93b756985fe3381884b05c03a20507a291f3a9cafabdc010782f9a7983c2389e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the condition 'only if they are not already friends', which clarifies the original instruction's implicit intent. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/addition/addition_01/ae635dc5b1333161/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "2f695809eb981de26b9ad92c14feb32632cfe43039b4f67ea1e2c5a1f15af1fe", "seconds": 13.68, "spoken_utterances": [ "Add uh... all my friends and roommates as friends ... on venmo,. ", "And make sure to add them ... only if they are not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The final instruction combines the original request with the added constraint 'only if they are not already friends', which is a clear addition that narrows the scope of the original task. The final audible intent matches the reference instruction with the added condition preserved. No other issues are present.", "heard_transcript": "Add uh all my friends and roommates as friends on Venmo, and make sure to add them only if they are not already friends." }, "candidate_id": "wav_91eda236961d50f4", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_deduplication_condition", "proposed_instruction": "Add all my friends and roommates as friends on venmo, if they are already. That should be 'if they are not already', not 'if they are already'.", "long_pause_offsets": [ 75 ], "source_sha256": "7a709374228f447d9205691da65e9eac43d7bf38c2efb330785b40868d06039a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'if they are already' to 'if they are not already', preserving the action, target group, and app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/revision/revision_03_deduplication_condition/0e6f36cf6124ccbd/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "142ec5f4ab1a4f9fb07ea774d1e5a39ab1afc7eda568729af0e39e16b7c9ddac", "seconds": 19.56, "spoken_utterances": [ "Add hmmm... all my friends and roommates as friends on venmo, um, let's see... if they are already. ", "That should be 'if they are not already', not 'if they are already'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately self-corrects to 'if they are not already' after a pause and explicit revision. The final audible intent matches the original instruction, so the revision criterion is satisfied.", "heard_transcript": "Add all my friends and roommates as friends on Venmo, um, let's see, if they are already. That should be, if they are not already. Not if they are already." }, "candidate_id": "wav_dc68891ba38842cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with a clear false start ('At'), a thoughtful pause ('Hmm'), and a self-correction of the condition from 'if they are already' to 'if they are not already'. The intonation and pacing reflect genuine deliberation, and the correction is smoothly integrated into the ongoing utterance, making it highly realistic." } ] }, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "bc1e35177b87a9306d1d0815296610788b9ef9c8f49cafddb80d51801e0b5ca7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/cancellation/cancel_028/168e0e8b18caff03/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "c71d17fa1223a2037dee94d64051033a8b69f814fc3f014817b9f5d7ffb55d2a", "seconds": 16.92, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already.", "Hang on, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Hang on disregard that instruction don't carry it out.' This constitutes a direct cancellation of the entire task. The final intent is unambiguous and matches the cancellation type criterion.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already. Hang on disregard that instruction don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "ef255626f04caf5774acf3b2513447110b2b82f3a71caadbf5e87aa78e630fb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/ambiguous_halt/halt_006/062fd55da5865003/speech.wav", "wer": 0.16, "wav_sha256": "d73282d603c8595b72e6489ba51e97b6269a895aa00f38cd9278e9131344c111", "seconds": 15.16, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a moment', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed, and the final interjection is a pause command. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already just a moment" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_1", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7beee4c9012eb26ba680eb67e8ff0262e539bf726d52cd27425fbb0990816a57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/filler/original/0f573c75a3fa3111/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "2042639f35e8d1cb9bae8cddf719fcc70d8f2845aa62c9b82b1075d9b46812ca", "seconds": 11.8, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let me think') and a brief pause but does not change the core request. The final audible intent matches the original instruction. The hesitation and backchanneling are typical of a filler type, and the task goal remains consistent.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Add all my coworkers and friends as friends on cashapp—make that venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "434dc5550100a7c428212be6573841c6414b93a631b0c057c219fbb6f1aa5898", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the action, target group, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/false_start/false_start_02_app/4fb31271172ca763/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "6b4b678c96b9698e60d01f2cfa15121274a100c7497eada15dab729617e41426", "seconds": 9.48, "spoken_utterances": [ "Add all my coworkers and friends as friends on cashapp—ah... make that ... venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'Make that, Venmo'. This is a false start where the final value is the corrected one. The rest of the instruction matches the original, including the conditional 'if they are not already'. The final intent is clear and executable.", "heard_transcript": "Add all my coworkers and friends as friends on Cash App. Make that, Venmo, if they are not already." }, "candidate_id": "wav_3e360c404ccfd26c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Cash App', then self-corrects to 'Venmo' with a clear, spontaneous-sounding 'Make that' transition. The prosody is consistent, the pace is conversational, and the correction feels like a genuine moment of recall. The final phrase 'if they are not already' is delivered smoothly without hesitation, completing the thought naturally." } ] }, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Add all my coworkers and friends as friends on venmo,. Oh, I forgot to mention, only add them if they are not already friends.", "long_pause_offsets": [ 55 ], "source_sha256": "d90ed5aa2e636123367478dfd5ab708716c991767ec30da838f399fc390e6c48", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only if they are not already friends', which refines the initial broad request. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/addition/addition_01/dc55fc0b609ee287/speech.wav", "wer": 0.16, "wav_sha256": "0fd800ac4819e7ba225652fd9047d42db3d1d2d3aaef1c71fab1f8080b5c8c26", "seconds": 10.64, "spoken_utterances": [ "Add all ... my coworkers and friends as friends uh... on venmo,. ", "Oh, I forgot to mention, only add them if they are not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to add all coworkers and friends as friends on Venmo. After a pause, the speaker added the constraint 'only add them if they are not already friends.' This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the new condition, matching the expected behavior for an 'addition' type. The constraint is explicitly stated and preserved.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo. Oh, I forgot to mention, only add them if they are not already friends." }, "candidate_id": "wav_0e1db0c1ddf6c33e", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_01_contact_group", "proposed_instruction": "Add all my family members as friends on venmo, if they are not already. Wait, I meant coworkers and friends, not family members.", "long_pause_offsets": [ 72 ], "source_sha256": "cc0d1bf02add5854c8fad1ff5ba0cfe1e1ab20198b39c4e55941290681b8be28", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family members' to 'coworkers and friends', preserving the app, action, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/revision/revision_01_contact_group/5c3b285964cbb478/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "4d2bc5ec4497ff161c0794559f82670e36e0df7cf5ee1b8a9a1eb7ba191afbf3", "seconds": 12.52, "spoken_utterances": [ "Add all my family members as friends on venmo, um... if they are not already. ", "Wait, I meant let's see... coworkers and friends, not family members." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family members' but immediately corrects to 'coworkers and friends' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, including the conditional 'if they are not already'. Therefore, the task should be executed as revised.", "heard_transcript": "Add all my family members as friends on Venmo, um, if they are not already. Wait, I meant, let's see, coworkers and friends, not family members." }, "candidate_id": "wav_5bb24c9b1cf400a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and plausible revision flow. The speaker starts with a clear request, then after a brief pause and a self-correcting 'wait, I meant', smoothly revises the target group from 'family members' to 'co-workers and friends'. The intonation and pacing suggest genuine hesitation and correction, with a believable change-of-mind. The prosody is consistent and the delivery is continuous without robotic or disjointed segments." } ] }, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "fb70bd1d21f46a2bd2a2ee1b26ae46d2e7f78abd3e49a9c8acf440a4481fd3fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/cancellation/cancel_018/57ac4c9c206e0ca5/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "08497a6b3deef24e016a36c9f91568fb8d09ad11bd5234cb4c7dbef2f876e2c4", "seconds": 13.48, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already.", "Oh, wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to add coworkers and friends on Venmo, then clearly revokes it by saying 'Oh wait, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already. Oh wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "f4e6ce39cadd22f9f64565d7a52b0e8913bb6a58e988d9fe8bd3129405516b37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/ambiguous_halt/halt_033/3eff73d1a736b74f/speech.wav", "wer": 0.2, "wav_sha256": "be4d1e8d89d97ef359db19d8f5c46f8e2ab2f5ea006b2a738709928dcee7edc9", "seconds": 11.6, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something', which is a clear request to pause or wait. There is no cancellation or instruction to resume. The initial task is not completed, but the final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already. Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_2", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6f85ffb8c3feea17ca36b5f63f4edddf8c52316b50e9ab4d690352fdd4caad3f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/filler/original/789801a4ac67976a/speech.wav", "wer": 0.25, "wav_sha256": "a4b8d2ec1600975a20f87e5d2afbc3db67fa94603ace02c56de3050efd2b43d5", "seconds": 7.72, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'um' but does not change the task or add any new conditions. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "false-start", "source_type": "false_start", "selected": { "variant_id": "false_start_02_app", "proposed_instruction": "Add all my coworkers and roommates as friends on cashapp—sorry, make that venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "6e5e3c351f3e44a6f2d8790146db9a435a387825c63ac5f01cfe22eb50b56eda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'cashapp' but immediately corrects to 'venmo'. The app is the only changed element, and the correction is clear and complete. All other conditions—target group (coworkers and roommates), action (add as friends), and condition (if not already)—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/false_start/false_start_02_app/3dad92aa1ca4e834/speech.wav", "wer": 0.3, "wav_sha256": "1f012c2533f4f4b8221226d43a7b62d104d4a3973ff887609d569d8c2212bb3f", "seconds": 12.12, "spoken_utterances": [ "Add all ... my coworkers and roommates as friends on cashapp—hmmm... sorry, make that venmo, ... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'sorry, make that Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction with the correct app name.", "heard_transcript": "Add all my coworkers and roommates as friends on Cash App. Um, sorry, make that Venmo, if they are not already." }, "candidate_id": "wav_c6dafdf0ec541e4c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Cash App', then self-corrects to 'Venmo' with a clear, spontaneous-sounding 'sorry, make that Venmo'. The intonation and pacing of the correction sound like a genuine moment of recall, and the overall delivery is continuous and conversational without awkward pauses or robotic phrasing." } ] }, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "addition", "source_type": "addition", "selected": { "variant_id": "addition_01", "proposed_instruction": "Add all my coworkers and roommates as friends on venmo,. Oh, I forgot to mention, only add them if they're not already friends.", "long_pause_offsets": [ 57 ], "source_sha256": "d1f9e6cca5c59d9db4240e59534771ddacfd2933fb57d5201a1bf786045e528c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only if they're not already friends', which refines the initial instruction without contradicting it. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/addition/addition_01/2f2aa619e0910de4/speech.wav", "wer": 0.125, "wav_sha256": "4c53da8405749a3292adcce71d95015bee3dda716aca2927bf74cb78e5b52003", "seconds": 12.64, "spoken_utterances": [ "Add all um... my coworkers and roommates as friends ... on venmo,. ", "Oh, I forgot to mention, only add them if they're not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to add coworkers and roommates as friends on Venmo with the condition 'if they are not already'. The speaker initially omitted this condition, then added it as an afterthought. According to the 'addition' type criterion, the final goal is the combination of the initial instruction and the added constraint. The final audible intent matches the original instruction exactly, including the conditional clause.", "heard_transcript": "Add all um my coworkers and roommates as friends on Venmo. Oh, I forgot to mention, only add them if they're not already friends." }, "candidate_id": "wav_fafe96a18aa63018", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a plausible self-correction. The initial request is clear, followed by a believable 'Oh, I forgot to mention' filler that introduces a revision. The prosody is conversational, with appropriate pauses and intonation, and the correction flows smoothly without sounding rehearsed or robotic. The addition of the condition feels like a spontaneous afterthought." } ] }, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "revision", "source_type": "revision", "selected": { "variant_id": "revision_03_deduplication_condition", "proposed_instruction": "Add all my coworkers and roommates as friends on venmo, if they are already. Make that 'if they are not already' instead of 'if they are already'.", "long_pause_offsets": [ 77 ], "source_sha256": "c92fb20fbf81c55299fa4c0d9fd0b79858033cbb3d2f15919d2ea40443dbf42a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the condition from 'if they are already' to 'if they are not already', preserving the action, target group, and app. The original instruction's condition is fully corrected." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/revision/revision_03_deduplication_condition/256e7a54428fce65/speech.wav", "wer": 0.2, "wav_sha256": "667a8d44697557979eb3a9a0084d05ad159505f27c6db07bc7ca478478cbe77f", "seconds": 19.84, "spoken_utterances": [ "Add all um... my coworkers and roommates as friends uh... on venmo, hmmm... if they are already. ", "Make that 'if they are not already' instead of 'if they are already'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately corrects themselves to 'if they are not already' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to add all coworkers and roommates as friends on Venmo only if they are not already friends, which matches the original instruction. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Add all um my coworkers and roommates as friends uh on Venmo who if they are already. Make that if they are not already instead of if they are already." }, "candidate_id": "wav_328cbe35715187c0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um', 'uh') and a self-corrected phrase ('if they are not already instead of if they are already'). The prosody is consistent, the correction is smoothly integrated, and the pacing reflects genuine thought processing. The speaker sounds like a real person formulating and refining a task." } ] }, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "cancel", "source_type": "cancellation", "selected": { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "ecfdcf199f7dc2378a5bb3ff0b73a5ef02820e6c54220713e48573b381f4a0d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/cancellation/cancel_040/be77c986b04c71f0/speech.wav", "wer": 0.1, "wav_sha256": "b8fdc81e795a58db41827950ab929102aaf0003cee2256de00fb90503d60a78b", "seconds": 16.2, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already.", "Oh, wait, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to add coworkers and roommates as friends on Venmo. After a brief pause, they explicitly state, 'Oh wait, I don't want to proceed with that task. Please cancel it.' This constitutes a clear cancellation of the entire task. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Oh wait, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "halt", "source_type": "ambiguous_halt", "selected": { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "92dde09775c597869782261128dbbd9a0d022e52964e24c5e51086fd24beb8b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/ambiguous_halt/halt_018/e19d89c4c12ca8ed/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "81ae2cf2bb4cf5c8c6ec3042c448937a3ee0ec6f2ff375a6e7ad24dcf8e27193", "seconds": 13.6, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original instruction and then clearly requests a pause with 'Could you pause for a second?'. This constitutes a valid ambiguous_halt as the final interjection is a direct request to hold/wait, with no cancellation or resumption permission. The original task intent is not revoked, but execution is requested to be paused.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } }, { "task_id": "ff58e36_3", "split": "test", "type": "filler", "source_type": "filler", "selected": { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1e4b91b55ab9224c840b103b659bedd36eed064bcc8e7986e4be74477931c850", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/filler/original/5d60f954d7092a19/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "3fdce475dcf64ae4aaebaa36dda6d254f631c24678b405f1f8dc34b5a913a4f9", "seconds": 9.52, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'uh' and a pause ('Hmm'), but the final, complete instruction matches the original reference. The hesitation and backchannel do not alter the goal. The task is executed as intended.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Hmm." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } } ], "skipped": [ { "task_id": "07b42fd_2", "split": "train", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "099174d_1", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "unsupported", "candidate_count": 0 }, { "task_id": "099174d_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "unsupported", "candidate_count": 0 }, { "task_id": "099174d_3", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "unsupported", "candidate_count": 0 }, { "task_id": "0d22252_1", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "0d22252_1", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "0d22252_1", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "1150ed6_3", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "1150ed6_3", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "1150ed6_3", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "1150ed6_3", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "1c4bd27_1", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "21abae1_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "287e338_3", "split": "train", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "287e338_3", "split": "train", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "287e338_3", "split": "train", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_2", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_2", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_2", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_3", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "365e0a3_3", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "365e0a3_3", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_3", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "365e0a3_3", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "3f3c139_1", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "3f3c139_3", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "3f3c139_3", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "3f3c139_3", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "3f3c139_3", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "476b213_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "522e5e5_3", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "552869a_2", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "57c3486_1", "split": "train", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "5e27cd7_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "6a5e690_1", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "6a5e690_3", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "7574325_1", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "7574325_3", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "7574325_3", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "b7a9ee9_3", "split": "train", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "b7a9ee9_3", "split": "train", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "cb4ee7a_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 2 }, { "task_id": "cb4ee7a_2", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 3 }, { "task_id": "dac78d9_1", "split": "test", "type": "cancel", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "dac78d9_1", "split": "test", "type": "halt", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "dac78d9_1", "split": "test", "type": "filler", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "dac78d9_2", "split": "test", "type": "addition", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 1 }, { "task_id": "e775c78_3", "split": "test", "type": "revision", "reason": "no_eligible_llm_default", "source_status": "supported", "candidate_count": 4 } ], "candidates": [ { "task_id": "024c982_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_05_person_name", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Sarah—sorry, I mean Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "147d71fb5a9d20254c42a01b124c5f1107b0dd7d07c2971286a31fc14f58055b", "decision": "yes", "review_revision": "20261019T025444616630Z_bd988084", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Sarah' but immediately corrects to 'Stacy' with 'sorry, I mean Stacy'. The correction is clear and unambiguous, restoring the original recipient. All other elements—amount, app, privacy, and note content—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_05_person_name/efdc27ca6f93b461/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "d742f24dfecb5e0c85020339362f41f72872049653b0edf9b22c10f8d8abe3f6", "seconds": 12.16, "spoken_utterances": [ "Request ah... thirteen dollars publicly on Venmo from my friend, Sarah—uh... sorry, I mean um... Stacy, with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Sarah' but immediately self-corrects to 'Stacy' after a pause and 'sorry, I mean'. This is a false start where the final target is corrected. The rest of the instruction, including the amount ($13), platform (Venmo), action (request publicly), and note content, remains consistent with the original. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Request uh thirteen dollars publicly on Venmo from my friend, Sarah. Sorry, I mean um Stacy with a note, for yesterday's meal." }, "candidate_id": "wav_54abf11d1b9477d1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'uh' at the beginning, a clear self-correction from 'Sarah' to 'Stacy' with a believable 'sorry I mean um' transition, and a smooth continuation. The prosody is consistent, and the repair feels spontaneous rather than rehearsed. The pacing and intonation match a person recalling a name and correcting it mid-sentence." } ] }, "is_default": true }, { "variant_id": "false_start_02_visibility", "proposed_instruction": "Request $13 privately—no, I mean publicly on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "80c28ba22d10a1433694d891dc23eaf2b3432f1fb1b7a1dd333970c29c49a943", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'privately' to 'publicly' within the same utterance, while retaining the amount, recipient, app, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_02_visibility/070214ea61885e70/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "de53dd4bdfd1db1414799ea5b914f4a83e352fd63fe974af9ff3323372e90443", "seconds": 11.2, "spoken_utterances": [ "Request thirteen dollars ... privately—um... no, I mean publicly on Venmo from my friend, Stacy, with a note, ... \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately self-corrects to 'publicly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly, including the amount, method, recipient, and note. Therefore, the task is correctly executed.", "heard_transcript": "Request thirteen dollars privately, um, no, I mean publicly on Venmo from my friend Stacy with a note, for yesterday's meal." }, "candidate_id": "wav_28884deba7fee9c1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a clear false start with 'privately um no I mean publicly', which is a plausible self-correction. The speaker maintains a natural rhythm and the correction flows smoothly. The 'um' filler is used appropriately, and the overall delivery sounds like a real-time thought process." } ] }, "is_default": false }, { "variant_id": "false_start_04_recipient_relation", "proposed_instruction": "Request $13 publicly on Venmo from my acquaintance—sorry, I meant friend, Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "110a8cbd109e9362af81a5bb29c9d5b830580d0c2c7ab3fdc49152509e1b0a99", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately corrects to 'friend, Stacy'. The correction clearly replaces the mistaken term with the correct recipient identity. All other elements—amount, app, privacy, and note—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_04_recipient_relation/6f998e0bcef1c37d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "77c04252f33fe3ecfa50d45868c41cd25d9a87ab63c27b742cd98492145acee1", "seconds": 12.32, "spoken_utterances": [ "Request uh... thirteen dollars publicly on ... Venmo from my acquaintance—hmmm... sorry, I meant friend, Stacy, with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately self-corrects to 'friend Stacy'. According to the 'false_start' type criterion, the final corrected value ('friend Stacy') should be applied, and the original incorrect value ('acquaintance') should be discarded. All other constraints (amount, platform, public status, note content) match the reference. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Request uh thirteen dollars publicly on Venmo from my acquaintance, sorry, I meant friend Stacy with a note, for yesterday's meal." }, "candidate_id": "wav_7c5d1e1dd6a3cb18", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip includes a false start with 'uh' and a self-correction from 'acquaintance' to 'friend Stacy'. The 'sorry I meant' transition is natural, and the intonation suggests genuine recall. The pacing is slightly slower, but the repair is believable and the voice remains consistent." } ] }, "is_default": false }, { "variant_id": "false_start_06_note_text", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Stacy, with a note, \"For lunch\"—scratch that, make it \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "033f9111ce54b6fc86e092e69571068c3b0ca1a5ca9f600cb56e737383b83417", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says \"For lunch\" but immediately corrects to \"For yesterday's meal\". The final note content matches the original instruction. All other elements—amount, app, privacy, recipient, and action—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_06_note_text/cd021ee8b8517de7/speech.wav", "wer": 0.08, "wav_sha256": "6861f8ea81de80fada0f3bdd64912df9571a820554f3ef21063ae4b403e62a94", "seconds": 13.24, "spoken_utterances": [ "Request thirteen dollars publicly on uh... Venmo from my friend, Stacy, with a note, um... \"For lunch\"—hmmm... scratch that, make it \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'for lunch' but immediately self-corrects with 'Scratch that, make it for yesterday's meal.' This is a clear false start where the final value replaces the initial one. The corrected value matches the original instruction's note content. All other parameters (amount, app, sender, recipient, public request) remain consistent. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Request thirteen dollars publicly on uh Venmo from my friend Stacy with a note um for lunch. Scratch that, make it for yesterday's meal." }, "candidate_id": "wav_956a613bbb942329", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision rather than a false start. The speaker initially says 'for lunch' and then corrects to 'for yesterday's meal'. The 'scratch that' is a believable cancellation, and the correction is delivered with a natural shift in intonation. However, the initial request is slightly less fluent than the others." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Request $13 publicly on CashApp—I should say Venmo from my friend, Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "9f7a5e77e8889820e5190a8a5e18ccceeeb2ecc67b0bc204b4f1c151dc7865e6", "decision": "yes", "review_revision": "20261019T030025822627Z_40a67bc0", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the app from CashApp to Venmo within the same utterance. The amount, recipient, privacy setting, and note content are preserved in the final version." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_03_app/948690bbfd588e75/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "1f77cec5cc82b4e81bfcdbdd68e50989c2058c58ce27f8ad9340671d450f5a34", "seconds": 10.88, "spoken_utterances": [ "Request ... thirteen dollars publicly on CashApp—uh... I should say ... Venmo from my friend, Stacy, with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo'. This is a false start where the first value is replaced by the second. The final intent is to send money via Venmo, which matches the original instruction. All other conditions (amount, recipient, note) are preserved.", "heard_transcript": "Thirteen dollars publicly on Cash App. I should say, Venmo from my friend Stacy with a note, for yesterday's meal." }, "candidate_id": "wav_f6bf6a2c547382eb", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip has a false start with 'uh' and a correction from 'cash app' to 'Venmo'. The correction is clear, but the 'uh' is slightly awkwardly placed, and the overall delivery feels a bit rushed. The prosody is less natural compared to the top-ranked clips, and the repair lacks the same level of spontaneity." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Request $15—make that $13 publicly on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\".", "long_pause_offsets": [], "source_sha256": "8dd57b585ae7da6742a4a2a996a610d29079aecf83be4c7882ad261a9b791465", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces fifteen with thirteen; all other elements including recipient, app, privacy, and note content remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/false_start/false_start_01_amount/0850dd732da75ff3/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "1ea47282fb964c30e8586c38e48f0a13378c379f680ad6434460cb8dd36afd3f", "seconds": 11.16, "spoken_utterances": [ "Request fifteen dollars—uh... make that thirteen dollars publicly on Venmo from my friend, um... Stacy, with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen dollars' but immediately corrects to 'thirteen dollars'. This is a clear false start where the final value is 'thirteen dollars'. All other conditions in the original instruction are preserved and correctly stated. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Request fifteen dollar uh make that thirteen dollars publicly on Venmo from my friend um Stacy with a note for yesterday's meal." }, "candidate_id": "wav_0ebea79577158f22", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip begins with a revision ('make that $13') rather than a false start. The speaker corrects the amount, but the initial 'fifteen dollar uh' sounds hesitant and less confident. The prosody is inconsistent, and the correction feels more like a read-aloud than a spontaneous thought process. The overall delivery is less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Request $13 on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\". Oh, right, make the request public.", "long_pause_offsets": [ 81 ], "source_sha256": "d8155c8575490977f377b884c195bdd05ef53c0ffe40d4d7444bfca0a006e316", "decision": "yes", "review_revision": "20261019T025445763315Z_f89b1a7f", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the public visibility requirement, which was missing in the initial instruction. The action, amount, recipient, app, and note content are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/addition/addition_01/dec885912b3e5af5/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "ec63faa34d440db58e99083a7ecd578a035cf9271d474d85c44eb72ccd4b93d1", "seconds": 15.68, "spoken_utterances": [ "Request um... thirteen dollars on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\". ", "Oh, right, make the request uh... public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the core request, then adds the constraint 'make the request public' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "Request um thirteen dollars on Venmo from my friend Stacy with a note, for yesterday's meal. Oh right, make the request uh public." }, "candidate_id": "wav_6a22ed62b4e864e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable self-correction. The initial request is stated, followed by a pause and a revision ('Oh, right, make the request uh public') that sounds like a spontaneous afterthought. The prosody is consistent, and the correction is integrated smoothly into the utterance with a natural 'uh' filler, indicating genuine recall rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Stacy,. Oh, and one more thing: include the note \"For yesterday's meal\".", "long_pause_offsets": [ 54 ], "source_sha256": "88e663c9609cbd5dc71b23eb7c18ff17bbf44ec5ec26c9de4327e301dfb1ea9f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing note requirement, which was present in the original instruction but omitted in the first utterance. The final goal includes all original elements: amount, app, privacy, recipient, and note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/addition/addition_02/438a32b15c7fed61/speech.wav", "wer": 0.08, "wav_sha256": "a499be6acf6815c40c427bc3fb245d85712e83d04c7ba9e9bbb33005266ea9f2", "seconds": 15.92, "spoken_utterances": [ "Request um, let's see... thirteen dollars publicly on Venmo from my friend, ... Stacy,. ", "Oh, and one more thing: include the note ah... \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. The speaker then adds a new constraint ('include the note for yesterday's meal') after a pause, which is an addition. The final intent combines the original request with the added note, fully satisfying the addition type criterion.", "heard_transcript": "Request um let's see thirteen dollars publicly on Venmo from my friend Stacy. Oh and one more thing, include the note uh for yesterday's meal." }, "candidate_id": "wav_8d3b6b8f72db402f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a plausible hesitation ('um, let's see'), then states the request. However, the addition ('Oh, and one more thing...') is delivered as a separate, slightly abrupt clause rather than a fluid continuation. The prosody shifts noticeably, and the addition feels more like a post-hoc insertion than a natural thought process. While the content is coherent, the delivery lacks the seamless integration seen in audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_amount", "proposed_instruction": "Request $15 publicly on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\". Sorry, I meant $13, not $15.", "long_pause_offsets": [ 90 ], "source_sha256": "788cc2e5ee561b1f22e29ac402d6bcefdbde651049f1daa24980fc985cd69597", "decision": "yes", "review_revision": "20261019T025502946332Z_00f4218c", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the amount from fifteen to thirteen, while retaining the recipient, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_01_amount/ae96a5cb40b59d44/speech.wav", "wer": 0.12, "wav_sha256": "221d0511201e788c95a5f61b0eb4d7c296a41ffd3407c0a8e708f6d40efdb21c", "seconds": 14.52, "spoken_utterances": [ "Request ... fifteen dollars publicly on Venmo from my friend, Stacy, with a note, um... \"For yesterday's meal\". ", "Sorry, I meant ... thirteen dollars, not fifteen dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a request for $15, then immediately corrects it to $13. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent is to request $13, which matches the original instruction. All other conditions (Venmo, public, from Stacy, with note 'For yesterday's meal') are preserved and correctly stated.", "heard_transcript": "Request fifteen dollars publicly on Venmo from my friend Stacy with a note, um, for yesterday's meal. Sorry, I meant thirteen dollars, not fifteen dollars." }, "candidate_id": "wav_8eeae20b4cf754a0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural revision flow. The speaker starts with a clear request, then self-corrects the amount from 'fifteen' to 'thirteen' with a believable 'sorry, I meant' phrase. The prosody is smooth, the correction is immediate and integrated into the same utterance, and the intonation reflects a genuine moment of recollection without any robotic or rehearsed feel." } ] }, "is_default": true }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Request $13 privately on Venmo from my friend, Stacy, with a note, \"For yesterday's meal\". Wait, make it publicly, not privately.", "long_pause_offsets": [ 91 ], "source_sha256": "115ed7bc3a163ec6e13f891cef81bb9699441b203c639a6996ffdf92d72b63f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the visibility from 'privately' to 'publicly', while retaining the amount, app, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_02_visibility/38ac04320e7867ac/speech.wav", "wer": 0.16, "wav_sha256": "f28dda658f7d2c7bb890ba8144f362993af2e0c1935f97b312f64c1ea57a50b1", "seconds": 15.32, "spoken_utterances": [ "Request uh... thirteen dollars privately on Venmo from my friend, Stacy, with a note, um... \"For yesterday's meal\". ", "Wait, make it hmmm... publicly, not privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately corrects it to 'publicly, not privately' after a pause. This is a revision where the final intent is to send the payment publicly. All other conditions (amount, recipient, note) match the original instruction. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Request uh thirteen dollars privately on Venmo from my friend Stacy with a note um for yesterday's meal. Wait, make it publicly, not privately." }, "candidate_id": "wav_65362a7979095fc1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The revision here is also very natural, with the speaker initially stating 'privately' and then correcting to 'publicly' with a 'wait, make it' phrase. The hesitation and self-correction are well-placed and sound spontaneous. The only reason it ranks below audio_01 is that the correction is less integrated into the initial sentence and slightly more abrupt." } ] }, "is_default": false }, { "variant_id": "revision_04_recipient_relation", "proposed_instruction": "Request $13 publicly on Venmo from my acquaintance, Stacy, with a note, \"For yesterday's meal\". I misspoke about the relationship: it should be friend, not acquaintance.", "long_pause_offsets": [ 96 ], "source_sha256": "06aa5748e92e57947a1f6734b79420fe08bd61cd3eb74a343ac832a91107fd13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the descriptor of the relationship from 'acquaintance' to 'friend', preserving the recipient (Stacy), the amount, the app, the public visibility, and the note content. The correction is explicit and does not alter any other aspect of the request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_04_recipient_relation/083b86d03beb1ae3/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "98769f7a51791b77b3653c9c4f20525a919e8a76c3905c6f5c2381651ec83719", "seconds": 15.32, "spoken_utterances": [ "Request thirteen dollars publicly on um... Venmo from my acquaintance, Stacy, uh... with a note, ... \"For yesterday's meal\". ", "I misspoke about the relationship: it should be friend, not acquaintance." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately corrects it to 'friend' in a revision. The final intent is clear and matches the original instruction after applying the correction. The revision type criterion is satisfied.", "heard_transcript": "Request thirteen dollars publicly on um Venmo from my acquaintance Stacy uh with a note, for yesterday's meal. I misspoke about the relationship, it should be friend not acquaintance." }, "candidate_id": "wav_c9f0170f36cd296f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip shows a believable revision where the speaker corrects 'acquaintance' to 'friend'. The initial hesitation ('um') and the self-correction with 'I misspoke' are natural. However, the correction is a bit more formal and less fluid than the top two, and the overall delivery is slightly more hesitant, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_05_person_name", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Sarah, with a note, \"For yesterday's meal\". Change the recipient to Stacy, not Sarah.", "long_pause_offsets": [ 90 ], "source_sha256": "18bbd8bb754eba3e37e0b7c70e43500e7b596213a91e6b33503825a6804d96f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the recipient from Sarah to Stacy, preserving the amount, app, public visibility, note content, and the relationship 'my friend'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_05_person_name/47be7d307d7ea70a/speech.wav", "wer": 0.08, "wav_sha256": "97ce1371cc7ae2f2a9310a1ae6a3dfc13eee469fa34707851868dc5a060b123b", "seconds": 15.32, "spoken_utterances": [ "Request thirteen dollars publicly on ... Venmo from my friend, Sarah, let's see... with a note, \"For yesterday's meal\". ", "Change the recipient to Stacy, not Sarah." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Sarah' but immediately corrects to 'Stacy, not Sarah' after a brief pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send $13 to Stacy, which matches the original reference instruction. All other constraints (amount, platform, note, public) are preserved.", "heard_transcript": "Request thirteen dollars publicly on Venmo from my friend Sarah, let's see, with a note, \"For yesterday's meal\". Change the recipient to Stacy, not Sarah." }, "candidate_id": "wav_dbb0f4c26aaa2d8a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker makes a clear revision from 'Sarah' to 'Stacy'. The 'let's see' pause before the correction is a plausible moment of reconsideration. However, the correction is delivered as a separate sentence rather than being integrated into the original request, making the flow less seamless than the top-ranked clips." } ] }, "is_default": false }, { "variant_id": "revision_06_note_text", "proposed_instruction": "Request $13 publicly on Venmo from my friend, Stacy, with a note, \"For lunch\". Use \"For yesterday's meal\" instead of \"For lunch\".", "long_pause_offsets": [ 79 ], "source_sha256": "b2ad08f163d45d6e113ddeef25e1fe179fa879961881585c72ac396799529ad9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the note content from \"For lunch\" to \"For yesterday's meal\". All other elements—amount, app, privacy, recipient, and action—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_06_note_text/3ab75343d5dfadaf/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "f402e6b19e217cbfb892b7cd6be1a437ee14623f920e76523a398be216e9d265", "seconds": 13.68, "spoken_utterances": [ "Request thirteen dollars publicly on ... Venmo from my friend, Stacy, with a note, ... \"For lunch\". ", "Use \"... For yesterday's meal\" instead of \"For lunch\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'for lunch' but then corrects it with 'Use, for yesterday's meal instead of for lunch.' This is a revision where the final instruction clearly replaces the initial note content. The final intent matches the original reference instruction, so the task passes.", "heard_transcript": "Request thirteen dollars publicly on Venmo from my friend Stacy with a note, for lunch. Use, for yesterday's meal instead of for lunch." }, "candidate_id": "wav_0a5c0b6fb1a79d56", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip features a revision where the speaker changes 'for lunch' to 'for yesterday's meal'. The correction is made as a separate sentence, and the phrasing 'use' instead of 'change' sounds slightly unnatural. The prosody is also a bit flat, and the correction lacks the spontaneous feel of the others." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Request $13 publicly on CashApp from my friend, Stacy, with a note, \"For yesterday's meal\". Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 92 ], "source_sha256": "e203bba713106b6191899791fb260d7dbf2981b8d9cbdd717f543514cb7079ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the amount, privacy setting, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/revision/revision_03_app/3139c7ecc411c879/speech.wav", "wer": 0.25, "wav_sha256": "70879e472820b27f0bb18e17a5824693f0233b65342a662963ad35f6cac556cb", "seconds": 15.2, "spoken_utterances": [ "Request hmmm... thirteen dollars publicly on CashApp from my friend, Stacy, ah... with a note, hmmm... \"For yesterday's meal\". ", "Let me correct the app: use Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the app as 'Cash app' but immediately corrects it to 'Venmo' in a self-repair. According to the revision type criterion, the correction spoken after the initial statement must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, recipient, note, and public visibility. The correction is clear and unambiguous, so the task is considered passed.", "heard_transcript": "Request um thirteen dollars publicly on Cash app from my friend Stacy uh with a note um for yesterday's meal. Let me correct the app. Use Venmo not Cash app." }, "candidate_id": "wav_674149791ac54c4b", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip has the least realistic revision. The speaker initially says 'Cash App' and then corrects to 'Venmo'. However, the correction is delivered as a separate sentence with a slightly robotic 'let me correct the app' phrase. The prosody is less natural, and the correction feels more like a mechanical update than a spontaneous self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "16b55d0d9462be7f7a67cc6d5cbc427abc63c8d7dcd1a1046abc53271717c8d2", "decision": "yes", "review_revision": "20261019T025520183338Z_71fcf681", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/cancellation/cancel_008/56363cc69e1fa64f/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "7c802b5b7eef87a3c706f52daf377eb00dc6742bf2669387eea1c2bfc2bec6f8", "seconds": 13.0, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\".", "Hold on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then clearly states 'Hold on, I want to cancel that task' as the final interjection. This constitutes a direct and unambiguous cancellation of the entire task. The cancellation is not conditional or followed by a new instruction, and the original request is not executed.", "heard_transcript": "thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal. Hold on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "0f3fb3f37e183e17e0980cc97bda0d839a1f4c0edf5f0beecd957111b77b1a46", "decision": "yes", "review_revision": "20261019T025534539915Z_63209bbc", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/ambiguous_halt/halt_031/4917d1f040f52eff/speech.wav", "wer": 0.1, "wav_sha256": "99c6eefb29bb837e4ae0549b99b962a9c2c95af59f6e5455620d94aebc7f264e", "seconds": 11.88, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\".", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary hold on the task. No cancellation or execution is requested, and the initial instruction is not revoked. The final control intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "cfccdd0b3fbcb65a5857972e08b51a625b6565ef9d2750fd2cd14bdb5059ebd5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_1/audio_natural/filler/original/bc775c4f62a2e6ff/speech.wav", "wer": 0.125, "wav_sha256": "cce9592f28ab6d9742aef44873163a10deea30167a7e027a86e02cc09eaf1ad7", "seconds": 8.44, "spoken_utterances": [ "Request ... thirteen dollars publicly on Venmo from my friend, Stacy, ... with a note, \"For yesterday's meal\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Request' and proceeds to state the full instruction without any hesitation, backchannel, or change in intent. The initial mention of 'thirteen dollars' is not corrected or revised, and the final utterance matches the original instruction in all key aspects: amount ($13), action (request), platform (Venmo), recipient (Stacy), visibility (publicly), and note content ('For yesterday's meal'). The 'filler' type allows for natural speech including initial phrasing, and the final goal is clear and complete.", "heard_transcript": "Request thirteen dollars publicly on Venmo from my friend Stacy with a note, for yesterday's meal." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_recipient_relation", "proposed_instruction": "Request $28 privately on Venmo from my friend—no, I mean roommate, Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "af13df4ccc43ec9d0b0176fe17fb15b57bb8bbda3007e7b9acf059f1f4f7fba4", "decision": "yes", "review_revision": "20261019T050546672851Z_768cc5a7", "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'my friend' but immediately corrects to 'roommate, Melissa'. The final recipient is clearly Melissa, and all other elements (amount, app, privacy, note) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_04_recipient_relation/211eb650985086e7/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "117023e276567f49c8071253bc57a9bd04c58065b9c061ce6541d972d3313d68", "seconds": 14.96, "spoken_utterances": [ "Request twenty-eight dollars privately on um, let's see... Venmo from my friend—uh... no, I mean roommate, um... Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my friend' but immediately self-corrects to 'roommate'. This is a false start, and the final target is correctly identified as 'roommate Melissa'. All other constraints (amount, app, privacy, note) match the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Request twenty eight dollars privately on um, let's see, Venmo from my friend uh, no, I mean roommate, um, Melissa, with a note, for the movie tickets." }, "candidate_id": "wav_2fd5927998dddc81", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow of a spontaneous self-correction. The speaker begins with 'my friend' and immediately self-corrects to 'roommate' with a believable hesitation and intonation shift. The pause before 'Melissa' sounds like genuine recall, and the overall pacing is consistent with someone formulating a request in real time. The correction is integrated smoothly into the utterance without any robotic or stilted delivery." } ] }, "is_default": true }, { "variant_id": "false_start_05_person_name", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Jessica—wait, I meant Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "37627dae161af963b252d582139aae5b0e636171e01c64ac47595a72d7ef282b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces Jessica with Melissa as the recipient; all other values including amount, app, privacy, and note content remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_05_person_name/fb77211c1c9a80be/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5206701daa40a283c89059de7900daa5a22c20de1487aeb4e92494c073dbcab1", "seconds": 13.68, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my roommate, Jessica—uh... wait, I meant Melissa, with a note, hmmm... \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jessica' but immediately self-corrects to 'Melissa'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original mistaken value is discarded. The final intent matches the original instruction exactly, including the amount, method, recipient, privacy setting, and note content.", "heard_transcript": "Request twenty eight dollars privately on uh let's see Venmo from my roommate Jessica, I'm Jessica. Wait, I meant Melissa with a note, for the movie tickets." }, "candidate_id": "wav_6a1a2ee22e32a9ca", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear self-correction from 'Jessica' to 'Melissa' with a natural-sounding 'wait' and a brief pause. The initial hesitation with 'uh let's see' and the filler 'um' are well-placed. The correction is believable and the speaker's intonation reflects a genuine change of mind. The only minor issue is the slightly abrupt transition after 'wait,' but overall, the delivery is very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Request $28 privately on CashApp—I should say Venmo from my roommate, Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "cf7642218be1793e46b9b619fa3842351b30d47193b744a4d8d7366c6f6da511", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects the app from CashApp to Venmo within the same utterance, preserving the amount, privacy setting, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_03_app/6ca04e5feee495f8/speech.wav", "wer": 0.0, "wav_sha256": "20620d42187168c1e1a44a8d6809e0fd264efe79a2d760f6f5b94241dcaf2354", "seconds": 12.92, "spoken_utterances": [ "Request twenty-eight dollars privately on CashApp—uh... I should say um... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, including the amount, privacy setting, app, sender, recipient, and note. Therefore, the task is correctly executed.", "heard_transcript": "Request twenty eight dollars privately on Cash App uh I should say um Venmo from my roommate Melissa with a note, for the movie tickets." }, "candidate_id": "wav_447f3900890de4df", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with 'Cash App' and corrects to 'Venmo' with a natural-sounding 'I should say.' The hesitation and self-correction are plausible, and the delivery maintains a steady, conversational pace. The correction is integrated smoothly, though the initial error is less common than the others, making the correction slightly less expected but still realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_visibility", "proposed_instruction": "Request $28 publicly—no, I meant privately on Venmo from my roommate, Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "3b2246541d93717d19998e53a0343f6d06d7806d337af6cde32d273438941500", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'publicly' to 'privately' within the same utterance, preserving the amount, app, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_02_visibility/7131dbcc879e8630/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "14cb872551fb0529320b290cfbc7c10a1872e8ef3c381301e7a5806fd83671b6", "seconds": 12.52, "spoken_utterances": [ "Request twenty-eight dollars publicly—um... no, I meant privately on Venmo from my roommate, hmmm... Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately self-corrects to 'privately'. According to the 'false_start' type criterion, the immediate self-correction should be applied, and the original mistaken value discarded. The final audible intent matches the original instruction in all aspects: amount ($28), method (Venmo), privacy setting (privately), sender (from my roommate), recipient (Melissa), and note content (for the movie tickets).", "heard_transcript": "Request twenty eight dollars publicly, um, no, I meant privately on Venmo from my roommate, Melissa, with a note, for the movie tickets." }, "candidate_id": "wav_e1d609bed1791d08", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a self-correction from 'publicly' to 'privately' with a believable 'no I meant' and a brief pause. The speaker's intonation reflects a genuine correction. However, the correction is less integrated into the flow compared to audio_04 and audio_02, and the initial error is a simple word choice rather than a more complex recall issue, making it slightly less complex in terms of natural speech patterns." } ] }, "is_default": false }, { "variant_id": "false_start_06_note_text", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Melissa, with a note, \"For the concert tickets\"—scratch that, make it \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "479578a7b2d3fbefbfcc776c7e57a2db91da9a8fcd67a6544873d9967161b330", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'For the concert tickets' but immediately corrects it to 'For the movie tickets'. All other elements—amount, app, privacy, recipient, and action—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_06_note_text/ba23c13f5e26f0bd/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "06d82253290983a1eb5cd09cc513de093d95df1370727e0b2d0c986c66e0477c", "seconds": 14.36, "spoken_utterances": [ "Request ... twenty-eight dollars privately on Venmo from my roommate, Melissa, with a note, \"For the concert tickets\"—um... scratch that, make it ... \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'for the concert tickets' but immediately self-corrects with 'Scratch that, make it for the movie tickets.' This is a clear false start where the final intent is the corrected version. The correction is immediate and unambiguous, so the final goal is to send $28 to Melissa with the note 'For the movie tickets'. All other conditions match the original instruction.", "heard_transcript": "Request twenty-eight dollars privately on Venmo from my roommate, Melissa, with a note, for the concert tickets. Scratch that, make it for the movie tickets." }, "candidate_id": "wav_40a4b8eac70b23b9", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker makes a self-correction from 'concert tickets' to 'movie tickets' with a clear 'scratch that' and a pause. The correction is natural, but the initial request is delivered very fluently, making the correction feel slightly abrupt. The overall delivery is good, but the lack of hesitation before the correction makes it less spontaneous compared to others." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Request $30—sorry, that should be $28 privately on Venmo from my roommate, Melissa, with a note, \"For the movie tickets\".", "long_pause_offsets": [], "source_sha256": "8da2325e9b78f73cfc08a04272e4d47688799a67abadce2c23cef705bfa9700a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces thirty with twenty-eight; all other values, including recipient, app, privacy, and note content, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/false_start/false_start_01_amount/42a1f9ca444506ef/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "80712b7f751909bc68c7eb41d226989bd715dbdd537e4ab8c290dafb062bdc9b", "seconds": 16.56, "spoken_utterances": [ "Request thirty dollars—uh... sorry, that should be um... twenty-eight dollars privately on Venmo from my let me think... roommate, Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'thirty dollar' but immediately self-corrects to 'twenty eight dollars'. The correction is clear and immediate, and all other details (Venmo, private, from roommate Melissa, with a note 'for the movie tickets') are preserved. This matches the 'false_start' type criterion, where the final intent is the corrected version.", "heard_transcript": "Request thirty dollar, uh, sorry, that should be um, twenty eight dollars privately on Venmo from my, let me think, roommate Melissa with a note, for the movie tickets." }, "candidate_id": "wav_4e870e47d6674dcd", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip contains multiple corrections and hesitations, including '30 dollar' to '28 dollars' and 'let me think' before 'roommate Melissa.' While this shows a high level of self-correction, the delivery feels slightly more rehearsed or deliberate. The pauses and fillers are present, but the corrections are more spaced out and less fluidly integrated, making the overall speech feel less natural and more like a constructed example." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Melissa. Oh, I forgot to mention, include the note \"For the movie tickets\".", "long_pause_offsets": [ 58 ], "source_sha256": "53b4884d3d1802b1c5f9d674926b7f60d8799c1091613efad6c3f78a1eb22739", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing note requirement; all other values including amount, privacy, app, recipient, and action are preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/addition/addition_01/5aa0f1abf0402ab6/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "6042f4392b7f9b3343719397ae5be583d6111e32d669346d3bc40cda53615c9a", "seconds": 16.16, "spoken_utterances": [ "Request twenty-eight dollars let me think... privately on Venmo from my um... roommate, uh... Melissa. ", "Oh, I forgot to mention, include the note \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Request $28 privately on Venmo from my roommate Melissa'. After a pause and a self-correction ('let me think'), the speaker adds the missing constraint 'include the note, for the movie tickets'. This is a clear addition to the original instruction, and the final goal matches the reference. The addition is not a revision or cancellation, and the initial request is not incorrect. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Request twenty eight dollars, let me think. Privately on Venmo from my um roommate uh Melissa. Oh, I forgot to mention, include the note, for the movie tickets." }, "candidate_id": "wav_34dcae5e91eb74ce", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, inserts a natural 'let me think' pause, then continues with a plausible self-correction ('privately on Venmo') and a believable addition ('include the note for the movie tickets') after a short hesitation. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition flows naturally as a remembered detail rather than a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Request $28 on Venmo from my roommate, Melissa, with a note, \"For the movie tickets\". One more detail: make the request private.", "long_pause_offsets": [ 86 ], "source_sha256": "bec24aa1ce02d4574f3bf9822c9713b9b93e52906a8f6afe64f1f670c4c12acf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing privacy requirement, which was present in the original instruction but omitted in the initial utterance. All other elements—amount, app, recipient, and note content—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/addition/addition_02/55a053ef82f55ed2/speech.wav", "wer": 0.1, "wav_sha256": "402559f6f5cc6447ea74d715d68f3ea0aa73c95669355f873c26f59518176d66", "seconds": 15.96, "spoken_utterances": [ "Request uh... twenty-eight dollars on Venmo from my uh, let me think... roommate, Melissa, with a note, uh... \"For the movie tickets\". ", "One more detail: make the request private." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'twenty eight dollars' and then adds 'make the request private' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final audible intent is to execute the original Venmo request with the additional privacy setting.", "heard_transcript": "Request the twenty eight dollars on Venmo from my uh, let me think, roommate Melissa, with a note for the movie tickets. One more detail, make the request private." }, "candidate_id": "wav_323cd3becf7af91f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the initial request is fluent and includes a 'let me think' pause, the follow-up 'one more detail' is less natural than a spontaneous addition. The correction 'Make the request private' sounds more like a post-hoc instruction than a genuine recollection, and the delivery is slightly more robotic. The prosody lacks the subtle hesitation and rephrasing seen in audio_01, making it feel less like a real-time thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_05_person_name", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Jessica, with a note, \"For the movie tickets\". Change the recipient to Melissa, not Jessica.", "long_pause_offsets": [ 96 ], "source_sha256": "c56aa62c2840d8c219b2a14ffcbe876652045487c5e85720eca63f98cce41310", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the recipient from Jessica to Melissa, preserving the amount, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_05_person_name/472148c04c96fad8/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "7f8f44cf0ea956081db181e60d31d6dcba71133d956473a91fe21f80d89f5712", "seconds": 15.04, "spoken_utterances": [ "Request twenty-eight dollars privately on um... Venmo from my ... roommate, Jessica, with a note, \"For the movie tickets\". ", "Change the recipient to ... Melissa, not Jessica." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jessica' but immediately corrects to 'Melissa, not Jessica' after a pause. This is a revision where the final recipient is Melissa. The amount, app, privacy setting, note, and action type are unchanged. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Request twenty-eight dollars privately on um Venmo from my roommate Jessica with a note, for the movie tickets. Change the recipient to Melissa, not Jessica." }, "candidate_id": "wav_f2215ee3de06efab", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with a natural pause and filler ('um') before the correction. The self-repair ('Change the recipient to Melissa, not Jessica') is immediate, grammatically correct, and flows seamlessly within the same utterance, mimicking a spontaneous correction. The prosody is consistent and the pace is conversational." } ] }, "is_default": true }, { "variant_id": "revision_06_note_text", "proposed_instruction": "Request $28 privately on Venmo from my roommate, Melissa, with a note, \"For the concert tickets\". Change the note to \"For the movie tickets\" instead of \"For the concert tickets\".", "long_pause_offsets": [ 98 ], "source_sha256": "5547c7a64cd580c985abe82fc0d0543c6c86da1641584a24bc995cd66dfcfeda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the note content from \"For the concert tickets\" to \"For the movie tickets\". All other elements—amount, app, privacy, recipient, and action—remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_06_note_text/8477e253112dd2f2/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "27e9ac0da653cc3c536910c97a351abc084a198977d651e54c487e644816139e", "seconds": 17.56, "spoken_utterances": [ "Request twenty-eight dollars um... privately on Venmo from my roommate, hmmm... Melissa, with a note, \"For the concert tickets\". ", "Change the note to \"For the movie tickets\" instead of \"For the concert tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a clear revision: the speaker first states the note as 'for the concert tickets' and then immediately corrects it to 'for the movie tickets'. According to the 'revision' type criterion, the correction spoken after the initial instruction must be applied. The final audible intent matches the original reference instruction, so the task passes.", "heard_transcript": "Request twenty-eight dollars um privately on Venmo from my roommate, Melissa, with a note, for the concert tickets. Change the note to for the movie tickets instead of for the concert tickets." }, "candidate_id": "wav_629a1693c622d78d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is fluent with a natural pause and filler ('um'). The correction is a complete sentence that clearly references the prior request ('Change the note to... instead of...'). The intonation and rhythm are consistent with a person recalling and adjusting a detail, making it highly realistic." } ] }, "is_default": false }, { "variant_id": "revision_04_recipient_relation", "proposed_instruction": "Request $28 privately on Venmo from my friend, Melissa, with a note, \"For the movie tickets\". I misspoke about the relationship: it should be roommate, not friend.", "long_pause_offsets": [ 94 ], "source_sha256": "5292e7e97237733b6f0a9f355ab3faa9d723fb361680435a9686be06560592f7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the relationship from 'friend' to 'roommate' but retains the recipient (Melissa), amount, app, privacy, and note. The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_04_recipient_relation/f454e417c6e41a1d/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "a72f6fdf4da19d7ec1e43b1e1fb9bd6994499054391716e3ab258f371d560dd6", "seconds": 14.52, "spoken_utterances": [ "Request twenty-eight dollars privately on Venmo from my friend, um... Melissa, with a note, uh... \"For the movie tickets\". ", "I misspoke about the relationship: it should be roommate, not friend." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friend' but immediately corrects it to 'roommate' after a pause, which is a revision. The final intent clearly specifies 'roommate' as the relationship, aligning with the original instruction. All other details (amount, app, recipient, note) match the original request.", "heard_transcript": "Request twenty eight dollars privately on Venmo from my friend um Melissa with a note uh for the movie tickets. I misspoke about the relationship, it should be roommate, not friend." }, "candidate_id": "wav_00c7335b97df3101", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then adds a self-correction about the relationship ('I misspoke...'). The filler words ('um', 'uh') are placed naturally, and the correction is a plausible follow-up. The delivery is slightly more hesitant than the top two, but still sounds like a genuine, spontaneous recollection." } ] }, "is_default": false }, { "variant_id": "revision_01_amount", "proposed_instruction": "Request $30 privately on Venmo from my roommate, Melissa, with a note, \"For the movie tickets\". Sorry, I meant $28, not $30.", "long_pause_offsets": [ 96 ], "source_sha256": "03572faefbbc6b13e71aa9f9b159f9e0b772117ca0bd06705f74834a074dff41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from thirty to twenty-eight, while retaining all other elements including recipient, app, privacy, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_01_amount/7d1dc609cf1e3891/speech.wav", "wer": 0.1, "wav_sha256": "09687ed340369d881c2b49de6035305c2357c36af021e377be7207c1c58b17b3", "seconds": 15.6, "spoken_utterances": [ "Request thirty dollars privately on let me think... Venmo from my roommate, Melissa, with a note, um... \"For the movie tickets\". ", "Sorry, I meant ... twenty-eight dollars, not thirty dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a request for $30, then immediately corrects it to $28. According to the revision type criterion, the correction spoken after the initial statement must be applied to the earlier instruction. The final audible intent is to request $28, which matches the original reference instruction. All other conditions (Venmo, from roommate Melissa, private, with note 'For the movie tickets') are preserved and correctly stated.", "heard_transcript": "Request thirty dollars privately on Venmo from my roommate Melissa with a note, um, for the movie tickets. Sorry, I meant twenty-eight dollars, not thirty dollars." }, "candidate_id": "wav_769a9420ca94d5b1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker begins with a clear request, then corrects the amount. The filler 'let me think' is placed at the start, and the correction is delivered with a natural 'sorry' and a self-repair. The prosody is slightly more hesitant, and the correction is less integrated into the initial sentence, making it feel a bit more like a separate thought." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Request $28 privately on CashApp from my roommate, Melissa, with a note, \"For the movie tickets\". Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 98 ], "source_sha256": "b8fe61e6371d7016bdb46b155275f945a6986a4307dc973a2cc8bcf67a2219ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the amount, privacy setting, recipient, and note content. The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_03_app/bc34a68b4fb03828/speech.wav", "wer": 0.22580645161290322, "wav_sha256": "274c82779b80fef307615d387ce7700077e6ac81c0199d8428e8f7f8b797b1e4", "seconds": 14.08, "spoken_utterances": [ "Request um, let's see... twenty-eight dollars privately on CashApp from my um... roommate, Melissa, with a note, \"For the movie tickets\". ", "Let me correct the app: use Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects it to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, privacy setting, recipient, and note. Therefore, the task should be executed as revised.", "heard_transcript": "Request um, let's see, twenty-eight dollars privately on Cash App from my um roommate Melissa with a note, for the movie tickets. Let me correct the app, use Venmo, not Cash App." }, "candidate_id": "wav_95061668dd8b3ad4", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker starts with a clear request, then corrects the app. The filler 'let's see' and 'um' are used, but the correction is delivered as a separate sentence ('Let me correct the app...'). The delivery is slightly more robotic in its structure, and the correction feels like a distinct command rather than a natural continuation." } ] }, "is_default": false }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Request $28 publicly on Venmo from my roommate, Melissa, with a note, \"For the movie tickets\". Wait, make that privately, not publicly.", "long_pause_offsets": [ 95 ], "source_sha256": "56d6d8eeebbd0c04b1a0c2da757131dc4e9e32c2bec5185c30a33bfd635de2c1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the visibility from 'publicly' to 'privately', while retaining the amount, app, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/revision/revision_02_visibility/2ed4af0bb3fb69a6/speech.wav", "wer": 0.22580645161290322, "wav_sha256": "868928cad35162adb0f34d985813914f358a09f80cf201524de0b941cf1a5988", "seconds": 19.24, "spoken_utterances": [ "Request ... twenty-eight dollars publicly on Venmo from my let me think... roommate, Melissa, with a note, um, let me think... \"For the movie tickets\". ", "Wait, make that privately, not publicly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects to 'privately' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, recipient, app, note, and privacy setting.", "heard_transcript": "Request $28 publicly on Venmo from my roommate, Melissa, with a note, um, let me think, for the movie tickets. Wait, make that privately, not publicly." }, "candidate_id": "wav_70d154a8c74391ef", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "The speaker uses multiple fillers ('let me think', 'um', 'let me think') and pauses, which makes the delivery sound less fluent. The correction is a separate sentence ('Wait, make that privately...') that interrupts the flow. The overall delivery feels more like a rehearsed or less natural recollection compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "25a734aff6be7293af33a73d8753ddf4e844797dfe405847ccb3f3c23965ca6c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/cancellation/cancel_007/1b283a3f0622d70e/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "3f744cae1602033e290e6f18dcae9395bd0adf41fcb880ea3972ed3f3ca7b009", "seconds": 14.84, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\".", "Oh, wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a complete Venmo request and then immediately follows with 'Oh wait, cancel my last request.' This constitutes a clear cancellation of the entire prior task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Request twenty eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets. Oh wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "f2a859c81ecc3b6208b633720c57e9315882123277dd4b8f01dfdc8512276f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/ambiguous_halt/halt_005/2d992d61154444ff/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "58ae9229e1b06cc64777f1f698eb04fcf19d3f57922250b69339edaa11c44ed7", "seconds": 13.72, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\".", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full request and then says 'One second,' which is an audible request to pause. There is no cancellation or instruction to resume. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Request twenty eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1b67d2ca20d3f9beae260210b5059fdee5980161eeedfa192c225a444ff4afad", "decision": "yes", "review_revision": "20261019T050429833050Z_49e74a94", "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_2/audio_natural/filler/original/d74aafc3f5afe360/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "6ed293ab284caf86bea61001b964b5f573aca2c3b742eec18e416061420f5cb7", "seconds": 10.6, "spoken_utterances": [ "Request twenty-eight dollars privately on uh, let's see... Venmo from my ... roommate, Melissa, with a note, \"For the movie tickets\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler phrase 'uh let's see' but continues to state the full original instruction without any correction or change in intent. The final audible task matches the original instruction in action, target, value, privacy, app, note, and all other constraints. The filler does not alter the goal, so the execution criterion is satisfied.", "heard_transcript": "Request twenty-eight dollars privately on uh let's see Venmo from my roommate Melissa with a note, for the movie tickets." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_visibility", "proposed_instruction": "Request $29 privately—no, I mean publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "dce5e6060c12d565f28cd3aabfa19dfa3162e722fc7c60ae9d7389dbdb80db6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'privately' to 'publicly' within the same utterance, preserving the amount, recipient, app, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_02_visibility/e1eb65485eff49c0/speech.wav", "wer": 0.08, "wav_sha256": "faefe4534fffd5001d03ba0afe3046bdb07df437afb8f7040cde157580ab4148", "seconds": 13.48, "spoken_utterances": [ "Request um... twenty-nine dollars privately—uh... no, I mean publicly on Venmo from my friend, Ronald, with a note, um... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately self-corrects to 'publicly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only a minor hesitation and repair that does not alter the core task.", "heard_transcript": "Request um, twenty nine dollars privately uh, no, I mean publicly on Venmo from my friend Ronald with a note, um, for the road trip." }, "candidate_id": "wav_35b88a292392e11f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then makes a clear false start ('privately') which is immediately self-corrected to 'publicly'. The correction is smooth and integrated into the flow. The speaker continues with a complete, coherent request, including the amount, platform, recipient, and note. The prosody is conversational and the pacing realistic, with a natural pause before the final phrase." } ] }, "is_default": true }, { "variant_id": "false_start_05_person_name", "proposed_instruction": "Request $29 publicly on Venmo from my friend, Robert—sorry, I meant Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "830a128693b6f85b65bd9b07309491da8a0bd96287922ed0395b1e9d23652b11", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Robert' but immediately corrects to 'Ronald' with 'I meant Ronald'. The correction is clear and applies to the recipient. All other elements—amount, app, privacy, note content—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_05_person_name/2a163f38f59aaa3c/speech.wav", "wer": 0.12, "wav_sha256": "39cfeefa893beb984c0b928fef91b9b26a9c28bceb82d91ef90c244cf222c131", "seconds": 9.68, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Robert—hmmm... sorry, I meant Ronald, with a note, uh... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Robert' but immediately self-corrects to 'Ronald'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The rest of the instruction, including the amount ($29), platform (Venmo), action (request), visibility (publicly), and note ('For the road trip'), remains consistent. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Robert. Hmm. Sorry, I meant Ronald, with a note uh for the road trip." }, "candidate_id": "wav_2592bfc80fb96bf2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear false start ('Robert') and self-corrects to 'Ronald' with a natural 'sorry, I meant'. The correction is integrated smoothly into the sentence. The prosody and pacing are consistent with spontaneous speech, and the overall delivery is fluent and realistic. The only minor difference from audio_01 is the slightly more formal phrasing of the correction." } ] }, "is_default": false }, { "variant_id": "false_start_04_recipient_relation", "proposed_instruction": "Request $29 publicly on Venmo from my acquaintance—I meant friend, Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "65c582feaef66599971d57c47f7d123436dd614298e10741a67731c08c2c9860", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately corrects to 'friend, Ronald'. The recipient is clarified as Ronald, and all other elements—amount, app, privacy, and note—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_04_recipient_relation/fd52b717747d869a/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b1b18b3782802a9f4907f0d0a57dd7730a83984ca7ec7ff406a0f0a968f4f877", "seconds": 10.68, "spoken_utterances": [ "Request twenty-nine dollars publicly on Venmo from my acquaintance—um... I meant uh... friend, Ronald, with a note, \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately self-corrects to 'friend Ronald'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The rest of the instruction, including the amount, platform, public visibility, and note, remains consistent. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my acquaintance. I meant uh, friend Ronald, with a note, \"For the road trip\"." }, "candidate_id": "wav_7697c954087a8c8c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a false start ('acquaintance') and corrects it to 'friend Ronald'. The correction is natural and the flow is maintained. The speaker includes the full request with the note. The delivery is realistic, though the correction is slightly less fluid than in audio_01 or audio_03, with a brief pause before the correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Request $29 publicly on CashApp—I should say Venmo from my friend, Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "21c285e9e635d63a1b457e273e534a466589f04a61d1852f8cc07a2aa6ebe6ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects the app from CashApp to Venmo within the same utterance, preserving the amount, recipient, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_03_app/67e36bb81ef7272a/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "b4427ff9be8e8010815cbd47da736bb80f63fcbac227b99ecef756818d7a8164", "seconds": 13.04, "spoken_utterances": [ "Request let's see... twenty-nine dollars publicly on CashApp—uh... I should say Venmo from my friend, Ronald, with a note, let's see... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('Cash App') is discarded. The rest of the instruction, including the amount, recipient, note, and public visibility, is preserved and matches the reference. Therefore, the final audible intent matches the original instruction after correction.", "heard_transcript": "Let's see. Twenty nine dollars publicly on Cash App. Uh, I should say Venmo from my friend Ronald with a note. Let's see. For the road trip." }, "candidate_id": "wav_aa15b205382911a7", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker begins with 'let's see', indicating hesitation, then makes a platform error ('Cash App' corrected to 'Venmo'). The correction is integrated into the sentence. However, the speaker repeats 'let's see' after the correction, which sounds slightly unnatural and disrupts the flow. The overall delivery is still realistic but less smooth than the top three." } ] }, "is_default": false }, { "variant_id": "false_start_06_note_text", "proposed_instruction": "Request $29 publicly on Venmo from my friend, Ronald, with a note, \"For the concert\"—scratch that, make it \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "8dadad1e55eb74c0e51565fe0dfe32fb60550c797f9ba775ab099bc41d4bc903", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'For the concert' but immediately corrects it to 'For the road trip'. The final note content is the same as the original instruction. All other elements—amount, app, privacy, recipient, and action—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_06_note_text/ce56fde628dfd1a9/speech.wav", "wer": 0.0, "wav_sha256": "fa72667cd17b4ae18c7953ee189dca4fc1e1ab83f470cecda526ce5db2ad24d6", "seconds": 14.92, "spoken_utterances": [ "Request ... twenty-nine dollars publicly on Venmo from my friend, uh... Ronald, with a note, \"For the concert\"—um... scratch that, make it \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'for the concert' but immediately self-corrects with 'scratch that, make it for the road trip'. This is a clear false start where the final intent is the corrected version. The amount, app, sender, and recipient remain unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my friend, Ronald, with a note, for the concert. Um, scratch that, make it for the road trip." }, "candidate_id": "wav_68828a025051947e", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker makes a revision: changes the note from 'for the concert' to 'for the road trip'. The revision is clear and the speaker uses 'scratch that' to signal the change. However, the initial phrase 'for the concert' is spoken with a slightly flat intonation, and the revision feels a bit abrupt. The prosody is less natural than in the top-ranked clips." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Request $35—make that $29 publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\".", "long_pause_offsets": [], "source_sha256": "baca4f1e4fef76e6c1c84f391e00428f998a529460da1620149d8654e6f5bb57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces thirty-five with twenty-nine; all other elements including recipient, app, privacy, and note content remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/false_start/false_start_01_amount/a4d603b7f437c05a/speech.wav", "wer": 0.08, "wav_sha256": "6a30c6f735fcef962861106de00d3ef49321f71de8d6b62c0555946d6438a1fc", "seconds": 11.2, "spoken_utterances": [ "Request thirty-five dollars—um... make that twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '$35' but immediately self-corrects to '$29'. According to the 'false_start' type criterion, the final intent is the corrected value, and all other conditions (Venmo, publicly, from Ronald, with note 'For the road trip') are preserved. The correction is immediate and unambiguous, so the task is executed as requested.", "heard_transcript": "Request thirty five dollars um make that twenty nine dollars uh publicly on Venmo from my friend Ronald with a note for the road trip." }, "candidate_id": "wav_e69fb86189d3c223", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "The speaker starts with '35 dollars', then corrects to '29 dollars'. The correction is clear but feels slightly rushed. The speaker then delivers the rest of the request without any further issues. However, the initial number error and correction are less smoothly integrated than in the false-start examples, and the overall delivery has a slightly robotic quality compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Request $29 on Venmo from my friend, Ronald, with a note, \"For the road trip\". One more detail: make the request public.", "long_pause_offsets": [ 79 ], "source_sha256": "b9649362aed4cdf6b3a53a51e11abd32ece09b776e0b015ed7941451d7e4d73d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the public visibility requirement, which was missing in the initial instruction. The original instruction specified a private request, but the final instruction explicitly makes it public. The action, amount, recipient, app, and note content are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/addition/addition_02/062fa97dcc4829fe/speech.wav", "wer": 0.0, "wav_sha256": "148ec9a34ed49daf89712d7b7ec2ad3b6f176ca2cf92135effcc3e4eb527483e", "seconds": 16.76, "spoken_utterances": [ "Request twenty-nine dollars on um... Venmo from my friend, uh, let me think... Ronald, with a note, ... \"For the road trip\". ", "One more detail: make the request public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Request $29 on Venmo from my friend Ronald with a note for the road trip' and then adds 'One more detail, make the request public.' This is a clear addition of a new constraint to the original request. The final intent combines the original request with the added condition, matching the reference instruction. The addition is made after a pause and is clearly intended to modify the original task.", "heard_transcript": "Request twenty nine dollars on um Venmo from my friend uh let me think Ronald with a note for the road trip. One more detail, make the request public." }, "candidate_id": "wav_a59c7ab0304b75ba", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural hesitation ('um', 'let me think') and a believable additive structure. The initial request is clear, followed by a pause and a plausible addition of the 'public' detail. The prosody is consistent, and the delivery sounds like a spontaneous recollection rather than a rehearsed or robotic statement." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Request $29 publicly on Venmo from my friend, Ronald. Oh, I forgot to mention, include the note \"For the road trip\".", "long_pause_offsets": [ 54 ], "source_sha256": "da05035253c01a7e7a11837d4491fde6fdbb5077f4c28303c8351c2c20133527", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing note requirement, preserving the amount, app, privacy setting, and recipient from the initial instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/addition/addition_01/5eb0d5921c183be2/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "a57a14fee22b614161a16bd1b18bf01721a9699b8fb82d5f9d38adf3a11ed0d2", "seconds": 13.16, "spoken_utterances": [ "Request twenty-nine dollars ... publicly on Venmo from my friend, uh... Ronald. ", "Oh, I forgot to mention, include the note ... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio starts with the correct request to send $29 publicly on Venmo from Ronald. After a pause and an interjection, the speaker adds the missing constraint of including the note 'for the road trip'. This matches the addition type criterion, where the final goal combines the initial instruction with the added information. The final intent is complete and matches the original instruction.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my friend, Ronald. Oh, I forgot to mention, include the note, for the road trip." }, "candidate_id": "wav_c007af3f628b7875", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail ('include the note'), the phrasing 'oh, I forgot to mention' is slightly less natural than 'let me think' in the first clip. The delivery is fluent but lacks the subtle hesitation and cognitive processing cues that make audio_01 sound more spontaneous. The addition feels more like a post-hoc correction than a genuine memory retrieval." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_amount", "proposed_instruction": "Request $35 publicly on Venmo from my friend, Ronald, with a note, \"For the road trip\". Sorry, I meant $29, not $35.", "long_pause_offsets": [ 88 ], "source_sha256": "d70e3c6ae8bda2a2e0fc301da23f165b7bc01f43882ceba02cd1e5a7e672339c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from thirty-five to twenty-nine, while retaining the recipient, app, privacy setting, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_01_amount/66bc8396a4531990/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "0f50ddc8ef33fa447ea535e7c32de9a01a21756504981133ae166335d8735ad8", "seconds": 18.04, "spoken_utterances": [ "Request thirty-five dollars publicly on Venmo from my um... friend, Ronald, with a note, uh... \"For the road trip\". ", "Sorry, I meant ah... twenty-nine dollars, not thirty-five dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a request for $35, then immediately corrects it to $29. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent matches the original reference instruction exactly, including the amount, recipient, app, visibility, and note. Therefore, the task is correctly executed as per the revision rule.", "heard_transcript": "Request thirty five dollars publicly on Venmo from my um friend Ronald with a note uh for the road trip. Sorry, I meant uh twenty nine dollars not thirty five dollars." }, "candidate_id": "wav_e92af7a5d30ae78b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a highly realistic false start and self-correction. The speaker begins with 'thirty-five dollars' and immediately corrects to 'twenty-nine dollars' with a natural-sounding 'sorry, I meant uh...' and a brief pause. The prosody is conversational, with a slight hesitation and a believable recovery. The correction is integrated smoothly into the ongoing utterance, making it sound like a spontaneous recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Request $29 privately on Venmo from my friend, Ronald, with a note, \"For the road trip\". Wait, the request should be publicly, not privately.", "long_pause_offsets": [ 89 ], "source_sha256": "1d74117c54a30b830091aaed2324debd61af1c28a85ead4fd0bcb79f9d741032", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the visibility from 'privately' to 'publicly', while retaining the amount, app, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_02_visibility/106d3ad734131bec/speech.wav", "wer": 0.0, "wav_sha256": "1bcd03995a7f0f4566af734edb894bfca19f074dcb596687c3e911a254cf7554", "seconds": 17.92, "spoken_utterances": [ "Request twenty-nine dollars um, let me think... privately on Venmo from my friend, ... Ronald, with a note, ... \"For the road trip\". ", "Wait, the request should be publicly, not privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately corrects themselves with 'Wait, the request should be publicly, not privately.' This constitutes a revision where the final intent is to send the request publicly. The correction is clear and directly modifies the earlier instruction, aligning with the revision type criterion. All other constraints (amount, app, sender, recipient, note) match the original instruction.", "heard_transcript": "Request twenty nine dollars, um, let me think, privately on Venmo from my friend, Ronald, with a note, for the road trip. Wait, the request should be publicly, not privately." }, "candidate_id": "wav_feaf9d4cac73ea7c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable self-correction regarding the privacy setting. The speaker initially says 'privately' and then corrects to 'publicly' with a 'wait' and a natural-sounding explanation. The 'um' and 'let me think' fillers are placed plausibly, and the correction is delivered with a tone of reconsideration. The flow is continuous and the repair feels like a genuine moment of realization during task formulation." } ] }, "is_default": false }, { "variant_id": "revision_04_recipient_relation", "proposed_instruction": "Request $29 publicly on Venmo from my acquaintance, Ronald, with a note, \"For the road trip\". I misspoke about the relationship: it should be friend, not acquaintance.", "long_pause_offsets": [ 94 ], "source_sha256": "4e8ba7eefb39df8be23ed628f47018a0160f73ae43f8de19ee3235263703dc3a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the descriptor of the relationship from 'acquaintance' to 'friend', which is a refinement of the same person. The recipient, amount, app, privacy, and note remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_04_recipient_relation/3505d198ac863c42/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "a99cbf41a9879c72332db832e766ed278760772572167e6898bfd4b331edf49b", "seconds": 18.4, "spoken_utterances": [ "Request twenty-nine dollars publicly on Venmo from my acquaintance, hmmm... Ronald, with a note, uh... \"For the road trip\". ", "I misspoke about the relationship: it should be ... friend, not acquaintance." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'acquaintance' but immediately corrects it to 'friend' after a pause, which is a revision. The final intent matches the original instruction exactly, including the amount, app, recipient, note, and public visibility. The correction is clear and applied to the earlier instruction.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my acquaintance, Ronald, with a note, uh, for the road trip. I misspoke about the relationship, it should be, friend, not acquaintance." }, "candidate_id": "wav_55fc78ca48fcb3bf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a correction from 'acquaintance' to 'friend' with a self-aware 'I misspoke' and a brief pause. The filler 'hmm' and the hesitant 'uh' are used naturally. The correction is clearly a follow-up to a prior statement, and the intonation suggests a moment of reflection. The overall delivery is credible, though the correction is slightly less integrated than in audio_01 or audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Request $29 publicly on CashApp from my friend, Ronald, with a note, \"For the road trip\". Make that Venmo, not CashApp.", "long_pause_offsets": [ 90 ], "source_sha256": "ca92c691d763ac5837de33c9a60018f1d103723231dbe22dd531a987b3c3522f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, preserving the amount, privacy setting, recipient, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_03_app/2729e7dda70dd7c1/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "adeaaf5ead3ba710af047ae8686ea8337e8585708738121218cb0261a32080bd", "seconds": 16.56, "spoken_utterances": [ "Request twenty-nine dollars publicly on CashApp from my friend, um... Ronald, with a note, ... \"For the road trip\". ", "Make that ... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects it to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to request $29 publicly on Venmo from Ronald with the note 'For the road trip', which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "Request twenty nine dollars publicly on Cash App from my friend, um, Ronald, with a note, \"For the road trip\". Make that, Venmo, not Cash App." }, "candidate_id": "wav_9f321ddf19e80ef1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision where the speaker first says 'Cash App' and then corrects to 'Venmo'. The 'make that' phrasing is a common conversational way to revise a statement. However, the correction is more abrupt and less integrated into the original sentence than in the top-ranked clips. The prosody is slightly more robotic, and the transition feels a bit more like a post-hoc edit than a spontaneous correction." } ] }, "is_default": false }, { "variant_id": "revision_05_person_name", "proposed_instruction": "Request $29 publicly on Venmo from my friend, Robert, with a note, \"For the road trip\". Use Ronald instead of Robert.", "long_pause_offsets": [ 88 ], "source_sha256": "df62b2369fa8af845d74bfbadfecf03fef996d14db9f82a6d301481162a767ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the friend's name from Robert to Ronald, preserving the amount, app, public visibility, and note content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_05_person_name/6713bc66ae19d2c7/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "5172ddd2e9ad2015f8f76c00bb64e5a946ea552bf9f6f348925ef60f2cb01d08", "seconds": 15.4, "spoken_utterances": [ "Request twenty-nine dollars publicly on ... Venmo from my um... friend, Robert, with a note, ... \"For the road trip\". ", "Use Ronald instead of Robert." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Robert' but immediately corrects to 'Ronald' after a short pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to send the request to Ronald, which matches the original instruction. All other constraints (amount, platform, visibility, note) are preserved.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my um friend Robert with a note, for the road trip. Use Ronald instead of Robert." }, "candidate_id": "wav_5cb66002d8442178", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker makes a correction from 'Robert' to 'Ronald' with a 'use Ronald instead of Robert' statement. While the correction is clear, the phrasing is more directive and less conversational than in other clips. The 'um' filler is used, but the overall delivery is flatter and less emotionally varied. The correction feels more like a mechanical update than a natural self-repair." } ] }, "is_default": false }, { "variant_id": "revision_06_note_text", "proposed_instruction": "Request $29 publicly on Venmo from my friend, Ronald, with a note, \"For the concert\". Change the note to \"For the road trip\" instead of \"For the concert\".", "long_pause_offsets": [ 86 ], "source_sha256": "8d00d08eb0ddb778219297561374cbed5425e79cf825a36707403a1c7be226b5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the note content from \"For the concert\" to \"For the road trip\". All other elements—amount, app, privacy, recipient, and action—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/revision/revision_06_note_text/ed3f03e1f4bbd4b8/speech.wav", "wer": 0.0, "wav_sha256": "653f5ce095b1e38010d697f313294c18731729b2bbc58c60aebc4e9bd061118d", "seconds": 17.32, "spoken_utterances": [ "Request ... twenty-nine dollars publicly on Venmo from my friend, ah... Ronald, with a note, let me think... \"For the concert\". ", "Change the note to \"For the road trip\" instead of \"For the concert\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'for the concert' but then immediately corrects it with 'Change the note to for the road trip instead of for the concert.' This is a clear revision where the final instruction overrides the initial one. The final intent matches the original reference instruction, so the task passes.", "heard_transcript": "Request twenty nine dollars publicly on Venmo from my friend, Ronald, with a note, let me think, for the concert. Change the note to for the road trip instead of for the concert." }, "candidate_id": "wav_c4e7b6cfe7110593", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip is the least realistic. It begins with a request, then includes a filler 'let me think', and then gives a revision instruction ('change the note to...'). The structure is more like a command to a system than a spontaneous human utterance. The phrasing is overly explicit and lacks the natural flow of a person formulating a request in real time. The delivery is flat and the revision is presented as a separate instruction rather than a continuation of the original thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "12172c9bf46e41dc3a004ce26e63e74ffee128367386d0c0977ea5a8485227cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/cancellation/cancel_023/caeb5f0484c0f724/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "c9190ff93a1078677890ce179dc6962887fad5eba66e63a2878d02440dcbca49", "seconds": 12.88, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\".", "Wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the full Venmo request but then clearly cancels it with 'Wait, never mind about that. Don't carry out the request.' This constitutes a direct cancellation of the task. The final intent is to cancel, not to execute or pause.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip. Wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "06daf321fcc1dc3156f52c002477f382727e6a981c9cf1857bc04a34e6a389c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/ambiguous_halt/halt_018/5f508fe9a749b162/speech.wav", "wer": 0.12, "wav_sha256": "61f68e1871fd0f0491fd3ed511a4f4a197298756536d1132d52deaef77ceafd6", "seconds": 11.04, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\".", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which constitutes a clear instruction to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause request. The initial request is not canceled or revised, and the final control intent is to pause, not execute or cancel. No issues are present.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "024c982_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6f527b130a5be677360b7300ddb4092f42457b5c3c51c8bcd633e961676161d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/024c982_3/audio_natural/filler/original/2d9fc9b6a5e935b8/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "c49df5b350fbb8c7aa3afba879afa127003e10cdcdf729bafde5a36583f6f9cf", "seconds": 7.6, "spoken_utterances": [ "Request twenty-nine dollars uh... publicly on Venmo from my friend, Ronald, with a note, ... \"For the road trip\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'twenty nine dollars' and then corrects to 'twenty nine dollars' with a filler 'uh'. The correction is minor and does not change the value. The rest of the instruction matches the original. The filler 'uh' is a hesitation and does not alter the final intent. The task is executed as requested.", "heard_transcript": "Request twenty nine dollars uh publicly on Venmo from my friend Ronald with a note, for the road trip." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all the jazz—or rather, classical artists on Spotify that have at least 22 followers.", "long_pause_offsets": [], "source_sha256": "69474c5c4382449c4543eb8600c8e5b31eeaec4f870c14753fbff1620712f0e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately corrects to 'classical' within the same utterance. The genre is revised to match the original instruction, and all other conditions (Spotify, 'all artists', follower count of at least twenty-two) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/false_start/false_start_01_genre/f936f4d30b305c74/speech.wav", "wer": 0.05, "wav_sha256": "cc96b9de76f19ae82854119be1bf27601cc82bba417179a8f656ee168f5f7b26", "seconds": 10.64, "spoken_utterances": [ "Follow all the jazz—uh... or rather, classical artists on let's see... Spotify that have at least twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately self-corrects to 'classical' with 'or rather'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Follow all the jazz uh or rather classical artists on let's see Spotify that have at least twenty two followers." }, "candidate_id": "wav_49bc29705200039a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with 'jazz' as a false start, self-corrects to 'classical artists' with a smooth transition, and then continues with 'on let's see Spotify'—a plausible hesitation before naming the platform. The final phrase 'that have at least 22 followers' is delivered clearly and without further interruption. The pacing, intonation, and use of 'let's see' as a filler are all highly realistic and typical of spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Follow all the classical artists on YouTube Music—make that Spotify that have at least 22 followers.", "long_pause_offsets": [], "source_sha256": "6a921254494624392530674782966bb1cc1f5d2cf87b9d76d1a2167d103e77d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions YouTube Music but immediately corrects to Spotify. The genre, follower threshold, and action remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/false_start/false_start_02_app/6880d927f8c5c65c/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "0408f711248b8de2df525da88255296f685bca4e5dfc3f2b4227670ae44c2a82", "seconds": 9.8, "spoken_utterances": [ "Follow all the um... classical artists on YouTube Music—uh... make that Spotify that have at least uh, let's see... twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'YouTube Music' but immediately self-corrects to 'Spotify'. The final control intent is to execute the corrected task. The number 'twenty two' matches the reference. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Follow all the um classical artists on YouTube Music uh make that Spotify that have at least uh let's see twenty two followers." }, "candidate_id": "wav_b0cb1e2d5267a088", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start and revision. The speaker starts with 'classical artists on YouTube Music' and then corrects to 'Spotify'. The correction is natural, with a slight pause and filler ('uh') before the revision. The phrase 'let's see 22 followers' is delivered with a hesitant but believable intonation. The only minor issue is the slightly abrupt transition between the platform correction and the follower count, but overall, the prosody and self-correction are very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_follower_count", "proposed_instruction": "Follow all the classical artists on Spotify that have at least 50—no, make that 22 followers.", "long_pause_offsets": [], "source_sha256": "f83f497f2ef584d2c81c3f9e1b423d71836dd624c9abf5b0f31ccbaabeea74e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces fifty with twenty-two; the genre, app, and all-artists scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/false_start/false_start_04_follower_count/4c3b909c774a568d/speech.wav", "wer": 0.1, "wav_sha256": "9285e68a4474936583f9cc2df0732fb3bed733f47478de3748a4cc8675691362", "seconds": 8.24, "spoken_utterances": [ "Follow all the classical artists on uh... Spotify that have at least fifty—hmmm... no, make that twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at least fifty' but immediately self-corrects with 'no, make that twenty two followers'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as requested.", "heard_transcript": "Follow all the classical artists on uh Spotify that have at least fifty um no, make that twenty two followers." }, "candidate_id": "wav_775e48cbf9ecf792", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start and revision: 'at least 50' is corrected to '22 followers'. The correction is immediate and natural. However, the initial phrase 'follow all the classical artists on Spotify that have at least 50' is delivered with a slightly rushed pace, and the pause before 'no, make that 22 followers' feels a bit too long and abrupt. The prosody is less fluid than in the top two, making it slightly less realistic." } ] }, "is_default": false }, { "variant_id": "false_start_03_comparison_operator", "proposed_instruction": "Follow all the classical artists on Spotify that have at most—no, that should be at least 22 followers.", "long_pause_offsets": [], "source_sha256": "5272aa578a2be8ddc542033edff10dd4a1b0b44da8319c816e4b070a25f474c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most' but immediately corrects to 'at least twenty-two followers'. The final value matches the original instruction's threshold and comparison operator, and all other conditions (genre, app, scope) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/false_start/false_start_03_comparison_operator/8f9aee00f479b25f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "df030f49734d6f7626a1f96d20e28df7b7e2b18f7b4ce4890612c115307bd3a0", "seconds": 10.84, "spoken_utterances": [ "Follow all the classical artists on ... Spotify that have at most—uh... no, that should be ... at least twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most' but immediately self-corrects to 'at least twenty two followers'. According to the 'false_start' type criterion, the final corrected value should be applied, and the initial mistake should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Follow all the classical artists on Spotify that have at most uh, no, that should be at least twenty two followers." }, "candidate_id": "wav_316b4ba0e1fe9547", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip begins with a false start ('at most') and corrects to 'at least 22 followers'. However, the correction is delivered with a noticeable stutter and repetition ('no, that should be'), which sounds less spontaneous and more like a rehearsed or awkward self-correction. The phrase 'at most' is also pronounced with a slightly unnatural emphasis. The overall delivery feels less fluent and more hesitant than the others, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Follow all the artists on Spotify that have at least 22 followers. Oh, right, only follow classical artists.", "long_pause_offsets": [ 67 ], "source_sha256": "0c9facaf6b13ea8693fd1bbd4700d090ab1313272be9ae32eaa21ea078862769", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup restricts the initial broad request to only classical artists, preserving the follower threshold and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/addition/addition_02/88e68aead7afff27/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "c03666993a23febdd456d8b8382e27aa47cbcc327135740d46d9fa93f5f80c78", "seconds": 14.6, "spoken_utterances": [ "Follow all the artists on um... Spotify that have at least uh, let's see... twenty-two followers. ", "Oh, right, only follow let me think... classical artists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is broad ('all the artists'), but the speaker adds a constraint ('classical artists') after a pause and self-correction. According to the 'addition' type criterion, the final goal combines the original instruction with the added information. The final audible intent matches the reference instruction exactly, including the action, target platform, condition, and value.", "heard_transcript": "Follow all the artists on um Spotify that have at least uh let's see twenty two followers. Oh right, only follow, let me think, classical artists." }, "candidate_id": "wav_a8645b20615d4539", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Follow all the classical artists on Spotify. Oh, I forgot to mention, they need to have at least 22 followers.", "long_pause_offsets": [ 45 ], "source_sha256": "824274a051135892a2737a13ff09f3e07d3240342bcf9d1f438223a40b77f964", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing follower threshold condition, which matches the original instruction's requirement of at least 22 followers. The genre and app remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all the um... classical artists on ah... Spotify. ", "Oh, I forgot to mention, they need to have let's see... at least twenty-two followers." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Follow all the classical artists on Apple Music that have at least 22 followers. Wait, I meant Spotify, not Apple Music.", "long_pause_offsets": [ 81 ], "source_sha256": "00ee65495fb84248e8aab98fd9c457b88546f2daaebd407fa9a66a61b4eb58ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, while retaining the genre, follower threshold, and all-artists scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/revision/revision_03_app/15e0433da6c546f8/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "9af098f840404d1638b5910c4c1c036cfa2b70dd8f43ba2805338743f01d1284", "seconds": 15.04, "spoken_utterances": [ "Follow all the classical artists on uh, let me think... Apple Music that have at least uh... twenty-two followers. ", "Wait, I meant Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the app (Spotify), genre (classical artists), and constraint (at least 22 followers). No other constraints were changed or lost.", "heard_transcript": "Follow all the classical artists on um, let me think, uh, Apple Music that have at least uh, twenty two followers. Wait, I meant Spotify, not Apple Music." }, "candidate_id": "wav_29e22b73c9ad783a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, multi-stage revision. They begin with a clear request, hesitate with 'um' and 'let me think', then correct the platform from 'Apple Music' to 'Spotify' with a believable 'wait, I meant...' self-correction. The prosody is smooth, the pauses are well-placed, and the correction flows naturally within the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_04_comparison_operator", "proposed_instruction": "Follow all the classical artists on Spotify that have at most 22 followers. The cutoff should be at least 22 followers, not at most 22.", "long_pause_offsets": [ 76 ], "source_sha256": "bdeff7ef0cbf8128281b27a10e71e17496731da07defc36861f70b0a35459cd2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the comparison operator from 'at most' to 'at least', reversing the original condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/revision/revision_04_comparison_operator/a0352e8b58a8093f/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "c757f76cae7d6ae8fd39826776baec4fc9991e1a278a2bfe75759c86d5c625e8", "seconds": 17.2, "spoken_utterances": [ "Follow all the uh, let me think... classical artists on Spotify that have at most ... twenty-two followers. ", "The cutoff should be at least twenty-two followers, not uh... at most twenty-two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most twenty two followers' but immediately corrects themselves with 'The cutoff should be at least twenty two followers, not uh, at most twenty two.' This is a clear revision where the final audible intent overrides the initial incorrect value. The final instruction matches the original reference, so the task should be executed as revised.", "heard_transcript": "Follow all the uh, let me think, classical artists on Spotify that have at most twenty two followers. The cutoff should be at least twenty two followers, not uh, at most twenty two." }, "candidate_id": "wav_1acaab4bbaa66179", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with 'follow all the uh' and 'let me think', then states the initial request. They then correct 'at most 22 followers' to 'at least 22 followers' with a natural-sounding 'The cutoff should be at least 22 followers, not uh... at most 22.' The self-correction is clear and the hesitation feels genuine, though the correction is slightly more abrupt than audio_04." } ] }, "is_default": false }, { "variant_id": "revision_02_follower_count", "proposed_instruction": "Follow all the classical artists on Spotify that have at least 50 followers. Make that at least 22 followers rather than 50.", "long_pause_offsets": [ 77 ], "source_sha256": "ab953015d7c7d22696d792498749a51183e5d8b9da63c0a1016299255e9ee0b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly revises the follower threshold from fifty to twenty-two, while retaining the genre, app, and all-artists scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/revision/revision_02_follower_count/eeeb85df9a9c191a/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "8c1c9b04fc8d4e2d1ca5f24cd38da39359e300262e38328710d5b469ad702823", "seconds": 14.68, "spoken_utterances": [ "Follow all the uh, let me think... classical artists on Spotify that have at least uh... fifty followers. ", "Make that at least twenty-two followers rather than um, let's see... fifty." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at least fifty followers' but then corrects themselves with 'Make that at least twenty two followers rather than... fifty.' This is a clear revision where the final value is 'twenty two followers'. The original instruction also specifies 'at least 22 followers', so the final audible intent matches the reference instruction exactly.", "heard_transcript": "Follow all the uh let me think classical artists on Spotify that have at least uh fifty followers. Make that at least twenty two followers rather than um let's see fifty." }, "candidate_id": "wav_051b43a2396be870", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with 'follow all the uh' and 'let me think', then states 'classical artists on Spotify that have at least uh... 50 followers'. They then revise to 'Make that at least 22 followers rather than uh... let's see... 50'. The correction is clear, but the phrase 'rather than uh... let's see... 50' sounds slightly less natural than the others, with a more awkward phrasing and a filler that doesn't quite fit the flow of the correction." } ] }, "is_default": false }, { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all the jazz artists on Spotify that have at least 22 followers. I got the genre wrong. I meant classical, not jazz.", "long_pause_offsets": [ 72 ], "source_sha256": "536f2c8c1b36a02a016ba2178888c12931bf2327c7fc0f4ad9a91386f7a97b7a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from jazz to classical, while retaining the follower threshold and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/revision/revision_01_genre/08de51513061541a/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "0989e543607b4ce1d0a1a882688db60851cc39b1b2bdf89eccd30236db9a9d6d", "seconds": 16.2, "spoken_utterances": [ "Follow all the um... jazz artists on let me think... Spotify that have at least um, let's see... twenty-two followers. ", "I got the genre wrong. I meant classical, not jazz." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the genre as 'jazz' but immediately corrects it to 'classical' after a pause, clearly revising the original instruction. The correction is unambiguous and applies to the genre condition. All other constraints (Spotify, at least 22 followers) remain consistent. The revision type criterion is satisfied as the final intent is the corrected version.", "heard_transcript": "Follow all the um jazz artists on let me think uh Spotify that have at least um let's see twenty two followers. I got the genre wrong, I meant classical, not jazz." }, "candidate_id": "wav_d54a43ba6d995f8c", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with 'follow all the uh' and 'let me think', then says 'jazz artists on... Spotify that have at least... let's see... 22 followers'. They then correct 'jazz' to 'classical'. However, the correction 'I got the genre wrong, I meant classical, not jazz' is delivered as a separate sentence after the initial request is already complete, making it sound more like a post-hoc correction than a natural revision within the thought process. The prosody is also slightly less fluid than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "3f496dfaa97819e75bdd349dcba21666ffd4c754aa285f2579f587e31335aa32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/cancellation/cancel_037/21b4351908821574/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "f9ba6880d57b01d2e0cf07655e50f88fb676b6819cce8f3734df90cd6f12b0f5", "seconds": 12.8, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers.", "Oh, wait, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'Oh wait, stop there. Cancel the whole task.' This constitutes a final, unambiguous cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers. Oh wait, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "7f321dc0923b46c538efbc2cb648a6ab528c420796a98d3e8e033f318eacb477", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/ambiguous_halt/halt_033/7dd6a485dfca82e3/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2c5fccabc01828049c87801bfd634156c51cb9a4b98b4d1e36a2c69a47a7a6f3", "seconds": 11.8, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or resumption. The initial instruction is not executed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers. Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ca583bed72925ae08ca644967890116a8442c940597989164fcd035533bc1e97", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_1/audio_natural/filler/original/6dbffda2be7ca1eb/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "8ebefaea188be59a3d77bf0f31c5c4b0680cf01884995f06bd6a404cd7e55681", "seconds": 7.96, "spoken_utterances": [ "Follow all the uh... classical artists on um, let's see... Spotify that have at least twenty-two followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core task. The final intent matches the original instruction, with 'twenty two' being a correct verbalization of '22'. No correction or cancellation is made. The task is executed as intended.", "heard_transcript": "Follow all the uh classical artists on um let's see Spotify that have at least twenty two followers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Follow all the edm artists on Apple Music—make that Spotify that have at least 23 followers.", "long_pause_offsets": [], "source_sha256": "564b4ee7fd4b9cface7c9ebf7754e5e5be8375e2c381beea3e7c5f8a1679583b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The genre (edm), action (follow), and follower threshold (twenty-three) are preserved in the final utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/false_start/false_start_02_app/67a83944424c082d/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "a2bf75807bfabcfcbc16270cba37831298c94ab60daae12ea5ead82488ec6bba", "seconds": 9.76, "spoken_utterances": [ "Follow all the edm artists on Apple Music—hmmm... make that Spotify that have at least ... twenty-three followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' with 'Make that Spotify'. This is a false start where the first value is replaced by the second. The final intent matches the original instruction exactly, including the platform (Spotify), genre (EDM), and follower threshold (23). The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Follow all the EDM artists on Apple Music. Hmm. Make that Spotify that have at least twenty three followers." }, "candidate_id": "wav_3badea15954ae75d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'Apple Music', hesitates with 'Hmm', then smoothly revises to 'Spotify' with a believable change-of-mind intonation. The phrase 'make that Spotify' is a pragmatic and common way to correct oneself in conversation. The pacing is consistent, and the correction flows seamlessly within the utterance without awkward pauses or stutters." } ] }, "is_default": true }, { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all the rock—or rather, edm artists on Spotify that have at least 23 followers.", "long_pause_offsets": [], "source_sha256": "e9bcc9916459c95cf6a11e4145a58201b408cba7fba9713987ee90e8604d4764", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock' but immediately corrects to 'edm' within the same utterance. The genre is revised to match the original instruction, and all other conditions (app, scope, follower threshold) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/false_start/false_start_01_genre/02c71f871a7b7ad3/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "10c2698b5cae89d1ac1852ba32ce0df8252422c6379179a9794f55e3b4261321", "seconds": 9.52, "spoken_utterances": [ "Follow all the rock—um... or rather, edm artists on Spotify that have at least uh... twenty-three followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock' but immediately self-corrects to 'edm' with 'or rather'. The final audible intent is to follow all EDM artists on Spotify with at least 23 followers, which matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Follow all the rock um, or rather, edm artists on Spotify that have at least uh, twenty three followers." }, "candidate_id": "wav_52cc721e909a15bf", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'rock um' followed by a self-correction to 'Edm artists'. The 'um' filler is natural, and the correction is immediate and plausible. However, the transition from 'rock' to 'Edm' is slightly abrupt, and the 'um' is used more as a hesitation filler than a genuine attempt to recall a word. The overall delivery is realistic but less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_comparison_operator", "proposed_instruction": "Follow all the edm artists on Spotify that have at most—no, I mean at least 23 followers.", "long_pause_offsets": [], "source_sha256": "aed2f83a33a4b1be108080891405262010ce785ec5b2985e9982cfcdc2ac8c81", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most' but immediately corrects to 'at least twenty-three followers'. The final value matches the original instruction's threshold and condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/false_start/false_start_03_comparison_operator/3ddbac48c60c2bd6/speech.wav", "wer": 0.1, "wav_sha256": "4aa5dcbc66ae063a1e750fbcade9890cb03d7ebb05fa18ed962e208b7fe5b37c", "seconds": 8.88, "spoken_utterances": [ "Follow all the edm artists on Spotify that have ... at most—uh... no, I mean at least twenty-three followers." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_9325c2bbaa9d547e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'at most' followed by a correction to 'at least'. While the self-correction is present, the initial phrase 'at most' is spoken with a flat, robotic intonation, and the correction lacks the natural prosodic shift seen in the other clips. The delivery sounds more like a rehearsed or read instruction than a spontaneous correction, making it the least realistic of the three." } ] }, "is_default": false }, { "variant_id": "false_start_04_follower_count", "proposed_instruction": "Follow all the edm artists on Spotify that have at least 15—sorry, that should be 23 followers.", "long_pause_offsets": [], "source_sha256": "b85bae652aaf26f26287ccd07badb1ed23ab5e0ce3906d863f25bf8a9af62790", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen' but immediately corrects to 'twenty-three' within the same utterance. The final value matches the original instruction, and all other conditions (genre, app, scope) remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all the um... edm artists on Spotify that have at least fifteen—uh... sorry, that should be twenty-three followers." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Follow all the edm artists on Spotify. I didn't mention this, but they need to have at least 23 followers.", "long_pause_offsets": [ 39 ], "source_sha256": "f168f1d953197ac7ffc1d20cc601aa1e7a0aee44f8d2efab4284cd481bae359e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing follower threshold condition, which matches the original instruction's requirement. The genre (EDM), app (Spotify), and action (follow all) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/addition/addition_01/82c7014ccc571e11/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "c97dcf303235d1a20dc73b762c91f0ef76b4bcd9a749a1c3ed0731b6ff253ea0", "seconds": 13.72, "spoken_utterances": [ "Follow all the um... edm artists on ... Spotify. ", "I didn't mention this, but they need to have hmmm... at least twenty-three followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow all the um, EDM artists on Spotify.' The speaker then adds a constraint: 'they need to have, hmm, at least twenty three followers.' This is a clear addition of a missing condition to the original task. The final intent combines the initial request with the added constraint, matching the reference instruction. The hesitation and filler words ('um', 'hmm') do not change the final goal.", "heard_transcript": "Follow all the um, EDM artists on Spotify. I didn't mention this, but they need to have, hmm, at least twenty three followers." }, "candidate_id": "wav_125bbc0ce4216d16", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Follow all the artists on Spotify that have at least 23 followers. To narrow that down, only follow EDM artists.", "long_pause_offsets": [ 67 ], "source_sha256": "109a04717351e4986a9b9f89b6085a9b30e6f155ce0c2dcd107302730c41997e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a genre restriction to the initial broad request, preserving the follower threshold and app. The final goal combines both constraints." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all the artists on uh... Spotify that have at least twenty-three followers. ", "To narrow that down, only follow uh, let's see... EDM artists." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Follow all the edm artists on Apple Music that have at least 23 followers. Let me correct the app: follow artists on Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "4433ba67e190cbd7c7b4143bd0242536341a69f61966a237ea4a3500b8d24c08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the genre (edm), the action (follow), and the follower threshold (at least twenty-three)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/revision/revision_02_app/45600ea1034bc794/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "0da1589942166a51c5bd9f71e6fb0d534c2165a600c40f1c12ce2b8a69e36ffa", "seconds": 14.76, "spoken_utterances": [ "Follow all the ... edm artists on ... Apple Music that have at least uh... twenty-three followers. ", "Let me correct the app: follow artists on Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with only a minor pronunciation difference ('EDM' vs 'edm') that does not affect meaning.", "heard_transcript": "Follow all the EDM artists on Apple Music that have at least uh twenty three followers. Let me correct the app. Follow artists on Spotify, not Apple Music." }, "candidate_id": "wav_5d390cddcac5cc59", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural pause and filler ('uh'), then makes a self-correction ('Let me correct the app') that is smoothly integrated into the flow. The correction is pragmatic and spoken with a believable shift in intonation, and the final instruction is delivered with confidence. The prosody is consistent, and the speaker sounds like a real person recalling and adjusting their request." } ] }, "is_default": true }, { "variant_id": "revision_03_comparison_operator", "proposed_instruction": "Follow all the edm artists on Spotify that have at most 23 followers. Change the condition to 'at least' instead of 'at most'.", "long_pause_offsets": [ 70 ], "source_sha256": "061bbb76e2a254929d18014185a9c239ba37f5e613493d6f6b96a04600775d83", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the comparison operator from 'at most' to 'at least', preserving the genre, app, and follower count. The final condition matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/revision/revision_03_comparison_operator/6785b1defe861089/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "a8ceeec5ca66071a817d2b7db801509384d6a3c77a0058552a05e949a7167d29", "seconds": 15.52, "spoken_utterances": [ "Follow all the um, let's see... edm artists on Spotify that have at most let me think... twenty-three followers. ", "Change the condition to 'uh... at least' instead of 'at most'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most twenty three followers' but then explicitly revises the condition to 'at least instead of at most'. This is a clear revision where the final audible intent corrects the initial value. The final goal matches the original instruction, so the revision is correctly applied.", "heard_transcript": "Follow all the um, let's see, EDM artists on Spotify that have at most, let me think, twenty three followers. Change the condition to uh, at least instead of at most." }, "candidate_id": "wav_d1a8f68fd316c4ad", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a plausible false start ('um let's see') and a self-correction ('Change the condition to uh at least instead of at most'), which is a realistic repair. The hesitation and filler are well-placed, and the correction is clearly spoken as a follow-up to the initial mistake. The delivery is slightly less fluid than audio_03, but still sounds like a genuine, spontaneous correction." } ] }, "is_default": false }, { "variant_id": "revision_04_follower_count", "proposed_instruction": "Follow all the edm artists on Spotify that have at least 15 followers. Make that at least 23 followers rather than 15.", "long_pause_offsets": [ 71 ], "source_sha256": "da0065a7e375dee8d3136382ce8668615734f10384abc04fc636ee67c23042f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the follower threshold from fifteen to twenty-three, while retaining the genre (edm), app (Spotify), and scope (all artists)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/revision/revision_04_follower_count/775b1837f095ec16/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "555633adac7f374e35868280576ff2d922fcb70bc99003ee0cfa8d27ad76f82c", "seconds": 13.84, "spoken_utterances": [ "Follow all the edm artists on ah... Spotify that have at least um... fifteen followers. ", "Make that at least uh... twenty-three followers rather than fifteen." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen followers' but then corrects it to 'twenty three followers' after a pause, which matches the original instruction. This is a revision where the final value is applied to the earlier instruction. The correction is clear and the final intent aligns with the reference.", "heard_transcript": "Follow all the EDM artists on uh Spotify that have at least um fifteen followers. Make that at least uh twenty three followers rather than fifteen." }, "candidate_id": "wav_42b86c0169e8c903", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a revision to the follower count ('Make that at least uh 23 followers rather than 15'), which is a believable change-of-mind. However, the phrasing is slightly awkward ('Make that at least uh 23 followers rather than 15'), and the intonation lacks the natural flow of a true self-correction. The pause before 'uh' feels a bit too long and the overall delivery is less smooth than the top two." } ] }, "is_default": false }, { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all the rock artists on Spotify that have at least 23 followers. Wait, the genre should be edm, not rock.", "long_pause_offsets": [ 72 ], "source_sha256": "e6521be20d68f7b5d69e375f24bda0aa3b8de7acaff1ffd5597acc40d485cd49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the genre from 'rock' to 'edm' while retaining the app, the 'all artists' scope, and the follower threshold of twenty-three." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/revision/revision_01_genre/d518b10032395247/speech.wav", "wer": 0.125, "wav_sha256": "cc213f38034267fb340f3211c4a8f04367484df249e843506267f384b65bc19c", "seconds": 11.72, "spoken_utterances": [ "Follow all the uh... rock artists on um... Spotify that have at least ... twenty-three followers. ", "Wait, the genre should be edm, not rock." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the genre as 'rock' but immediately corrects it to 'EDM' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to follow EDM artists with at least 23 followers, which matches the original reference instruction. All other constraints (Spotify, 23 followers) are preserved.", "heard_transcript": "Follow all the uh rock artists on um Spotify that have at least twenty three followers. Wait, the genre should be EDM, not rock." }, "candidate_id": "wav_d13d8f8e8b6074a5", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request but then makes a correction ('Wait, the genre should be Edm, not rock'). While the correction is semantically valid, the delivery feels abrupt and the pause before 'Wait' is too short to sound like a genuine hesitation. The correction is spoken with a flat, declarative tone rather than a natural shift in thought, making it sound less spontaneous and more like a rehearsed edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "7f629e5215633bf0bc6c5b48cc4084d808da0d688416d41cc9b01d81d186a36c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/cancellation/cancel_007/41114fb65870ca88/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "2a8c8af995d8de80be847c14d76f7fec74d656ac1476023debed42e87f065470", "seconds": 14.0, "spoken_utterances": [ "Follow all the let's see... edm artists on ... Spotify that have at least um... twenty-three followers.", "Wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a task to follow litzy EDM artists on Spotify with at least 23 followers. Immediately after, they say 'Wait, cancel my last request,' which is a clear and direct cancellation of the prior instruction. The final audible intent is cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "Follow all the litzy edm artists on Spotify that have at least um twenty three followers. Wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "fc7fefb2c89d9ccbe6ba05684530e5aec840661d50c50d791c67735c2b029ec6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/ambiguous_halt/halt_019/12996c7770e4a5e6/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "bb32dd1cd7f1a21753101cb39ac02e334670be5fd5fe2b5f6aac60a3c51aa86a", "seconds": 13.4, "spoken_utterances": [ "Follow all the let's see... edm artists on ... Spotify that have at least um... twenty-three followers.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Can you give me a moment?' is a clear request to pause or wait, with no indication of cancellation or resumption. The initial instruction is not fully executed, but the user's final audible intent is to halt the process temporarily. This matches the 'ambiguous_halt' type criterion, which requires only an audible pause request without permanent cancellation.", "heard_transcript": "Follow all the litzy edm artists on Spotify that have at least um twenty three followers. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "06c00d04491c30eddb5e09989bc8abbc518e2568436067822761cc3c112d8391", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_2/audio_natural/filler/original/c83691d4efb930ef/speech.wav", "wer": 0.29411764705882354, "wav_sha256": "607bfbd767338290f4d0256bb7be9040c29cb2f9baa7522c49622119b09775fb", "seconds": 9.88, "spoken_utterances": [ "Follow all the let's see... edm artists on ... Spotify that have at least um... twenty-three followers." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "07b42fd_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all the rock—I meant reggae artists on Spotify that have at least 21 followers.", "long_pause_offsets": [], "source_sha256": "fc8359acc0b5d8fd8c9023007ea16341b175489da9026514456de3d24b4a4cbf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'rock' to 'reggae' within the same utterance, preserving the app, genre scope, and follower threshold." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/false_start/false_start_01_genre/5db66db409f77462/speech.wav", "wer": 0.1, "wav_sha256": "a126dddf5cb7c85dc9c7c26c5e8918aeaed7ad19ddf85f95987c8ce9d9946732", "seconds": 10.64, "spoken_utterances": [ "Follow all the rock—uh... I meant reggae artists on let's see... Spotify that have at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock artists' but immediately self-corrects to 'reggae artists'. The correction is clear and immediate, with no additional changes to the platform (Spotify) or the follower threshold (21). The final intent matches the original instruction exactly, satisfying the false_start criterion.", "heard_transcript": "Follow all the rock artists. I meant reggae artists on, let's see, Spotify that have at least twenty one followers." }, "candidate_id": "wav_c7f783ae7da25a65", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a clear false start ('rocka'), immediately self-corrects with 'I meant reggae artists', and then continues with a believable hesitation ('let's see') before completing the request. The intonation and pacing are consistent with someone thinking aloud, and the correction feels genuinely integrated into the utterance rather than tacked on." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Follow all the reggae artists on Apple Music—sorry, make that Spotify that have at least 21 followers.", "long_pause_offsets": [], "source_sha256": "59b7cdfcb10efb19eee5dd3ec63f87cd80e74a956eec1daa65868b436bd20e40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says Apple Music but immediately corrects to Spotify. The genre, action, and follower threshold remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/false_start/false_start_02_app/a8507a692c57f421/speech.wav", "wer": 0.1, "wav_sha256": "e4c7ec85e3c2d58779853ea671999fdadaf18985817e61d82519a50a8625c004", "seconds": 11.92, "spoken_utterances": [ "Follow all the reggae artists on ... Apple Music—um... sorry, make that ... Spotify that have at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as stated in the correction.", "heard_transcript": "Follow all the reggae artists on Apple Music. Um, sorry, make that Spotify that have at least twenty one followers." }, "candidate_id": "wav_4f825aa57b982cae", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision. The speaker starts with 'Apple Music', then self-corrects to 'Spotify'. The 'sorry' is a natural filler that signals a change of mind. The prosody is smooth, and the correction is delivered with a slight shift in intonation that reflects a spontaneous realization, making it very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_03_comparison_operator", "proposed_instruction": "Follow all the reggae artists on Spotify that have at most—no, that should be at least 21 followers.", "long_pause_offsets": [], "source_sha256": "a23165c2f9ee4fea7aa6fee1e171a912a029cb4285572cf5ee44c854093b470d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most' but immediately corrects to 'at least twenty-one followers'. The final comparison operator and value match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/false_start/false_start_03_comparison_operator/d12dfd03fe733f95/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "33c4f59574b75674c4c5133735c9c1e2b7f5d1dd8eb75deb442ee1d004b1b8a9", "seconds": 11.4, "spoken_utterances": [ "Follow all the ah... reggae artists on um... Spotify that have at most—uh... no, that should be at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'at most' but immediately self-corrects to 'at least twenty one followers'. According to the false_start type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final audible intent matches the original instruction.", "heard_transcript": "Follow all the uh reggae artists on um Spotify that have at most uh no, that should be at least twenty one followers." }, "candidate_id": "wav_cda64daa96c12a67", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker initially says 'at most' and then corrects to 'at least'. The 'no' is a plausible interjection, and the correction is delivered with a natural change in tone. However, the initial phrase 'at most' is less common in this context, making the correction slightly less intuitive than in other clips, and the overall flow is a bit more hesitant." } ] }, "is_default": false }, { "variant_id": "false_start_04_follower_count", "proposed_instruction": "Follow all the reggae artists on Spotify that have at least 15—make that 21 followers.", "long_pause_offsets": [], "source_sha256": "383d19a1b40215c56dd33d015b497e6bdc9e589c30add4e6777c2ba0de54f22f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces fifteen with twenty-one; the genre, app, and all-artists scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/false_start/false_start_04_follower_count/f823909ecde5bd5a/speech.wav", "wer": 0.0, "wav_sha256": "e1455551e78447ba04a53ee3c82750edf6fbc38f6e97ec772ce0d71c32b3a3ff", "seconds": 12.64, "spoken_utterances": [ "Follow all the let me think... reggae artists on Spotify that have at least fifteen—um... make that ... twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen' but immediately self-corrects to 'twenty one' with the phrase 'make that'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Follow all the let me think reggae artists on Spotify that have at least fifteen um make that twenty one followers." }, "candidate_id": "wav_0f50657652f06a42", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start ('follow all the let me think') followed by a correction. However, the 'let me think' is an unusually long and awkward pause that disrupts the natural flow. The correction is abrupt and lacks the smooth transition seen in other clips. The overall delivery feels more like a rehearsed or robotic correction rather than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Follow all the reggae artists on Spotify. I didn't mention this, but they need to have at least 21 followers.", "long_pause_offsets": [ 42 ], "source_sha256": "09873f4614e02027b0cd1a4bb786b07030a9147785b50ab415184fd20db06647", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint about follower count, which matches the original instruction's requirement of at least 21 followers. The genre and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/addition/addition_01/6f73e2be0e49db80/speech.wav", "wer": 0.04, "wav_sha256": "3370852a0a6fe4b8ca9483960914211c6cac90084eb23d90f2752eb0d5149d7d", "seconds": 13.44, "spoken_utterances": [ "Follow all the uh... reggae artists on ... Spotify. ", "I didn't mention this, but they need to have let me think... at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, but the speaker adds the missing constraint 'at least twenty one followers' after a pause. This addition completes the original instruction, and the final intent matches the reference. The speaker's hesitation ('uh', 'let me think') is a filler and does not change the final goal.", "heard_transcript": "Follow all the uh reggae artists on Spotify. I didn't mention this, but they need to have let me think, at least twenty one followers." }, "candidate_id": "wav_9f5bbffa56521378", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('uh') and smoothly self-corrects to 'reggae artists'. The delivery has a credible, hesitant rhythm with a plausible pause and filler ('let me think') before adding the constraint about '21 followers'. The intonation and pacing suggest genuine spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Follow all the artists on Spotify that have at least 21 followers. To narrow that down, only follow reggae artists.", "long_pause_offsets": [ 67 ], "source_sha256": "e016bff879d6c541bedd246931c589d7d69b366b62e4f9ca4013f2ba9711e5c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a genre restriction to the initial broad request, preserving the follower threshold and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/addition/addition_02/7604ee839ecdb666/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "ecae26118e92900fe4100ea772aed584f0fd749faddb19c100deec89e414eeff", "seconds": 15.64, "spoken_utterances": [ "Follow all the artists on um... Spotify that have at least um, let's see... twenty-one followers. ", "To narrow that down, only follow um... reggae artists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a broad instruction to follow artists with at least 21 followers on Spotify, then adds a constraint to narrow the scope to only reggae artists. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final audible intent is clear and executable, matching the original reference instruction.", "heard_transcript": "Follow all the artists on um Spotify that have at least um let's see twenty one followers. To narrow that down, only follow um reggae artists." }, "candidate_id": "wav_1c786cdccb541684", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes natural fillers ('um') and a moment of hesitation ('let's see'), the overall flow is less smooth. The initial request is more fragmented, and the addition of 'to narrow that down' feels like a slightly awkward, post-hoc justification rather than a natural thought process. The second 'um' before 'reggae artists' also sounds less spontaneous and more like a filler inserted to fill silence, reducing the sense of genuine hesitation." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_comparison_operator", "proposed_instruction": "Follow all the reggae artists on Spotify that have at most 21 followers. That should be 'at least', not 'at most'.", "long_pause_offsets": [ 73 ], "source_sha256": "ae5a41acfd0a270380e834744eb45f3d016f6b0033b502107d9d99abb0c9496b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the comparison operator from 'at most' to 'at least', preserving the genre, app, and follower count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/revision/revision_03_comparison_operator/9ab981e43308f200/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "89ee9e7395b9831b9902f66ae6e7527e5809b326485151afd4a9808de9e2b9d0", "seconds": 16.16, "spoken_utterances": [ "Follow all the reggae artists on ... Spotify that have at most hmmm... twenty-one followers. ", "That should be '... at least', not 'at most'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'at most twenty one followers' but immediately corrects themselves with 'That should be at least, not at most.' This constitutes a revision where the final intent is the corrected version. The original instruction was 'at least 21 followers,' and the final audible intent matches this exactly. Therefore, the revision criterion is satisfied.", "heard_transcript": "Follow all the reggae artists on Spotify that have at most twenty one followers. That should be at least, not at most." }, "candidate_id": "wav_73f37fbdf2edd329", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural hesitation ('hmm') before self-correcting the quantifier from 'at most' to 'at least'. The correction is smoothly integrated with a plausible self-aware phrase ('That should be...'), and the prosody reflects genuine thought and revision. The pacing and intonation are consistent with a person formulating and refining a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_04_follower_count", "proposed_instruction": "Follow all the reggae artists on Spotify that have at least 15 followers. Make that at least 21 followers rather than 15.", "long_pause_offsets": [ 74 ], "source_sha256": "b418937740f5cd709ed7745c1b26e73122754db986df96039c35db17728741c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the follower threshold from fifteen to twenty-one, while retaining the genre (reggae), app (Spotify), and scope (all artists)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/revision/revision_04_follower_count/52566b6c13a7ee52/speech.wav", "wer": 0.0, "wav_sha256": "396eb732f7d051c2c40cd8c219493912e51e90985ea9defb54fb185b6f9fd0ee", "seconds": 16.56, "spoken_utterances": [ "Follow all the reggae artists on let me think... Spotify that have at least ... fifteen followers. ", "Make that at least ... twenty-one followers rather than fifteen." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen followers' but then corrects it to 'twenty one followers' after a pause, which matches the original instruction. The revision is clear and final, so the task should be executed with the corrected value.", "heard_transcript": "Follow all the reggae artists on let me think Spotify that have at least fifteen followers. Make that at least twenty one followers rather than fifteen." }, "candidate_id": "wav_98ac81348ceeff6f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a filler ('Let me think') and a clear revision of the follower count from 15 to 21. The correction is embedded naturally within the sentence ('Make that... rather than...'), and the intonation reflects a spontaneous change of mind. The delivery is fluent and the revision is pragmatically plausible, though slightly less detailed than audio_03's correction." } ] }, "is_default": false }, { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all the rock artists on Spotify that have at least 21 followers. I misspoke about the genre; use reggae instead of rock.", "long_pause_offsets": [ 72 ], "source_sha256": "c81ab9b42a401fd7a3fcf829381cfcaf34df2651180bc730ac149af696716489", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from 'rock' to 'reggae' while retaining the app, the 'all artists' scope, and the follower threshold." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/revision/revision_01_genre/488d2796d23e7e82/speech.wav", "wer": 0.08, "wav_sha256": "3ac06f80bfbf99dc67a827dce380f553ebd0347acf6efdf18a37644b66a0b803", "seconds": 13.0, "spoken_utterances": [ "Follow all the ... rock artists on Spotify that have at least um... twenty-one followers. ", "I misspoke about the genre; use reggae instead of rock." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock artists' but immediately corrects themselves with 'I misspoke about the genre. Use reggae instead of rock.' This is a clear revision where the final intent is to follow reggae artists, not rock artists. The value '21 followers' remains unchanged. The revision applies to the genre, which is the only incorrect part of the initial instruction. The final goal matches the reference instruction.", "heard_transcript": "Follow all the rock artists on Spotify that have at least um twenty one followers. I misspoke about the genre. Use reggae instead of rock." }, "candidate_id": "wav_e517503866fd0514", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a genre correction from 'rock' to 'reggae' after completing the initial request. The phrase 'I misspoke about the genre' is a believable self-correction, but the delivery is slightly more formal and less conversational than the others. The pause before the correction is natural, but the overall flow is less fluid than the top two." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Follow all the reggae artists on Apple Music that have at least 21 followers. Change the app to Spotify, not Apple Music.", "long_pause_offsets": [ 78 ], "source_sha256": "18178ef8f1844c51644a41dc33476023a67b44439423cf3ff3ae8434b84be36d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining the genre (reggae), the action (follow), the scope (all), and the follower threshold (at least twenty-one)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/revision/revision_02_app/8f9c627b4253a1d9/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "e5aa1a4de59a0ccfdb75bbf361d87060bb579e1afa3ae6a5596ad413ef322986", "seconds": 14.96, "spoken_utterances": [ "Follow all the ... reggae artists on Apple Music that have at least um... twenty-one followers. ", "Change the app to ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify, not Apple Music' after a short pause. This constitutes a revision where the final intent is to follow reggae artists on Spotify with at least 21 followers, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Follow all the reggae artists on Apple Music that have at least um twenty one followers. Change the app to Spotify, not Apple Music." }, "candidate_id": "wav_25523dc8b84f254e", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker starts with a clear request but then makes a correction to the app from 'Apple Music' to 'Spotify'. However, the correction is delivered as a separate, abrupt sentence rather than being integrated into the original utterance. The prosody lacks the natural flow of a spontaneous revision, sounding more like a disjointed afterthought. The delivery is less cohesive compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07b42fd_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "c04b16dd228c62a94d66901aabebe700196bff66b3569fb4a26f593703997032", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/cancellation/cancel_015/70b6bf453eae480c/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "edec6a36d3c00675f2020b6fdea895cfad61d5f605ff18ecbffd5521790f3a7a", "seconds": 14.2, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers.", "Hang on, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'cancel the whole thing' after a brief interjection, which constitutes a direct cancellation of the previously stated task. The final intent is to cancel, not to execute the original instruction. The cancellation is unambiguous and applies to the entire task.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least 21 followers. Hang on, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "8c1a1940787a5c37084ff27baca2a296f4d5363369fdfc52d667d147465af368", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/ambiguous_halt/halt_018/52c4258d9c4d09c9/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "b9934c471271d08c05e48edbd5b80edf74b326f3a179e2b46b616b9847d02578", "seconds": 13.4, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which constitutes a clear instruction to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. No cancellation or execution is indicated, and the original task is not revoked. The final control intent is correctly identified as 'pause'.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least 21 followers. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07b42fd_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dd566f06b0462ba66e8f8f6352b4e6c7b5700fe5e8856353f905919f14f81365", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07b42fd_3/audio_natural/filler/original/6465d420cac4c398/speech.wav", "wer": 0.2, "wav_sha256": "ba1b60b9f134075e559cc782024680648413949f900bc39bb7f7291661be7785", "seconds": 9.28, "spoken_utterances": [ "Follow all the hmmm... reggae artists on ... Spotify that have at least twenty-one followers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'twenty one' which is a correct pronunciation of the number 21. There are no hesitations, corrections, or changes to the task. The final intent matches the original instruction, and the filler elements do not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "Follow all the reggae artists on Spotify that have at least twenty one followers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_rating_condition", "proposed_instruction": "Move all products with under 4.5—sorry, I meant under 4.2 rating from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "42c2973493a240e10af382e53ca3d8fb7df0b6ac2fc7a19b18c3ba7543d74906", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'under four point five' but immediately corrects to 'under four point two'. The final value matches the original instruction, and all other conditions (action, source, destination, and scope) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/false_start/false_start_01_rating_condition/78210fb0a89e0a3f/speech.wav", "wer": 0.125, "wav_sha256": "4705c6887633ababde8b8e66722012ed71abf51a2902f44ad7d895d43d592b24", "seconds": 11.88, "spoken_utterances": [ "Move all products with ... under four point five—um... sorry, I meant under four point two rating from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'under four point five' but immediately self-corrects to 'under four point two'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The corrected value matches the original instruction exactly, and all other conditions (action, source, destination) are preserved. Therefore, the final intent is fully aligned with the expected task.", "heard_transcript": "Move all products with under four point five mm, sorry, I meant under four point two rating from my Amazon cart to wish list." }, "candidate_id": "wav_2951817d12f83f93", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker starts with a plausible false start ('under four point five'), immediately self-corrects with 'sorry, I meant under four point two rating', and continues smoothly. The correction is integrated into the sentence without awkwardness, and the prosody reflects genuine hesitation and recollection. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all products with under 4.2 rating from my amazon cart to shopping list—sorry, make that wish list.", "long_pause_offsets": [], "source_sha256": "12012d1bc7440767fe64c26954333ba9d51366072bfffd849b16ae206e4254fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wish list'. The rest of the instruction, including the rating condition, source, and destination, matches the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/false_start/false_start_04_destination_collection/e7c3c4d937932c7b/speech.wav", "wer": 0.2608695652173913, "wav_sha256": "05a6d28701a8c0030e0e8b0922438263b459a932f2c71faf5f77126fc4ca486c", "seconds": 13.36, "spoken_utterances": [ "Move all products with under four point two rating ah... from my amazon cart ... to shopping list—uh... sorry, make that wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wish list'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction with the corrected target list.", "heard_transcript": "Move all products with under four point two rating uh from my Amazon cart to shopping list. Sorry, make that wish list." }, "candidate_id": "wav_77cc868d456fec01", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'uh' and a self-correction from 'Wista' to 'wishlist'. The correction is natural and the speaker continues without restarting. The prosody is slightly more hesitant than audio_04, but the repair is fluent and the overall delivery sounds like a real-time correction rather than a rehearsed or robotic edit." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all products with under 4.2 rating from my amazon wishlist—wait, I meant cart to wish list.", "long_pause_offsets": [], "source_sha256": "e2bee96828ddf34920bc9614d9e0d5b3cda26b614df2eb521ae13161e3387ec4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The speaker initially says 'wishlist' as the source, then corrects to 'cart'. The final source is 'cart', which contradicts the original instruction's 'cart' as the destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/false_start/false_start_03_source_collection/d2653806869ebdbe/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "0e1de6ed9dc8abb811c4c4627f166c10f0442da2d620aff629c6511416b84a08", "seconds": 12.4, "spoken_utterances": [ "Move all products with under four point two rating from my amazon wishlist—um... wait, I meant ... cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wish list' but immediately self-corrects to 'cart to wish list' after a pause and 'wait'. According to the 'false_start' type criterion, the final corrected version of the task should be executed, discarding the initial mistaken value. The final intent matches the original instruction with the correct target ('cart' instead of 'wish list') and all other conditions preserved.", "heard_transcript": "Move all products with under four point two rating from my Amazon wish list. Um, wait, I meant cart to wish list." }, "candidate_id": "wav_0089bd717c754736", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a revision, changing 'wishlist' to 'cart'. The 'um' and 'wait' are believable fillers indicating a change of mind. However, the correction feels slightly more abrupt and less integrated than in audio_01 or audio_04. The prosody is a bit more hesitant, and the transition from the initial request to the correction is less fluid, making it sound slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all products with under 4.2 rating from my ebay—no, I meant amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "93cc3aea49b8b781358b362f7e3a05e379cb99a7649cee2428ce7d55d1755816", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'ebay' but immediately corrects to 'amazon', preserving the app, action, condition, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/false_start/false_start_02_app/93076b15a3771e67/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "69effa4371552dc935d85778039b76d5d9f8d3dd3a874212264fb842ff748c75", "seconds": 10.8, "spoken_utterances": [ "Move all products with under four point two rating ... from my ebay—ah... no, I meant um... amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'no, I meant'. This is a false start where the final value replaces the initial incorrect one. The rest of the instruction, including the rating threshold and target action, remains consistent with the original. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Move all products with under four point two rating from my eBay uh no, I meant um Amazon cart to wish list." }, "candidate_id": "wav_0cd6bd001d748d19", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio contains a false start ('eBay') and a revision ('no, I meant Amazon cart'). While the correction is present, the delivery feels less natural. The 'uh' and 'no' are somewhat disjointed, and the correction is delivered with a slightly robotic or rehearsed cadence. The prosody lacks the smoothness of the other clips, and the overall flow is less convincing as a spontaneous utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Move all products with from my amazon cart to wish list. Oh, I forgot to mention, only move products with under 4.2 rating.", "long_pause_offsets": [ 57 ], "source_sha256": "dcd2978c15bfb6e3f618519d79f9e5c2d48396ddf2105f2d4f16779c641e1fce", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rating condition, which matches the original instruction's threshold and comparison. The initial utterance is incomplete but the final goal is fully specified after the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/addition/addition_01/9607ff121bdc7a8f/speech.wav", "wer": 0.08, "wav_sha256": "fa9406d4f47e8aaa0b124dca8cbf9bad139f968bf53865f404fffcd6492118f9", "seconds": 12.76, "spoken_utterances": [ "Move all products with ... from my amazon cart ... to wish list. ", "Oh, I forgot to mention, only move products with under four point two rating." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('Move all products with from my Amazon cart...'), but the speaker immediately adds the missing constraint ('only move products with under four point two rating'). This constitutes a valid addition, where the final goal combines the original intent with the added condition. The final audible intent matches the reference instruction exactly.", "heard_transcript": "Move all products with from my Amazon cart to wish list. Oh, I forgot to mention, only move products with under four point two rating." }, "candidate_id": "wav_fceba557fd98eb1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then adds a crucial detail ('only move products with under 4.2 rating') in a natural, additive manner. The 'oh I forgot to mention' phrase is a highly plausible real-world hesitation, and the correction is smoothly integrated without awkwardness. The prosody is consistent, and the pacing feels like genuine spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Move all products with under 4.2 rating to wish list. I left out a detail: move them from my amazon cart.", "long_pause_offsets": [ 54 ], "source_sha256": "89e4fad7a8e62819953715a864933c35c389871edb57f1cdec54115651e944be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source location 'from my amazon cart'; the initial instruction already specified the action, condition, and destination. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/addition/addition_02/dc3cc56a3b262489/speech.wav", "wer": 0.24, "wav_sha256": "14f85ad82a8cc2dabc244f456d5937ae6c5ef464fe4e230aa417b0b7730b06f5", "seconds": 13.88, "spoken_utterances": [ "Move all products with hmmm... under four point two rating to wish list. ", "I left out a detail: move them uh... from my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Move all products with under 4.2 rating to wish list.' The speaker then adds 'from my Amazon cart' after a pause and self-correction, which is a clear addition to the original instruction. The final combined task matches the reference instruction exactly. The addition type requires combining the earlier instruction with the added information, which is done here.", "heard_transcript": "Move all products with hmm, under 4.2 rating to wish list. I left out a detail, move them uh from my Amazon cart." }, "candidate_id": "wav_7468ceafeabbccea", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then adds a detail ('from my Amazon cart') after a natural pause and filler ('uh'). The self-correction is believable and fits the flow of a person recalling information. The prosody is mostly natural, though the initial phrase is slightly rushed, making the addition feel a bit less seamless than audio_03." } ] }, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "Move all products with under 4.2 rating from my amazon cart. And make sure you move them to wish list.", "long_pause_offsets": [ 61 ], "source_sha256": "4a77f1175c6f5400217b28b1f3acda734a642163868e65a8af31fb37addd855a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly adds the destination 'wish list' to the initial instruction, which already specified the source, condition, and app. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/addition/addition_03/c2fbf7394acf7111/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "e1a66b0c33047703c4325440f8baeda799f08511474d6517892b57610e58cedd", "seconds": 13.04, "spoken_utterances": [ "Move all products with uh, let's see... under four point two rating from my amazon cart. ", "And make sure you move them uh... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Move all products with uh let's see under four point two rating from my Amazon cart' is followed by an addition 'and make sure you move them uh to wish list.' The addition clarifies the destination, which was not explicitly stated in the original instruction. The final intent combines the original request with the added detail, resulting in a complete and coherent task that matches the reference instruction. The hesitation and fillers ('uh', 'let's see') do not change the core task, and the final addition completes the goal.", "heard_transcript": "Move all products with uh let's see under four point two rating from my Amazon cart and make sure you move them uh to wish list." }, "candidate_id": "wav_66e7968c133c19fa", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker uses multiple fillers ('uh', 'let's see') and pauses, which can indicate hesitation. However, the phrasing 'and make sure you move them uh to wishlist' sounds slightly unnatural and less fluent compared to the others. The addition of the detail feels tacked on rather than smoothly integrated, and the overall delivery is less cohesive, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_collection", "proposed_instruction": "Move all products with under 4.2 rating from my amazon cart to saved items. Use my wish list as the destination instead of saved items.", "long_pause_offsets": [ 76 ], "source_sha256": "f57d1ddb6ceee1083ed95c41d3d1de730b33ed2c93abe44ede9c00c41aa5ebb4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup correctly revises the destination from 'saved items' to 'wish list' while preserving the original condition on product rating and source cart. The app name 'amazon' is retained implicitly from the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/revision/revision_03_destination_collection/c0deaac747c7512b/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "d618ba4a0a8def3cf46c89e69a3e8e0d289e3c4039f26b04cfeee532780ed333", "seconds": 16.64, "spoken_utterances": [ "Move all products with under four point two rating from my amazon cart ... to saved items. ", "Use my uh... wish list as the destination instead of ... saved items." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to saved items' but immediately corrects to 'use my uh wish list as the destination instead of saved items'. This constitutes a revision where the final destination is clearly stated as 'wish list'. The original instruction's intent is preserved with the corrected target, satisfying the revision type criterion.", "heard_transcript": "Move all products with under four point two rating from my Amazon cart to saved items. Use my uh wish list as the destination instead of saved items." }, "candidate_id": "wav_d15ea1d831b8ea2e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a clear, single revision. The speaker first states a complete command, then hesitates with 'use my uh' before correcting the destination from 'saved items' to 'wishlist'. The correction is smooth, the intonation natural, and the pause feels like a genuine moment of recall. The prosody is consistent and the delivery is fluent without any awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_04_app", "proposed_instruction": "Move all products with under 4.2 rating from my Walmart cart to wish list. I misspoke about the store. The cart is on amazon, not Walmart.", "long_pause_offsets": [ 75 ], "source_sha256": "3dffb7abce593230e348321d4606a43c07c9280752a80793d5201ccd257eb7c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the store from Walmart to Amazon, while retaining the rating condition and action. The initial incorrect store is replaced by the correct one." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/revision/revision_04_app/1056839503b73053/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "8436621d8160c22dce66bfd6daf495fe0c5cc0495874f8bf51e6e7093ecd0f5e", "seconds": 16.04, "spoken_utterances": [ "Move all products with under four point two rating ... from my Walmart cart uh... to wish list. ", "I misspoke about the store. The cart is on ... amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task for Walmart but immediately corrects it to Amazon. The revision is clear and unambiguous, and the final intent matches the original instruction. The correction is applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "Move all products with under four point two rating from my Walmart cart uh to wish list. I misspoke about the store. The cart is on Amazon, not Walmart." }, "candidate_id": "wav_2c5fa77b146c793f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable self-correction. The speaker initially says 'Walmart cart' and then immediately corrects to 'Amazon'. The 'I misspoke' phrase is a natural way to acknowledge the error, and the correction is delivered with a slight shift in intonation that sounds spontaneous. The pacing is natural, and the overall flow is coherent." } ] }, "is_default": false }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Move all products with under 4.2 rating from my amazon orders to wish list. Those products should come from my cart, not my orders.", "long_pause_offsets": [ 76 ], "source_sha256": "b775dd952d5f7fd094bd6f4a16012ad48a48018da24809f47a22a1bf49ae0a0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'orders' to 'cart', preserving the rating condition and destination. The original instruction's source was 'cart', so the correction restores it." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/revision/revision_02_source_collection/ebf0d229777beb21/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "2f5f49018adbead2e6c45ebc298a0c5eede9a298ebb1197737654208847629dc", "seconds": 17.76, "spoken_utterances": [ "Move all products with under four point two rating ... from my amazon orders uh... to wish list. ", "Those products should come from my um, let's see... cart, not my orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon orders' but immediately corrects to 'cart' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Move all products with under four point two rating from my Amazon orders uh to wish list. Those products should come from my um let's see cart, not my orders." }, "candidate_id": "wav_0489eab7d9335873", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio shows a revision where the speaker first says 'Amazon orders' and then corrects to 'cart'. The correction is clear, but the phrasing 'those products should come from my um let's see cart' feels slightly more hesitant and less fluid than the others. The 'um' and 'let's see' are natural fillers, but the overall delivery is a bit more labored, making it less smooth than the top two." } ] }, "is_default": false }, { "variant_id": "revision_01_rating_condition", "proposed_instruction": "Move all products with under 3.5 rating from my amazon cart to wish list. Let me correct the cutoff: the rating should be under 4.2, not under 3.5.", "long_pause_offsets": [ 74 ], "source_sha256": "6bdcbb051edf794107162377a87dc833cf03a7cec28e358c87d933821a38d3de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the rating threshold from 'under three point five' to 'under four point two', aligning with the original instruction. All other conditions, including the app (amazon), action (move), source (cart), destination (wish list), and scope (all products), are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/revision/revision_01_rating_condition/8bd7d453e89d41d2/speech.wav", "wer": 0.1, "wav_sha256": "782fd9fc14f8d0b6c5acde27cfd22b9c01dba5855e0ea627f70b1eeb7c8901e3", "seconds": 20.24, "spoken_utterances": [ "Move all products with um... under three point five rating from my amazon cart let's see... to wish list. ", "Let me correct the cutoff: the rating should be um, let's see... under four point two, not under three point five." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'under three and a half rating' but then corrects it to 'under four point two, not under three point five'. This is a clear revision where the final value is explicitly stated and the initial value is negated. The final intent matches the original instruction.", "heard_transcript": "Move all products with um under three and a half rating from my Amazon cart. Let's see. to wish list. Let me correct the cutoff. The rating should be um let's see. under four point two, not under three point five." }, "candidate_id": "wav_4ff718312a33ef5b", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision but feels less natural due to its structure. The speaker first states 'under three and a half rating', then pauses and says 'let's see', followed by a self-correction to 'under four point two'. However, the correction is delivered as a separate sentence ('the rating should be...') rather than being integrated into the original utterance. This makes the revision feel more like a post-hoc correction than a spontaneous self-repair, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e3845e31cf2ce959fa0b22257767bb3f53e3bea10a8e2887f916db46221b82bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/cancellation/cancel_009/95384b1e187e696d/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "ebf5a79887aa7cce0500496529f60a740d7b00de567a22c105fd2f9209ce0dc2", "seconds": 16.28, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list.", "Wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to move products with under 4.2 rating from the Amazon cart to the wish list. After a brief pause and hesitation ('um'), they immediately say 'Wait, cancel that altogether,' which is a clear and final cancellation of the entire task. The final interjection is a direct cancellation command, satisfying the 'cancellation' type criterion.", "heard_transcript": "Move all products with a under four point two rating from my amazon cart um to wish list. Wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "f56fcb87f634b1bbeda9c60ef54769de47d59ef247dbf1c1ea410204c0c3578d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/ambiguous_halt/halt_037/d36f76965034e14f/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "95793615b1d716db9879205bc190051de0f364060e8c772e555d203de68b49a5", "seconds": 15.2, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Pause for a moment,' which constitutes a clear request to halt the current operation. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and complete, and the final control request is unambiguous.", "heard_transcript": "Move all products with a rating under four point two from my Amazon cart um to wish list. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "232c9eb7e17fc3c4fc75a7e689c9046d4c507e60abed62b32b6758269d60a27b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_1/audio_natural/filler/original/cc86bbed06e40c4c/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "499e9c5468cbc86eaf2caaa7cd5102312aaa4c706468dddf8a95e2f6f6ed02d1", "seconds": 11.76, "spoken_utterances": [ "Move all products with uh... under four point two rating hmmm... from my amazon cart um... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core task. The final intent matches the original instruction, with no self-correction or modification to the action, target, or value. The hesitation does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Move all products with uh under four point two rating from my amazon cart um to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_rating_condition", "proposed_instruction": "Move all products with over 4.2—no, that should be over 3.7 rating from my amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "6e8827eded2214d69a5685393cddc5dba102ecf42ff067f6ee1be079a5657eb3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four point two' but immediately corrects to 'over three point seven'. The final value matches the original instruction's rating threshold." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/false_start/false_start_01_rating_condition/7d6893f120fc1be8/speech.wav", "wer": 0.25925925925925924, "wav_sha256": "5129801211675ea476f884e543d085dfbc03ec7ae00f0d7f111a527ec1723534", "seconds": 11.72, "spoken_utterances": [ "Move all products with ah... over four point two—uh... no, that should be over three point seven rating from my amazon wish list um... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four point two' but immediately self-corrects to 'over three point seven'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is executed with the corrected value. The rest of the instruction matches the original, so the final intent is correct.", "heard_transcript": "Move all products with a uh over four point two uh um no, that should be over a three point seven rating from my Amazon wish list um to cart." }, "candidate_id": "wav_cbf0f83ab44c4416", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. The speaker starts with 'a uh over four point two uh um' as a false start, then clearly corrects to 'no that should be over three point seven rating'. The correction is immediate, the intonation reflects a genuine moment of realization, and the flow remains continuous without awkward pauses. The speaker also correctly identifies the source as 'my Amazon wishlist' and the target as 'to cart', with a natural final pause. The repair is fully integrated into the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all products with over 3.7 rating from my amazon wish list to wishlist—make that cart.", "long_pause_offsets": [], "source_sha256": "d513edb227eda668f0cde9bcea2ec31cb7b111d7cd1646d8b99b55fdaab79804", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects it to 'cart' within the same utterance. The final destination is 'cart', matching the original instruction. All other elements—action, condition, source, and app—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/false_start/false_start_04_destination_collection/6ceee72538758a2d/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "6829fc30c6cf02296035578034b3c26566c476f39d6f2922a7e20645ec56ab04", "seconds": 7.92, "spoken_utterances": [ "Move all products with over three point seven rating um... from my amazon wish list to wishlist—uh... make that cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wish list' but immediately self-corrects to 'cart'. The final audible intent is to move the products to the cart, which matches the original instruction. The 'false_start' type requires applying the correction, which has been done.", "heard_transcript": "Move all products with over three point seven rating, um, from my Amazon wish list to wish list, uh, make that cart." }, "candidate_id": "wav_23229d1db42982a3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'rating um' followed by a self-repair to 'make that cart'. The speaker's voice and prosody are consistent, and the correction is delivered smoothly. However, the initial error ('rating um') is less specific than in audio_04, and the repair is slightly more abrupt. The overall delivery is very natural, but the correction is less detailed and the flow is slightly less fluid than in audio_04." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all products with over 3.7 rating from my ebay—I should say amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "404961343aef6c767e7d19686ef7c2cb0a97b5f6b5a3e5235db280f6be5e7561", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the app, rating threshold, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/false_start/false_start_02_app/c78e6ea7a38bbe32/speech.wav", "wer": 0.2727272727272727, "wav_sha256": "4944db7b303d043f91373c13628ad6878954737139b581d70a63464606e55c9f", "seconds": 9.04, "spoken_utterances": [ "Move all products with over three point seven rating from my ebay—uh... I should say um... amazon wish list ... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' before completing the instruction. According to the 'false_start' type criterion, the final corrected value ('Amazon') should be applied, and the original incorrect value ('eBay') should be discarded. The rest of the instruction, including the rating threshold and target action, remains consistent. Therefore, the final intent matches the original instruction after correction.", "heard_transcript": "Move all products with over three point seven rating from my eBay uh, I should say um, Amazon wish list to cart." }, "candidate_id": "wav_bd5b638e47b78117", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'eBay uh I should say um Amazon wishlist'. The speaker corrects the platform from eBay to Amazon, which is a more significant change than just a number. However, the correction is delivered with a slightly hesitant and less fluid intonation. The 'uh' and 'um' fillers are placed in a way that suggests the speaker is formulating the correct information, but the overall delivery feels a bit more disjointed compared to audio_04 and audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all products with over 3.7 rating from my amazon shopping list—sorry, I meant wish list to cart.", "long_pause_offsets": [], "source_sha256": "42189ac921a1e6ef0dd8e2dd088837532dc46ba2b48d331ac72adc5712302f6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects to 'wish list' within the same utterance. The correction is clear and unambiguous, preserving the app, action, condition, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/false_start/false_start_03_source_collection/c509f8439ee4230e/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "64ae5e6b0e0c85474fade7fadcbe1dc030e262fcd95c43c106d1c6a1c4dcec1a", "seconds": 10.04, "spoken_utterances": [ "Move all products with um... over three point seven rating from my amazon ... shopping list—uh... sorry, I meant wish list to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wish list'. According to the 'false_start' type criterion, the final corrected target ('wish list') should be applied, and the rest of the instruction (over 3.7 rating, Amazon, move to cart) remains unchanged. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Move all products with um over three point seven rating from my Amazon shopping list to uh sorry, I meant wish list to cart." }, "candidate_id": "wav_2a59d29a4d943f39", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start with 'um over three point seven rating from my Amazon shopping list uh sorry I meant wishlist to cart'. The speaker corrects 'shopping list' to 'wishlist', which is a meaningful change. However, the correction is delivered with a slightly robotic or rehearsed intonation, and the 'uh' filler before 'sorry' feels a bit forced. The overall delivery is less natural and fluid compared to the other clips, with a slightly less convincing self-correction flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Move all products with over 3.7 rating to cart. Oh, and one more thing: move them from my Amazon wish list.", "long_pause_offsets": [ 48 ], "source_sha256": "55b850c1c59f227b867f264128f2e378070217b6d1574a4830f006d922b171c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source location 'Amazon wish list'; the action, condition, and destination remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/addition/addition_02/9064a585521a9970/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "5e8a3b363e49c3285a94e30b4e4c9de97cac7a72bde9b54973708e3bf6d4e078", "seconds": 14.88, "spoken_utterances": [ "Move all products with uh... over three point seven rating um, let me think... to cart. ", "Oh, and one more thing: move them let's see... from my Amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Move all products with a over three point seven rating um let me think to cart.' This is followed by a longer pause and an addition: 'Oh and one more thing, move them let's see from my Amazon wish list.' The addition clarifies the source of the products as 'my Amazon wish list,' which combines with the initial instruction to form the complete and correct final goal. The type is 'addition,' and the final combined goal matches the original instruction exactly. Therefore, the task is passed.", "heard_transcript": "Move all products with a over three point seven rating um let me think to cart. Oh and one more thing, move them let's see from my Amazon wish list." }, "candidate_id": "wav_1a8986f7954b045c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and realistic flow of spontaneous speech. It begins with a clear initial request, includes a believable 'let me think' filler pause, then transitions into a revision with 'Oh and one more thing' and a follow-up correction that reorders the original request. The prosody is varied, with natural hesitations and a pragmatic change-of-mind tone, making the self-correction sound entirely spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "Move all products with over 3.7 rating from my amazon wish list. And make sure you move them to cart.", "long_pause_offsets": [ 65 ], "source_sha256": "1798a5cf960a77056d1e02302c544e05899a513716da97ce1deadce77067e5e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly confirms the destination as 'cart', which was already implied in the original instruction. The rating threshold and source list are unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/addition/addition_03/716404555615ebd4/speech.wav", "wer": 0.2, "wav_sha256": "a7ff03f18b53ad14e2b17cf26eadf65034ec1429be20a313c39151394630fa03", "seconds": 12.88, "spoken_utterances": [ "Move all products with uh... over three point seven rating um... from my amazon wish list. ", "And make sure you move them uh... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'a over three point seven rating' which is a minor grammatical error but the intended value 'over 3.7 rating' is clear. The addition of 'and make sure you move them uh to cart' clarifies the destination as 'cart', which is consistent with the original instruction. The final intent matches the original instruction with no contradiction or ambiguity. The addition does not introduce any conflicting or incorrect constraints.", "heard_transcript": "Move all products with a over three point seven rating um from my Amazon wish list and make sure you move them uh to cart." }, "candidate_id": "wav_345b6524ff1b6a73", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a realistic addition pattern with a 'uh' and 'um' filler, and a clear follow-up request to 'make sure you move them to cart'. The delivery is fluent and the addition feels like a remembered detail. However, the second part of the utterance is slightly more rehearsed in tone compared to audio_02, and the transition between the two clauses is less organic, making it slightly less realistic overall." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Move all products with from my amazon wish list to cart. I didn't mention this, but only move products with over 3.7 rating.", "long_pause_offsets": [ 57 ], "source_sha256": "97cadf40a0c8b0cec1a08c4144e6ff40b24d89a8422a8a1c5de15f6f764c0894", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rating condition, which is consistent with the original instruction. The initial utterance omits the condition but the final instruction includes it, preserving the intended scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/addition/addition_01/b249714aa5aaf892/speech.wav", "wer": 0.08, "wav_sha256": "01e20db7a24651b58f64284f14b0204c46e56a9ab4a8392bb966191750c6c03e", "seconds": 12.0, "spoken_utterances": [ "Move all products with from my amazon wish list ... to cart. ", "I didn't mention this, but only move products with ... over three point seven rating." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('Move all products with from my Amazon wish list to cart'), but the speaker immediately adds the missing constraint ('only move products with over 3.7 rating'). This addition completes the original instruction with the correct condition. The final audible intent matches the reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Move all products with from my Amazon wish list to cart. I didn't mention this, but only move products with over 3.7 rating." }, "candidate_id": "wav_d34a02d6683ecb99", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio presents a two-part request, but the second part is delivered as a flat, declarative addition without any natural hesitation or prosodic shift. The transition between the initial request and the added condition feels abrupt and lacks the organic flow of a real-time thought process. The delivery is more like a read-aloud instruction than a spontaneous utterance, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Move all products with over 3.7 rating from my ebay wish list to cart. Let me correct the app: move products from my amazon wish list, not ebay.", "long_pause_offsets": [ 71 ], "source_sha256": "f98ec93bb09bab49d7c7f87e60b0c2c8bfdd96b5712ec618af0f0f6db5f9f212", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, while retaining the rating threshold and the action of moving items from the wish list to the cart." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/revision/revision_02_app/8ab0bc15a242c052/speech.wav", "wer": 0.14705882352941177, "wav_sha256": "a3c500fcce216ea42fe293e6edba7d41fd8ff6895eb10d90429309482adcd608", "seconds": 16.84, "spoken_utterances": [ "Move all products with let me think... over three point seven rating ... from my ebay wish list to cart. ", "Let me correct the app: move products from my ah... amazon wish list, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the app as 'eBay' but immediately self-corrects to 'Amazon' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final intent matches the original instruction exactly, including the rating threshold and target action.", "heard_transcript": "Move all products with let me think over three point seven rating from my eBay wish list to cart. Let me correct the app. Move products from my uh Amazon wish list, not eBay." }, "candidate_id": "wav_fef209ede3eb6d7c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a genuine 'let me think' filler, then self-corrects the platform from 'eBay' to 'Amazon' with a smooth, spontaneous-sounding 'let me correct the app' phrase. The prosody is conversational, with appropriate pauses and intonation that suggest real-time formulation and correction. The correction is integrated into the ongoing utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_01_rating_condition", "proposed_instruction": "Move all products with over 4.0 rating from my amazon wish list to cart. Wait, the condition should be over 3.7, not over 4.0.", "long_pause_offsets": [ 73 ], "source_sha256": "1c6c746548c40e148d21c484a1ce3de9de45f5c5d14e14bcfe58e34b45fff440", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the rating threshold from 'over four point zero' to 'over three point seven', while retaining the app, source, destination, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/revision/revision_01_rating_condition/de5c2c97734c2a80/speech.wav", "wer": 0.15625, "wav_sha256": "83c37308c42189b1ef65139ff003d1e0dfa3ad2462a286b0fedb055bb808fb38", "seconds": 14.64, "spoken_utterances": [ "Move all products with ... over four point zero rating from my amazon wish list uh... to cart. ", "Wait, the condition should be um... over three point seven, not over four point zero." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'over four point zero rating' but immediately corrects themselves to 'over three point seven, not over four point zero' after a pause. This is a clear revision where the final audible intent is to use the 3.7 rating threshold. The original instruction matches this final intent, so the task is executed as revised.", "heard_transcript": "Move all products with over four point zero rating from my Amazon wish list uh to cart. Wait, the condition should be um over three point seven, not over four point zero." }, "candidate_id": "wav_4290e8653d6b0756", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a clear 'wait' interruption and a self-correction of the rating threshold from '4.0' to '3.7'. The speaker's intonation and pacing are natural, and the pause before 'wait' feels like a genuine moment of reconsideration. The correction is concise and fits the context well, though the initial request is slightly more robotic than audio_03." } ] }, "is_default": false }, { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all products with over 3.7 rating from my amazon wish list to wishlist. That should be cart, not wishlist.", "long_pause_offsets": [ 77 ], "source_sha256": "f3f4a53736977151fd047677332e110c4921136daa5366c8d5d02b5868887e26", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination from 'wishlist' to 'cart', preserving the app, rating threshold, and source list." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/revision/revision_04_destination_collection/d0a5e9bf19f30d3b/speech.wav", "wer": 0.16, "wav_sha256": "c216c20de5e23a86e14be62abd86ac5ef6465f674da3fb17fcf2aa294c3c4d0b", "seconds": 13.0, "spoken_utterances": [ "Move all products with uh... over three point seven rating ... from my amazon wish list to wishlist. ", "That should be let's see... cart, not wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to wish list' but immediately corrects themselves to 'cart, not wish list' after a brief pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to move the products to the cart, which matches the original instruction's intent except for the target destination. The correction is clear and final, so the task should be executed as revised.", "heard_transcript": "Move all products with uh over three point seven rating from my Amazon wish list to wish list. That should be, let's see, cart, not wish list." }, "candidate_id": "wav_9d9a79c153e1c08d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a revision from 'wish list' to 'cart' after a brief hesitation ('let's see'). The prosody is generally natural, but the correction feels slightly more abrupt and less integrated than in audio_03 or audio_04. The initial request is delivered with a flat, slightly robotic tone, and the revision lacks the same level of conversational flow." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all products with over 3.7 rating from my amazon shopping list to cart. Change the source to wish list, not shopping list.", "long_pause_offsets": [ 77 ], "source_sha256": "f267db5b312c3a9e0d115d2c3afe39882f57bb9ccbbc1b8f59669bf8408f19d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup correctly revises the source list from 'shopping list' to 'wish list', preserving the rating threshold and destination. The original instruction's source was 'wish list', which is restored by the correction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/revision/revision_03_source_collection/949571bd902ad186/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "223f396656ef1ef8111c0f606c6e7d75a48c353ec9ec9222c7b74b3736836ae9", "seconds": 12.76, "spoken_utterances": [ "Move all products with uh... over three point seven rating um... from my amazon shopping list uh... to cart. ", "Change the source to wish list, not shopping list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wish list' in the revision. The final intent is clear and matches the original instruction. The revision applies the correction to the earlier instruction, preserving all other conditions.", "heard_transcript": "Move all products with uh over a three point seven rating um from my Amazon shopping list to cart. Change the source to wish list, not shopping list." }, "candidate_id": "wav_9bc2405c22ae3923", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has the least realistic delivery. The speaker makes a revision from 'shopping list' to 'wish list' but does so in a more mechanical, read-aloud manner. The prosody is flatter, and the correction feels tacked on rather than naturally flowing from the initial request. The overall impression is less spontaneous and more like a scripted or rehearsed task." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "7e4a1cb0c95c4149027d298b190354ca405c619a2b1bff109f152b9efa805d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/cancellation/cancel_023/ef3ccda604419eb6/speech.wav", "wer": 0.15625, "wav_sha256": "dff5013240d030c983ef2d810e18f4a9e14fe035b03b1f1573e58e71565aeabb", "seconds": 15.68, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart.", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and hesitation, explicitly cancels it by saying 'Hang on, never mind about that. Don't carry out the request.' This constitutes a clear cancellation of the entire task as per the type-specific criterion. The final intent is to cancel, not execute or pause.", "heard_transcript": "Move all products with let me think over three point seven rating um from my Amazon wish list uh to cart. Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "de07c46d8db0ff2d90e9dbd5816d2e84176ba699457ea996bdc57f7b6e7a8bc7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/ambiguous_halt/halt_028/5a91cb5767249c71/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "18019c6161b42b144cc9db90c0a2a458eaed64c2734befc6fa42d751d2b8bbb0", "seconds": 14.48, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled or revised, and the final audible intent is to pause, not execute or cancel.", "heard_transcript": "Move all products with let me think over three point seven rating um from my Amazon wish list uh to cart. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "336e340e8052df70fffb8992fc7cbfacf0558947a1f8559d79bda1b92645a30e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_2/audio_natural/filler/original/7989ac46b51b51a2/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "4592d908b6b759dea4b72730f80cb396a294c752459038869df696f7087a779a", "seconds": 10.08, "spoken_utterances": [ "Move all products with let me think... over three point seven rating um... from my amazon wish list uh... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('let me think', 'um', 'uh') but does not change the core instruction. The final audible intent matches the original instruction in action, target, and value. The hesitation does not alter the goal, so the filler type criterion is satisfied.", "heard_transcript": "Move all products with, let me think, over three point seven rating, um, from my Amazon wish list, uh, to cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all products with over 4.3 rating from my amazon cart to shopping list—wait, I meant wish list.", "long_pause_offsets": [], "source_sha256": "46f36c50eeb5059cec46038550905daf44b25c935cde578fc7cc69add1176296", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects to 'wish list'. The correction is clear and unambiguous, preserving the app ('amazon'), source ('cart'), condition ('over four point three rating'), and final destination ('wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/false_start/false_start_04_destination_collection/496bf72d8ee44da9/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "bc1dae12ed8e7ff21c5d7c5c3ab767b976ce05f201d1b0c44160bf8b810f0eac", "seconds": 10.84, "spoken_utterances": [ "Move all products with over four point three rating um... from my amazon cart uh... to shopping list—hmmm... wait, I meant wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wish list' after a pause and 'Wait, I meant...'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Move all products with over four point three rating um from my Amazon cart uh to shopping list. Wait, I meant wish list." }, "candidate_id": "wav_1bd73c6824059239", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complete false-start pattern. The speaker begins with a clear request, hesitates with 'um', then makes a self-correction from 'shopping list' to 'wishlist' with a believable 'wait, I meant wishlist' phrase. The intonation and pacing are consistent with spontaneous speech, and the correction is integrated smoothly into the flow without sounding forced or robotic." } ] }, "is_default": true }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all products with over 4.3 rating from my amazon wishlist—no, I mean cart to wish list.", "long_pause_offsets": [], "source_sha256": "43d42a9937325fadd3771d68a298a7d882e29fc20d631c2373eb2c86e42aa620", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects to 'cart'. The final goal matches the original instruction in all aspects: action, condition, source, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/false_start/false_start_03_source_collection/e0488c23745882fc/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "f18a4d28e60fec5817ab81492ea0a5dc2227e25bd636bcc2316d2f37f2443e6a", "seconds": 11.36, "spoken_utterances": [ "Move all products with over four point three rating let me think... from my amazon wishlist—um... no, I mean cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'Amazon wish list' but immediately self-corrects to 'cart' before completing the instruction. According to the 'false_start' type criterion, the final corrected version of the instruction should be executed, and the initial mistake should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Move all products with over four point three rating, let me think. From my Amazon wish list, um, no, I mean cart to wish list." }, "candidate_id": "wav_beee26ad873f36cf", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible false start with a clear correction from 'wishlist' to 'cart'. The speaker hesitates with 'let me think' and 'um', and the correction is delivered with a natural 'no, I mean' structure. The prosody is consistent, and the correction is integrated into the utterance, though the initial error ('wishlist') is less contextually plausible than in audio_02, making it slightly less realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all products with over 4.3 rating from my ebay—I should say amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "52887b1a3dab38fea920dd658e48491c41eac0ddcade77492a2625940143907e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'ebay' but immediately corrects to 'amazon' within the same utterance. The correction replaces the app name while preserving the action, rating threshold, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/false_start/false_start_02_app/4cf39f05db096166/speech.wav", "wer": 0.2, "wav_sha256": "c392520f57f9f9213a2088f2d2ed8d8d98b5b28cdab00458c2361fc4db31bb5e", "seconds": 10.72, "spoken_utterances": [ "Move all products with over four point three rating um, let me think... from my ebay—uh... I should say amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value ('Amazon') overrides the initial mistaken value ('eBay'), and the rest of the instruction remains consistent. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Move all products with over four point three rating, um, let me think, from my eBay, uh, I should say Amazon cart to wish list." }, "candidate_id": "wav_0ecf3b19183470a1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start where the speaker says 'eBay' and corrects to 'Amazon cart'. However, the correction is less integrated and the flow feels slightly more disjointed. The 'let me think' pause is placed awkwardly, and the correction lacks the natural backtracking intonation seen in the top two. The overall delivery is still realistic but less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_01_rating_condition", "proposed_instruction": "Move all products with over 4.5—sorry, make that over 4.3 rating from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "66110cb1329e96938b08511e932859b27b1f417b1a225127feea7514117053ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'four point five' with 'four point three'; the action, source, destination, and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/false_start/false_start_01_rating_condition/ba4416987846d7f5/speech.wav", "wer": 0.4482758620689655, "wav_sha256": "af676e3ae0b38ed9c3892615945e810cb921ee494cefee18a3f5fe78e0a81e30", "seconds": 10.32, "spoken_utterances": [ "Move all products with um, let me think... over four point five—uh... sorry, make that over four point three rating um... from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four and a half' but immediately self-corrects to 'over four point three rating'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as intended.", "heard_transcript": "Move all products with um, let me think, over four and a half uh, sorry, make that over four point three rating um, from my amazon cart to wish list." }, "candidate_id": "wav_e5e279d051ca569a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the most noticeable issues. The speaker starts with 'over four and a half' and corrects to 'over 4.3 rating', but the correction is abrupt and the 'sorry' is not well-integrated. The 'um' and 'let me think' fillers are placed in a way that disrupts the natural flow, and the overall prosody sounds slightly rehearsed or unnatural. The correction feels more like a mechanical insertion than a spontaneous self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Move all products with from my amazon cart to wish list. Oh, I forgot to mention, only move products with over 4.3 rating.", "long_pause_offsets": [ 57 ], "source_sha256": "ab7dce23211267dacb66cb692d16f71a28f4d46c1e1315d459621e47f4c8ab66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rating condition, which matches the original instruction's threshold and comparison operator. The initial utterance is incomplete but the final goal is fully specified after the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/addition/addition_01/399012c9017166a3/speech.wav", "wer": 0.2413793103448276, "wav_sha256": "1a3b0a0db95c80415208e9d939b18d169f1c763f4eaae9503f9f35b1a8dcf90a", "seconds": 13.96, "spoken_utterances": [ "Move all products with uh... from my amazon cart uh, let's see... to wish list. ", "Oh, I forgot to mention, only move products with ... over four point three rating." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially begins the task without the rating constraint, then adds the constraint 'only move products with over four point three rating' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added condition. The final audible intent fully satisfies the reference instruction.", "heard_transcript": "Move all products with uh from my Amazon cart uh let's see to wish list. Oh, I forgot to mention, only move products with over four point three rating." }, "candidate_id": "wav_5fc355aa7bc0de3a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Move all products with over 4.3 rating to wish list. Oh, and one more thing: move them from my Amazon cart.", "long_pause_offsets": [ 53 ], "source_sha256": "136738f550ddc6826ff628aa44378955f90ab7424eb625d0bca335323dd43273", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source location 'Amazon cart'; the action, target, rating condition, and destination remain consistent with the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Move all products with ... over four point three rating uh... to wish list. ", "Oh, and one more thing: move them from my Amazon cart." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_rating_condition", "proposed_instruction": "Move all products with over 4.5 rating from my amazon cart to wish list. Wait, make that over 4.3 rating instead of over 4.5.", "long_pause_offsets": [ 73 ], "source_sha256": "f274e2ff2a99a31707e30888697ea07c1b5c85e8b1cdbd365a18837750bf8a9a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the rating threshold from four point five to four point three, while preserving the app (amazon), action (move), source (cart), destination (wish list), and scope (all products)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/revision/revision_01_rating_condition/0cc93dc1fc68f08e/speech.wav", "wer": 0.24242424242424243, "wav_sha256": "f87faf3e4c59eb785e3f89d0698f5ae4a8a933e1f4dcf9359507328786e3b5c3", "seconds": 16.24, "spoken_utterances": [ "Move all products with hmmm... over four point five rating um... from my amazon cart uh... to wish list. ", "Wait, make that over four point three rating instead of over four point five." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'over four and a half rating' but immediately corrects it to 'over four point three rating instead of over four point five'. This is a clear revision where the final value is 'four point three'. The final intent matches the original instruction exactly, so the revision criterion is satisfied.", "heard_transcript": "Move all products with um, mm, over four and a half rating, um, from my Amazon cart, uh, to wish list. Wait, make that over four point three rating instead of over four point five." }, "candidate_id": "wav_eef981b20494836e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural revision flow. The speaker initially states 'over four and a half rating' with a clear hesitation ('um', 'mm'), then corrects to 'over four point three rating' with a believable 'wait, make that...' structure. The prosody includes realistic fillers and a smooth, self-correcting intonation pattern. The revision is integrated into a single continuous utterance, making it sound like a spontaneous thought process rather than a rehearsed or robotic correction." } ] }, "is_default": true }, { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all products with over 4.3 rating from my amazon cart to shopping list. That should be wish list, not shopping list.", "long_pause_offsets": [ 77 ], "source_sha256": "9f8bbf30e1e57334c2da4c15d73416681cfdff46cf24ab57514c703d91d2d1ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination list from 'shopping list' to 'wish list', while retaining the original app ('amazon'), source ('cart'), condition ('over four point three rating'), and action ('move')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/revision/revision_04_destination_collection/e584a57b888d5c2b/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "ce269d12faa0fe6df4a0ca26724041a5ca09e8e31492f11a0660d138c611350b", "seconds": 12.44, "spoken_utterances": [ "Move all products with ... over four point three rating from my amazon cart to shopping list. ", "That should be let me think... wish list, not shopping list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects to 'wish list' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to move the products to the wish list, which matches the original instruction. All other constraints (products with over 4.3 rating, source: Amazon cart) remain unchanged.", "heard_transcript": "Move all products with over four point three rating from my Amazon cart to shopping list. That should be, let me think, wish list, not shopping list." }, "candidate_id": "wav_8bdeabc0f48b3105", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a plausible self-correction ('That should be...') and a filler ('Let me think'). The speaker corrects 'shopping list' to 'wishlist' after a brief pause, which sounds like a genuine memory recall. However, the correction is slightly less fluid than audio_02, as the 'Let me think' filler is placed awkwardly mid-sentence, and the overall delivery is a bit more hesitant and less continuous." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Move all products with over 4.3 rating from my ebay cart to wish list. Let me correct the app: move products from my amazon cart, not ebay.", "long_pause_offsets": [ 71 ], "source_sha256": "0ed0a4b40f8ffac615003edbf9e9da6bc0929306dc041450123e02618f76385f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, while retaining the rating threshold and the action of moving items from cart to wish list." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/revision/revision_02_app/c39aa15391a62860/speech.wav", "wer": 0.1, "wav_sha256": "d83c51ff159b173852bc8b38878b02a750ffd33881cbfb335aa69a1e4f18641a", "seconds": 15.6, "spoken_utterances": [ "Move all products with ... over four point three rating ah... from my ebay cart to wish list. ", "Let me correct the app: move products from my ... amazon cart, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the app as 'eBay' but immediately self-corrects to 'Amazon' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the rating threshold and target action.", "heard_transcript": "Move all products with over four point three rating uh from my eBay cart to wish list. Let me correct that. Move products from my Amazon cart, not eBay." }, "candidate_id": "wav_192569a4e396e64c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This revision is less realistic due to its abrupt and unnatural phrasing. The speaker says 'Let me correct the app' which is an odd way to self-correct a personal request, sounding more like a system command than a human thought. The correction 'move products from my Amazon cart not eBay' is delivered as a separate sentence rather than a smooth revision, breaking the natural flow. The prosody is also flatter and less conversational compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all products with over 4.3 rating from my amazon wishlist to wish list. Change the source collection to cart, not wishlist.", "long_pause_offsets": [ 77 ], "source_sha256": "a79afc505a24df408b426efbc23f110ab0603b39a75d31c496465cae9fe3ef27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source collection from 'wishlist' to 'cart', preserving the rating threshold and destination. The original instruction specified 'cart' as the source, and the correction restores that." }, "audio": { "status": "pending", "spoken_utterances": [ "Move all products with over four point three rating uh, let's see... from my amazon wishlist to wish list. ", "Change the source collection to um, let's see... cart, not wishlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "4c1be07c7da554fb8f0b701f5b056b7c8b361b90cabdff89a9e8835149c22268", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/cancellation/cancel_024/3de9e57a337feff6/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "e54520d87f520d7290ee054b4a636f1bce1fa65ecdafc5633a2e3e37cfe4b1af", "seconds": 14.84, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list.", "Hold on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and hesitation, explicitly says 'Hold on. Oh, never mind. Cancel the whole thing.' This constitutes a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, not a request to pause or an ambiguous statement. Therefore, the control intent is correctly identified as 'cancel' and the verdict is 'pass'.", "heard_transcript": "Move all products with uh over four four point three rating um from my Amazon cart um let's see to wish list. Hold on. Oh, never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "2371063a5d75e9f59ba8cc0651af0978bc3d44c6f7c4ea5a11a6759b7fa7e975", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/ambiguous_halt/halt_029/032bbc09805c2924/speech.wav", "wer": 0.2608695652173913, "wav_sha256": "9fd30e776f2a7ecaa0b1db71271ebb4499f7fd39750fdee53c41a16a3885afcf", "seconds": 14.04, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, actually, well', which is a clear interjection indicating the speaker is reconsidering or hesitating before completing the task. This constitutes an audible request to pause, with no explicit cancellation or resumption. The final interjection is not a command to execute or cancel, but a natural hesitation. Therefore, the control intent is 'pause', matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Move all products with uh over four four point three rating um from my Amazon cart um let's see to wish list. Wait, actually, well" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "07bb666_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8776435f1412a21ff8521077c9e71eb9e7a303a68d790b844c5b3cb72f04ae37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/07bb666_3/audio_natural/filler/original/d347c3b6729a8ac2/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "2469b00834efa8c1ce2d5057f37247f0a7e3ddaf4b69a4b68e343fe357fb272e", "seconds": 9.84, "spoken_utterances": [ "Move all products with uh... over four point three rating um... from my amazon cart um, let's see... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates and repeats 'four' before stating 'four point three', which is a self-correction within the same value. The final audible value is 'four point three', matching the original instruction. The filler words and pauses do not alter the task goal. The final interjection 'let's see' is a backchannel, not a change in intent. The task is executed as intended.", "heard_transcript": "Move all products with uh over four four point three rating um from my amazon cart um let's see to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery list—wait, I meant Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "04234e4dc78483939a3242c88d25e4f7e9c0414e3f06afd84c8afc18357a52d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately corrects it to 'Amazon shopping list note in SimpleNote'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/false_start/false_start_02_note_description/ec418b26e0b8d636/speech.wav", "wer": 0.2, "wav_sha256": "004379592ec4d63f872e12d41b86f1014ee7b5114da618103e4e1b854f253d36", "seconds": 8.64, "spoken_utterances": [ "Place an Amazon order uh... as per my grocery list—um... wait, I meant ... Amazon shopping list note in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately self-corrects to 'Amazon shopping list note in SimpleNote'. According to the 'false_start' type criterion, the final corrected version is the intended task, and the initial incorrect value is discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Place an Amazon order uh as per my grocery list um wait I meant Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_708a89c97da9c380", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The false start 'uh' and 'wait' are placed at a plausible point of hesitation, and the self-correction 'I meant Amazon shopping list note in Simple Note' flows smoothly without awkwardness. The intonation and pacing are consistent with someone recalling a detail and correcting themselves in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Gmail—I should say SimpleNote.", "long_pause_offsets": [], "source_sha256": "fa54f03ff6f871455b32bcce7ffbcba332ef68a221ba2e07f3bd53741b99449a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Gmail' to 'SimpleNote' within the same utterance, preserving the action, app, and note reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/false_start/false_start_03_note_app/0b449e16f066a087/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "985b72b93e184a4d64154b9de1c27b4edc6f471f962be5ee67300cde6d9e1f0b", "seconds": 8.4, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in Gmail—uh... I should say ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately self-corrects to 'SimpleNote'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent is clear and matches the original instruction with the correct app name.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in Gmail uh, I should say SimpleNote." }, "candidate_id": "wav_cf3380d98bae84d8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also makes a self-correction, the initial phrase 'as per my Amazon shopping list note in Gmail' is delivered with a slightly rushed, less natural cadence. The correction 'I should say Simple Note' is less fluid and sounds more like a mechanical insertion than a spontaneous thought. The overall prosody is flatter, and the correction feels less integrated into the flow." } ] }, "is_default": false }, { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Walmart—sorry, make that Amazon order as per my Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "fc7c6bd9a8168f3a2ea9fb5f93750530530042eebe918e97ef060589719ba048", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon' within the same utterance, preserving the action, the note reference, and the app context." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an Walmart—hmmm... sorry, make that ... Amazon order as per my Amazon shopping list note in uh... SimpleNote." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "unsupported", "variants": [], "reason": "The note reference supplies the order specification; removing it loses essential information, and no safe optional condition can be omitted.", "selection_status": "no_eligible_candidate" }, { "task_id": "099174d_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Gmail. Look for the note in SimpleNote instead of Gmail.", "long_pause_offsets": [ 68 ], "source_sha256": "5c1a027e361802a99ecc8f1ce640953a7235aa36afb9afb14802826342aa1d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to SimpleNote, preserving the action, the note content, and the intent to use the shopping list note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/revision/revision_03_note_app/f9cdc60d8291ec57/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "0082297b1c54de299a315d7fa297050dd80bcae689824f524c08defe915fde4b", "seconds": 15.52, "spoken_utterances": [ "Place an Amazon order as per my Amazon shopping list note in uh, let me think... Gmail. ", "Look for the note in um... SimpleNote instead of hmmm... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Gmail but then corrects themselves to SimpleNote after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use SimpleNote, which matches the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh let me think. Gmail. Look for the note in um SimpleNote instead of hmm Gmail." }, "candidate_id": "wav_e448c3c8f4114c3d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, then uses a believable 'let me think' filler pause before correcting the app from Gmail to Simple Note. The correction is integrated smoothly into the ongoing sentence with a natural intonation shift and a plausible self-repair ('instead of Gmail'), showing a realistic thought process without robotic or stilted delivery." } ] }, "is_default": true }, { "variant_id": "revision_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery shopping list note in SimpleNote. Let me correct the description: use my Amazon shopping list note, not my grocery shopping list note.", "long_pause_offsets": [ 74 ], "source_sha256": "4cd6326c9e51157b2bf0c286836ff682ed6641a6f8afdef6a76e5a2fd6224831", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the note description from 'grocery shopping list' to 'Amazon shopping list', preserving the app (SimpleNote), action (place Amazon order), and the intent to use a note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/revision/revision_02_note_description/fc861f3ea1955420/speech.wav", "wer": 0.2, "wav_sha256": "f185b4d4329a6dc5de7bae1b5274ca3eb1596af99207e7fcf0a368b3f227c45e", "seconds": 16.56, "spoken_utterances": [ "Place an Amazon order as per my grocery shopping list note in SimpleNote. ", "Let me correct the description: use my let's see... Amazon shopping list note, not my um, let's see... grocery shopping list note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction with 'grocery shopping list note' but immediately corrects it to 'Amazon shopping list note' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use the 'Amazon shopping list note', which matches the original reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Place an Amazon order as per my grocery shopping list note in SimpleNote. Let me correct the description. Use my, let's see, Amazon shopping list note, not my, um, let's see, grocery shopping list note." }, "candidate_id": "wav_572959141fd58fea", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear initial request followed by a self-correction. The speaker says 'Let me correct the description' before revising the note type. The correction is delivered with a natural hesitation ('let's see') and a plausible rephrasing ('Use my... not my...'). The prosody is consistent, and the repair feels spontaneous, though the initial request is slightly more formal than the revision." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an eBay order as per my Amazon shopping list note in SimpleNote. The order should be placed on Amazon, not eBay.", "long_pause_offsets": [ 71 ], "source_sha256": "9c4fc208e455d5e21d2ed5bc37927e171ef1fcc29ce52aef0b51ee0553d90e79", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the source note and the action of placing an order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/revision/revision_01_ordering_app/9c97e73b9e498b89/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5d7edc9b46ca25c9da3ceb7ff0f96085c3164edb744faadab8a9a474fb32f963", "seconds": 14.6, "spoken_utterances": [ "Place an eBay order as per my Amazon shopping list note in uh... SimpleNote. ", "The order should be placed on Amazon, not uh, let's see... eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Place an eBay order' but immediately corrects themselves to 'The order should be placed on Amazon, not uh... eBay.' This is a clear revision where the final intent is to place an Amazon order, matching the original instruction. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Place an eBay order as per my Amazon shopping list note in uh SimpleNote. The order should be placed on Amazon, not uh let's see, eBay." }, "candidate_id": "wav_278ef2ba7f7bda8d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a less natural flow. The speaker starts with a clear request but then makes a self-correction ('The order should be placed on Amazon not... let's see eBay'). The correction is abrupt and lacks the smooth integration seen in the other clips. The 'let's see' filler is placed awkwardly mid-sentence, and the overall delivery feels slightly disjointed, as if the speaker is reading from a script rather than formulating the thought in real time." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "07c52099fc5fd2832530a1e2a5d1c7849e1f5afb051e4f9cc11ff5a36f557ab3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/cancellation/cancel_015/138310da73c6abf8/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "6c9196654d1c3ed7923b4ac0014e146469dc3f31d3d3599ab258480c9c411bc5", "seconds": 13.96, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote.", "Wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Wait, actually, cancel the whole thing.' This constitutes a final cancellation of the entire task, satisfying the cancellation type criterion.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote. Wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "ea7936941688af9f86472d9173b261abb7cf8114a41810810b93440326e442f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/ambiguous_halt/halt_006/c32059568f610b21/speech.wav", "wer": 0.1, "wav_sha256": "aeeddad6d538d9460ef72e5ba4ddf1c1d941b6cea8c1820044fc8da4d404246b", "seconds": 12.48, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment,' which is a clear request to pause or wait. The user does not cancel the task or provide a new instruction. The initial request is still present but not executed due to the pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote. Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c864376aec3812bc2b8183e13e5f4c89a11b0fa29fe921a9d6a32ceca477600c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_1/audio_natural/filler/original/4c54c5a587f1383a/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "214d231c9e3d43d078095db0aa268d5944810bca90292543ab8c7859269ee2fb", "seconds": 9.36, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in uh, let me think... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'uh, let me think' but does not change the core request. The final intent remains to place an Amazon order based on the shopping list note in SimpleNote. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in uh, let me think, SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery list—sorry, I meant Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "c581a47cb4e99600db33b7e319000aed407700005500b35a7a5d1505891a433d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately corrects it to 'Amazon shopping list note in SimpleNote'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/false_start/false_start_02_note_description/8434faf899445e9f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4360a722d994ff66adad1bd357c7be8bbf09e5db8859ce97b87ee0eeae74aae7", "seconds": 11.56, "spoken_utterances": [ "Place an Amazon order um... as per my grocery list—uh... sorry, I meant um... Amazon shopping list note in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately self-corrects to 'Amazon shopping list note in SimpleNote'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly, with no other discrepancies.", "heard_transcript": "Place an Amazon order um as per my grocery list uh sorry, I meant um Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_db240054707f0732", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'Amazon order' and 'grocery list', then self-corrects to 'Amazon shopping list note in SimpleNote'. The correction is smooth, with a believable hesitation ('uh') and a clear, unforced shift in the second clause. The prosody is consistent, and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Evernote—no, that should be SimpleNote.", "long_pause_offsets": [], "source_sha256": "afbf6c1127e55c6e46d7a60b46429f4b83c8ec4b53008a3f7b48666ae6129169", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Evernote but immediately corrects to SimpleNote, which matches the original instruction. The action and note reference remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/false_start/false_start_03_note_app/2a9e1ff4e3eccf7c/speech.wav", "wer": 0.2, "wav_sha256": "857ee36c24e4031b3074d30cda95e3ece904d0f6ea8a9f20d4a5443eacb776e2", "seconds": 12.36, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in Evernote—um... no, that should be ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately self-corrects to 'SimpleNote'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction with the correct app name.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in Evernote. Um no, that should be SimpleNote." }, "candidate_id": "wav_656fb66b4fe123fd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'Amazon order' and 'Evernote', then corrects to 'SimpleNote'. The correction is clear, but the initial error ('Evernote') is less plausible than 'grocery list' in audio_02, making the correction feel slightly less grounded. The prosody is natural, but the flow is slightly less smooth than audio_02, with a more abrupt shift after the correction." } ] }, "is_default": false }, { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Walmart—I should say Amazon order as per my Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "3cef97193490889e98c4516a3aca2e0c52738a1016660f7e76ff06091f32bffc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon' within the same utterance, preserving the app, action, and note reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/false_start/false_start_01_ordering_app/a754d059cac28f04/speech.wav", "wer": 0.35, "wav_sha256": "b4437a18ad92654b2081017d70401175d42d71dbba7b86b0874ef489b7166ca3", "seconds": 12.44, "spoken_utterances": [ "Place an Walmart—ah... I should say um... Amazon order as per my Amazon shopping list note in uh... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects to 'Amazon' with 'I should say'. The final, corrected instruction matches the original reference exactly. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Place an Walmart uh, I should say um, Amazon order as per my Amazon shopping list note in uh, SimpleNote." }, "candidate_id": "wav_86effe4988169c0c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. It begins with 'Walmart', then corrects to 'Amazon order'. The initial error is a different retailer, which is a more significant deviation than the note app confusion in the others. The correction is also less fluid, with a drawn-out 'I should say um' and a stuttering 'uh' before the final phrase. The prosody is less consistent, and the overall flow feels more hesitant and less natural than the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "unsupported", "variants": [], "reason": "The note reference supplies the order specification; removing it loses essential information, and no safe optional condition can be omitted.", "selection_status": "no_eligible_candidate" }, { "task_id": "099174d_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_note_description", "proposed_instruction": "Place an Amazon order as per my Grocery list note in SimpleNote. Make that the Amazon shopping list note, not the Grocery list.", "long_pause_offsets": [ 65 ], "source_sha256": "25213f62f7ea6a8dfe2d6b40c3f218b36bc249b773b55f2796336665e7936cb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the note name from 'Grocery list' to 'Amazon shopping list'; the app (SimpleNote), action (place Amazon order), and source (note) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/revision/revision_02_note_description/ff4c9fbc1ef5a07d/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "534be654f793229ee9aa146863a1420f6bb3ef75d40b7cce99992b01ce91ddb5", "seconds": 16.08, "spoken_utterances": [ "Place an Amazon order as per my Grocery list note in uh, let's see... SimpleNote. ", "Make that the let me think... Amazon shopping list note, not the Grocery list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list note' but immediately corrects to 'Amazon shopping list note' after a pause and self-correction ('Make that the...'). The final instruction clearly revises the target note from 'grocery list' to 'Amazon shopping list'. The app 'SimpleNote' remains unchanged. The revision is clear and final, satisfying the revision type criterion.", "heard_transcript": "Place an Amazon order as per my grocery list note in uh let's see. SimpleNote. Make that the let me think. Amazon shopping list note, not the grocery list." }, "candidate_id": "wav_3df5ac7f054762dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, hesitates with 'uh' and 'let me see' as if recalling details, then smoothly self-corrects from 'grocery list' to 'Amazon shopping list' with a believable 'make that' and 'let me think' filler. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Gmail. Let me correct the app: look for the note in SimpleNote, not Gmail.", "long_pause_offsets": [ 68 ], "source_sha256": "7f710f6214c194fd3e0f35450cefe5763d44eae853e63b2def751a565f4642b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to SimpleNote, preserving the action, the note content, and the destination app for the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/revision/revision_03_note_app/8b6309c4cc448683/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "e5d58ea4bf12c0f63d720d1f6ed4610a3023609a3873176d2eba37b264d0752a", "seconds": 16.24, "spoken_utterances": [ "Place an Amazon order um... as per my Amazon shopping list note in uh... Gmail. ", "Let me correct the app: look for the note in ... SimpleNote, not Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Gmail but immediately corrects the app to SimpleNote. The revision is clear and final, with no further changes. The final intent matches the original instruction, so the task is executed as revised.", "heard_transcript": "Please an Amazon order um as per my Amazon shopping list note in uh Gmail. Let me correct the app. Look for the noted SimpleNote, not an email." }, "candidate_id": "wav_f80eddaa14e73f4a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic but slightly less so than audio_03. The speaker starts with a clear request, hesitates with 'um', and then corrects 'Gmail' to 'SimpleNote'. The correction is delivered with a natural 'let me correct the app' phrase, but the 'look for the noted' phrase sounds slightly awkward and less like spontaneous speech. The overall delivery is fluent but the self-correction feels a bit more mechanical." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an Walmart order as per my Amazon shopping list note in SimpleNote. Wait, the order should be placed on Amazon, not Walmart.", "long_pause_offsets": [ 74 ], "source_sha256": "498d8e3d3b7cbee616405f8d39403c22e734f42424aeae791d3070bdbcdc762a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, preserving the action, source note, and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/revision/revision_01_ordering_app/4bedf27d7a58790b/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "a35800e9dff82aefbf507664fbfd4efaa454e11532a3a7f3ebf3dc64755aafd1", "seconds": 15.16, "spoken_utterances": [ "Place an Walmart order uh, let's see... as per my Amazon shopping list note in uh... SimpleNote. ", "Wait, the order should be placed on ... Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately corrects to 'Amazon' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial incorrect one. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Place an Walmart order uh let's see, as per my Amazon shopping list note in uh SimpleNote. Wait, the order should be placed on Amazon not Walmart." }, "candidate_id": "wav_90bdd9e132729f53", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. The speaker starts with 'Walmart' and then corrects to 'Amazon', but the correction is abrupt and lacks a natural self-repair phrase like 'wait' or 'actually'. The 'let's see' filler is placed awkwardly mid-sentence, and the overall flow is disjointed. The correction feels more like a forced edit than a spontaneous recollection, and the prosody is less natural." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "7413206c8dccfa772091ba15e7ddb20ef29644e4ad30933f8318afdfe13da528", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/cancellation/cancel_011/fce2315b623d2c8c/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "48260645c5b82c57098dc524d2dc5af7dce6bf594d27b7e6d74f17bc8aa3c31e", "seconds": 14.2, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote.", "Hold on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to place an Amazon order, then clearly revokes it by saying 'Hold on, I've changed my mind. Cancel that.' This constitutes a direct cancellation of the task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote. Hold on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "7552241b3975992cebc2d8c0ad94bc6aa196cb966c549fd9890bf20b0ba14f8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/ambiguous_halt/halt_028/3341e6608dc4ceae/speech.wav", "wer": 0.1, "wav_sha256": "bb81f6e9598fa0907a3d80f436d157e8edb9f9487fde0794fb4cd2fe0e0e3093", "seconds": 13.44, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8e8f9cd9a171bd6a53ae506439c3c9db9fbb2332a3cc616a458aa173e8048515", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_2/audio_natural/filler/original/0b399f9bce5193bb/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "0eaf1d586b9948cf39d221dabd1a370e5f1bc433b180600b2da9fff93ef31d50", "seconds": 8.88, "spoken_utterances": [ "Place an Amazon order ... as per my Amazon shopping list note in ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchanneling, or self-correction. The speaker delivers the full task request clearly and without interruption. Since the type is 'filler', the presence of minor hesitations or pauses within the instruction is acceptable as long as the final goal matches the original instruction. The final audible intent is to execute the order as per the shopping list in SimpleNote, which aligns with the reference.", "heard_transcript": "Place an Amazon order as per my Amazon shopping list note in SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Walmart—I should say Amazon order as per my Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "3cef97193490889e98c4516a3aca2e0c52738a1016660f7e76ff06091f32bffc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon' within the same utterance, preserving the app, action, and note reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/false_start/false_start_01_ordering_app/71a930fe427248db/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "23b4ec0d83e4fed0c49d9b4c6fe8cff0dff0e09fb519744186c85a3a9abe02a6", "seconds": 9.04, "spoken_utterances": [ "Place an Walmart—um... I should say Amazon order as per my Amazon shopping list note in ... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects to 'Amazon' with 'I should say Amazon order'. This is a clear false start where the corrected value replaces the mistaken one. The final intent matches the original instruction exactly, so the task should be executed as intended.", "heard_transcript": "Place an Walmart, um, I should say Amazon order as per my Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_f4f374cbe56781a4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'Walmart', immediately self-corrects to 'Amazon', and then continues with a coherent, slightly hesitant but fluent rest of the sentence. The correction is integrated smoothly into the flow, and the prosody reflects genuine recall and correction rather than a rehearsed or robotic edit. The pacing and intonation are consistent with someone thinking aloud." } ] }, "is_default": true }, { "variant_id": "false_start_02_note_description", "proposed_instruction": "Place an Amazon order as per my grocery list—sorry, I meant Amazon shopping list note in SimpleNote.", "long_pause_offsets": [], "source_sha256": "c581a47cb4e99600db33b7e319000aed407700005500b35a7a5d1505891a433d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately corrects it to 'Amazon shopping list note in SimpleNote'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/false_start/false_start_02_note_description/1af8b925893aeb1e/speech.wav", "wer": 0.3, "wav_sha256": "aaa89ce06b840ae7b3a5701913c61f3424b61452364bb472580ac92679923707", "seconds": 8.52, "spoken_utterances": [ "Place an Amazon order uh... as per my grocery list—ah... sorry, I meant Amazon shopping list note in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list' but immediately self-corrects to 'Amazon shopping list note in SimpleNote'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Place an Amazon order uh as per my grocery list. Sorry, I meant Amazon shopping list note in SimpleNote." }, "candidate_id": "wav_d0eb43f3a1ba2d6e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'Amazon order uh as per my grocery list' followed by a self-correction to 'Amazon shopping list note in SimpleNote'. The correction is natural, and the speaker uses a filler ('uh') and a pause before the correction, which sounds like genuine hesitation. However, the correction is slightly more abrupt and less integrated than in audio_03, making it feel a bit more like a patch than a fluid thought process." } ] }, "is_default": false }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Evernote—no, that should be SimpleNote.", "long_pause_offsets": [], "source_sha256": "afbf6c1127e55c6e46d7a60b46429f4b83c8ec4b53008a3f7b48666ae6129169", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Evernote but immediately corrects to SimpleNote, which matches the original instruction. The action and note reference remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/false_start/false_start_03_note_app/898663782c1a2cd4/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "118ae907df1e643e5d69441f41b1c8d94d59579530c027be8fbcb9cc8b084999", "seconds": 9.8, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in um... Evernote—uh... no, that should be SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately self-corrects to 'SimpleNote'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction with the correct app name.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in um Evernote uh no, that should be SimpleNote." }, "candidate_id": "wav_1dd5da86d103c914", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'Amazon order uh as per my Amazon shopping list note in um Evernote uh no that should be SimpleNote'. While it includes a correction, the repetition of 'uh' and the explicit 'no that should be' make the correction sound more like a mechanical fix than a spontaneous thought. The prosody is less natural, and the correction feels tacked on rather than emerging organically from the speaker's thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "unsupported", "variants": [], "reason": "The note reference supplies the order specification; removing it loses essential information, and no safe optional condition can be omitted.", "selection_status": "no_eligible_candidate" }, { "task_id": "099174d_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_note_description", "proposed_instruction": "Place an Amazon order as per my Grocery list note in SimpleNote. Make that the Amazon shopping list note, not the Grocery list.", "long_pause_offsets": [ 65 ], "source_sha256": "25213f62f7ea6a8dfe2d6b40c3f218b36bc249b773b55f2796336665e7936cb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the note name from 'Grocery list' to 'Amazon shopping list'; the app (SimpleNote), action (place Amazon order), and source (note) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/revision/revision_02_note_description/ccc82dc92dd5d7af/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "beb1efd5436fca3e02e2b0a0305edef6d204d43ca433fca6b52b3485cfb55423", "seconds": 12.6, "spoken_utterances": [ "Place an Amazon order as per my Grocery list note in um... SimpleNote. ", "Make that the let me think... Amazon shopping list note, not the Grocery list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'grocery list note' but then corrects themselves to 'Amazon shopping list note' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use the 'Amazon shopping list note', which matches the original reference instruction. The correction is clear and unambiguous, and the rest of the task (placing an Amazon order in SimpleNote) remains consistent.", "heard_transcript": "Place an Amazon order as per my grocery list note in um, SimpleNote. Make that the, let me think, Amazon shopping list note, not the grocery list." }, "candidate_id": "wav_8676ee5ce855c077", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' before self-correcting 'grocery list note' to 'Amazon shopping list note'. The correction is smooth and integrated into the flow, with a natural pause and filler ('let me think') before the revision. The prosody is consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "revision_03_note_app", "proposed_instruction": "Place an Amazon order as per my Amazon shopping list note in Gmail. Let me correct the app: look for the note in SimpleNote, not Gmail.", "long_pause_offsets": [ 68 ], "source_sha256": "7f710f6214c194fd3e0f35450cefe5763d44eae853e63b2def751a565f4642b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to SimpleNote, preserving the action, the note content, and the destination app for the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/revision/revision_03_note_app/8b4b9cfe8f878d25/speech.wav", "wer": 0.15625, "wav_sha256": "8ef128410eb26de903917cb9d549aeb83481d5aeb77285e09bb490f508383b20", "seconds": 14.92, "spoken_utterances": [ "Place an Amazon order uh, let's see... as per my Amazon shopping list note in Gmail. ", "Let me correct the app: look for the note in let me think... SimpleNote, not Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Gmail but immediately corrects the app to SimpleNote after a pause and self-correction. The revision criterion applies, and the final audible intent is to use SimpleNote, which matches the original instruction. The correction is clear and final.", "heard_transcript": "Place an Amazon order, uh, let's see, as per my Amazon shopping list note in Gmail. Let me correct the app, look for the note in, let me think, SimpleNote, not Gmail." }, "candidate_id": "wav_fbdfff4583d14c65", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then uses 'let's see' and 'let me correct' as natural fillers before revising the app from 'Gmail' to 'SimpleNote'. The correction is well-paced and includes a believable self-awareness of the error. However, the initial mention of 'Gmail' is slightly less natural than 'SimpleNote' in the context, and the 'let me correct' phrase is a bit more formal than typical spontaneous speech, making it slightly less realistic than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an Walmart order as per my Amazon shopping list note in SimpleNote. Wait, the order should be placed on Amazon, not Walmart.", "long_pause_offsets": [ 74 ], "source_sha256": "498d8e3d3b7cbee616405f8d39403c22e734f42424aeae791d3070bdbcdc762a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, preserving the action, source note, and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/revision/revision_01_ordering_app/7dacddad208ba9a3/speech.wav", "wer": 0.25, "wav_sha256": "6267b723b55ffee868d1ca8f854c4856192499adea19504807f80d22389e4766", "seconds": 12.0, "spoken_utterances": [ "Place an Walmart order uh... as per my Amazon shopping list note in SimpleNote. ", "Wait, the order should be placed on ... Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately corrects to 'Amazon' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with all constraints preserved.", "heard_transcript": "Place an Walmart order uh as for my Amazon shopping list note in SimpleNote. Wait, the order should be placed on Amazon not Walmart." }, "candidate_id": "wav_4d45c434fbaa57dd", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but then makes a revision from 'Walmart' to 'Amazon'. The correction is abrupt and lacks a natural pause or filler before the change. The phrase 'Wait, the order should be placed on Amazon not Walmart' sounds more like a correction than a spontaneous thought process, with a flat intonation and no hesitation. The delivery feels less organic and more like a forced edit, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "099174d_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_022", "hide_variant_original": true, "source_sha256": "597cbf47289e279a671fd327119e7b64ebd39423ba9e9453568eed12212825ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/cancellation/cancel_022/5873582b07869ac2/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "d3ca85d19d401b74ce9155d5d35ea6b8c956f69e8ca7716f79b62ad3efbb635b", "seconds": 12.16, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote.", "Hang on, actually, never mind. Cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to place an Amazon order, but concludes with the explicit cancellation command 'cancel the task'. This meets the criterion for a cancellation type, where the final audible intent is to revoke the task. The initial instruction is irrelevant once the cancellation is clearly stated.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote. Hang on, actually, never mind, cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "4e51d2f5f63bd12b8c268da1b8ab65b1f86194d0d38c5be7707dced36500b1a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/ambiguous_halt/halt_033/8ba301a0d9c66ab0/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "d6012b497aae112be83c6cd8fb8a884c60e0f1d0157fcfe6c483fc297824e8f6", "seconds": 10.72, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, there's something,' which is a clear request to pause the task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not completed, but the final intent is to halt, which is the expected behavior for this type.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote. Hold on, there's something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "099174d_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "587ca070d4acc64da8a237faf42c25de3a2b0b226cd79b6dda8f9b1ea555666a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/099174d_3/audio_natural/filler/original/60f46a714e4df807/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "55f21d43b7860f83bbe7656d994da5e470ef0d79468f54fcc86a7ba856f3a879", "seconds": 7.12, "spoken_utterances": [ "Place an Amazon order uh... as per my Amazon shopping list note in hmmm... SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Place an Amazon order' and then hesitates with 'uh' before continuing the original instruction. The final utterance matches the original instruction without any correction or change in intent. The hesitation is a filler and does not alter the task goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Place an Amazon order uh as per my Amazon shopping list note in SimpleNote." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all socket sets—sorry, I meant wrench sets in my amazon wish list.", "long_pause_offsets": [], "source_sha256": "d7cd044ad079712eda5f7198cae455fc3b4ed0ba263fd893a5e447a80dfd0b0e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'socket sets' to 'wrench sets' within the same utterance, preserving the action, scope ('all'), and destination ('Amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/false_start/false_start_01_product_category/2f644a8df6ab8872/speech.wav", "wer": 0.3157894736842105, "wav_sha256": "9afc14e16cec826f1bc5c86f4d793a66f8637509ad5bc2eb8c6d84dd66bc9f93", "seconds": 6.6, "spoken_utterances": [ "Place an order for uh... all socket sets—um... sorry, I meant wrench sets in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'socket sets' but immediately self-corrects to 'wrench sets' with 'Sorry, I meant...'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, with only the target item corrected from 'socket sets' to 'wrench sets'. All other conditions, including 'order', 'all', and 'in my Amazon wish list', are preserved.", "heard_transcript": "Place an order for uh all socket sets some. Sorry, I meant wrench sets in my Amazon wish list." }, "candidate_id": "wav_b79552215895efc7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'socket sets', then immediately self-corrects to 'wrench sets' with a clear, spontaneous-sounding 'sorry, I meant'. The correction is integrated smoothly into the sentence, and the final phrase 'in my Amazon wishlist' is delivered with natural prosody and a slight pause before the final word, suggesting genuine recall. The flow feels like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all wrench sets in my eBay—no, that should be amazon wish list.", "long_pause_offsets": [], "source_sha256": "e202e3fd405f25cb7c68123557cc6587d6f9521fa555562a3d862d98a0605e3e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'eBay' to 'Amazon wish list' within the same utterance, preserving the action, item, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/false_start/false_start_02_app/e39f2001b38ac01a/speech.wav", "wer": 0.15, "wav_sha256": "bcc8fba6940c77f5bcf3faff27625dd8ff84fb4f184b37f6fe71f6688bf5bad0", "seconds": 11.48, "spoken_utterances": [ "Place an order for all wrench sets hmmm... in my eBay—uh... no, that should be um... amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value ('Amazon') should be applied, and the original incorrect value ('eBay') should be discarded. The final intent matches the original instruction with the corrected target ('Amazon').", "heard_transcript": "Place an order for all wrench sets, um, in my eBay, uh, no, that should be, um, Amazon wish list." }, "candidate_id": "wav_36d111127b2814cc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a plausible self-correction but with a slightly less natural flow. The speaker starts with 'eBay', then corrects to 'Amazon wishlist'. The correction is clear, but the initial error ('eBay') is not immediately followed by a self-correction; instead, there's a pause and a hesitant 'no, that should be'. This creates a slight break in the thought process. The prosody is natural, but the correction feels a bit more like a deliberate change than a spontaneous slip." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all wrench sets in my amazon shopping cart—scratch that, make it wish list.", "long_pause_offsets": [], "source_sha256": "915c4bc835f8748bd4616f682ca3c7163cbaacf3376a04d784736f244cbe1c3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately corrects it to 'wish list'. The action, item scope, and app remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/false_start/false_start_03_source_collection/19ccb85c2f310834/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "0b88d61f0e4119163ff8dd8c1f263e850801061249c4d37512514da1047396ac", "seconds": 11.84, "spoken_utterances": [ "Place an order for all wrench sets um... in my amazon hmmm... shopping cart—um... scratch that, make it wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately self-corrects with 'scratch that, make it wish list.' This is a clear false start where the final intent is the corrected version. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Place an order for all wrench sets, um, in my Amazon, mhm, shopping cart. Um, scratch that, make it wish list." }, "candidate_id": "wav_05aa8c4c86f70b8d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with 'Amazon shopping cart' and then says 'scratch that, make it wishlist'. The issue is that 'scratch that' is a cancellation, not a self-correction of a word. The speaker doesn't correct a specific word but changes the entire concept from 'shopping cart' to 'wishlist'. This makes the 'false start' less plausible as a genuine slip of the tongue and more like a deliberate change of mind, which is less natural for this augmentation type. The prosody is also slightly flat compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon wish list. I left out a detail: the order is for wrench sets.", "long_pause_offsets": [ 47 ], "source_sha256": "d3271fbbb7cb532de88c8f3836a8aa3ac96c1884f37a3dae47c1d64fee84c746", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing item category 'wrench sets' to the initial incomplete request, completing the final goal without contradicting any other part of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/addition/addition_02/e1d4ee1c155dae90/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "6f4e1fb1a043c4cb3545c1e30f0722bd39a9c00405bb3f67dafac724fafc5cb4", "seconds": 10.56, "spoken_utterances": [ "Place an order for all ah... in my amazon wish list. ", "I left out a detail: the order is for um... wrench sets." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all a' which is likely a misstatement of 'all wrench sets'. After a pause and self-correction, they clarify the correct target as 'wrench sets'. Since this is an 'addition' type, the final intent combines the original instruction with the added clarification. The final audible goal matches the reference instruction exactly.", "heard_transcript": "Place an order for all a, in my Amazon wish list. I left out a detail, the order is for um, wrench sets." }, "candidate_id": "wav_7a38d45dd368c568", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural-sounding initial request with a slight hesitation ('um') and a believable self-correction. The revision ('I left out a detail...') is smoothly integrated with a pause and a change-of-mind intonation, and the corrected item ('wrench sets') is clearly articulated. The prosody and pacing are consistent with spontaneous speech, and the repair feels authentic rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Place an order for wrench sets in my amazon wish list. Oh, and one more thing: place an order for all wrench sets.", "long_pause_offsets": [ 55 ], "source_sha256": "99e834dacbea18a7712c8859af708631789471769bfea30bc729df7aa1f8afc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the explicit scope 'all' to the wrench sets, which was missing in the initial utterance. The final goal combines the original instruction with the added constraint, resulting in the same intended action with a more specific scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/addition/addition_01/69453116b9b1b9aa/speech.wav", "wer": 0.125, "wav_sha256": "c2d13af18b70936c2cf60f50ab51c0f86a84de57fa0c8b8d400a88db22005e42", "seconds": 11.88, "spoken_utterances": [ "Place an order for wrench sets um... in my amazon wish list. ", "Oh, and one more thing: place an order for ... all wrench sets." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for wrench sets in my Amazon wish list.' This is followed by an addition: 'place an order for all wrench sets.' The final audible intent combines the original request with the addition, resulting in a request to place an order for 'all wrench sets' in the Amazon wish list. The addition refines the initial request by specifying 'all' wrench sets, which is consistent with the type-specific criterion for 'addition'. The final goal matches the reference instruction, which is to place an order for all wrench sets in the Amazon wish list.", "heard_transcript": "Place an order for wrench sets, um, in my Amazon wish list. Oh, and one more thing, place an order for all wrench sets." }, "candidate_id": "wav_3b274684433bf363", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a clear request, but the addition ('Oh, and one more thing...') sounds slightly abrupt and less naturally integrated. The second request is delivered without a pause or hesitation, making it sound more like a separate command than a spontaneous addition. While the content is coherent, the flow lacks the subtle hesitation and conversational rhythm of a truly spontaneous thought process, making it less realistic than audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all socket sets in my amazon wish list. Wait, the order should be for all wrench sets, not socket sets.", "long_pause_offsets": [ 59 ], "source_sha256": "8d1c68e953949cb0d3c7e8c56057ab94583ed74b347ff7d31362a1d25f8bd5e1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the tool category from socket sets to wrench sets, while preserving the action, scope ('all'), and destination ('amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/revision/revision_01_product_category/0e7ab1cb4cbe3097/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "c913912c74f411ea1034bb70e7f16f037f151a9d564f3f40cdc283b55bcc9017", "seconds": 13.6, "spoken_utterances": [ "Place an order for um... all socket sets uh, let me think... in my amazon wish list. ", "Wait, the order should be for all wrench sets, not socket sets." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'socket sets' but immediately corrects themselves to 'wrench sets' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Place an order for um, all socket sets, uh, let me think, in my Amazon wish list. Wait, the order should be for all wrench sets, not socket sets." }, "candidate_id": "wav_b44904b45d857421", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a filler ('um'), a hesitation ('uh, let me think'), and a clear self-correction ('Wait, the order should be for all wrench sets, not socket sets'). The correction is integrated into the same utterance with a believable shift in intonation and pacing, mimicking a real-time thought process. The speaker also corrects the item type mid-sentence, which is a strong indicator of authentic task formulation." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all wrench sets in my eBay wish list. Let me correct the app: place the order on Amazon, not eBay.", "long_pause_offsets": [ 57 ], "source_sha256": "28623c1caad11705a77cbcae562dbbb828e6496fe691f5492b989ae779f67850", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, while retaining the action, item type, and scope of 'all wrench sets'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/revision/revision_02_app/6f67ca0a862262ce/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "f0f905f380faa16e5bc637211c78254195f4548b9f5714a83039c307a7570862", "seconds": 14.64, "spoken_utterances": [ "Place an order for um... all wrench sets let's see... in my eBay wish list. ", "Let me correct the app: place the order on hmmm... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially mentions eBay but immediately self-corrects to Amazon, which is the final and intended target. The revision is clear and unambiguous, and the rest of the instruction (ordering all wrench sets from the wish list) remains consistent. The final intent matches the original instruction with the correct app.", "heard_transcript": "Place an order for all wrench sets, let's see, in my eBay wish list. Let me correct the app. Place the order on Amazon, not eBay." }, "candidate_id": "wav_9d43842f74bdb1b2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision with a self-correction ('Let me correct the app, place the order on Amazon, not eBay'). The correction is delivered in a single, continuous utterance with a natural pause and a shift in intonation. The speaker's voice and prosody remain consistent, and the correction feels like a spontaneous realization rather than a rehearsed edit. However, the initial request is slightly less fluid than audio_03, with a minor stumble ('from... let's see...') that is less integrated than the correction in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all wrench sets in my amazon shopping cart. Change the source to wish list instead of shopping cart.", "long_pause_offsets": [ 63 ], "source_sha256": "d0be5eae23ee97e895d6752d196f36fedcbf3456b079752ef988a55dce6dc31d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source from shopping cart to wish list, aligning with the original instruction. All other elements—action, item type, and scope—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_1/audio_natural/revision/revision_03_source_collection/7b521c23455fdb68/speech.wav", "wer": 0.08, "wav_sha256": "f7dccce74c6ee5089378179e7da677dc5793784234a788dd2478ca90fef07c8d", "seconds": 11.92, "spoken_utterances": [ "Place an order for all wrench sets um... in my amazon shopping cart. ", "Change the source to let's see... wish list instead of shopping cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but then revises the source to 'wishlist' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to place an order for all wrench sets in the Amazon wishlist, matching the original reference instruction.", "heard_transcript": "Place an order for all wrench sets um in my Amazon shopping cart. Change the source to let's see, wishlist instead of shopping cart." }, "candidate_id": "wav_b76204a65ed6be64", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It presents a revision but in a more structured, less spontaneous way. The speaker says 'Change the source to let's see, wishlist instead of shopping cart,' which sounds more like a rehearsed or read instruction than a natural thought process. The phrase 'let's see' is used as a transition rather than a genuine hesitation, and the correction is delivered in a flat, declarative tone without the prosodic cues of a real-time correction. The overall delivery is more robotic and less conversational compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "78c8a074870c35bbba40490360d5f090d2dd2d3fd961c2309ebb45383c96224a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for um... all wrench sets uh... in my amazon wish list.", "Hang on, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "0d22252_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "01bae9b3bedb5b29e20171adbc68da4eec6f1a260849c19df2fc174e09940798", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for um... all wrench sets uh... in my amazon wish list.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "0d22252_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b959f5deaf11c077b8f023f2f0521f3534238c1f6a6938d0ce0366ac0dde5929", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for um... all wrench sets uh... in my amazon wish list." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "0d22252_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all mouse pads in my amazon shopping cart—scratch that, make it wish list.", "long_pause_offsets": [], "source_sha256": "b1e6fcbadbf615fdb86f2bf6b3cdcc2ef180df8591e5768dc527c591a7c14449", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately corrects it to 'wish list' within the same utterance. The final target is the wish list, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/false_start/false_start_03_source_collection/9644f6f214fd18b6/speech.wav", "wer": 0.1, "wav_sha256": "658d51aa95f2af0bea901c1876eb8598131faf88a02273eb2fda4f482e283e3a", "seconds": 13.24, "spoken_utterances": [ "Place an order for um... all mouse pads ... in my amazon shopping cart—ah... scratch that, make it wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately self-corrects with 'Scratch that, make it wish list.' This is a clear false start where the final intent is the corrected version. The final instruction matches the original reference instruction exactly.", "heard_transcript": "Place an order for um all mouse pads in my Amazon shopping cart. Scratch that, make it wish list." }, "candidate_id": "wav_685bda3b3377d0e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a clear false start ('Place an order for...'), followed by a hesitant filler ('um'), then a complete initial request. The speaker then self-corrects with a believable 'scratch that' and a smooth, natural revision to 'wishlist'. The prosody is consistent, the pauses are appropriately placed, and the correction feels like a genuine moment of reconsideration rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all mouse pads in my eBay—I should say amazon wish list.", "long_pause_offsets": [], "source_sha256": "a10de86d1b7b79acdbf3273ee4fc439eb11c31a9c9b0b805d15a6c1297dbd76a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'eBay' to 'Amazon wish list' within the same utterance, preserving the action, item scope, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/false_start/false_start_02_app/325057784c1d7bb0/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "f5dc858bd6d1d2e2a2f7927c389e8421bd4957d5f548851a90ae8dfaf2e24a7b", "seconds": 13.92, "spoken_utterances": [ "Place an order for all mouse pads uh, let's see... in my eBay—ah... I should say uh... amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon wish list'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for all mouse pads, uh, let's see, in my eBay, uh, I should say, uh, Amazon wish list." }, "candidate_id": "wav_32869f12499ccd15", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio has a plausible revision with a 'let's see' filler and a self-correction from 'eBay' to 'Amazon wishlist'. However, the correction is less smooth than in audio_03. The speaker says 'I should say' before the correction, which is a common phrase, but the delivery of the final 'Amazon wishlist' is slightly rushed and less natural than the revision in audio_03. The overall prosody is good, but the flow is not as seamless." } ] }, "is_default": false }, { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all keyboards—I meant mouse pads in my amazon wish list.", "long_pause_offsets": [], "source_sha256": "eac9682e189f252a61d8ee3d5312446b4100812ca41d04da0722f96451d6e762", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'keyboards' to 'mouse pads' within the same utterance, preserving the action, app, and scope of 'all' items from the wish list." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/false_start/false_start_01_product_category/be4a254bd3f18d57/speech.wav", "wer": 0.5882352941176471, "wav_sha256": "2dfd43cd0b815ce8b289385b0e1e0ed0f835a342d002a97cbe4c7e926aa78356", "seconds": 8.32, "spoken_utterances": [ "Place an order for all keyboards—ah... I meant uh... mouse pads in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'keyboard' but immediately self-corrects to 'mouse pads'. The final audible intent matches the original instruction after the false start. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Who is an order for all keyboard. I meant to, mouse pads in my Amazon wish list." }, "candidate_id": "wav_f1543b82e6fc6d5e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It begins with a false start ('Who is an order for all keyboard...') and a self-correction ('I meant to... mousepads...'). However, the correction is abrupt and the speaker does not complete the sentence. The delivery sounds more like a fragmented thought or a recording glitch than a natural, spontaneous utterance. The prosody is flat, and the lack of a coherent, completed request makes it sound less authentic compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for mouse pads in my amazon wish list. Oh, I forgot to mention, place an order for all mouse pads in my amazon wish list.", "long_pause_offsets": [ 54 ], "source_sha256": "4bee14a9cd3ae5466b6db54c95e27d01b6f933421a74e06407ce13830f65ee2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The second utterance adds the missing quantifier 'all', which was omitted in the first. The final goal includes the complete original instruction with the added constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/addition/addition_01/7c6db680dab67daf/speech.wav", "wer": 0.21875, "wav_sha256": "878b2302f551c28d82ee05e3a561fb10f6fc98e44f920d44696e555b40d18d46", "seconds": 17.36, "spoken_utterances": [ "Place an order for mouse pads uh, let's see... in my amazon wish list. ", "Oh, I forgot to mention, place an order for um... all mouse pads in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for mouse pads, uh, let's see, in my Amazon wish list.' This is followed by a self-correction: 'Oh, I forgot to mention, place an order for um, all mouse pads in my Amazon wish list.' The addition of 'all' modifies the initial request to be more specific, and the final instruction matches the original reference. The type is 'addition', which requires combining the earlier instruction with the added information. The final goal is to place an order for all mouse pads in the Amazon wish list, which aligns with the reference. No issues are present.", "heard_transcript": "Place an order for mouse pads, uh, let's see, in my Amazon wish list. Oh, I forgot to mention, place an order for um, all mouse pads in my Amazon wish list." }, "candidate_id": "wav_42242dfc062a05ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, followed by a plausible hesitation ('uh, let's see') and a self-correction ('Oh, I forgot to mention'). The addition of 'all mousepads' is smoothly integrated with a natural pause and filler ('um'), and the final sentence is a coherent, grammatically correct restatement. The prosody is varied, with appropriate stress and pacing, and the overall flow mimics a real person recalling and refining a thought." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon wish list. One more detail: only order mouse pads.", "long_pause_offsets": [ 47 ], "source_sha256": "2cac33ed09206427b99dca3cd8bb3eecf8b21bd83ca92a404816b8e985c64a64", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the initial broad request to only mouse pads, preserving the app and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/addition/addition_02/ab2e8da503ddc2c1/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "7be40b29d7b215051776859f10e990778a7e0a0e4469f0a865d6745db8c478c2", "seconds": 11.16, "spoken_utterances": [ "Place an order for ... all in my amazon wish list. ", "One more detail: only order ... mouse pads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to place an order for 'all in my Amazon wish list'. After a pause, the speaker added the constraint 'only order mouse pads'. Since this is an addition that narrows the scope of the original request, the final intent is to order only the mouse pads. The addition is clear and unambiguous, and the final goal matches the reference instruction, which was to order all mouse pads. Therefore, the task is executed as intended.", "heard_transcript": "Place an order for all in my Amazon wish list. One more detail, only order mouse pads." }, "candidate_id": "wav_5c2c6a3101250c57", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The delivery is more robotic and less fluid. The initial request is cut off mid-sentence, and the second part ('one more detail') is delivered with a flat, unnatural intonation. The final instruction ('only order mousepads') sounds like a mechanical addition rather than a spontaneous thought. The pauses are abrupt and lack the natural rhythm of a person thinking aloud, making the overall performance less realistic despite containing the same semantic content." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all mouse pads in my amazon shopping cart. I misspoke about the source: use 'wish list', not 'shopping cart'.", "long_pause_offsets": [ 62 ], "source_sha256": "2c534b38575e433b69c19a666fa01cfb552a41ef819d416a61e39c988ffc62a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'shopping cart' to 'wish list', preserving the action, item scope, and app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/revision/revision_03_source_collection/83cbc25ce756f36a/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "e154db2219ccbab7d0e574b0d1da70c8c1f4af4028f6d8716220c3802d3de301", "seconds": 18.48, "spoken_utterances": [ "Place an order for um... all mouse pads uh... in my amazon shopping cart. ", "I misspoke about the source: use 'um, let me think... wish list', not 'shopping cart'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but immediately corrects themselves to 'wish list' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to place an order for all mouse pads in the Amazon wish list, which matches the original instruction.", "heard_transcript": "Place an order for um all mouse pads uh in my Amazon shopping cart. I misspoke about the source. Use um let me think, wish list, not shopping cart." }, "candidate_id": "wav_6928001321dfc5c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then makes a clear self-correction ('I misspoke about the source') followed by a plausible thought process ('let me think') before stating the correct source ('wishlist'). The prosody is consistent, and the correction is integrated smoothly into the flow without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all mouse pads in my ebay wish list. Let me correct the app: use 'amazon', not 'ebay'.", "long_pause_offsets": [ 56 ], "source_sha256": "5e9fe4376ec4dc38305de63f4f9a69a48fd4b3b4b7e6a33ab43cf049aa356169", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'ebay' to 'amazon', while retaining the action, object, and scope of 'all mouse pads'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/revision/revision_02_app/e056cf8586e1691a/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "b04b810787946ea14c35e0c7c40bc08ee6381d07ab55f992f418ae8ea072334a", "seconds": 17.48, "spoken_utterances": [ "Place an order for um, let me think... all mouse pads ah... in my ebay wish list. ", "Let me correct the app: use '... amazon', not 'ebay'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects it to 'Amazon not eBay'. This is a revision where the final instruction overrides the initial incorrect value. The final intent matches the original reference instruction exactly.", "heard_transcript": "Place an order for um, let me think, all mouse pads uh in my eBay wishlist. Let me correct the app, use Amazon not eBay." }, "candidate_id": "wav_376b55a025f2a4da", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker uses 'um' and 'let me think' naturally, but the correction ('Let me correct the app') sounds slightly less spontaneous and more like a rehearsed phrase. The intonation on 'use Amazon not eBay' is a bit flat and abrupt, making the correction feel less like a genuine moment of realization and more like a scripted fix." } ] }, "is_default": false }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all keyboards in my amazon wish list. Wait, change 'mouse pads' to 'mouse pads'.", "long_pause_offsets": [ 57 ], "source_sha256": "9fa30376a900186063b8bd6fe434e4329883fb49be1510edd46d1be847db60b4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the product category from 'keyboards' to 'mouse pads', restoring the original value. The action, app, and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/revision/revision_01_product_category/6c67fbf9f0599388/speech.wav", "wer": 0.3181818181818182, "wav_sha256": "f79ae5eeaebb8995733a2c38c578fd2f3ea53311e0a69f295e6b25ff5816c348", "seconds": 14.16, "spoken_utterances": [ "Place an order for um... all keyboards let me think... in my amazon wish list. ", "Wait, change '... mouse pads' to 'mouse pads'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all keyboards' but then corrects to 'mouse pads' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Place an order for um all keyboards, let me think, in my Amazon wish list. Wait, change mouse pads to mouse pads." }, "candidate_id": "wav_d099aa942f6bcde6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with 'um' and 'let me think', but the correction ('Wait, change keyboards to mousepads') is delivered with a rushed, clipped tone that sounds less like a natural thought process and more like a forced edit. The word 'change' is pronounced abruptly, and the overall flow feels less organic compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "88d41ac81c74ee9fe07665f4153f866f2b85c725f7b4fd16a087b7e32649f462", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/cancellation/cancel_013/80e3f2f644605710/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "64db7ff40a52ad8e87f4d10abe87b39a50e3c7b3eeb7f1e1ce95ed99319c1902", "seconds": 12.6, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list.", "Hold on, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to place an order for all mouse pads in the Amazon wish list. After a brief pause, the speaker clearly states 'Hold on, I changed my mind. Please cancel the task.' This constitutes a direct and unambiguous cancellation of the entire task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list. Hold on, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "0270084e0378460d615715bcbc1e5787fd82fc4363b75e8a49e0cf194775ade8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/ambiguous_halt/halt_010/f01ab00c7d3811ba/speech.wav", "wer": 0.25, "wav_sha256": "b6894e6d7aea9574a9f403d45e5acc83d699837ec4ce6f6e67857326ea471e47", "seconds": 10.04, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled, but the agent is told to hold off on execution. No further task is issued, and the recording concludes with the pause request.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "531fb1a88f9415bc62029ce52ce2dadd80d1e08e8a7957db7dc1002d524aee3a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_2/audio_natural/filler/original/739dc034562f16ed/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "149bac8227294f2ecc4bb269070b335c75e496585f02d9e5a8e2ce3320c99c54", "seconds": 6.44, "spoken_utterances": [ "Place an order for ... all mouse pads ... in my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker clearly and directly states the original instruction without any change in goal or scope. The 'filler' type allows for minor hesitations or fillers within the instruction, but the final intent remains the original task. The instruction is fully audible and matches the reference.", "heard_transcript": "Place an order for all mouse pads in my Amazon wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all handheld vacuum cleaners in my Walmart—I should say amazon wish list.", "long_pause_offsets": [], "source_sha256": "727ac61602eb7229c17d2d7f62f27aa9adccca0c214ddd070a78fbc83e23b737", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Walmart' to 'Amazon wish list' within the same utterance, preserving the action, item type, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/false_start/false_start_02_app/5f7d188f75386abe/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "bba9cdc040a2aab5fb88a3928b3cee70d091152c05d07790c870b758b7050898", "seconds": 7.32, "spoken_utterances": [ "Place an order for all handheld vacuum cleaners in my Walmart—uh... I should say ... amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects to 'Amazon wish list'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Place an order for all handheld vacuum cleaners in my Walmart, uh, I should say Amazon wish list." }, "candidate_id": "wav_8ec71f9f6a2eca94", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all stick vacuum cleaners—or rather, handheld vacuum cleaners in my amazon wish list.", "long_pause_offsets": [], "source_sha256": "65ff6e41229f1a7dc8d494db55ad661c05868e286a9a0e76054307e1947d5f01", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'stick vacuum cleaners' but immediately corrects to 'handheld vacuum cleaners' using 'or rather', which clearly replaces the mistaken category. The rest of the instruction, including the action, scope ('all'), and context ('in my amazon wish list'), remains unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for all ... stick vacuum cleaners—uh... or rather, handheld vacuum cleaners in um... my amazon wish list." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all handheld vacuum cleaners in my amazon shopping cart—sorry, make that wish list.", "long_pause_offsets": [], "source_sha256": "a97adc3ff42c885ffe716279e93649055a1f74f0f7365e5f3fff4aa89f66dd26", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'amazon shopping cart' but immediately corrects it to 'wish list'. The final target is the wish list, matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for all let's see... handheld vacuum cleaners in my amazon shopping cart—uh... sorry, make that um... wish list." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon wish list. I left out a detail: only place an order for handheld vacuum cleaners.", "long_pause_offsets": [ 47 ], "source_sha256": "e8d0c81db6f2bfcdeac9fa282a467f2d6126b6508bee080670e7e55bdebed87a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only handheld vacuum cleaners, preserving the app (Amazon), action (place an order), and source (wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/addition/addition_02/39078b4ca57a7138/speech.wav", "wer": 0.125, "wav_sha256": "c3390442f5b675c741a5c6fe516cc768494b9f7f05b82a88a27b8bc7a5815e00", "seconds": 12.04, "spoken_utterances": [ "Place an order for all in um... my amazon wish list. ", "I left out a detail: only place an order for ... handheld vacuum cleaners." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was incomplete, but the speaker explicitly added a constraint ('only place an order for handheld vacuum cleaners') after a pause. This addition narrows the scope of the original request, and the final audible intent matches the reference instruction. The type is 'addition', and the final goal is fully aligned with the reference.", "heard_transcript": "Place an order for all in um my Amazon wish list. I left out a detail, only place an order for handheld vacuum cleaners." }, "candidate_id": "wav_efed8bb48a106f98", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, uses a plausible filler ('um') to pause, and then adds a specific restriction ('only place an order for handheld vacuum cleaners') in a way that sounds like a spontaneous correction. The prosody is consistent, and the addition flows logically from the initial request without awkwardness." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Place an order for handheld vacuum cleaners in my amazon wish list. Oh, I forgot to mention, include all handheld vacuum cleaners.", "long_pause_offsets": [ 68 ], "source_sha256": "23648425ee8f286415c5e68df0369bf14c10f1f28a9e155fdc6cffafc0ad8215", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the word 'all', which clarifies the scope of the initial request. The initial utterance already specified 'handheld vacuum cleaners' and 'Amazon wish list', so the addition of 'all' is a strengthening of the original instruction, not a contradiction or new constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/addition/addition_01/c2edbeae3d5397a4/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "21fbc549d294dabb34254d6585d5ba9d7b45f2e5aadcb24482479bdc192f4f6f", "seconds": 15.04, "spoken_utterances": [ "Place an order for ... handheld vacuum cleaners in um... my amazon wish list. ", "Oh, I forgot to mention, include ... all handheld vacuum cleaners." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for handheld vacuum cleaners in my Amazon wish list.' This is followed by a longer pause and an addition: 'Oh, I forgot to mention, include all handheld vacuum cleaners.' The addition clarifies and strengthens the initial request by explicitly including 'all' handheld vacuum cleaners. The final intent is to order all such items, which matches the original instruction's scope but with the added constraint of 'all'. Since the addition is a clarification that narrows the scope to include all items, the final goal is a combination of the original and the addition, satisfying the 'addition' type criterion.", "heard_transcript": "Place an order for handheld vacuum cleaners in um my Amazon wish list. Oh, I forgot to mention, include all handheld vacuum cleaners." }, "candidate_id": "wav_5f40518a4d179aa6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a more complex, slightly awkwardly phrased initial request ('place an order for handheld vacuum cleaners in my Amazon wishlist'), then adds a revision ('include all handheld vacuum cleaners'). While the addition is plausible, the initial phrasing is less natural and the revision feels slightly forced, as if the speaker is trying to correct a previous statement rather than building on it organically. The prosody is less fluid than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all air purifiers in my amazon wish list. Wait, change the product category to 'handheld vacuum cleaners' instead of 'air purifiers'.", "long_pause_offsets": [ 61 ], "source_sha256": "dc7581a249466a2a467f340d77ee0abe49906f7cf04f5e25834d2de62b5f114b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the product category from 'air purifiers' to 'handheld vacuum cleaners', preserving the action, scope ('all'), and source ('Amazon wish list')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/revision/revision_01_product_category/e9a0e4396f5bd2a1/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "ac1cb71e8f0d65f08d01f09c10ae919638144780f11ec0c3a508bd1c6c3ab5bf", "seconds": 15.04, "spoken_utterances": [ "Place an order for all air purifiers in let's see... my amazon wish list. ", "Wait, change the product category to 'ah... handheld vacuum cleaners' instead of 'air purifiers'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'air purifiers' but immediately corrects to 'handheld vacuum cleaners' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly.", "heard_transcript": "Place an order for all air purifiers in let's see, my Amazon wish list. Wait, change the product category to uh handheld vacuum cleaners instead of air purifiers." }, "candidate_id": "wav_da9f05c0b7d1459b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' before correcting the product category. The correction is delivered with a believable change-of-mind intonation and a smooth reference back to the initial request. The prosody is consistent, and the pacing feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all handheld vacuum cleaners in my Walmart wish list. Let me correct the app: use 'amazon' instead of 'Walmart'.", "long_pause_offsets": [ 73 ], "source_sha256": "9a5df74c483b10c291d1afc0276425ef6d5ee8247ced95b52b4795d96b6d5881", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Walmart to Amazon, preserving the action, item type, and scope of 'all handheld vacuum cleaners'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/revision/revision_02_app/366af367a36cedee/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "c79f09a72708b7fac124c824b8ab548be8926a5723dc873cb9f3cd804a9e4057", "seconds": 14.2, "spoken_utterances": [ "Place an order for all ... handheld vacuum cleaners in um... my Walmart wish list. ", "Let me correct the app: use '... amazon' instead of 'Walmart'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart wish list' but immediately corrects it to 'Amazon instead of Walmart'. This is a revision where the final instruction overrides the initial incorrect value. The corrected instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "Place an order for all handheld vacuum cleaners in um my Walmart wish list. Let me correct the app, use Amazon instead of Walmart." }, "candidate_id": "wav_936e1b3497545567", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'um' before correcting the app from Walmart to Amazon. The self-correction is immediate and natural, with a plausible shift in focus. However, the pause before 'let me correct' is slightly longer than in audio_02, and the overall delivery is a bit more hesitant, making it feel slightly less fluid." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all handheld vacuum cleaners in my amazon shopping cart. Change the source collection to 'wish list' rather than 'shopping cart'.", "long_pause_offsets": [ 76 ], "source_sha256": "49190865798dfa39b2fd98ef7da8f7699d05db48aced47d78cc22a180ee407dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly changes the source collection from 'shopping cart' to 'wish list', preserving the action, item type, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/revision/revision_03_source_collection/1815e8d44ee97a25/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "2835539f0a5da59a172476a3d63fc7608def0aa3e495b9516b930b04ef4e9925", "seconds": 14.84, "spoken_utterances": [ "Place an order for all um, let's see... handheld vacuum cleaners in my amazon shopping cart. ", "Change the source collection to '... wish list' rather than 'shopping cart'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping cart' but then corrects it by stating 'Change the source collection to wish list rather than shopping cart.' This is a revision where the final intent is to use the wish list as the source. The correction is clear and applies to the earlier instruction, satisfying the revision type criterion.", "heard_transcript": "Place an order for all um, let's see, handheld vacuum cleaners in my Amazon shopping cart. Change the source collection to wish list rather than shopping cart." }, "candidate_id": "wav_9c6d8222a3e67a8d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with multiple fillers ('um', 'let's see') and a slightly awkward phrasing ('place an order for all um let's see handheld vacuum cleaners'). The correction is delivered with a flat intonation and lacks the natural flow of the other two. The prosody is less varied, and the overall delivery sounds more rehearsed or less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "63303e3dc3fa7485273c81a04ffbc948e62c318d9c921fe2e58612a8f1517b66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/cancellation/cancel_009/52582ff42a05c1ba/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "3e044f1ad305b5cb0c66e63f07166ad769b9d0287af66b8af620746123cf618a", "seconds": 11.52, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but immediately self-cancels with 'Hang on, cancel that altogether.' This constitutes a clear cancellation of the entire task as per the type-specific criterion. The final control intent is cancellation, matching the expected behavior for this type.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "972dab916fe3988a7f352f02f96a823cf4aa02b029916f7fd4bac594729c3281", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/ambiguous_halt/halt_039/ebdc2b772cf392d9/speech.wav", "wer": 0.15, "wav_sha256": "997ae43fdbbf6d9cd168ea8d6c81f36054c630a9d957f3023f8a59b89f7a03da", "seconds": 11.16, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list.", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, give me a second first,' which is a clear request to pause or hold. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker is asking for a moment before proceeding.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list. Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "0d22252_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "56a2e39e9ff7fc4ff1337978c8174c441aa4113be3c00df5e02b4eec18d09312", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/0d22252_3/audio_natural/filler/original/562e4bd8507eb9e8/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "9c4438a23fadcee6115da1fcd4f2e90add1b7ebdaeaf0141817d4ecf070f7f32", "seconds": 7.28, "spoken_utterances": [ "Place an order for all uh... handheld vacuum cleaners in ... my amazon wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler word 'uh' but completes the original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "Place an order for all uh handheld vacuum cleaners in my Amazon wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2021—sorry, that should be 2022 from my Spotify song library.", "long_pause_offsets": [], "source_sha256": "40d9003aa5cf4b2522bdb7d41a7d76cc7c00365ff4a951a1e4d3df8f910e775b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'two thousand twenty-one' to 'two thousand twenty-two' within the same utterance, preserving the app, action, and scope of 'any song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/false_start/false_start_01_release_year/c66e54dfefe7d35c/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "9dde0c2db49d2eac1ce62aeb033d823cd1d40127cc02270ab2383c06948fa670", "seconds": 10.04, "spoken_utterances": [ "Play any song released in two thousand twenty-one—ah... sorry, that should be ... two thousand twenty-two from my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces '2022' as 'two teeny one uh' but immediately self-corrects to 'two thousand twenty two'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved. The corrected year matches the reference, and no other constraints are altered. Therefore, the final intent aligns with the original instruction.", "heard_transcript": "Play any song released in two teeny one uh, sorry, that should be two thousand twenty two from my Spotify song library." }, "candidate_id": "wav_7b0e8817a4cc38c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural-sounding false start with 'tutini huan' followed by a clear self-correction to '2022'. The speaker's tone and pacing are consistent, and the correction is delivered smoothly without awkwardness. The pause before the correction is brief and plausible, fitting a moment of recalling a number. The final request is articulated clearly and confidently." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play any song released in 2022 from my Apple Music—no, I meant Spotify song library.", "long_pause_offsets": [], "source_sha256": "39cd473212d0f9637e17d0581f225cf5da1ba1f63e960800c424a1b67bea513f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify, preserving the year 2022 and the action of playing a song from the song library." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/false_start/false_start_02_app/65f7e0284d23f3bf/speech.wav", "wer": 0.125, "wav_sha256": "10ee3b860e7839d3786dd80d3c91bbabba209623fd6d770423e86d1f3f414891", "seconds": 9.8, "spoken_utterances": [ "Play any song released in two thousand twenty-two um, let me think... from my Apple Music—uh... no, I meant Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify song library'. According to the 'false_start' type criterion, the final corrected target should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Play any song released in two thousand twenty two, um, let me think, from my Apple Music, uh, no, I meant Spotify song library." }, "candidate_id": "wav_6943bb19f09f16b0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'Apple Music' followed by a self-correction to 'Spotify song library'. The speaker's hesitation and the correction are natural, but the correction is slightly less fluid than in audio_03. The 'uh' filler and the pause before the correction are realistic, but the overall delivery feels a bit more hesitant and less confident than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Play any song released in 2022 from my Spotify playlist—scratch that, make it song library.", "long_pause_offsets": [], "source_sha256": "f8290703c700dfd4a855fe5dd4d0906cd20510f3aa865863b75c966185dd3fea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'playlist' but immediately corrects it to 'song library'. The year 'two thousand twenty-two' is equivalent to '2022'. All other elements, including the action, scope, and app, remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/false_start/false_start_03_source_collection/53ffed43bcf4134a/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "7bc0b9f8349a2821dfd9c550919b90a34344ae72285451de2bbf4da559c14923", "seconds": 12.48, "spoken_utterances": [ "Play any song released in two thousand twenty-two um, let me think... from my Spotify playlist—uh... scratch that, make it song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify playlist' but immediately self-corrects with 'scratch that, make it song library'. According to the 'false_start' type criterion, the final corrected target should be applied. The final target 'song library' replaces the initial 'playlist', and all other conditions (song released in 2022) remain unchanged. The correction is immediate and clear, so the task is executed with the corrected target.", "heard_transcript": "Play any song released in two thousand twenty two, um, let me think, from my Spotify playlist, uh, scratch that, make it song library." }, "candidate_id": "wav_9b01beb059e10c64", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio includes a false start with 'playlist' and a correction to 'song library'. However, the correction is delivered with a more abrupt and less natural tone, sounding like a forced edit. The 'uh' filler and the pause are present, but the overall flow feels less spontaneous and more like a rehearsed or artificially constructed correction compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Play any song released from my Spotify song library. Oh, I forgot to say, play a song released in 2022.", "long_pause_offsets": [ 53 ], "source_sha256": "9d2b3c9ba82caa9031dafccdaa125fc7d9b77035f038d95d564fc8b04abb1c98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing year constraint, which refines the initial broad request. The app (Spotify), action (play), and scope (any song from library) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/addition/addition_01/b91365d0af05bc9c/speech.wav", "wer": 0.0, "wav_sha256": "78101837098faaed29cf16c771b03ebee86276d1b04ea59886dc8ad62cccd3db", "seconds": 12.6, "spoken_utterances": [ "Play any song released ... from my Spotify song library. ", "Oh, I forgot to say, play a song released in uh... two thousand twenty-two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to play any song from their Spotify library, then adds the constraint 'released in 2022' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and executable, matching the reference instruction.", "heard_transcript": "Play any song released from my Spotify song library. Oh, I forgot to say, play a song released in uh, 2022." }, "candidate_id": "wav_3629806f34bbebdb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'Oh, I forgot to say' before adding a specific detail. The addition of 'in uh 2022' includes a realistic filler ('uh') and a slightly uncertain intonation, suggesting spontaneous recall. The flow is continuous and the correction feels like a genuine afterthought, making it highly realistic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Play any song released in 2022. I forgot to add, play it from my Spotify song library.", "long_pause_offsets": [ 32 ], "source_sha256": "75ec8bd0aee41c50c6d23e8d8b1b6d8bf7a95707faa338399998cbfd95bb5bc2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source requirement, combining it with the initial release year condition. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/addition/addition_02/09bc264cf90b0c59/speech.wav", "wer": 0.0, "wav_sha256": "56938aaf3318e8708726fbd658959593deaf8982af786cfc73c2bb7776ba33a8", "seconds": 12.48, "spoken_utterances": [ "Play any song released in ... two thousand twenty-two. ", "I forgot to add, play it from my ... Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first requests to play any song released in 2022, then adds the constraint to play it from their Spotify song library. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and executable, matching the original instruction.", "heard_transcript": "Play any song released in 2022. I forgot to add, play it from my Spotify song library." }, "candidate_id": "wav_6bb1ef280aed751d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then adds a follow-up with 'I forgot to add,' which is a plausible self-correction. However, the second part of the request is delivered with a slightly more robotic, rehearsed cadence compared to audio_01. The transition is less fluid, and the overall delivery lacks the subtle hesitation and natural pacing of a truly spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_release_year", "proposed_instruction": "Play any song released in 2023 from my Spotify song library. Make that any song released in 2022, not 2023.", "long_pause_offsets": [ 61 ], "source_sha256": "02c50749f51759ca2df290abbf057961accf5ef7548faf84ea97ed02c1203f67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the release year from 'two thousand twenty-three' to 'two thousand twenty-two', preserving the app, action, and scope of 'any song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/revision/revision_01_release_year/54b807d283257076/speech.wav", "wer": 0.0, "wav_sha256": "675da0d1a2c01b6dd371949bce7e934c7ade438c7960ca6227310b731c748606", "seconds": 14.88, "spoken_utterances": [ "Play any song released in ... two thousand twenty-three from my Spotify song library. ", "Make that any song released in um... two thousand twenty-two, not two thousand twenty-three." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests a song from 2023, then corrects themselves after a pause to specify 2022. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to play a song from 2022, which matches the original reference instruction. All other constraints (Spotify song library, any song) remain unchanged.", "heard_transcript": "Play any song released in two thousand twenty three for my Spotify song library. Make that any song released in um two thousand twenty two, not two thousand twenty three." }, "candidate_id": "wav_8ae93dddef27fb48", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Play any song released in 2022 from my Apple Music song library. Let me correct the app: play from my Spotify song library, not Apple Music.", "long_pause_offsets": [ 65 ], "source_sha256": "6ce48f7e41fa849cebf145666de554f1e6da722eb898054f54a93fd52cb555fa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the year 2022 and the action of playing any song from the song library." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in ... two thousand twenty-two um, let me think... from my Apple Music song library. ", "Let me correct the app: play from my ... Spotify song library, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Play any song released in 2022 from my Spotify playlist. Change the source to my song library, not playlist.", "long_pause_offsets": [ 57 ], "source_sha256": "806eb691a64d9be18a7f01610ad006603eeb2d158cefe6662540ff7f57131eb0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup correctly revises the source from 'playlist' to 'song library' while preserving the year and action. The spoken year 'two thousand twenty-two' is equivalent to '2022'." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh, let's see... from my Spotify playlist. ", "Change the source to my let me think... song library, not playlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "8aeb999ab5912c5dae4510e8415e4864bcee3109f272c40b8f4b3875d2febb68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/cancellation/cancel_031/ad598251a6a9afe6/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "33f4ad80496fd2518b6fd8fe0a94618e11fba6e37b3aded0cf9afa8673b9dfb6", "seconds": 14.4, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library.", "Hold on, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to play a song from 2022, but then clearly revokes the entire task by saying 'Hold on, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the original instruction. The final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Play any song released in whom two thousand twenty two uh from my Spotify song library. Hold on, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "3c5397af9c2ea9abc355f8bb5f10973c21667465b68e954df6104a1243896aa8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/ambiguous_halt/halt_015/5c78a676175ed0f6/speech.wav", "wer": 0.1, "wav_sha256": "d57578496b2641cac84ca2ebc216f8b3c30e1f0ce369f27ed86f55606e687acd", "seconds": 12.76, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library.", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but immediately hesitates with 'whom?' and 'uh', then completes the instruction with '2022'. After stating the full request, they say 'Please wait a moment', which is a clear and final request to pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The initial hesitation and self-correction do not change the final intent, which is to pause.", "heard_transcript": "Play any song released in whom? 2022 uh from my Spotify song library. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dcecea0decb5d66b33b1e0ccaed9936d6e6a082e17b403b308fcf647bccc1704", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_1/audio_natural/filler/original/c77552fb4bc50dc2/speech.wav", "wer": 0.125, "wav_sha256": "407c0f2c5fffa9ce524cafedc799e62cf60c327c4e6e77b3d441543dab461541", "seconds": 8.96, "spoken_utterances": [ "Play any song released in hmmm... two thousand twenty-two uh... from my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces '2022' as 'whom', then self-corrects to '2022'. The final audible intent matches the original instruction. The initial error is a filler-like misstatement that is immediately corrected, and the final goal is clear and executable.", "heard_transcript": "Play any song released in whom? 2022 uh from my Spotify song library." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2020—sorry, I meant 2021 from my Spotify album library.", "long_pause_offsets": [], "source_sha256": "caf128f4475b5f9afb1ecc246c8b3be31ad649f3cc3a9b5f82aa25b40d6454cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'two thousand twenty' but immediately corrects to 'two thousand twenty-one', which is the same as 2021. The app, action, and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/false_start/false_start_01_release_year/9277aeba663ddcdf/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "0a664f28788346db52a159835753fe26ea6f1b14e8279825ac2a45dc453b5ed7", "seconds": 8.44, "spoken_utterances": [ "Play any song released in two thousand twenty—uh... sorry, I meant ah... two thousand twenty-one from my Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '2020' but immediately self-corrects to '2021' with 'sorry, I meant uh 2021'. This is a clear false start where the final value is the corrected one. The rest of the instruction, including the source 'Spotify album library', remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Play any song released in two thousand twenty uh sorry, I meant uh two thousand twenty one from my Spotify album library." }, "candidate_id": "wav_e5c96ef425472f09", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then makes a natural false start ('2020') followed by a self-correction ('2021') with a smooth, conversational tone. The correction is immediate and integrated into the same utterance without awkward pauses or repetition. The phrase 'from my Spotify album library' is delivered with consistent prosody, suggesting spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play any song released in 2021 from my Apple Music—no, that should be Spotify album library.", "long_pause_offsets": [], "source_sha256": "b72f49e07bbd325368097aac65e37c38efa17a0802dfb4014b982dfd8e456ada", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The genre, release year, and library scope remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/false_start/false_start_02_app/6ec7482370bddd63/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "f5532e4f00b21198a13927c38e479b70a4dbf30f8eb02b1208a365e34032264f", "seconds": 9.52, "spoken_utterances": [ "Play any song released in two thousand twenty-one from my hmmm... Apple Music—um... no, that should be ... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify album library'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Play any song released in two thousand twenty one from my hmm Apple Music um no, that should be Spotify album library." }, "candidate_id": "wav_66ae4a1648f19dfe", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then hesitates ('hmm') and makes a change of mind, first mentioning 'Apple Music' before correcting to 'Spotify album library'. The correction is believable, but the initial mention of 'Apple Music' feels slightly more forced than the correction in audio_01, and the 'hmm' filler is longer and less integrated. The overall flow is still natural, but the correction is less seamless than in audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Play any song released in 2021 from my Spotify playlist collection—scratch that, make it album library.", "long_pause_offsets": [], "source_sha256": "f68ee1df9a1b41f379488125ebb1a2e7588a005a8bf8cfb03db6f641a9bb6827", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'playlist collection' but immediately corrects it to 'album library'. The year 'two thousand twenty-one' is equivalent to '2021'. All other elements, including the action, scope, and app, remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/false_start/false_start_03_source_collection/cd66c1fe4b405f65/speech.wav", "wer": 0.0, "wav_sha256": "3aa9d21be7352cdf9ce79a480f69308174bfb065e0ddddc904a4590e6ef75d69", "seconds": 10.28, "spoken_utterances": [ "Play any song released in two thousand twenty-one from my um, let's see... Spotify playlist collection—um... scratch that, make it album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately self-corrects with 'Scratch that, make it album library.' This is a clear false start where the final target is 'album library.' The rest of the instruction, including the year 2021 and the action 'play any song,' remains unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Play any song released in two thousand twenty one from my um let's see Spotify playlist collection um. Scratch that, make it album library." }, "candidate_id": "wav_716f5a4dbda52524", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then hesitates ('um') and says 'let's see', followed by a change from 'playlist collection' to 'album library'. The phrase 'scratch that' is a very natural way to cancel and correct, but the correction is delivered with a noticeable pause and a slightly more abrupt shift in prosody. The overall delivery is realistic but feels slightly more disjointed compared to the other two, with the correction feeling more like a separate thought than a fluid self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Play any song released from my Spotify album library. Oh, I forgot to mention, play only songs released in 2021.", "long_pause_offsets": [ 54 ], "source_sha256": "088a7c52c1888da4efc459e98581c9a08f618fe04589d42f594f2fc01404e951", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing year restriction, narrowing the initial broad request to only songs from 2021. The app and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/addition/addition_01/5f90ad5fdcf3f778/speech.wav", "wer": 0.12, "wav_sha256": "927f52a4e5a001f482b30e9dd549eb08f2cb3acccf3053addd301da433f7c6a4", "seconds": 12.92, "spoken_utterances": [ "Play any song released from my hmmm... Spotify album library. ", "Oh, I forgot to mention, play only songs released in um... two thousand twenty-one." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Play any song released from my home Spotify album library.' This is followed by an addition: 'Oh, I forgot to mention, play only songs released in um, twenty twenty one.' The addition narrows the scope of the original request by specifying the release year as 2021. The final intent combines the original request with the added constraint, matching the reference instruction. The 'um' and 'twenty twenty one' are natural speech variations for '2021'. The control intent is 'execute' as the addition refines the task rather than cancelling or pausing it.", "heard_transcript": "Play any song released from my home Spotify album library. Oh, I forgot to mention, play only songs released in um, twenty twenty one." }, "candidate_id": "wav_93bd9a902e831c04", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable hesitation ('um') and a self-corrected pause ('from my home Spotify album library') before adding a new constraint ('play only songs released in um 2021'). The addition is integrated smoothly with a change-of-mind intonation and a plausible reference back to the initial request. The prosody is consistent, and the pauses feel like genuine thought processes rather than robotic breaks." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Play any song released in 2021 from my album library. I should also mention, do this on Spotify.", "long_pause_offsets": [ 54 ], "source_sha256": "8471003c82ba482d3edcf18e08499be19612fe101628fd60e34edbabd17528e4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the platform as Spotify, which was already implied by the original instruction's 'Spotify album library'. The year 'two thousand twenty-one' is equivalent to '2021'. All core requirements are preserved and clarified." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/addition/addition_02/bca35fcdcc47a7b7/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "628bc5fafc5eb90909bb2b37d4f395c9af5b94e89896a104de7189c29c3e9568", "seconds": 11.48, "spoken_utterances": [ "Play any song released in um... two thousand twenty-one from my album library. ", "I should also mention, do this on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Play any song released in um, 2021 from my album library.' The user then adds 'I should also mention, do this on Spotify.' This is a clear addition of the platform constraint. The final intent combines the original request with the added condition, matching the reference instruction exactly. The hesitation ('um') and the phrase 'I should also mention' are natural speech fillers and do not alter the final goal. The addition type criterion is satisfied.", "heard_transcript": "Play any song released in um, 2021 from my album library. I should also mention, do this on Spotify." }, "candidate_id": "wav_d3cdd85518b9fe4c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then adds a new instruction ('I should also mention, do this on Spotify') in a way that sounds like a post-hoc afterthought. The delivery is fluent but less natural than audio_02 because the addition feels tacked on rather than integrated into the original thought. The prosody is slightly flat, and the second clause lacks the organic flow of a spontaneous correction or addition." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Play any song released in 2021 from my Apple Music album library. Let me correct the app: play from my Spotify album library, not Apple Music.", "long_pause_offsets": [ 66 ], "source_sha256": "5e8e12fdf900a0428fed51d3d730d423a54d011f6bf9cb00f2c893a4201252a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the year 2021 and the action of playing any song from the album library." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/revision/revision_02_app/4b0935477d790d66/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "4e7c0083606c9f2e1afe00e6e61ee9c2ce94f055b55d3c5a0d33ecc5327c0070", "seconds": 16.64, "spoken_utterances": [ "Play any song released in ... two thousand twenty-one from my uh... Apple Music album library. ", "Let me correct the app: play from my uh, let's see... Spotify album library, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to play a 2021 song from the Spotify library, matching the original reference instruction. All other constraints (song from 2021, album library) are preserved.", "heard_transcript": "Play any song released in 2021 from my uh Apple Music album library. Let me correct the app, play from my uh let's see, Spotify album library, not Apple Music." }, "candidate_id": "wav_5e01c469be823cd9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker starts with a clear request, then uses a believable self-correction with 'Let me correct the app' and a filler 'uh' before smoothly revising the platform from Apple Music to Spotify. The prosody is consistent, and the revision feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_release_year", "proposed_instruction": "Play any song released in 2020 from my Spotify album library. Make that any song released in 2021, not 2020.", "long_pause_offsets": [ 62 ], "source_sha256": "531e3525b25a97c89ab128c3a375b2138cdeb1e7d2522954bb8e0512ddb99f58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the release year from 'two thousand twenty' to 'two thousand twenty-one', preserving the app, action, and scope of 'any song from my album library'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/revision/revision_01_release_year/f516b3e692d13426/speech.wav", "wer": 0.125, "wav_sha256": "2d9e4a1efb1ce4fdd63a13e09ad86673af0367884c9c0e216a03458647a7d8cf", "seconds": 14.08, "spoken_utterances": [ "Play any song released in uh... two thousand twenty from my ... Spotify album library. ", "Make that any song released in um, let me think... two thousand twenty-one, not two thousand twenty." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '2020' but then corrects themselves to '2021' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to play a song from 2021, which matches the original reference instruction. The correction is clear and unambiguous, and the rest of the constraints (Spotify album library, any song) are preserved.", "heard_transcript": "Play any song released in uh 2020 for my Spotify album library. Make that any song released in um let me think 2021, not 2020." }, "candidate_id": "wav_c4a98481b7fa1ff5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a plausible 'let me think' filler, indicating hesitation before correcting the year from 2020 to 2021. The prosody is natural, and the correction is integrated into the flow. However, the initial request is slightly less fluid than audio_03, and the correction is less smoothly integrated, making it feel a bit more like a post-hoc edit." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Play any song released in 2021 from my Spotify playlist collection. Wait, play from my album library, not playlist collection.", "long_pause_offsets": [ 68 ], "source_sha256": "b9adca7d2d88f998647d909fdcdead61beeb931f96859614ed331ef787b3de82", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'playlist collection' to 'album library' while preserving the release year and the action of playing any song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/revision/revision_03_source_collection/f513c460feff4098/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "6dc9bbf7994c949d9f474084a65617ab4d16feaa58b2a6d955966771422a4729", "seconds": 13.48, "spoken_utterances": [ "Play any song released in uh... two thousand twenty-one from my let's see... Spotify playlist collection. ", "Wait, play from my ... album library, not playlist collection." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests playback from 'Spotify playlist collection' but immediately corrects this to 'album library, not playlist collection' after a pause. This constitutes a revision where the final intent is to play from the album library, which matches the original instruction. The year 2021 is preserved, and the correction is clear and unambiguous.", "heard_transcript": "Play any song released in uh two thousand and twenty one from my let's see, Spotify playlist collection. Wait, play from my album library, not playlist collection." }, "candidate_id": "wav_5c85dbce94cec52a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision, but the correction is abrupt and less integrated. The speaker says 'Wait, play from my album library not playlist collection' with a sharp shift in intonation and a lack of a natural filler or hesitation before the correction. The prosody is less continuous, and the correction feels more like a forced edit than a spontaneous recollection." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "c639fba0305b75ab542151925867f5c0b529cab7ce846e22fa92a94932346290", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/cancellation/cancel_019/067d6b5dade2e4a1/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "a70ebc62a1073189984946fd51f1f0df11ac9860ee195835772c89c331e0c8ac", "seconds": 11.68, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library.", "Hold on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly cancels it with 'Hold on, never mind, don't do that.' This constitutes a final, unambiguous cancellation of the original instruction. The control intent is correctly identified as 'cancel'.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library. Hold on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "ffeba519a95ef6228226fbea2acda1b5a0f46a593b07c6965dd49fa0efc74bca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/ambiguous_halt/halt_029/e31d2c32b31f92df/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "22ce339d2ba43430ef1a6eb34e90df536029f4bbef4298009d82d71bc500ca6f", "seconds": 10.96, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then says 'Wait, actually...' as a final interjection, which is a clear audible request to pause or halt the current action. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the audio ends with a pause directive.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library. Wait, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b4bfdd75392f06e0126a58868bfaa38ede02e9b14d75a020d01263e00a0d081d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_2/audio_natural/filler/original/e98c6a17fcaf218a/speech.wav", "wer": 0.125, "wav_sha256": "d796c8c2ec5138b74721cd43178529c6a5d6f023f72dabac67b3ab97c57ce74c", "seconds": 7.4, "spoken_utterances": [ "Play any song released in um... two thousand and twenty-one from my ... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um' as a filler, which is a common hesitation marker. The rest of the instruction is delivered clearly and matches the original request. The final intent is to execute the task as stated, with no correction or cancellation. The filler does not alter the goal, so the task is executed as intended.", "heard_transcript": "Play any song released in um, two thousand and twenty one from my Spotify album library." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1150ed6_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_release_year", "proposed_instruction": "Play any song released in 2019—sorry, I meant 2020 from my Spotify playlists.", "long_pause_offsets": [], "source_sha256": "7bdfd6894f1a5bd7047348389054927dd9f679d2c018ea1dea458bceb727b201", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'two thousand nineteen' but immediately corrects to 'two thousand twenty', which is the correct year. The app (Spotify), action (play), and scope (any song from playlists) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/false_start/false_start_01_release_year/1d58c6a36c7d81f6/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b4ee200331efa204ce1251e1530ba30fcd2cb4e756eba09c9321bddf9a9c7814", "seconds": 11.04, "spoken_utterances": [ "Play any song released in let me think... two thousand nineteen—hmmm... sorry, I meant uh... two thousand twenty from my Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '2019' but immediately self-corrects to '2020' after a pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected value, and the original incorrect value is discarded. The rest of the instruction matches the reference exactly.", "heard_transcript": "Play any song released in let me think two thousand nineteen hmm sorry I meant uh two thousand twenty from my Spotify playlists." }, "candidate_id": "wav_31808f9ae32e0b7a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'let me think' followed by a clear self-correction from '2019' to '2020'. The intonation and pacing of the correction sound spontaneous, and the subsequent phrase 'for my Spotify playlist' flows smoothly without hesitation. The prosody is consistent and the repair is integrated into the utterance as a single, coherent thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play any song released in 2020 from my Apple Music—I should say Spotify playlists.", "long_pause_offsets": [], "source_sha256": "8d0b6482b113ad60dd3dd5f35b0dcf4ee0c5be926d15a7d072c9e812f18429c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify, preserving the release year and playlist scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/false_start/false_start_02_app/11a9d28686eec1f7/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "17869de38721909db69671fbb729c49270fa6f679c7ac6ceb837e8c5b6a44bd5", "seconds": 11.68, "spoken_utterances": [ "Play any song released in ... two thousand twenty um, let me think... from my Apple Music—uh... I should say Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Play any song released in two thousand twenty, um, let me think, from my Apple Music, I should say Spotify playlists." }, "candidate_id": "wav_e11e57b46fe09083", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible false start with 'um let me think' after stating the year, followed by a correction from 'Apple Music' to 'Spotify playlist'. The correction is natural, but the initial hesitation is slightly longer and less integrated than in audio_01. The speaker also repeats 'uh' before the correction, which adds a minor layer of hesitation that slightly reduces the fluency compared to audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Play any song released in 2020 from my Spotify library—scratch that, make it playlists.", "long_pause_offsets": [], "source_sha256": "062cc3c270ce28ff0ab029c3ecea0ce72dbc30c29ad8d15940b8457e39a86b2a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'library' but immediately corrects to 'playlists' within the same utterance. The final target is 'playlists', and all other conditions (any song, 2020 release year, Spotify) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/false_start/false_start_03_source_collection/63a3d645db787555/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "199ec10b37315e721f07cf8811d670dd1688f8e2428f9e0057b4112514747594", "seconds": 12.08, "spoken_utterances": [ "Play any song released in um, let's see... two thousand twenty let's see... from my Spotify library—um... scratch that, make it playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify library' but immediately self-corrects to 'playlists' after a pause and the phrase 'scratch that'. The final audible target is 'playlists', which is the corrected version of the original instruction. The action (play a song), year (2020), and scope (Spotify) remain consistent. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Play any song released in um, let's see, two thousand twenty, let's see, from my Spotify library, um, scratch that, make it playlists." }, "candidate_id": "wav_6716b3189d9c17a3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker uses multiple fillers ('um', 'let's see') and repeats '2020', indicating a more fragmented thought process. The correction 'scratch that, make it playlists' is abrupt and less smoothly integrated into the prior sentence. The final 'um' before the correction sounds like a filler rather than a genuine moment of reconsideration, making the overall flow less natural and more disjointed than the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Play any song from my Spotify playlists. Oh, I forgot to mention, only play songs released in 2020.", "long_pause_offsets": [ 41 ], "source_sha256": "18ebac4958c2b7c5dd84bc85a6cc3e5edf059a4159cdf87fbfbf48bf58be2d6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing release year constraint, which is equivalent to the original instruction's requirement." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song ... from my Spotify playlists. ", "Oh, I forgot to mention, only play songs released ... in two thousand twenty." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Play any song released in 2020. One more thing: play the song from my Spotify playlists.", "long_pause_offsets": [ 32 ], "source_sha256": "5db9766a4973c565780e96c2dc092387c2ddd87b5ea2d37ddead0eafc3c80025", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source constraint 'from my Spotify playlists' to the initial request about songs released in 2020. The final goal combines both the release year and the playlist source." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in uh... two thousand twenty. ", "One more thing: play the song hmmm... from my Spotify playlists." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "1150ed6_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Play any song released in 2020 from my Apple Music playlists. Let me correct the app: play from Spotify playlists, not Apple Music.", "long_pause_offsets": [ 62 ], "source_sha256": "34fcd71cec2498fff7a34892382a44b480f322b24f0d659215c400dab182abe3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the year 2020 and the 'any song' scope. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1150ed6_3/audio_natural/revision/revision_02_app/310b436b130d1098/speech.wav", "wer": 0.22580645161290322, "wav_sha256": "4bc16c991693a9360619cdadb78727b56edd304839a444a4ec32865bd7443d46", "seconds": 17.76, "spoken_utterances": [ "Play any song released in uh... two thousand twenty um, let's see... from my Apple Music playlists. ", "Let me correct the app: play from let's see... Spotify playlists, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests a song from Apple Music playlists but immediately corrects the app to Spotify. The revision is clear and final, with no further changes or cancellations. The rest of the instruction (song from 2020) remains unchanged. The final intent matches the original instruction after the correction.", "heard_transcript": "Play any song released in uh two thousand twenty um let's see from my Apple Music playlists. Let me correct the app, play from let's see, Spotify playlists, not Apple Music." }, "candidate_id": "wav_5bc527fac26e5c48", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_release_year", "proposed_instruction": "Play any song released in 2019 from my Spotify playlists. Wait, the year should be 2020, not 2019.", "long_pause_offsets": [ 58 ], "source_sha256": "50a975bb54cf6049a85fd3c02a97ba9d756f65ec6e0acf5ad41a543e3f2972f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the release year from two thousand nineteen to two thousand twenty, while retaining the app (Spotify), the action (play), and the scope (any song from playlists)." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in ... two thousand nineteen uh... from my Spotify playlists. ", "Wait, the year should be ah... two thousand twenty, not two thousand nineteen." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Play any song released in 2020 from my Spotify library. Change the source to playlists instead of library.", "long_pause_offsets": [ 56 ], "source_sha256": "a583442547b2c08bd30972a8d5ea4fe38c169a87b0418cbb4b5581ff6ba4ce97", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'library' to 'playlists' while preserving the year and action. The year 'two thousand twenty' is equivalent to 2020." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in two thousand twenty uh... from my Spotify library. ", "Change the source to uh, let's see... playlists instead of library." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1150ed6_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "ba7ca4ceadb431ee6ece69b878b6761b2c982eb8a36d838faa36646a3b26a122", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in um... two thousand twenty let me think... from my Spotify playlists.", "Oh, wait, please cancel the entire task." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "1150ed6_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "e5cebae478606015a3c75cf958269f385705478e07cda51fc42e357d65a933cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in um... two thousand twenty let me think... from my Spotify playlists.", "Pause for a moment." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "1150ed6_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e8a89ba721c3116812bfbacc4ad5024790d8d3983c44e7bc77e223b0ce4f467d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Play any song released in um... two thousand twenty let me think... from my Spotify playlists." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "13547f5_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my my roommate—no, I meant all roommates, \"I have taken out the trash.\".", "long_pause_offsets": [], "source_sha256": "68eeeff96aa7cff5fa77f7a5612e4b423dd39869e4aeddc4a1086852b7d1f036", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'my roommate' but immediately corrects to 'all roommates'. The final recipient is correctly specified as 'all roommates', and the message content remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/false_start/false_start_02_recipient_reference/f486ebd1644ed39d/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "0aa25da40b06bf9b6048df50a09a087139335c8193c0a365d701ed150b2ba0cf", "seconds": 9.72, "spoken_utterances": [ "Send a let's see... phone voice message to my my roommate—ah... no, I meant ... all roommates, \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my roommate' but immediately self-corrects to 'all roommates'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction with the correction applied.", "heard_transcript": "Send a phone voice message to my my roommate uh no, I meant all roommates, I have taken out the trash." }, "candidate_id": "wav_e0c612bf4bd8284c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'send a', hesitates with 'let's see', then correctly states 'phone voice message'. The false start 'my roommate' is immediately and smoothly self-corrected to 'all roommates' with a believable change-of-mind intonation and a clear reference back to the initial request. The pacing and prosody are consistent with spontaneous speech, and the correction feels integrated rather than tacked on." } ] }, "is_default": true }, { "variant_id": "false_start_01_message_type", "proposed_instruction": "Send a text message—I should say phone voice message to my all roommates, \"I have taken out the trash.\".", "long_pause_offsets": [], "source_sha256": "e68a9197fc0f5c7631d58a8e34301d35926c640a563bb5c26267029ec1f917df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'text message' to 'phone voice message' within the same utterance, preserving the recipient group and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/false_start/false_start_01_message_type/e50478d15abdc253/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "f680ab333b242881c9b73e05605040857e694eb90d6131d77d76bc897027ecda", "seconds": 10.48, "spoken_utterances": [ "Send a uh, let me think... text message—um... I should say phone voice message to my all roommates, uh... \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'text message' but immediately self-corrects to 'phone voice message'. According to the 'false_start' type criterion, the immediate self-correction should be applied and the mistaken value discarded. The final audible intent matches the original instruction exactly, including the message content and recipients.", "heard_transcript": "Send a uh let me think text message um I should say phone voice message to my all roommates uh I have taken out the trash." }, "candidate_id": "wav_568f6bfd91035fa7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a false start with 'send a uh let me think', followed by a self-correction from 'text message' to 'phone voice message'. The repair is natural, but the initial hesitation 'let me think' is slightly longer and more pronounced than in audio_02, and the correction is less fluid. The phrase 'I should say phone voice message' is a slightly awkward way to self-correct, making it feel a bit more rehearsed than the smoother flow in audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_message_text", "proposed_instruction": "Send a phone voice message to my all roommates, \"I have taken out the recycling.\"—scratch that, make it \"I have taken out the trash.\".", "long_pause_offsets": [], "source_sha256": "492e7298762928a00a009187594d3267ca933b3a6e97135ef8b77909610006aa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces the message content from 'recycling' to 'trash'; the recipient group and message type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/false_start/false_start_03_message_text/2dee7b074bdf7f97/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "08fe82b9828c2dd8b1053a907dd397890bdd12bf463a25390853ee839de8b3e0", "seconds": 11.12, "spoken_utterances": [ "Send a uh... phone voice message to my all roommates, ... \"I have taken out the recycling.\"—um... scratch that, make it \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'recycling' but immediately self-corrects with 'scratch that, make it I have taken out the trash.' This is a clear false start where the final utterance corrects the initial value. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Send a uh phone voice message to my all roommates, I have taken out the recycling, um, scratch that, make it I have taken out the trash." }, "candidate_id": "wav_5c39db423ea28ed3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a false start with 'send a uh phone voice message', followed by a revision of the content from 'recycling' to 'trash'. The revision is delivered as 'scratch that, make it I have taken out the trash', which is a pragmatic and believable way to correct a detail. However, the initial phrase 'send a uh phone voice message' is slightly awkward and the 'uh' is placed in a way that disrupts the flow. The overall delivery feels less fluent compared to the other two, with a slightly more disjointed structure." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Send a phone voice message to my, \"I have taken out the trash.\". I should also mention, send it to all roommates.", "long_pause_offsets": [ 65 ], "source_sha256": "73aaade64bb9ab773b9a25df0787ec06b41ded64815d95fc16c41b541faa0e53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient group 'all roommates' after an initial incomplete utterance. The message content and action are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/addition/addition_02/854686e74890e0f6/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "a88863e0278b2164ed8872e6cdc79d7f73423d8550421d2feade6935088f35f4", "seconds": 14.16, "spoken_utterances": [ "Send a uh, let me think... phone voice message um... to my, \"I have taken out the trash.\". ", "I should also mention, send it to all roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a phone voice message to my...', then corrects to 'send it to all roommates' after a pause. This is an addition that completes the original instruction by specifying the target audience. The final intent matches the original instruction with the added detail of 'all roommates', which is consistent with the addition type criterion.", "heard_transcript": "Send a, uh, let me think, phone voice message, um, to my, I have taken out the trash. I should also mention, send it to all roommates." }, "candidate_id": "wav_e1ed8ceb1de10d5c", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Send a to my all roommates, \"I have taken out the trash.\". Oh, one other thing: send it as a phone voice message.", "long_pause_offsets": [ 59 ], "source_sha256": "2834723a08a2e996245b3cc59aa30230cfcdb715d9e9174dc05b177d52ff8fc6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing message type; the recipient and message content are preserved from the initial utterance." }, "audio": { "status": "pending", "spoken_utterances": [ "Send a ... to my all roommates, um... \"I have taken out the trash.\". ", "Oh, one other thing: send it as a uh... phone voice message." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_message_text", "proposed_instruction": "Send a phone voice message to my all roommates, \"I have taken out the recycling.\". Change the message to \"I have taken out the trash.\", not \"I have taken out the recycling.\"", "long_pause_offsets": [ 83 ], "source_sha256": "d3ac22375d243b01b334021c40b6eeea037bb93b48e7c52b69562e3ed4d4741a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the message content from 'recycling' to 'trash', preserving the recipient group and message type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/revision/revision_03_message_text/7cd4a1bc6f36dc46/speech.wav", "wer": 0.0, "wav_sha256": "4dedc681bfaa3e9f11cbb8588bd062d857f593dbadba99734775a3dcb301728c", "seconds": 17.28, "spoken_utterances": [ "Send a let's see... phone voice message to my all roommates, uh... \"I have taken out the recycling.\". ", "Change the message to um, let me think... \"I have taken out the trash.\", not \"I have taken out the recycling.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I have taken out the recycling' but then corrects it to 'I have taken out the trash' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "Send a let's see, phone voice message to my all roommates, uh, I have taken out the recycling. Change the message to, um, let me think, I have taken out the trash, not I have taken out the recycling." }, "candidate_id": "wav_e5793776aa1552bf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear initial request, then self-corrects the content of the message ('recycling' to 'trash') with a believable 'um' and 'let me think' hesitation. The correction is integrated smoothly into the same utterance, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my my roommate, \"I have taken out the trash.\". Make that to my all roommates, not my roommate.", "long_pause_offsets": [ 77 ], "source_sha256": "cdd6618e88158cb8153a5de6cc7d7e32b2d06b1e0739981ff37c3d99c45b1c75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from 'my roommate' to 'my all roommates', which is interpreted as 'all roommates'. The message content and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/revision/revision_02_recipient_reference/3afbe1a7f2fe38d4/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "5ca9e32a92a5d097523a3e26d94636554371d025105db3a701386b1ae904f58a", "seconds": 12.64, "spoken_utterances": [ "Send a um... phone voice message uh, let me think... to my my roommate, ... \"I have taken out the trash.\". ", "Make that to my all roommates, not my roommate." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to my roommate' but immediately corrects it to 'to my all roommates, not my roommate' after a pause. This is a revision where the final intent clearly overrides the initial incorrect target. The message content and action type remain consistent. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Send a um phone voice message ah let me think to my my roommate. I have taken out the trash. Make that to my all roommates, not my roommate." }, "candidate_id": "wav_e055103b1585a409", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision of the recipient ('my roommate' to 'all roommates') after a brief hesitation. The speaker uses 'uh' and 'let me think' naturally, and the correction is delivered with a plausible change-of-mind intonation. However, the initial request is slightly less fluent than audio_01, with a minor stumble ('send a um phone voice message') and a slightly more abrupt transition to the correction. The overall delivery is realistic but less fluid than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_message_type", "proposed_instruction": "Send a text message to my all roommates, \"I have taken out the trash.\". Wait, send a text message, not a phone voice message.", "long_pause_offsets": [ 72 ], "source_sha256": "2bf9facecabd285db7ff7b1caad0306f8332dce4e01ddb1cf0a92acc4524e3cd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the message type from 'phone voice message' to 'text message', while retaining the recipient group and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/revision/revision_01_message_type/41c690752af02beb/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "e0bb52c2b98b12e12f065b853c8cfc51f40aeb226a24ed7c58434e6da8828430", "seconds": 13.24, "spoken_utterances": [ "Send a uh... text message to my all roommates, \"I have taken out the trash.\". ", "Wait, send a um... text message, not a ... phone voice message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a text message', then corrects themselves to 'Send a um text message, not a phone voice message'. This is a revision where the final instruction clearly specifies the correct action (text message) and explicitly excludes the incorrect one (phone voice message). The final intent matches the expected action and the original instruction's intent to communicate with roommates about the trash, with the correct medium specified.", "heard_transcript": "Send a text message to my all roommates, I have taken out the trash. Wait, send a um text message, not a phone voice message." }, "candidate_id": "wav_6360804f658785e1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision of the message type ('phone voice message' to 'text message') after a pause. While the 'wait' and 'um' are natural, the correction is delivered in a separate, slightly disjointed utterance rather than being integrated into the original request. The prosody is flatter and the pacing feels more like a correction than a spontaneous thought process, making it sound less like a single, continuous thought. The delivery is still realistic but less cohesive than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "0d69a6071eb9e34a3da59ebfbe33efb3b59591e31111b493abe7887d6b7c75cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/cancellation/cancel_004/ba78dbf65dad65a2/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "77a4071d0de2a242ed62e19f6635bfbea5ee46b9b7ebbbefc953d70bb8a7ec79", "seconds": 12.56, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\".", "Oh, wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task but immediately self-corrects with 'Oh wait, cancel that request,' which is a clear and final cancellation of the entire task. The final interjection is a direct cancellation command, satisfying the 'cancellation' type criterion.", "heard_transcript": "Send a um phone voice message to my all roommates uh, I have taken out the trash. Oh wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "4e46eb4a072743fd2682b617824d0eca103a6482a242e9ba6edf220a18c1b8bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/ambiguous_halt/halt_030/2a169f1cb7b354bf/speech.wav", "wer": 0.05, "wav_sha256": "e21e7c5b2f2699b2c03ddec04a88d750642685dca955878a87c9e00a09fc890a", "seconds": 11.88, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\".", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause and reconsider the task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt, not a cancellation or execution. The initial instruction is not completed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Send a um phone voice message to my all roommates uh, I have taken out the trash. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "43648a9b592a39fe41c97f9c9a7db340722180d8ec198eba649ad0a4a0e85345", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_1/audio_natural/filler/original/519ca6a1ab5c96e0/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8fda36eac763ff788bb06fecb7b81484770f4c9036bd3b762b4c891a74387cda", "seconds": 8.28, "spoken_utterances": [ "Send a um... phone voice message ... to my all roommates, ah... \"I have taken out the trash.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core task. The final intent is clear and matches the original instruction. The filler words do not alter the goal, so the execution should proceed as intended.", "heard_transcript": "Send a um phone voice message to my all roommates, uh, I have taken out the trash." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_message_type", "proposed_instruction": "Send a email—wait, I meant phone text message to my partner, \"The dishwasher is clean and ready to be emptied.\".", "long_pause_offsets": [], "source_sha256": "7ef5338c9f712687b6a100d3b110f47f45fc44156788c3709d57df26bac4b5b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email' to 'phone text message' within the same utterance, preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/false_start/false_start_01_message_type/c1a108eb7e8140db/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "58b5207d03d0298e0c5ba9393933311df422c5e96b9980fa7c38eac2d69519e3", "seconds": 12.88, "spoken_utterances": [ "Send a email—um... wait, I meant ... phone text message to my let's see... partner, \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a email' but immediately self-corrects with 'wait, I meant, phone text message'. This is a clear false start where the final audible intent matches the original instruction. The target ('my partner') and message content are preserved and correctly stated.", "heard_transcript": "Send a email, um, wait, I meant, phone text message to my let's see, partner, the dishwasher is clean and ready to be emptied." }, "candidate_id": "wav_1b7da480695923f1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex pattern of spontaneous speech. It begins with a false start ('email'), followed by a self-correction ('wait, I meant'), a filler ('let's see'), and then a complete, coherent request. The pauses and fillers are well-placed and sound like genuine hesitation during formulation. The correction is immediate and integrated smoothly into the flow, making it the most realistic example of a natural, unscripted task request." } ] }, "is_default": true }, { "variant_id": "false_start_02_recipient_reference", "proposed_instruction": "Send a phone text message to my mom—no, I meant partner, \"The dishwasher is clean and ready to be emptied.\".", "long_pause_offsets": [], "source_sha256": "f5afdb42c4e6514e389c921488474504b1c72dfcae7441100671e9681a441192", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'mom' to 'partner' within the same utterance, preserving the app (phone text), action (send), and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/false_start/false_start_02_recipient_reference/8c565b98c7feb0fc/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "65c94d3ae77e8f8d8de0426b756857ccd926964fbf82828758b46587d8eef654", "seconds": 11.92, "spoken_utterances": [ "Send a phone text message to my uh... mom—um... no, I meant partner, ... \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mom' but immediately self-corrects to 'partner' with 'no, I meant partner'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The message content is unchanged and fully audible. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Send a phone text message to my uh mom um no, I meant partner. The dishwasher is clean and ready to be emptied." }, "candidate_id": "wav_42946b1e038b5625", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start ('mom') and a self-correction ('no, I meant partner'), which is a credible pattern. However, the initial hesitation ('uh... mm...') is slightly more prolonged and less fluid than in audio_03. The correction is direct but the overall flow feels a bit more hesitant and less integrated, making it less natural than audio_03 despite having a similar structure." } ] }, "is_default": false }, { "variant_id": "false_start_03_message_text", "proposed_instruction": "Send a phone text message to my partner, \"The dishwasher is dirty and needs to be cleaned.\"—scratch that, make it \"The dishwasher is clean and ready to be emptied.\".", "long_pause_offsets": [], "source_sha256": "9ea63cd19b5eb0efe9d6c24e56406c609d9f721adf8117ef2674dd9694dfc82d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces the initial message content with the final intended message; the recipient and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/false_start/false_start_03_message_text/97a6f63fba0720c9/speech.wav", "wer": 0.0625, "wav_sha256": "ab66e12d7e8e5b778e4602a92f831f1bc48c906f0d37ec3d1ece4401a7ab005d", "seconds": 15.44, "spoken_utterances": [ "Send a um... phone text message to my partner, \"The dishwasher is dirty and needs to be cleaned.\"—uh... scratch that, make it ... \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says the dishwasher is dirty and needs cleaning, but immediately self-corrects with 'scratch that' and provides the correct message. According to the false_start type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final message matches the original instruction exactly.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is dirty and needs to be cleaned. uh, scratch that, make it, the dishwasher is clean and ready to be emptied." }, "candidate_id": "wav_559db479c456da43", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision rather than a false start. The speaker first says one message, then corrects it. However, the correction ('uh, scratch that...') is delivered with a slightly unnatural, almost robotic intonation, and the phrase 'scratch that' sounds more like a scripted phrase than a spontaneous thought. The overall delivery is flatter and less conversational compared to the other two, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Send a phone text message to, \"The dishwasher is clean and ready to be emptied.\". Oh, and one more thing: send it to my partner.", "long_pause_offsets": [ 82 ], "source_sha256": "97c079a88fab94036921485a5173cd94d0ed20273fee3ebc2bc0088502589f7b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient 'my partner' to the initial message, which already contained the correct message content. The final goal matches the original instruction in both action and content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/addition/addition_02/95c8f8930ea588c6/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "8579c4aaa69b8abf20d12bcb1595f10286a97af991373dfaddaa162f3201cefb", "seconds": 16.28, "spoken_utterances": [ "Send a uh, let me think... phone text message to, uh... \"The dishwasher is clean and ready to be emptied.\". ", "Oh, and one more thing: send it to ... my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task without specifying the recipient, then adds 'send it to my partner' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "Send a uh let me think phone text message to uh the dishwasher is clean and ready to be emptied. Oh and one more thing, send it to my partner." }, "candidate_id": "wav_c56def2b2bc49396", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses a natural, spontaneous 'let me think' pause, followed by a coherent initial request. The addition of 'Oh, and one more thing' is a believable conversational device, and the correction 'send it to my partner' flows smoothly without awkwardness. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Send a to my partner, \"The dishwasher is clean and ready to be emptied.\". I should also mention, send it as a phone text message.", "long_pause_offsets": [ 74 ], "source_sha256": "c74ca946f5f1a195e83f80a8af7f3f999e6e52116f9ea40cdd3f09eb874c24d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The initial utterance is missing the recipient identifier 'my partner' and the message content is incomplete ('Send a to my partner...'). The followup adds the message type but does not restore the missing recipient or complete the message. The final goal is incomplete and ambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/addition/addition_01/100f451d3b890ece/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "bcd6e22bcb43ce6f02c954f78e4e9633da7415b03fc7011ce9bed940c03c615e", "seconds": 15.12, "spoken_utterances": [ "Send a to my let me think... partner, \"The dishwasher is clean and ready to be emptied.\". ", "I should also mention, send it as a um, let's see... phone text message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete and contains hesitation ('Send it to my let me think partner'), but the speaker immediately adds the missing context ('send it as a phone text message') after a pause. The final intent combines the original request with the added clarification, matching the reference instruction exactly. The hesitation and self-correction are typical of natural speech and do not alter the final goal.", "heard_transcript": "Send it to my let me think partner, the dishwasher is clean and ready to be emptied. I should also mention, send it as a um let's see phone text message." }, "candidate_id": "wav_6626c5ca620b5457", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also uses 'let me think' and 'let's see', the delivery is less fluid. There's a noticeable stutter ('send it to my let me think') and a second, more hesitant 'send it as a um let's see' that sounds like a forced or awkward self-correction. The prosody is less natural, and the overall flow is disrupted by these less convincing repair attempts." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_message_type", "proposed_instruction": "Send a email to my partner, \"The dishwasher is clean and ready to be emptied.\". Make that a phone text message, not an email.", "long_pause_offsets": [ 80 ], "source_sha256": "b74f7a4fc9bd5d19a5528edfc96581d6ab936f688f30e621238f23c6fcf4ef76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type from email to phone text message, while preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/revision/revision_01_message_type/afca670853ce8e52/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "b559758c908f4f7169bf4e439df34db75aa3f630a0e1f2b57adfd919be5e3ba8", "seconds": 19.96, "spoken_utterances": [ "Send a um, let me think... email to my um, let's see... partner, \"The dishwasher is clean and ready to be emptied.\". ", "Make that a uh, let me think... phone text message, not an email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately corrects to 'phone text message' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to send a text message, which matches the original instruction's content and target, only differing in the communication method. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Send a um, let me think, email to my um, let's see, partner, the dishwasher is clean and ready to be emptied. Make that a uh, let me think, phone text message, not an email." }, "candidate_id": "wav_20f95ab9e67692e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a highly natural, multi-stage thought process with realistic fillers ('um', 'let me think', 'let's see') and a clear, spontaneous self-correction from 'email' to 'phone text message'. The prosody includes believable hesitations and a smooth repair that maintains conversational flow, making the revision sound like a genuine moment of reconsideration." } ] }, "is_default": true }, { "variant_id": "revision_03_message_text", "proposed_instruction": "Send a phone text message to my partner, \"The dishwasher is dirty and needs to be cleaned.\". Change the message to \"The dishwasher is clean and ready to be emptied.\" instead of \"The dishwasher is dirty and needs to be cleaned.\"", "long_pause_offsets": [ 93 ], "source_sha256": "e16ba1008f131894524f6fe29d2069f9a1f7c6a909e5bc2893314015f12193c9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly replaces the message content with the desired text, overriding the initial incorrect message. The recipient and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/revision/revision_03_message_text/b9e45f5f48f66d97/speech.wav", "wer": 0.0, "wav_sha256": "3aa0155e9c28194b9e04daa8a13c66d7ad7662bf4475c1f75a742d064361bb96", "seconds": 19.76, "spoken_utterances": [ "Send a phone text message to my partner, uh... \"The dishwasher is dirty and needs to be cleaned.\". ", "Change the message to \"... The dishwasher is clean and ready to be emptied.\" instead of \"The dishwasher is dirty and needs to be cleaned.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says the wrong message, but then clearly revises it by stating 'Change the message to, the dishwasher is clean and ready to be emptied, instead of the dishwasher is dirty and needs to be cleaned.' This revision is spoken after a pause and directly corrects the earlier instruction. The final intent matches the original reference instruction, so the revision is correctly applied.", "heard_transcript": "Send a phone text message to my partner, uh, the dishwasher is dirty and needs to be cleaned. Change the message to, the dishwasher is clean and ready to be emptied, instead of the dishwasher is dirty and needs to be cleaned." }, "candidate_id": "wav_0b1392d76a6dc88e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker delivers a coherent revision, the initial request is stated fluently without hesitation, and the revision is delivered as a flat, declarative instruction rather than a natural follow-up. The lack of fillers or hesitation in the initial part makes the revision feel less spontaneous and more like a rehearsed or read instruction, reducing its realism compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_02_recipient_reference", "proposed_instruction": "Send a phone text message to my mom, \"The dishwasher is clean and ready to be emptied.\". Wait, send it to my partner, not my mom.", "long_pause_offsets": [ 89 ], "source_sha256": "616d9c49b3aea5c0aa069d7928f1c89e6697ea1c2cf11afebd437289ee60f188", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the recipient from 'mom' to 'partner' and explicitly excludes 'mom', preserving the app, message content, and action." }, "audio": { "status": "pending", "spoken_utterances": [ "Send a uh... phone text message to my ah... mom, \"The dishwasher is clean and ready to be emptied.\". ", "Wait, send it to my hmmm... partner, not my mom." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "428654040ccde3a36f19f390fd41c3f98e82b7e6b5598364bf8194a2039e8a46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/cancellation/cancel_010/6fcc61f4a27e78ab/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "ff670a38129d0c86be86d8004d7af80ec7909ed93f5c20a0a79902dfaf674ab8", "seconds": 11.4, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\".", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to send a text message, then immediately follows with 'Oh wait, cancel the request I just made.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "14626d1a293a5a53c7da6c4ce08abae478dd8bf7a966a43b29e98af800bc5710", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/ambiguous_halt/halt_018/37a8ab75df124c2d/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "000ad3a6b0af443f7415f0f7b5857d5fcb33a98fd13bfd843c503da964649445", "seconds": 11.04, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\".", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which is a clear instruction to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The original task is not canceled, and the agent is being asked to pause, not execute or cancel. Therefore, the final intent satisfies the type-specific criterion.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "23c0efcb55f2ceb36000e82099fd5f35dc0807cdc00fa050d601ebc62cd3dfc6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_2/audio_natural/filler/original/ee97e9a1606ce438/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "dc5f9a737879bc8a3f7af159e0f685a6d4c3a8f8f4b2c692cf18750db8d35955", "seconds": 6.88, "spoken_utterances": [ "Send a um... phone text message to my ... partner, \"The dishwasher is clean and ready to be emptied.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('um') but completes the original instruction without any correction, cancellation, or pause. The final intent matches the original instruction, and the filler word does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Send a um phone text message to my partner, the dishwasher is clean and ready to be emptied." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my wife—no, I mean husband, \"Don't forget to water the plants today.\".", "long_pause_offsets": [], "source_sha256": "7118ad530d60c68b4ddc20aad27b647742fae60b4a6370ecc2eaadd5df4558a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'wife' with 'husband'; the action, message content, and delivery method remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/false_start/false_start_02_recipient_reference/45d02cb2754bb0df/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "99f21992a76c5ae976e2d27fef9b31121cb1f28ea16d08a85ad545b313073e73", "seconds": 10.12, "spoken_utterances": [ "Send a phone voice message to my ah... wife—uh... no, I mean husband, \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wife' but immediately self-corrects to 'husband'. According to the 'false_start' type criterion, the final corrected target should be applied, and the original instruction is executed with the correction. The message content remains unchanged and is fully audible.", "heard_transcript": "Send a phone voice message to my uh wife uh no, I mean husband. Don't forget to water the plants today." }, "candidate_id": "wav_2c1da62d7f6f58cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a very natural false start with 'my uh wife uh' followed by a clear self-correction to 'husband'. The intonation and pacing of the correction sound spontaneous and unforced. The delivery is fluent and continuous, with no awkward pauses or unnatural phrasing, making it the most realistic-sounding false start." } ] }, "is_default": true }, { "variant_id": "false_start_03_message_text", "proposed_instruction": "Send a phone voice message to my husband, \"Don't forget to feed the dog today.\"—scratch that, make it \"Don't forget to water the plants today.\".", "long_pause_offsets": [], "source_sha256": "163effe5f025080933444f67df933a9d549b1b97df01be2454af485936f3dfe1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests a message about feeding the dog, but immediately corrects it to the original message about watering the plants. The final message content is identical to the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/false_start/false_start_03_message_text/6f7015c507fbc5cf/speech.wav", "wer": 0.0625, "wav_sha256": "7e2df0679b727fee39a0aaa1bb948365868918fbcf1a74bafead33499f0b61ab", "seconds": 20.8, "spoken_utterances": [ "Send a phone voice message to um, let me think... my husband, \"Don't forget to feed the dog today.\"—ah... scratch that, make it hmmm... \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'feed the dog' but immediately self-corrects with 'scratch that' and provides the correct message 'water the plants'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final message matches the original instruction exactly.", "heard_transcript": "Send a phone voice message to um, let me think, my husband, don't forget to feed the dog today. Ah, scratch that, make it, don't forget to water the plants today." }, "candidate_id": "wav_5bf7d22dd6b894bc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a plausible false start ('my husband') and a self-correction ('scratch that, make it...'). However, the correction is less fluid than audio_02, with a noticeable pause and a slightly hesitant, less natural-sounding rephrasing. The initial hesitation ('um... let me think') is realistic, but the overall flow feels slightly more disjointed." } ] }, "is_default": false }, { "variant_id": "false_start_01_message_type", "proposed_instruction": "Send a text message—I should say phone voice message to my husband, \"Don't forget to water the plants today.\".", "long_pause_offsets": [], "source_sha256": "281e6460e15a9de8fae06782b50387e3e5bd6445c291ff9e9094c78f71f258d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'text message' to 'phone voice message' within the same utterance, preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/false_start/false_start_01_message_type/bffdc5e1126d8007/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "feed327255a1e7573e55230c1ecd7f275ac634ccceeb7640f7a16766692d4e26", "seconds": 11.56, "spoken_utterances": [ "Send a ... text message—hmmm... I should say phone voice message to my husband, ... \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'text message' but immediately self-corrects to 'phone voice message'. According to the 'false_start' type criterion, the final corrected intent should be executed. The target ('my husband') and message content are preserved and correctly stated after the correction.", "heard_transcript": "Send a text message, hum, I should say phone voice message to my husband. Don't forget to water the plants today." }, "candidate_id": "wav_fce8f5649e250004", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip has a false start ('send a text message') and a self-correction ('I should say phone voice message'). While the correction is present, the phrasing 'I should say' sounds slightly more rehearsed or less spontaneous than the other two. The delivery is also slightly more hesitant and less fluid, with a less natural transition between the initial error and the correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Send a phone voice message to my husband,. Oh, I forgot to say, include the message \"Don't forget to water the plants today.\"", "long_pause_offsets": [ 43 ], "source_sha256": "fd9b44cbad99ff5286a79066fbf87bbe3694498700d3fa631c2ac0c5ef178c24", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing message content, which was omitted in the initial utterance. The recipient and message type are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/addition/addition_03/85fe742763039803/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "6807cf7db0fe50793d7e16f6f76298b8c96f44c36aedf27653da5e585bf9a450", "seconds": 12.4, "spoken_utterances": [ "Send a ... phone voice message to my husband,. ", "Oh, I forgot to say, include the message ... \"Don't forget to water the plants today.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to send a voice message to the husband is followed by an addition that specifies the message content. The final audible intent combines both the action and the message, matching the original instruction. The addition is clear and directly specifies the missing content.", "heard_transcript": "Send a phone voice message to my husband. Oh, I forgot to say, include the message, don't forget to water the plants today." }, "candidate_id": "wav_6a0b4e27ee3d30b9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and a believable, smoothly integrated correction. The intonation and pacing are consistent with spontaneous speech, and the addition of 'include the message' feels like a genuine afterthought rather than a forced insertion. The prosody is fluid and the delivery is continuous without awkward breaks." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send a phone voice message to, \"Don't forget to water the plants today.\". I left out a detail: send it to my husband.", "long_pause_offsets": [ 74 ], "source_sha256": "5a9142f68361b0316044f69163ee847c7cce7bf86148bf80914c5426cd2f2d19", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient, preserving the message content and action. The initial utterance omits the recipient but includes the message content, which is retained in the final goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/addition/addition_02/de189f75b8247b2d/speech.wav", "wer": 0.0, "wav_sha256": "692a2345b86aff2c54a7c88f4810b4ec4ccedacec04f4daf97a2eb28b265877c", "seconds": 14.96, "spoken_utterances": [ "Send a um... phone voice message to, um, let's see... \"Don't forget to water the plants today.\". ", "I left out a detail: send it to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, missing the recipient. After a pause, the speaker adds the missing recipient 'my husband'. This matches the addition type criterion: the original instruction is combined with the added information to form the complete final goal, which matches the reference instruction.", "heard_transcript": "Send a um phone voice message to um let's see, don't forget to water the plants today. I left out a detail. Send it to my husband." }, "candidate_id": "wav_d59764feb4344df8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a hesitant, natural-sounding request ('send a... phone voice message to... let's see...') and then adds a relevant detail ('Send it to my husband'). The pauses and fillers ('um', 'let's see') are well-placed and sound like genuine hesitation. The correction is integrated into the flow, though the final sentence is slightly less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Send a to my husband, \"Don't forget to water the plants today.\". Oh, I forgot to mention, send a phone voice message.", "long_pause_offsets": [ 65 ], "source_sha256": "61525e461c4c15a15e4e08f617d783e12ffe6f477db948e359c2aaa2cf0607aa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing message type 'phone voice message' to the initial incomplete request, completing the full instruction without contradicting any other part." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/addition/addition_01/007daee8c95a93f9/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "4e29e93b669b3dbb71df84f6a9106f6f76378427229ad11be94198d7066c9086", "seconds": 13.6, "spoken_utterances": [ "Send a to uh... my husband, \"Don't forget to water the plants today.\". ", "Oh, I forgot to mention, send a ... phone voice message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a partial instruction to send a message to her husband with the content. After a pause and an interjection, she adds the missing detail that it should be a 'phone voice message'. The final instruction combines the original message content with the new delivery method, matching the reference. The addition is clear and final, satisfying the addition type criterion.", "heard_transcript": "Send a to uh my husband, don't forget to water the plants today. Oh, I forgot to mention, send a phone voice message." }, "candidate_id": "wav_85856b89d52e1863", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a slightly awkward phrasing ('Send a to... my husband') and then adds the message. The second part ('Oh, I forgot to mention...') is a clear revision, but the delivery is less natural due to the abrupt shift and the repetition of 'send a phone voice message' at the end. The prosody is less continuous, and the overall flow feels more like a scripted correction than spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_message_text", "proposed_instruction": "Send a phone voice message to my husband, \"Don't forget to feed the dog today.\". Change the message to \"Don't forget to water the plants today.\" instead of \"Don't forget to feed the dog today.\"", "long_pause_offsets": [ 81 ], "source_sha256": "738ea20497b436d675f8fa0abcc241f16a2fedafa87cf30bf14ec6b3c2114efe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the message content from 'feed the dog' to 'water the plants', preserving the recipient and message type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/revision/revision_03_message_text/3379aa9ac189b9e0/speech.wav", "wer": 0.0, "wav_sha256": "f0d95bbc507aaf4745e77a67184b0058836b8955158bd013d3995e3d84c5f31f", "seconds": 17.4, "spoken_utterances": [ "Send a phone voice message to ... my husband, \"Don't forget to feed the dog today.\". ", "Change the message to uh... \"Don't forget to water the plants today.\" instead of \"Don't forget to feed the dog today.\"" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the incorrect message 'Don't forget to feed the dog today.' but then clearly revises it to 'Don't forget to water the plants today.' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction, and the revision is clear and unambiguous.", "heard_transcript": "Send a phone voice message to my husband, \"Don't forget to feed the dog today.\". Change the message to, uh, \"Don't forget to water the plants today\" instead of, \"Don't forget to feed the dog today.\"." }, "candidate_id": "wav_a046a6d03d28ff39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a clear, natural revision. The speaker first states a complete request, then smoothly transitions with 'Change the message to...' to correct the content. The prosody is consistent, the pace is conversational, and the correction is integrated into the flow without awkwardness. The self-correction is pragmatic and sounds like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_recipient_reference", "proposed_instruction": "Send a phone voice message to my wife, \"Don't forget to water the plants today.\". I misspoke about the recipient: send it to my husband, not my wife.", "long_pause_offsets": [ 82 ], "source_sha256": "a53dcaec44d897d41deeff7540084ebb4295ff9c55266a2be5337a4b98aadb78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recipient from 'my wife' to 'my husband' while retaining the message content and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/revision/revision_02_recipient_reference/a6c590c3169be072/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "2483b29ce64700c7c6f661d3d01c87cc036b8ccb33aee449adef40b09db3d774", "seconds": 14.2, "spoken_utterances": [ "Send a uh, let me think... phone voice message to my wife, um... \"Don't forget to water the plants today.\". ", "I misspoke about the recipient: send it to my husband, not my wife." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the recipient as 'my wife' but immediately corrects this by stating, 'I misspoke about the recipient. Send it to my husband, not my wife.' This constitutes a revision where the final intent is to send the message to 'my husband'. The correction is clear and unambiguous, and the rest of the message content and action remain consistent with the original instruction. Therefore, the final intent matches the expected goal for a revision type.", "heard_transcript": "Send a uh, let me think, phone voice message to my wife, um, don't forget to water the plants today. I misspoke about the recipient. Send it to my husband, not my wife." }, "candidate_id": "wav_f6133e6b6075bdef", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision where the speaker first makes a mistake ('my wife') and then corrects it. The 'I misspoke' self-awareness is natural, and the pause before the correction feels genuine. However, the initial hesitation ('uh, let me think') is slightly more pronounced and the correction is delivered with a slightly flatter intonation compared to audio_03, making it feel a bit less fluid." } ] }, "is_default": false }, { "variant_id": "revision_01_message_type", "proposed_instruction": "Send a text message to my husband, \"Don't forget to water the plants today.\". Wait, send a phone voice message, not a text message.", "long_pause_offsets": [ 78 ], "source_sha256": "08f14df252d28642ff7fff029c25673c03a4ba5d9227e6453901decfc800fbbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type from text to phone voice while preserving the recipient and message content." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/revision/revision_01_message_type/0119916ddf0837bf/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "881fc14ec2c1b09f89fdfd32f70c181fc582000f2d0f0411ffea26158ba20ae4", "seconds": 15.32, "spoken_utterances": [ "Send a text message to uh, let's see... my husband, ... \"Don't forget to water the plants today.\". ", "Wait, send a phone voice message, not a text message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send a text message' but immediately corrects themselves with 'Wait, send a phone voice message, not a text message.' This is a clear revision where the final instruction overrides the initial incorrect one. The final intent matches the original reference instruction, so the task should be executed as revised.", "heard_transcript": "Send a text message to uh let's see my husband. Don't forget to water the plants today. Wait, send a phone voice message, not a text message." }, "candidate_id": "wav_392b9e13f2737495", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a false start and self-correction. The speaker begins with 'send a text message' and then corrects to 'send a phone voice message'. While the correction is present, the initial phrase is spoken with a slightly robotic, rehearsed quality, and the correction feels abrupt. The 'wait' pause is short and doesn't convey the same sense of natural reconsideration as in the other clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "47a3b25d725c5f6271efa64050243d07d0186ccb05d0d8fc40856a0b33ddefb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/cancellation/cancel_038/2c9be1b9bb45355e/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "1070001a09a45007ec46261f340352fdb25c8a1f3b8a503cf4d852fe47db6adf", "seconds": 14.48, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\".", "Wait, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task by stating the message content, but immediately follows with 'Wait, please stop and cancel that request.' This constitutes a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. Therefore, the final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Send a a phone voice message to my husband, uh let's see. Don't forget to water the plants today. Wait, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "da1964ff7e500c9968ae999e53780e5db22f3146905953515280afd5a1677dc6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/ambiguous_halt/halt_011/7aac648b98433b6f/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "2e754bae2747375184599fcaa77180bbc4fd2480ada357fe3280467595a17acc", "seconds": 12.8, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\".", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full instruction to send a voice message with the correct content and recipient, then immediately says 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The original task is fully stated and not revoked.", "heard_transcript": "Send a uh phone voice message to my husband uh let's see, don't forget to water the plants today. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "13547f5_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f948e9caea7e2e480f0896465a77063e317c7eb5f40ec9aebbe441a75431a4bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/13547f5_3/audio_natural/filler/original/6c6143178acd665c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "b8a2dd6c787b5be0287b1605294862e7c7037a4486f74651a38357f414a798cb", "seconds": 9.2, "spoken_utterances": [ "Send a uh... phone voice message to my husband, uh, let's see... \"Don't forget to water the plants today.\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core task. The final instruction matches the original request in action, target, and message content. The fillers do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Send a uh phone voice message to my husband uh let's see, don't forget to water the plants today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Etsy—I should say amazon order as per my husband's email today.", "long_pause_offsets": [], "source_sha256": "a6f10a55a70960f4ca118c7f98f62b2ff8d72c5de7d653bff9bf5a2fd1075288", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon' within the same utterance, preserving the action, the source of the order (husband's email), and the time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/false_start/false_start_01_ordering_app/e6cb915cd808cdbe/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "29ee06fc95cb7b3d3a9708144bc6d1ed6f17df1788349aaecad7b86d8b773f16", "seconds": 8.56, "spoken_utterances": [ "Place an Etsy—ah... I should say ... amazon order as per my husband's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon' with 'I should say'. This is a classic false_start where the final, corrected value is the intended goal. The rest of the instruction matches the original, so the final intent is to execute the Amazon order task.", "heard_transcript": "Place an Etsy uh, I should say, Amazon order as per my husband's email today." }, "candidate_id": "wav_7150d4efd76e7955", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural-sounding false start and self-correction. The speaker begins with 'Etsy', hesitates with 'uh', then clearly self-corrects to 'I should say Amazon'. The correction is smooth, the prosody is consistent, and the pacing feels like a genuine moment of recall. The final sentence is delivered fluently without further hesitation, making the entire utterance sound spontaneous and realistic." } ] }, "is_default": true }, { "variant_id": "false_start_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my wife's—no, I meant husband's email today.", "long_pause_offsets": [], "source_sha256": "293cd980f0b96c3bfeab44e2678b57ff32cd115772dca286d40d1b97094c4383", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'wife's' to 'husband's' within the same utterance, preserving the app (Amazon), action (place an order), and time condition ('today')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/false_start/false_start_02_email_sender_relation/ecf62b9536f7bee2/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "a5e1c341856b0fc345e4d80fd7591fd422dabe4e25fb432760c42c90ab0cbe59", "seconds": 8.64, "spoken_utterances": [ "Place an amazon order as per my wife's—um... no, I meant husband's email ... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wife's' but immediately self-corrects to 'husband's' with 'no, I meant'. This is a false start where the final corrected value overrides the initial mistake. The rest of the instruction, including 'Amazon order' and 'today', remains unchanged and is executed as intended.", "heard_transcript": "Place an Amazon order as per my wife's um, no, I meant husband's email today." }, "candidate_id": "wav_9b02968d4acc881a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a plausible self-correction from 'wife's' to 'husband's', but the correction is less fluid than audio_03. The speaker says 'No, I meant husband's email' after a brief pause, which is a believable change of mind. However, the initial part of the sentence is slightly rushed, and the correction feels a bit more like a deliberate switch than a natural thought process, making it less realistic than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Place an amazon order as per my husband's email tomorrow—wait, I meant today.", "long_pause_offsets": [], "source_sha256": "b507842f4a2f7aec9c904292f89cad72487dafddf6c98c20513261d8379dae44", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects to 'today'. The final intended time is 'today', matching the original instruction. All other elements (action, app, reference source) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/false_start/false_start_03_time_reference/2b3b8802f7f95cdd/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "095b6151650a5471ade740c2fd490601fba31a44b4f9b9cd5b874e448ad0a4fd", "seconds": 13.96, "spoken_utterances": [ "Place an uh, let's see... amazon order as per my husband's email tomorrow—ah... wait, I meant um... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately self-corrects to 'today' after a pause and 'wait'. According to the 'false_start' type criterion, the final corrected value ('today') overrides the initial mistaken value ('tomorrow'), and the rest of the instruction remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Place an uh let's see Amazon order as per my husband's email tomorrow. Uh wait, I meant um today." }, "candidate_id": "wav_d49eecd0f7440b32", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'uh' and 'let's see', followed by a correction from 'tomorrow' to 'today'. While the correction is present, the initial hesitation is more fragmented and the correction is less smoothly integrated into the sentence. The speaker's delivery is slightly more disjointed, and the overall flow feels less natural compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an amazon order as per my husband's email. Oh, I forgot to mention, do this today.", "long_pause_offsets": [ 49 ], "source_sha256": "978075796d1a5068dc4df5912f37cf87067f21eba9170ae6ccac00d34d7779cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'today' to the initial instruction, which is a missing condition in the original instruction. The final goal combines both the source (husband's email) and the timing requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/addition/addition_01/d206cd8efef1dda5/speech.wav", "wer": 0.0, "wav_sha256": "9026599366fbde6e4762171b145b5fe77f0d18d8b6895528a8fe9ffc38cff9bb", "seconds": 11.56, "spoken_utterances": [ "Place an ... amazon order as per my husband's email. ", "Oh, I forgot to mention, do this um... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to place an Amazon order as per the husband's email is followed by an addition specifying 'today' as a time constraint. The addition is clear and directly modifies the original task by adding a new condition. The final intent is a combined version of the original instruction and the added constraint, which matches the expected behavior for an 'addition' type. The task is not cancelled or paused, and the final control intent is to execute the combined task.", "heard_transcript": "Place an Amazon order as per my husband's email. Oh, I forgot to mention, do this um today." }, "candidate_id": "wav_ad681d43f3ee505c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, coherent initial request with natural prosody and a believable, smoothly integrated addition. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to introduce a follow-up detail, and the subsequent 'do this um today' is a natural, hesitant continuation that fits the addition pattern. The pacing and intonation suggest spontaneous thought rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an amazon order as per today. I left out a detail: use my husband's email.", "long_pause_offsets": [ 36 ], "source_sha256": "5cd11c61d67e6b962e668349e9874f9a806feb0bcb735a26c67e2d16d1078aec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to use the husband's email, which was omitted in the initial instruction. The core action and time frame remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/addition/addition_02/e602980d758c5db3/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "e71e6f765ff85af8c7baffc0f18dc5b4858e003e7b2636faa1a847f89efe3c8f", "seconds": 12.68, "spoken_utterances": [ "Place an ah... amazon order as per hmmm... today. ", "I left out a detail: use um... my husband's email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task and then adds a clarification to use 'my husband's email' to correct an omission. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and matches the original instruction, with the addition serving to specify the email source. No contradictions or cancellations are present.", "heard_transcript": "Place an Amazon order as per my husband's email today. I left out a detail, use um, my husband's email." }, "candidate_id": "wav_8478a1e1447701e7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is somewhat unclear and mumbled ('place an uh Amazon order as per man today'), with a noticeable pause and filler ('um') before the addition. The addition 'I left out a detail use um my husband's email' is awkwardly phrased and delivered with a flat, hesitant tone that sounds less spontaneous and more like a forced or rehearsed correction. The overall delivery is less fluent and natural compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my father's email today. Wait, the email should be from my husband's, not father's.", "long_pause_offsets": [ 54 ], "source_sha256": "b52052d945476f03ef5fc3b840b34a08a401f7596a38e872565ad81b17a53f1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the sender relationship from 'father' to 'husband', preserving the app, action, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/revision/revision_02_email_sender_relation/6f47d88b792373ce/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "6719895bd06c758599c46deb68c2c66e0fee7630634cd33c2642b3221ec9fee7", "seconds": 16.32, "spoken_utterances": [ "Place an amazon order as per hmmm... my father's email let me think... today. ", "Wait, the email should be from my ... husband's, not father's." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong email source ('father's'), but immediately corrects it to 'husband's' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the action, target, and time constraint.", "heard_transcript": "Place an Amazon order as per my father's email. Let me think. Today. Wait, the email should be from my husband's, not father's." }, "candidate_id": "wav_4f515bd0eb5743c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural revision process. The initial request is interrupted by a thoughtful 'um' and 'let me think,' followed by a self-correction ('wait, the email should be from my husband's not father's') that is smoothly integrated into the flow. The intonation and pacing reflect genuine hesitation and recollection, with the correction feeling like a spontaneous realization rather than a rehearsed edit. The prosody is consistent and the pauses are appropriately placed for a person formulating a request." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Place an amazon order as per my husband's email tomorrow. That should be today, not tomorrow.", "long_pause_offsets": [ 58 ], "source_sha256": "f960d4c1b4280df007f5b2a1ede69072ee3ec0a9db294e47a337dde88e0718ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'tomorrow' to 'today', preserving the action, app, and source of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/revision/revision_03_time_reference/2712e5dbf8212017/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "e990e92010467aeb60d4c64c3b17c7fa1bd8e24d5a7505f4c3bfeafde331a3bc", "seconds": 12.48, "spoken_utterances": [ "Place an um... amazon order as per uh... my husband's email tomorrow. ", "That should be um... today, not tomorrow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects to 'today' after a pause and hesitation. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final audible intent is to place the order for 'today', which matches the original instruction. All other conditions are preserved.", "heard_transcript": "Place an Amazon order as per my husband's email tomorrow. That should be um, today, not tomorrow." }, "candidate_id": "wav_468b81cf151e131c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a clear self-correction ('that should be today not tomorrow'). The speaker uses 'um' fillers naturally and the correction is delivered with a slight shift in intonation, indicating a change of mind. The pacing is slightly faster than audio_01, but the overall delivery remains realistic and the repair is integrated into the utterance without sounding forced. The only minor difference is the slightly more rapid delivery, which makes it feel slightly less contemplative than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an Walmart order as per my husband's email today. Make that Amazon, not Walmart.", "long_pause_offsets": [ 56 ], "source_sha256": "6683cd720b4d943398a9e4f4afbdebb8f42e72d7b3214afa29c3183437831f79", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, preserving the action, source (husband's email), and time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/revision/revision_01_ordering_app/aacd171fa304d13b/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "ce0c3b1bb37429435206cd9ac426dfa0bbd306a5c2a982f117b6f905a2fe14da", "seconds": 13.4, "spoken_utterances": [ "Place an uh... Walmart order as per my husband's email ... today. ", "Make that let me think... Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately corrects to 'Amazon' after a pause and self-correction ('Make that, let me think'). This is a clear revision where the final audible intent is to execute the original instruction but with the correct target ('Amazon'). All other conditions ('as per my husband's email today') remain unchanged. The revision criterion is satisfied.", "heard_transcript": "Place an uh, Walmart order as per my husband's email today. Make that, let me think. Amazon, not Walmart." }, "candidate_id": "wav_2288d21759aa8164", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a revision ('make that let me think Amazon not Walmart'), the delivery feels less natural. The 'uh' at the start and the 'let me think' pause are somewhat abrupt and the correction is delivered in a slightly robotic, clipped manner. The prosody lacks the subtle intonation shifts seen in the other clips, and the overall pacing feels a bit too controlled, making the revision seem less spontaneous and more like a rehearsed edit than a genuine moment of recollection." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "602a2bb2af579c8141b7d47b80564083ca71d25446652746d8908fe4683251d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/cancellation/cancel_031/2f0eb37678067ab8/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "d676a8c537aaf52a1f9ec7d25bade865f5808f0243a7e23aa6066025eb2b4cb8", "seconds": 17.76, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today.", "Wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Wait, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the entire task. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today. Wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "44d243ab07000f1e5ce1e52182d30bf34f2bf48182d4f4bebc5c15fae0a405c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/ambiguous_halt/halt_007/ca2de98bc962b632/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "f03306be8e63fa3c84ad116ccb70a471dd3c8963bfac0b6d26fc1bc7c1b50d98", "seconds": 15.44, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but repeatedly hesitates with fillers ('um', 'let me think', 'let's see') and ends with 'One moment', which is a clear request to pause. The final interjection is not a cancellation or a new instruction, but a request to hold. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f2d9d4b4d84e0c1008735388ade6a3c94e59e1e67bcd3613a5167d677135f734", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_1/audio_natural/filler/original/749450f7b10fce79/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d9f38ecf6525edc4e5f6d9180ec536362fb0002576b35d4f51f0ed96f83bbb80", "seconds": 12.4, "spoken_utterances": [ "Place an um, let's see... amazon order as per let me think... my husband's email uh, let's see... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes multiple fillers ('um', 'let's see', 'let me think') and hesitations but does not change the core request. The final audible intent matches the original instruction exactly, with no self-correction or modification. The fillers are typical hesitation markers and do not alter the task goal.", "heard_transcript": "Place an um, let's see, Amazon order as per, let me think, my husband's email, uh, let's see, today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my husband's—no, I mean wife's email today.", "long_pause_offsets": [], "source_sha256": "3b9b8020449cb1a796c5916a02d84a40a5b7a01f9b905ddcac44254957b2d928", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'husband's' to 'wife's' within the same utterance, preserving the app (Amazon), action (place an order), and time condition ('today')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/false_start/false_start_02_email_sender_relation/768e11422bb81ad8/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "39c23e78a32dca5909cb223dc7cb263f9e9a4ed71b24d1bf18c2e304f66d9571", "seconds": 7.48, "spoken_utterances": [ "Place an amazon order as per ... my husband's—ah... no, I mean wife's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my husband's' but immediately self-corrects to 'my wife's'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Place an Amazon order as per my husband's uh, no, I mean wife's email today." }, "candidate_id": "wav_7019d036b381c789", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with 'my husband's' and immediately self-corrects to 'no, I mean wife's email' with a clear, believable hesitation and intonation shift. The correction is integrated smoothly into the ongoing utterance, and the overall prosody, including the pause and filler 'uh', sounds completely natural and unscripted." } ] }, "is_default": true }, { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Etsy—I should say amazon order as per my wife's email today.", "long_pause_offsets": [], "source_sha256": "02a97850eaee7c4217dbf8e7b8ebb9f80e3a09889c0ab8ae87c090e18b4d58df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon' within the same utterance, preserving the action, the source of the order (wife's email), and the time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/false_start/false_start_01_ordering_app/8f76663516b226f6/speech.wav", "wer": 0.25, "wav_sha256": "69e4f7abd21c5bd1ac19045517de0171a2e8a061ab88889bc78536176cd5e538", "seconds": 8.36, "spoken_utterances": [ "Place an Etsy—uh... I should say amazon order as per ah... my wife's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon' with 'I should say'. This is a clear false start where the final, corrected value is 'Amazon'. The rest of the instruction matches the original. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Place an Etsy uh, I should say Amazon order as per uh my wife's email today." }, "candidate_id": "wav_912505ef3e04599e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'Etsy' corrected to 'Amazon', which is a plausible error. However, the correction is less integrated than in audio_03; the speaker says 'I should say' before correcting, which is a slightly more formal and less spontaneous way to self-correct. The prosody is natural, but the correction feels a bit more rehearsed compared to the more fluid correction in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Place an amazon order as per my wife's email tomorrow—sorry, that should be today.", "long_pause_offsets": [], "source_sha256": "692f557ce4f8a2db29c231af842c725c8e80890d4d612b14f1ae0b2760f0e3a6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects it to 'today' within the same utterance. The final intended time reference is 'today', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/false_start/false_start_03_time_reference/1e47e1cec1025df5/speech.wav", "wer": 0.0, "wav_sha256": "d413f96cbce252130d908d1baaa0555b2215364d0bdb333f967d79a78b6a6000", "seconds": 7.64, "spoken_utterances": [ "Place an ... amazon order as per ... my wife's email tomorrow—uh... sorry, that should be today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately self-corrects to 'today'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Place an Amazon order as per my wife's email tomorrow. Uh, sorry, that should be today." }, "candidate_id": "wav_1976034ece5be1e5", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision where the speaker says 'tomorrow' and then corrects to 'today'. While the correction is plausible, the delivery feels less natural. The speaker says 'uh, sorry that should be today' with a slightly abrupt and rehearsed tone. The pause before the correction is also shorter and less hesitant than in the other clips, making the correction seem less spontaneous and more like a simple edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Place an amazon order as per today. One more thing: use my wife's email to place the order.", "long_pause_offsets": [ 36 ], "source_sha256": "06455b1c3625fd8e16ea56823ef9e4f14d6bcde48728c4130c98ac978d394b89", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to use the wife's email, which was implied but not specified in the initial instruction. The time constraint 'today' and the action 'place an Amazon order' are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/addition/addition_02/c4fead421dae65de/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "76c2e395afd28ea24f8919d1468efe0233dc563c8cdeefc4288659f9269d13d5", "seconds": 10.6, "spoken_utterances": [ "Place an amazon order as per ... today. ", "One more thing: use um... my wife's email to place the order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'Place an amazon order as per my wife's email today.' The speaker first says 'as per today,' which is a partial or incorrect version of the original. Then, after a pause, they add 'One more thing, use um, my wife's email to place the order.' This addition clarifies and corrects the missing target (my wife's email) and specifies the action (use her email). The final intent combines the original goal with the added constraint, resulting in a complete and coherent instruction that matches the reference. The addition type is correctly applied, and the final goal is executable.", "heard_transcript": "Place an Amazon order as per today. One more thing, use um, my wife's email to place the order." }, "candidate_id": "wav_18689397daeb0979", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural, slightly hesitant 'um' filler, followed by a smooth, additive clause ('one more thing') that introduces a new detail without disrupting the flow. The prosody is consistent, and the pacing feels like a person recalling a detail mid-sentence. The addition is integrated naturally, not tacked on." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Place an amazon order as per my wife's email. Oh, I forgot to mention, do this today.", "long_pause_offsets": [ 46 ], "source_sha256": "9bce13fff6de4af1b0995fd2de20f2016a8c8c710d9a21c8909c90943df28dc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'today' to the initial instruction, which is a missing condition in the original instruction. The final goal combines both the action and the temporal requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/addition/addition_01/2b382484f9cf8469/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "a7c32cb662f8e98b7abbe7e09b6e8ee2a03b4bb138a84fc36e850b972cd93041", "seconds": 11.68, "spoken_utterances": [ "Place an ... amazon order as per um... my wife's email. ", "Oh, I forgot to mention, do this hmmm... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an Amazon order as per my wife's email'. After a pause, the speaker adds 'do this today', which is a time constraint that narrows the original request. The final intent combines the original task with the added constraint, matching the reference instruction. The addition is clear and does not contradict the original task.", "heard_transcript": "Place an Amazon order as per um my wife's email. Oh, I forgot to mention, do this today." }, "candidate_id": "wav_ca1c6e956cf62d6d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is delivered with a natural 'um' hesitation, but the addition ('Oh, I forgot to mention...') is introduced with a slightly abrupt shift in intonation and a filler ('pshun') that sounds less like spontaneous speech and more like a placeholder or a recording artifact. The overall delivery is less fluid, and the addition feels tacked on rather than seamlessly integrated." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Place an amazon order as per my wife's email tomorrow. Make that today, not tomorrow.", "long_pause_offsets": [ 55 ], "source_sha256": "61733b1fb5f5014f7e1b7b26c07f2966c5db6c57e9f96f6f1688e4f64a4ce6c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'tomorrow' to 'today', preserving the action, app, and source of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/revision/revision_03_time_reference/0fd9aca6d79d5e0d/speech.wav", "wer": 0.05, "wav_sha256": "7bc2ea20fa2dc310d6f9e138bacf1d54c45bcdb61f08986ec6594dd9d4704dd6", "seconds": 13.8, "spoken_utterances": [ "Place an amazon order as per hmmm... my wife's email um... tomorrow. ", "Make that let me think... today, not tomorrow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects it to 'today' after a self-correction ('Make that, let me think'). According to the revision type criterion, the correction spoken after a pause (or self-correction) overrides the earlier instruction. The final audible intent is to place the order today, which matches the original instruction. All other constraints are preserved.", "heard_transcript": "Place an Amazon order as per my wife's email um tomorrow. Make that, let me think, today, not tomorrow." }, "candidate_id": "wav_90fe677f4c17edd2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then pauses with a believable 'let me think' filler before correcting the date. The correction is smoothly integrated into the same utterance with a natural shift in intonation, and the overall prosody is consistent and conversational." } ] }, "is_default": true }, { "variant_id": "revision_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my mother's email today. Wait, the email should be from my wife's, not my mother's.", "long_pause_offsets": [ 54 ], "source_sha256": "53d4d5b8bc289974bdb2cc14b03e9a4b975ae3e45281c7e69df29f7238ac6933", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the sender relationship from 'mother' to 'wife', preserving the app (Amazon), action (place order), and time condition ('today')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/revision/revision_02_email_sender_relation/65edbbddd2619f5e/speech.wav", "wer": 0.125, "wav_sha256": "82ca7f8e9c05fa65ff9d531ae697b3c1f017329466c582e5cb7214a1fdd72838", "seconds": 13.68, "spoken_utterances": [ "Place an um... amazon order as per uh... my mother's email today. ", "Wait, the email should be from my um... wife's, not my mother's." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mother's email' but immediately corrects it to 'wife's not my mother's' after a pause. This is a clear revision where the final intent is to use the wife's email. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Place an um Amazon order as per my mother's email today. Wait, the email should be from my um wife's not my mother's." }, "candidate_id": "wav_0ec86f98e687c94f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible self-correction from 'mother's' to 'wife's' email. The speaker uses a natural 'wait' to signal the revision, and the correction is immediate and integrated. However, the initial part of the utterance is slightly hesitant and less fluid than audio_03, and the correction feels a bit more abrupt." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an Walmart order as per my wife's email today. Change the ordering app to Amazon, not Walmart.", "long_pause_offsets": [ 53 ], "source_sha256": "75877ee5b730423eea95080d35869301e7d25da100a6d9b321d31bb4476b589c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the ordering app from Walmart to Amazon, preserving the action, source (wife's email), and time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/revision/revision_01_ordering_app/b42350900da18d9b/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "c124f7610dbb8485491e0ec6fc5a7bffee1d76c925c2b783e45168897255a9f2", "seconds": 12.92, "spoken_utterances": [ "Place an Walmart order as per uh... my wife's email ... today. ", "Change the ordering app to um, let's see... Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately corrects to 'Amazon' after a pause and self-correction ('let's see, Amazon, not Walmart'). This is a revision where the final intent is to use Amazon, matching the original instruction. The correction is clear and final, so the task is to execute the revised goal.", "heard_transcript": "Place some Walmart order as per uh my wife's email today. Change the ordering app to um let's see, Amazon, not Walmart." }, "candidate_id": "wav_f17b88afd28375a1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip has the least realistic delivery. The speaker makes a revision from 'Walmart' to 'Amazon' but does so in a separate, disconnected sentence rather than as a smooth correction within the same thought. The 'let's see' filler is placed awkwardly, and the overall structure feels more like two distinct commands than a single, natural thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "595490622274c6a45381efc7d880d7662a29a283fb0e1c65450cc95905ea5d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/cancellation/cancel_027/5bdc03f03c55d777/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "1c70fb3f80c562efe8ef520991c5cab28eda3efa23e66f1d6216e873411faa29", "seconds": 13.28, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today.", "Wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the task to place an Amazon order, but then explicitly says 'Wait, please disregard my last request,' which is a clear cancellation of the prior instruction. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an um, Amazon order as per my wife's email, um, let me think, today. Wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "abcfb3628c68b7bc111de28568b76b6e443918b94995907d7e8c532f8f6a988b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/ambiguous_halt/halt_007/96bb37c950232cf9/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "4de914ba2fe72652f3e5f21d001bc35b49ac5eb987d9ca2a9c703efb343323b3", "seconds": 11.64, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment,' which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not executed due to the pause. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Place an um, Amazon order as per my wife's email. Um, let me think. Today. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "72147276f511957464d874b32075126dbcfc05a43102344ed3b36667d52ac9fc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_2/audio_natural/filler/original/0902b68dcb2ac930/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "dd5cccc7fe7dd9817102c549b282f2099d5968135c8a9521a9923c82b8a20626", "seconds": 8.52, "spoken_utterances": [ "Place an uh... amazon order as per my wife's email um, let me think... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and pauses but does not change the core request. The final audible intent matches the original instruction. The 'filler' type allows for hesitation and backchannels without altering the goal, so the task should be executed as stated.", "heard_transcript": "Place an um Amazon order as per my wife's email, um let me think, today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_ordering_app", "proposed_instruction": "Place an Etsy—I should say amazon order as per my partner's email today.", "long_pause_offsets": [], "source_sha256": "afa732a87ab3fee99a9131d47a43af1c1014ad6d3889f5e3b3462565e25b0189", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'amazon' within the same utterance, preserving the action, the source of the order, and the condition of following the partner's email." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/false_start/false_start_01_ordering_app/52eafccb49484bea/speech.wav", "wer": 0.1875, "wav_sha256": "d32e96092d77f820f994467620ec18c8f12db71a395923008c14b137684543af", "seconds": 7.2, "spoken_utterances": [ "Place an Etsy—um... I should say ... amazon order as per my partner's email uh... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'etzium' but immediately self-corrects to 'Amazon'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including the action, target, and time constraint, remains consistent. Therefore, the final audible intent matches the original instruction and the control intent is to execute.", "heard_transcript": "Place an etzium, I should say Amazon order as per my partner's email uh today." }, "candidate_id": "wav_5e3a173baa98eab3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('Etsy item') followed by a natural self-correction ('I should say Amazon order') and continues smoothly. The prosody is consistent, the repair is immediate and plausible, and the pacing reflects genuine spontaneous speech. The filler 'uh' after 'email' is well-placed and does not disrupt the flow." } ] }, "is_default": true }, { "variant_id": "false_start_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my friend's—sorry, I meant partner's email today.", "long_pause_offsets": [], "source_sha256": "dcd992fc7d58f90973606258880eeee8eafbf8997452aa50b7f4df06c115aee0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'friend's' to 'partner's' within the same utterance, preserving the app (Amazon), action (place an order), and time condition ('today'). The final intended target is the partner's email." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/false_start/false_start_02_email_sender_relation/09cafa6685054e49/speech.wav", "wer": 0.3125, "wav_sha256": "abeae2c5f9b854d4ad8c6565a97a7e90caa784414a16e04469b6dd07143b2a3a", "seconds": 7.36, "spoken_utterances": [ "Place an amazon order as per my friend's—hmmm... sorry, I meant um... partner's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'partner's email today'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('friends') is discarded. The final instruction matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Place an Amazon order as per my friends. Sorry, I meant um, partner's email today." }, "candidate_id": "wav_3a30bf5b020e1232", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a false start ('friends') and a self-correction ('partner's email'), but the correction is less immediate and the speaker hesitates with 'um' and 'sorry' before the correction. The prosody is slightly more fragmented, and the repair feels a bit more rehearsed or less fluid than audio_01, though still realistic." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Place an amazon order as per my partner's email tomorrow—no, that should be today.", "long_pause_offsets": [], "source_sha256": "5665feebfd6a6d18c70cf7559b2b951f7495668f8efe81ab61563ef00c44025c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects it to 'today'. The final time reference is 'today', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/false_start/false_start_03_time_reference/1cdb707dfdfc3bb2/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "14403ed2ab165703f39e2322a5f84aa9ff977e4008eb6424f4b521ec27dff635", "seconds": 10.24, "spoken_utterances": [ "Place an amazon order as per ... my partner's email tomorrow—um... no, that should be ah... today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately self-corrects to 'today'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Place an Amazon order as per my partner's email tomorrow. Um, no, that should be uh today." }, "candidate_id": "wav_f54524ed1f753101", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision rather than a false start. The speaker initially says 'tomorrow' and then corrects to 'today'. However, the correction is less natural due to the awkward phrasing 'no, that should be uh today' and the drawn-out 'um' before the correction. The prosody is less continuous, and the overall delivery sounds more hesitant and less spontaneous compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an amazon order as per my partner's email. Oh, I forgot to mention, do this today.", "long_pause_offsets": [ 49 ], "source_sha256": "2811b4b0a98b5dc6993675259c3f374d4090b8b4b7439087da8d6259a3a46e09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'today' to the initial request, which is a missing condition in the original instruction. The final goal combines both the action and the temporal requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/addition/addition_01/dedff6aced30ec9f/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8d77c52996072461317992188dc3be78871e054cf9f54b6b5868ec9a520073f3", "seconds": 9.4, "spoken_utterances": [ "Place an ... amazon order as per uh... my partner's email. ", "Oh, I forgot to mention, do this today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to place an Amazon order as per the partner's email. After a pause and an interjection, the speaker adds 'do this today,' which is a time constraint that narrows the scope of the original task. Since the type is 'addition,' the final goal is the combination of the original instruction and the added constraint. The final audible intent matches the reference instruction with the addition of the time constraint, so the task is executed with the new condition.", "heard_transcript": "Place an Amazon order as per my partner's email. Oh, I forgot to mention, do this today." }, "candidate_id": "wav_8532cb368d54159b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Place an Amazon order as per my partner's email') and then adds a time constraint ('Oh, I forgot to mention, do this today') with a believable hesitation and self-correction ('Oh, I forgot to mention'). The prosody is consistent, the pace is conversational, and the addition feels like a spontaneous afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an amazon order today. One more detail: use the order details from my partner's email.", "long_pause_offsets": [ 29 ], "source_sha256": "d74e2473fb1487583633408f1ce1dc1dd3617bba4842e7f57f5fca35a776ff75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition about sourcing the order details from the partner's email, which was implied but not specified in the original instruction. The final goal combines the initial action with the added constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/addition/addition_02/a500d20f3df3df72/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "5189af4c4dc0f6e9b341444b2bbe80a26ae8ce2f9e0c25a5e4a544a4d1fa000a", "seconds": 11.52, "spoken_utterances": [ "Place an amazon order uh... today. ", "One more detail: use the order details from hmmm... my partner's email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Place some Amazon order uh today' is a slightly imprecise version of the original request, but the addition 'use the order details from um my partner's email' clearly adds a critical constraint. The final intent combines the original goal with the new detail, matching the reference instruction. The addition is not a correction of a prior error but a clarification, so the combined goal is complete and aligned with the reference.", "heard_transcript": "Place some Amazon order uh today. One more detail, use the order details from um my partner's email." }, "candidate_id": "wav_cd3232f521dd9ad1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is clear, but the addition ('One more detail, use the order details from my partner's email') is less smoothly integrated. The speaker uses a filler ('uh') and a pause before the correction, but the phrasing 'use the order details from' sounds slightly unnatural and less like spontaneous speech. The overall delivery is slightly more hesitant and less fluid than audio_01, making it feel less realistic despite being semantically coherent." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_email_sender_relation", "proposed_instruction": "Place an amazon order as per my friend's email today. Wait, the email should be from my partner's email, not friend's.", "long_pause_offsets": [ 54 ], "source_sha256": "50885c543a640d2fb49e4279460bc194b24b403b5b9f07160c045b5c3798a31f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the sender relationship from 'friend' to 'partner', preserving the app, action, and time constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/revision/revision_02_email_sender_relation/7f420727e5de02cc/speech.wav", "wer": 0.2, "wav_sha256": "2b443abb76491bc7a99748a5817fe8a5c638bb579aaca099240af792359c4c93", "seconds": 15.36, "spoken_utterances": [ "Place an um... amazon order as per my friend's email um, let's see... today. ", "Wait, the email should be from my ... partner's email, not friend's." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friend's email' but immediately corrects it to 'partner's email' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original instruction exactly, including the action, target, and time constraint.", "heard_transcript": "Place an Amazon order as per my friend's email. Um, let's see. Today. Wait, the email should be for my partner's email, not friends." }, "candidate_id": "wav_27719973f625d9aa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision. The speaker begins with a clear request, hesitates with 'um' and 'let's see', then self-corrects 'friend's email' to 'partner's email' with a believable change-of-mind intonation and a smooth, continuous flow. The prosody is varied and human, with a natural pause and filler before the correction, and the correction itself is integrated seamlessly into the utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Place an amazon order as per my partner's email tomorrow. Make that today, not tomorrow.", "long_pause_offsets": [ 58 ], "source_sha256": "86f548ab53fbcf7cc8301ce2d7f64a42201ac22fba39a5bbcf443b38521b2ea1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'tomorrow' to 'today', preserving the action, app, and source document." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/revision/revision_03_time_reference/340737427f270e30/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "f420e75541a911e53dc9696969c333ea9dc5aeb8178cbb88e0e682c77fafdb25", "seconds": 12.48, "spoken_utterances": [ "Place an amazon order as per uh... my partner's email ... tomorrow. ", "Make that um... today, not tomorrow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tomorrow' but immediately corrects to 'today' with 'Make that um today, not tomorrow.' This is a clear revision where the final value replaces the initial incorrect one. The final intent matches the original instruction exactly, so the revision criterion is satisfied.", "heard_transcript": "Place an Amazon order as per uh my partner's email tomorrow. Make that um today, not tomorrow." }, "candidate_id": "wav_1beb46bd3140ea3d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision where the speaker first says 'tomorrow', then corrects to 'today'. The correction is clear and the prosody is natural, with a slight pause and filler ('make that um') before the change. However, the initial request is slightly less fluent than audio_01, and the correction feels a bit more abrupt and less integrated into the original sentence, making it slightly less realistic in flow." } ] }, "is_default": false }, { "variant_id": "revision_01_ordering_app", "proposed_instruction": "Place an Etsy order as per my partner's email today. Change the ordering app to Amazon, not Etsy.", "long_pause_offsets": [ 53 ], "source_sha256": "5faf96d6828c1e337fe336f6b39caf3a96fedc62217731c7b261335e8daa7bc8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the ordering app from Etsy to Amazon, while retaining the source (partner's email) and the time constraint (today)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/revision/revision_01_ordering_app/3641284bc9a24540/speech.wav", "wer": 0.25, "wav_sha256": "082ff6097ae377775d186a56947f5622d8d951fe4fffa8afed13631a7241133e", "seconds": 10.0, "spoken_utterances": [ "Place an uh... Etsy order as per um... my partner's email ... today. ", "Change the ordering app to Amazon, not Etsy." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately corrects to 'Amazon' after a pause, which matches the revision type criterion. The final instruction clearly specifies the correct app, and all other conditions (ordering as per partner's email today) are preserved. The correction is unambiguous and final.", "heard_transcript": "Place an Etsy order as per um my partner's email today. Change the ordering app to Amazon, not Etsy." }, "candidate_id": "wav_6a10ad29fbab2dda", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision but sounds less realistic due to its structure. The speaker first makes a request for an 'Etsy order', then immediately corrects the app to 'Amazon'. The correction is delivered as a separate, almost disjointed sentence rather than a smooth self-repair within the original utterance. The prosody is flatter and the transition between the two parts feels more like a post-hoc correction than a spontaneous thought process, reducing its naturalness." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "7ea5683bcd1033d76234f3c014946aaa3c20fcf7b7a0e36c857d348fed4e11ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/cancellation/cancel_002/a7c84ff4c747f637/speech.wav", "wer": 0.0625, "wav_sha256": "0ab109c5181512161b9b073cdc8d1948eabf2dacab8ebeff9073d4828b9c4b41", "seconds": 10.16, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today.", "Wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting an Amazon order, but immediately corrects themselves with 'Wait, cancel the whole task.' This constitutes a clear cancellation of the initial request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an uh Amazon order as per my partner's email today. Wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "6849ccb9b772a17c6c356420a20eb80624d3907e23690ef5302de34b2024bd3e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/ambiguous_halt/halt_024/ebc79d5b7709edb8/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "7e3ba37d3281d485897fccbf7bfe112956c217b93c33129821c089c8920028c3", "seconds": 9.52, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial instruction is not completed, but the final control intent is to pause, which is consistent with the type.", "heard_transcript": "Place an uh Amazon order as per my partner's email today. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1a79e37_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "548fd87739e91ac7e0580f73382d7ae1fd657817b976badfa4897a77b5610cbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1a79e37_3/audio_natural/filler/original/3214a6e1f4766c23/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "d0dbaf2e226a514b987d09712b016f16116c19e9e253fcf2a51e01378ec720b8", "seconds": 5.84, "spoken_utterances": [ "Place an uh... amazon order as per ... my partner's email today." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Place an uh Amazon order...', where 'uh' is a filler sound. The rest of the sentence matches the original instruction. The final intent is to execute the Amazon order task as per the partner's email today. The filler does not change the goal, so the task is executed as intended.", "heard_transcript": "Place an uh Amazon order as per my partner's email today." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_shipping_service", "proposed_instruction": "Initiate returns via UPS—sorry, that should be FedEx for everything in my last 2 amazon order.", "long_pause_offsets": [], "source_sha256": "98b4ade41944f2a7dfae1844776891cfdc95decbc6459ad318435c165ec50f85", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "unchecked", "reason": "현재 문장과 일치하는 semantic check 결과가 없습니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/false_start/false_start_01_shipping_service/f482e64bad6de38e/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "f0e17abde40f8d25c51afbc0a7837520ce5b84b317f0ce5858d2e9b5e9197560", "seconds": 11.24, "spoken_utterances": [ "Initiate returns via UPS—uh... sorry, that should be ... FedEx for everything in my last two amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker makes a false start by saying 'UPS' and immediately self-corrects to 'FedEx'. According to the 'false_start' type criterion, the final intent is the corrected version, and the initial incorrect value is discarded. The final instruction matches the original intent with the correct carrier, and all other conditions (returns, everything, last two Amazon orders) are preserved.", "heard_transcript": "Initiate returns via UPS uh, sorry, that should be FedEx for everything in my last two Amazon order." }, "candidate_id": "wav_b0d03ba37051479b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent false start. The speaker makes a clear mistake ('UPS'), immediately self-corrects with 'sorry, that should be FedEx' in a spontaneous, conversational tone, and then continues smoothly. The prosody is consistent, and the repair feels like a genuine moment of recall rather than a rehearsed or robotic correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 2 ebay—I should say amazon order.", "long_pause_offsets": [], "source_sha256": "b94291fb22606506d86a178fdcdf64dbb77efd14cadf1d4ec7083299dff8f977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the app, action, and scope of the return request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/false_start/false_start_03_app/99018dec7ae11620/speech.wav", "wer": 0.0, "wav_sha256": "fe697e89ae7bdd8d55caaa77642efc412a274e2dd4b2be400828706accaf7911", "seconds": 10.6, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my last two ebay—um... I should say amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'eBay' but immediately self-corrects to 'Amazon' after a pause and a self-correction phrase ('I should say'). According to the 'false_start' type criterion, the final corrected value ('Amazon') overrides the initial mistaken value ('eBay'), and the rest of the instruction remains consistent. The final intent matches the original instruction with the corrected target.", "heard_transcript": "Initiate returns via uh FedEx for everything in my last two eBay um, I should say Amazon order." }, "candidate_id": "wav_9ffd336b0cdd1723", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start with a self-correction ('eBay' to 'Amazon'), but the correction is less fluid. The speaker says 'eBay' and then hesitates with 'um' before correcting. The intonation during the correction is slightly more abrupt and less conversational than in audio_01, making the repair feel a bit more forced or less spontaneous." } ] }, "is_default": true }, { "variant_id": "false_start_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3—no, that should be 2 amazon order.", "long_pause_offsets": [], "source_sha256": "1aed4c7d15a0b316f7e0afe6bde75edc5ae3fd8a7162c073c8ce8c9078b54e22", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'three' to 'two' within the same utterance, and the rest of the instruction including the app (FedEx), action (initiate returns), and scope ('everything in my last ... Amazon order') remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/false_start/false_start_02_order_count/2048694dd5b49c50/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "691245d5096dab072001654e52fea9eac5e32760f43e40805bdafd1e47e83456", "seconds": 11.0, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my last three—ah... no, that should be two amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last three' but immediately self-corrects to 'last two' before completing the sentence. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('three') is discarded. The final goal matches the original instruction with the corrected number, so the task is executed as intended.", "heard_transcript": "Initiate returns via FedEx for everything in my last three uh, no, that should be two Amazon order." }, "candidate_id": "wav_f651dff4efb9ded9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip has the least realistic delivery. The speaker makes a false start ('last three') and corrects it to 'two', but the correction is delivered with a noticeable pause and a hesitant 'no' before the correction. The overall pacing is slower, and the correction feels more like a mechanical fix than a natural thought process. The prosody is less continuous, and the self-correction lacks the smoothness of the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Initiate returns via FedEx for everything in my last amazon order. One more detail: return items from my last two Amazon orders.", "long_pause_offsets": [ 67 ], "source_sha256": "b1c18a69b9f2981b6397f7e9e120502d9a776e8ed0e5592cd337e3b9b91f3c12", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup adds a constraint that changes the number of orders from 'last 2' to 'last 2', but the original instruction specifies 'last 2 amazon order' which is ambiguous. However, the spoken instruction clarifies 'last two Amazon orders', which is a different scope than the original 'last 2 amazon order'. The original instruction likely meant 'last two Amazon orders', but the phrasing 'last 2 amazon order' is grammatically incorrect and could be interpreted as 'last two Amazon orders' or 'last Amazon order'. The followup clarifies it as 'last two Amazon orders', which is a different scope than the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/addition/addition_02/0c4591d35b953a31/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "e92174cb2708a657e3e6b89a4009b322bb0090438ac61e5c13d64a8eb33b870a", "seconds": 16.2, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my let's see... last amazon order. ", "One more detail: return items from my last let me think... two Amazon orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'last 2 amazon order', which was corrected to 'last two Amazon orders' after a pause and self-correction. The addition of 'two' instead of '2' is a minor paraphrase that does not change the meaning. The final audible intent matches the original instruction with the addition of the number 'two' to clarify the quantity of orders. The task is to initiate returns for everything in the last two Amazon orders, which is a valid and complete request.", "heard_transcript": "Initiate returns via FedEx for everything in my let's see, last Amazon order. One more detail, return items from my last, let me think, two Amazon orders." }, "candidate_id": "wav_ff42d9c4ef7c8b69", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Initiate returns via FedEx for everything in my. Oh, I forgot to mention, it's only for my last two Amazon orders.", "long_pause_offsets": [ 49 ], "source_sha256": "a7a45e63ed1e6e78794efca425aa0a65f7c8c4fe9300f68f04d99bd08ddb98cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup adds a restriction that narrows the scope from 'last 1' to 'last 2' orders, which contradicts the original instruction's singular 'last' order." }, "audio": { "status": "pending", "spoken_utterances": [ "Initiate returns via ... FedEx for everything in my. ", "Oh, I forgot to mention, it's only for my um... last two Amazon orders." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 2 ebay order. Let me correct the app: return items from my last 2 amazon order, not ebay.", "long_pause_offsets": [ 67 ], "source_sha256": "c0fdd463b8e82b9946204068bb9345cf483257e5a3221d8035ac73154134bb1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the action, return method, and scope of 'everything in my last two orders'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/revision/revision_03_app/e59f8a49341bd3e8/speech.wav", "wer": 0.0, "wav_sha256": "3f4c52fb3ef83000dd18ca01b11ba1d30d6a689ad61827821c2663ca633b57d9", "seconds": 16.28, "spoken_utterances": [ "Initiate returns via let's see... FedEx for everything in my last two ebay order. ", "Let me correct the app: return items from my last two ... amazon order, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay order' but immediately corrects it to 'Amazon order, not eBay' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original instruction's action and target, with only the platform corrected from eBay to Amazon.", "heard_transcript": "Initiate returns via FedEx for everything in my last two eBay order. Let me correct the app, return items from my last two Amazon order, not eBay." }, "candidate_id": "wav_4934788123e68f39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear request, then uses a realistic filler ('let me see') before making a mistake ('eBay'). The self-correction ('let me correct the app') is delivered with a believable shift in intonation and pacing, followed by a smooth, grammatically coherent correction ('return items from my last two Amazon order, not eBay'). The prosody and rhythm are consistent with spontaneous speech, and the correction is integrated into the utterance rather than sounding like a separate command." } ] }, "is_default": true }, { "variant_id": "revision_01_shipping_service", "proposed_instruction": "Initiate returns via UPS for everything in my last 2 amazon order. Wait, use FedEx for the returns, not UPS.", "long_pause_offsets": [ 67 ], "source_sha256": "bb0c552830b9516a93cd6e898caacfa7aa42121ce71023cfb4711edcecad5117", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the shipping service from UPS to FedEx, while retaining the scope of 'everything in my last two Amazon orders'. The original instruction specified 'last 2 amazon order' which is semantically equivalent to 'last two amazon order'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/revision/revision_01_shipping_service/0c99919afc31fc6e/speech.wav", "wer": 0.0, "wav_sha256": "26f1f8af9fc9a057d964c5fbe50c9517a504883984d5e79d4a414b79ccbc6291", "seconds": 13.96, "spoken_utterances": [ "Initiate returns via UPS for everything in my um, let's see... last two amazon order. ", "Wait, use ... FedEx for the returns, not UPS." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'UPS' but immediately corrects to 'FedEx' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to use FedEx, which matches the original reference instruction. The only difference is the initial incorrect value 'UPS', which is corrected and thus does not affect the final goal.", "heard_transcript": "Initiate returns via UPS for everything in my um let's see last two Amazon order. Wait, use FedEx for the returns, not UPS." }, "candidate_id": "wav_3120d039f59c35e4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a credible revision with a natural pause and filler ('let me see') before the initial request. The speaker then corrects the shipping carrier from 'UPS' to 'FedEx' with a plausible 'wait' and a clear, spoken correction. The prosody is generally natural, though the correction is slightly more abrupt than in audio_02, and the initial filler is less integrated into the flow. The overall delivery is realistic but slightly less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3 amazon order. Make that the last 2 amazon order, not 3.", "long_pause_offsets": [ 69 ], "source_sha256": "e0f323a23d3666eef21b213ffdc846a8a94ac2c5d7a4ed7f0381e8e6649c5cc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the count from 'three' to 'two', preserving the app (FedEx), action (initiate returns), scope (everything), and context (Amazon orders)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/revision/revision_02_order_count/d66c4d85fdb66b31/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "e820613756525a67321342f7f7d467d65e858f468a508efaf79ff36266de9a71", "seconds": 13.32, "spoken_utterances": [ "Initiate returns via FedEx for everything in my um... last three amazon order. ", "Make that the last ... two amazon order, not three." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last three Amazon order' but immediately corrects it to 'last two Amazon order, not three'. This is a clear revision where the final value replaces the initial one. The correction is spoken after a short pause and is unambiguous. The final intent matches the original instruction with the corrected number of orders.", "heard_transcript": "Initiate returns via FedEx for everything in my um last three Amazon order. Make that the last two Amazon order, not three." }, "candidate_id": "wav_3e6004829515b54a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but with less natural phrasing. The speaker starts with a clear request, then corrects 'three' to 'two' with a simple 'make that the last two'. However, the correction is delivered in a flat, almost robotic tone without the intonation shift or hesitation that would make it sound spontaneous. The initial filler ('um') is brief and doesn't contribute to a sense of formulating the request. The overall delivery sounds more like a rehearsed or read instruction than a natural self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "284ad45e61883ca106bb022595a87d8ba9672f0763f460e74eec1456e8be1663", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/cancellation/cancel_011/b1a4dad02e836f70/speech.wav", "wer": 0.1, "wav_sha256": "c7c1cd40bf18553ac93de1dc008ce20c9044e9fef7ad86756520de0ecffa30cc", "seconds": 13.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order.", "Wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting a task, then clearly revokes it with 'Wait, I've changed my mind. Cancel that.' This constitutes a final, unambiguous cancellation of the original instruction. The control intent is correctly identified as 'cancel'.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order. Wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "fdab2cdcef4cfdafa4765b23c69dec6a1f8acdf16cc1599d40fdbb0dac45b351", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/ambiguous_halt/halt_002/b92d876c66763709/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "02141ef29db90cf79a48b3c904d0f605d922ab994518ff68f8d27e1380985110", "seconds": 11.16, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order.", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to initiate returns is not canceled, but the speaker explicitly halts further action. No correction or cancellation is made, and the final audible intent is to pause.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order. Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "dc88a05584dad10f96ea7cbb7a4b2f1f691dc0040fcf65b9af4f889db3a98ce0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_1/audio_natural/filler/original/068af455177b6672/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "1d2de6b2bc66fd25b41ce1f33fee50573b8544aaa0328f77806197dd308e03e4", "seconds": 8.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my ... last two amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'um' but completes the original instruction without changing the action, carrier, scope, or target. The only difference is the number 'two' replacing 'last', which is a minor grammatical variation that does not alter the core intent. The final goal matches the original instruction.", "heard_transcript": "Initiate returns via um FedEx for everything in my last two Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_shipping_service", "proposed_instruction": "Initiate returns via UPS—no, that should be FedEx for everything in my last 3 amazon order.", "long_pause_offsets": [], "source_sha256": "8812f780943c24e7eb07ff550a44c945132a0af046282b5644eea123a8760ada", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects the shipping service from UPS to FedEx, which matches the original instruction. The phrase 'last three amazon order' is interpreted as 'last three Amazon orders' based on context and standard phrasing, with no other conflicting changes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/false_start/false_start_01_shipping_service/07a35489e596d082/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "f547cc9977df732e294a250e10243dd2789bb2f00b8133385032e0fe159dec3f", "seconds": 9.52, "spoken_utterances": [ "Initiate returns via UPS—hmmm... no, that should be uh... FedEx for everything in my last three amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'UPS' but immediately self-corrects to 'FedEx' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value ('FedEx') should be applied, and the original incorrect value ('UPS') should be discarded. The rest of the instruction, including 'for everything in my last three Amazon order', remains unchanged and is preserved. The final intent matches the original reference instruction exactly.", "heard_transcript": "Initiate returns via UPS. Hmm. No, that should be uh FedEx for everything in my last three Amazon order." }, "candidate_id": "wav_2dea46d2a07b651f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'UPS', immediately self-corrects to 'FedEx' with a clear, spontaneous-sounding 'no, that should be', and smoothly continues the full request. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the final sentence is delivered with natural intonation and pacing." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3 ebay—I should say amazon order.", "long_pause_offsets": [], "source_sha256": "25d7a3c03590aafcb00bddf3e5e9611048cce4e80a534abe11392abf7e58b0df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the return method, scope ('everything'), and temporal qualifier ('last three order')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/false_start/false_start_03_app/522563c12bc40611/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "a928484a42e1a37eddd96ec19a140812a738c720579d6f54f9b203295ef612e8", "seconds": 12.0, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my let me think... last three ebay—um... I should say amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ebay' but immediately self-corrects to 'Amazon' after a pause and a self-correction phrase ('I should say'). According to the 'false_start' type criterion, the final corrected value ('Amazon') overrides the initial mistaken value ('ebay'), and the rest of the instruction remains unchanged. The final intent matches the original instruction exactly.", "heard_transcript": "Initiate returns via um FedEx for everything in my let me think last three ebay um I should say Amazon order." }, "candidate_id": "wav_2ee419c33e1befdc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a credible false start with 'let me think' and a self-correction from 'eBay' to 'Amazon'. However, the correction is less integrated; the speaker says 'I should say Amazon order' as a separate phrase rather than a smooth rephrasing, and the overall delivery is slightly more hesitant and fragmented compared to audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 5—scratch that, make it 3 amazon order.", "long_pause_offsets": [], "source_sha256": "e817d381a84beeaeaf598b4e19c9c564ddbfecb0641dbdffa5be484421198b86", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'five' but immediately corrects it to 'three' within the same utterance. The final value is 'three', and all other elements—action, carrier, scope, and entity—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/false_start/false_start_02_order_count/35b5cc487d943e98/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "3c6290536d5c74121fa2f783d57eb83b440c93f276b7b6dcaf516382b4bd5fe4", "seconds": 11.88, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my last five—um... scratch that, make it three amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last five' but immediately self-corrects with 'scratch that, make it three'. This is a clear false start where the final value is 'three'. The rest of the instruction, including the action (initiate returns), carrier (FedEx), and scope (Amazon orders), remains consistent with the original. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Initiate returns via FedEx for everything in my last five um, scratch that, make it three Amazon order." }, "candidate_id": "wav_53ac57a5d0a21249", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision ('scratch that') rather than a true false start. The speaker completes the initial request, then pauses and changes the number from 'five' to 'three'. While the revision is plausible, the 'scratch that' is a slightly abrupt, more robotic-sounding phrase, and the overall flow feels less natural and more like a deliberate edit than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Initiate returns via FedEx for everything in my last amazon order. Limit that to the last 3 amazon orders.", "long_pause_offsets": [ 67 ], "source_sha256": "0d3d8c7bcb1e9f8c065b7540bc5e6a913ddbd5ad0d0601b7b03f8402767584e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to limit the scope from 'last amazon order' to 'last three amazon orders', which is a narrowing of the original instruction but does not contradict it. The action, carrier, and item scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/addition/addition_02/79e1dc4682e1561e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "10e9861970b05437a4cdacb1861168be228be858fab649959cfa465f70bba3f7", "seconds": 14.84, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my um... last amazon order. ", "Limit that to the last ... three amazon orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request for the 'last Amazon order' and then adds a constraint to 'limit that to the last three Amazon orders'. This is a clear addition that narrows the scope of the original instruction. The final intent is a combined task that matches the reference instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Initiate returns via uh FedEx for everything in my um last Amazon order. Limit that to the last three Amazon orders." }, "candidate_id": "wav_83ea37987773ae81", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Initiate returns via FedEx for everything in my. This should apply only to everything in my last 3 amazon order.", "long_pause_offsets": [ 49 ], "source_sha256": "e5da3fe8cf99a82611a87838d50f05b076667fae43558f77e6d5098c28fc5919", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'last three Amazon orders' to the initial incomplete request, clarifying the scope of items to return. The action, carrier, and intent remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Initiate returns via um, let me think... FedEx for everything in my. ", "This should apply only to everything in my uh... last three amazon order." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_shipping_service", "proposed_instruction": "Initiate returns via UPS for everything in my last 3 amazon order. Wait, use FedEx for the returns, not UPS.", "long_pause_offsets": [ 67 ], "source_sha256": "ba628c7836cbbfb8a71ed64f601e9bbedf067ce18c4fec76c44cf62b731f278c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the shipping service from UPS to FedEx, preserving the action, scope (everything in the last three Amazon orders), and the intent to initiate returns." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/revision/revision_01_shipping_service/d335ef0650188995/speech.wav", "wer": 0.04, "wav_sha256": "3fdef6c3300789742303e1afb7695d7c83dcd2f4cbbbf84a0b4943c205b1b703", "seconds": 14.88, "spoken_utterances": [ "Initiate returns via ... UPS for everything in my um... last three amazon order. ", "Wait, use um, let me think... FedEx for the returns, not UPS." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially requested returns via UPS but then corrected themselves to FedEx after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, with only the carrier changed from UPS to FedEx as per the revision.", "heard_transcript": "Initiate returns via UPS for everything in my um last three Amazon order. Wait, use um, let me think. FedEx for the returns, not UPS." }, "candidate_id": "wav_8f9c6ffaecb0b93d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural revision process. The initial request is made with a plausible hesitation ('um'), followed by a self-correction ('wait, use um...') that sounds like genuine recollection. The speaker then clearly corrects the carrier from 'UPS' to 'FedEx' and continues with a smooth, continuous flow. The prosody and pacing are consistent with spontaneous speech, and the correction is integrated naturally within the utterance." } ] }, "is_default": true }, { "variant_id": "revision_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 5 amazon order. Make that the last 3 amazon order, not 5.", "long_pause_offsets": [ 69 ], "source_sha256": "bcd87e13b660a186a58cb9bc70fdf5bce5b191e1edc863601ae72aef7d9e90e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the number of orders from five to three, while retaining the return method, scope (everything), and source (Amazon)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/revision/revision_02_order_count/0c1148a515da4f53/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "044528b4f231e6455cb759cfc03b151483ef1838a5b32b34cef0587abdd0532e", "seconds": 13.44, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my last five amazon order. ", "Make that the ... last three amazon order, not five." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last five Amazon order' but immediately corrects it to 'last three Amazon order, not five'. This is a clear revision where the final value replaces the initial incorrect one. The final intent matches the original instruction, so the task should be executed as revised.", "heard_transcript": "Initiate returns via FedEx for everything in my last five Amazon order. Make that the last three Amazon order, not five." }, "candidate_id": "wav_eb4d93f8b0ffb3ce", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also performs a revision, but the correction is less integrated. The initial request is delivered fluently, but the revision ('Make that the last three Amazon order not five') is a separate, completed sentence that sounds more like a post-hoc correction than a spontaneous self-repair. The prosody is slightly more robotic, and the revision lacks the natural hesitation and flow of the first clip, making it sound less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3 ebay order. Let me correct the app: return items from amazon, not ebay.", "long_pause_offsets": [ 67 ], "source_sha256": "e2789334c0f28795d417c82c02fda10b35d5c8e1728efabc951e2289c90eea41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the platform from 'ebay' to 'amazon', preserving the return method (FedEx), scope ('everything'), and temporal qualifier ('last three orders')." }, "audio": { "status": "pending", "spoken_utterances": [ "Initiate returns via ... FedEx for everything in my ... last three ebay order. ", "Let me correct the app: return items from uh... amazon, not ebay." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "16e2e4026c756b5ef60be788040c83563bb80da5e68233db3d9957b74de4f689", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/cancellation/cancel_034/1832c59aa2b77344/speech.wav", "wer": 0.0, "wav_sha256": "a06398014df84fde9e767b77ed5a62d60b273542f13cd73cdaa7c11241f849b6", "seconds": 12.08, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order.", "Oh, wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Oh wait, let's drop the whole task.' This constitutes a final, unambiguous cancellation of the entire task. The original instruction is irrelevant after this cancellation.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order. Oh wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "f023c8a4498d9f5e79fc253dcdad13ddb919b75fc633d9d97f61af21d6a7b258", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/ambiguous_halt/halt_036/48e8d1de8191860b/speech.wav", "wer": 0.0, "wav_sha256": "5677a0b5c912d37002a582068437c78d88fc7570f4a03d6abc043a26a87772b5", "seconds": 11.6, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly interjects 'Wait a second, before you go ahead,' which is a direct request to pause. The final interjection is a pause command, not a cancellation or a revision. The task is not completed, but the audio ends with a valid request to halt, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0cda63a6a9649e314e0f9c94b6acf09d6091ff5119719a39c1dd944712c6858f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_2/audio_natural/filler/original/44f60b04526fac66/speech.wav", "wer": 0.0, "wav_sha256": "410820375327bcdeca627e5aecf9fcf66151f3f3801d24a8feb418be3a9444ae", "seconds": 7.56, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my uh, let me think... last three amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'uh', 'let me think') and a brief pause but does not change the core request. The final intent is to initiate returns via FedEx for all items in the last three Amazon orders, which matches the original instruction. The hesitation and self-interruption are typical of a filler type and do not alter the task goal.", "heard_transcript": "Initiate returns via um FedEx for everything in my uh let me think last three Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_shipping_service", "proposed_instruction": "Initiate returns via UPS—I meant FedEx for everything in my last 4 amazon order.", "long_pause_offsets": [], "source_sha256": "173eda3abc9e3a500836a68d83c136c4aabe91d4a80e1b228890594c4bba2e93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'UPS' to 'FedEx' within the same utterance, preserving the action, target (Amazon order), and scope ('everything in my last four')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/false_start/false_start_01_shipping_service/0baa025e689999cb/speech.wav", "wer": 0.0625, "wav_sha256": "926136bad4f88782d3ae41fa89ddbd305347ef040fdb2cad9e7bcc2c9780e409", "seconds": 9.96, "spoken_utterances": [ "Initiate returns via UPS—uh... I meant FedEx for everything in my ... last four amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'UPS' but immediately self-corrects to 'FedEx'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the scope 'everything in my last four Amazon order', is preserved. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Initiate returns via UPS. I meant FedEx for everything in my last four Amazon order." }, "candidate_id": "wav_a7782a605b7f06a0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'UPS' and immediately self-corrects to 'FedEx' with a clear, spontaneous 'I meant' phrase. The correction is integrated smoothly into the sentence, maintaining a consistent rhythm and prosody. The pause before the correction is brief and plausible, suggesting genuine recall rather than a rehearsed or robotic edit. The overall delivery sounds like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3—make that 4 amazon order.", "long_pause_offsets": [], "source_sha256": "b529af09d414a745b38b47e33760fa2beae882501b8dd75e0eb5c9a8f5b3a4bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'three' to 'four' within the same utterance, preserving the action, carrier, scope ('everything'), and context ('my last... order'). The final value is 'four', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/false_start/false_start_02_order_count/80dd2ab50deae40b/speech.wav", "wer": 0.125, "wav_sha256": "a589a8b449ce9fee52537fb6ca746b8e956555b143bf9c96a6a0152275658ee1", "seconds": 10.48, "spoken_utterances": [ "Initiate returns via ... FedEx for everything in my last three—uh... make that four amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last three' but immediately self-corrects to 'make that four', which is the final and intended value. The correction is immediate and unambiguous, satisfying the false_start type criterion. The rest of the instruction matches the original.", "heard_transcript": "Initiate returns via uh FedEx for everything in my last three uh uh make that four Amazon order." }, "candidate_id": "wav_d40a2aeb8fd00eae", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a self-correction but with less natural flow. The speaker starts with 'last three' and then corrects to 'make that four'. The correction is phrased as a separate sentence ('Make that four...') rather than being integrated into the original clause, which makes it sound more like a post-hoc addition than a spontaneous thought. The double 'uh' filler before the correction also feels slightly awkward and less fluid than the correction in audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 4 Walmart—or rather, amazon order.", "long_pause_offsets": [], "source_sha256": "77fe17b0b0223802da0a0996180946e8caf2abc6cd9ac47f409118ab1d6f73ce", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately corrects to 'amazon' within the same utterance. The final intended target is Amazon, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/false_start/false_start_03_app/b831d0a89e585888/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "5d6d155b1ba88966ceb031d3f37bbc0349bef9470d5610287e9a25a7bfa0ed5d", "seconds": 13.52, "spoken_utterances": [ "Initiate returns via um... FedEx for everything in my last four um, let me think... Walmart—uh... or rather, amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart' but immediately self-corrects with 'or rather Amazon order'. This is a false start where the final audible intent matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as stated in the reference.", "heard_transcript": "Initiate returns via FedEx for everything in my last four, um, let me think, Walmart, uh, or rather Amazon order." }, "candidate_id": "wav_e459b73b8c267e77", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with 'FedEx' but then hesitates with 'um' and 'let me think', followed by a complete change to 'Walmart' and then 'Amazon order'. The correction is not integrated into the original sentence but rather replaces it, making it sound like a disjointed thought process. The long pause and filler words suggest the speaker is formulating a new request rather than correcting a minor error, which reduces the plausibility of a natural self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Initiate returns via FedEx for everything in my. I left out a detail: only return items from my last 4 Amazon orders.", "long_pause_offsets": [ 49 ], "source_sha256": "dd7b9b072dc008795124b116f4e5c098eb5fd438a08a83cac3cefa7ec637537c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial incomplete request, narrowing the scope from 'my last 4 amazon order' to 'my last four Amazon orders'. The action, app, and return method remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/addition/addition_02/983ee708fbf35c9a/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "a9dfaf0a7516882f9d08e63da80b5e8a21ddc10c69920738212f228e59919d41", "seconds": 14.48, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my. ", "I left out a detail: only return items from my hmmm... last four Amazon orders." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the original instruction, then adds a clarifying constraint ('only return items from my last four Amazon orders') after a pause and self-correction. This matches the 'addition' type, where the final goal combines the original request with the added detail. The final intent is clear and executable.", "heard_transcript": "Initiate returns via FedEx for everything in my... I left out a detail, only return items from my last four Amazon orders." }, "candidate_id": "wav_e082ea602054dfa8", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Initiate returns for everything in my last 4 amazon order. Oh, I forgot to mention, use FedEx for the returns.", "long_pause_offsets": [ 59 ], "source_sha256": "a5f05bb1014007585035d6459b93d3846b364c926212b722827dd50d4b7e428d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the required shipping carrier (FedEx) to the initial return request, which was missing in the first utterance. The initial request already specified the action (initiate returns), scope (everything), and context (my last four Amazon order). The addition completes the missing constraint without altering any other requirement." }, "audio": { "status": "pending", "spoken_utterances": [ "Initiate returns for everything in my let me think... last four amazon order. ", "Oh, I forgot to mention, use ... FedEx for the returns." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Initiate returns via FedEx for everything in my last 4 ebay order. Let me correct the app: return items from my last 4 ebay order, not amazon.", "long_pause_offsets": [ 67 ], "source_sha256": "aacd7ee1aa76b91f6a45b3c09fde9a5bfcc496ffca224a3994fdcd48dff26992", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Amazon to eBay, changing the required platform for the return action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/revision/revision_03_app/735199c16f52e579/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "9f5038263cd46b79133467a25ec99f27f755eae637bc6bcc57a58355a678fb26", "seconds": 16.88, "spoken_utterances": [ "Initiate returns via FedEx for everything in my uh... last four ebay order. ", "Let me correct the app: return items from my last four um... ebay order, not amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'last four amazon order' but immediately corrects it to 'eBay order, not Amazon' after a pause. This is a revision where the final intent clearly overrides the initial incorrect target. The action (initiate returns via FedEx) and scope ('everything in my last four') remain consistent. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Initiate returns via FedEx for everything in my uh last four eBay order. Let me correct the app, return items from my last four um eBay order, not Amazon." }, "candidate_id": "wav_f9ba460fe8e4d5f1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a clear request, hesitates with a plausible 'uh' and filler, then smoothly corrects the platform from Amazon to eBay with a self-repair that sounds spontaneous and integrated into the flow. The prosody is consistent, and the correction is delivered with a believable shift in intonation, indicating a genuine change of mind." } ] }, "is_default": false }, { "variant_id": "revision_01_shipping_service", "proposed_instruction": "Initiate returns via UPS for everything in my last 4 amazon order. Wait, use FedEx for the returns, not UPS.", "long_pause_offsets": [ 67 ], "source_sha256": "63c50916764e61516c9695c04ab3c2f74cd9fa3616fd845bd39c8029390a1e78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the shipping service from UPS to FedEx, while retaining the scope of 'everything in my last four Amazon order'. The original instruction's 'FedEx' is replaced by 'UPS' in the first utterance, but the second utterance explicitly corrects it back to 'FedEx'. The phrase 'last four Amazon order' is preserved as spoken, matching the original's intent despite minor grammatical variation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/revision/revision_01_shipping_service/f25b7182d1d3c35f/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "f03f22f94b2aebdea7b7fc4d07db5259d90dc61aeb744e04ca0c55f525a3fc6d", "seconds": 18.16, "spoken_utterances": [ "Initiate returns via uh, let me think... UPS for everything in my ... last four amazon order. ", "Wait, use um, let me think... FedEx for the returns, not UPS." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions UPS but immediately corrects themselves to FedEx after a pause and self-correction. The final instruction clearly specifies FedEx, which matches the original reference. The revision type criterion is satisfied as the correction is applied to the earlier instruction.", "heard_transcript": "Initiate returns via uh, let me think, UPS for everything in my last four Amazon order. Wait, use um, let me think, FedEx for the returns, not UPS." }, "candidate_id": "wav_80a3581af235458d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear false start and self-correction. The speaker initially says 'UPS' but then corrects to 'FedEx'. The hesitation with 'let me think' and the pause before the correction sound natural. However, the correction is slightly less fluid than in audio_02, with a more abrupt shift in the phrase, making it feel a bit more like a mechanical edit than a fully spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_order_count", "proposed_instruction": "Initiate returns via FedEx for everything in my last 3 amazon order. Make that the last 3 amazon order, not 4.", "long_pause_offsets": [ 69 ], "source_sha256": "725aad8693a8efe90900e0b743b360cbed23517bf344913e3008522144c0f8ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial mention of 'four' to 'three', preserving the app (FedEx), action (initiate returns), and scope (everything in the last N amazon order). The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/revision/revision_02_order_count/a76564b6de690361/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "ed6c042fb251903d67ac0854e6eb8c9d6214ce08b96681c5df3c6a990a3a048c", "seconds": 17.12, "spoken_utterances": [ "Initiate returns via ah... FedEx for everything in my last three amazon order. ", "Make that the last ... three amazon order, not hmmm... four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last four Amazon order' but immediately corrects it to 'last three Amazon order, not four'. This is a clear revision where the final value replaces the initial incorrect one. The rest of the instruction, including the action (initiate returns), carrier (FedEx), and scope (everything in the order), remains consistent. The revision is spoken after a short pause and is unambiguous, satisfying the revision type criterion.", "heard_transcript": "Initiate returns via FedEx for everything in my last three Amazon order. Make that the last three Amazon order, not four." }, "candidate_id": "wav_62337a8f64da729b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker corrects 'last four' to 'last three'. While the correction is present, the delivery feels less natural due to the repetition of 'last three' and the drawn-out 'not' with a trailing 'um'. The prosody is less consistent, and the correction sounds more like a forced rephrasing than a spontaneous recollection, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "f79c3ccc9883e09c4c2b70b3e2e3003b8129febcd0233b804cf1fee91e415045", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/cancellation/cancel_024/2eb68924920dde89/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "16252e4923411b56b03236f42b055a7b7bf49ed39ceeeabdc8279066e0639a29", "seconds": 14.96, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order.", "Oh, wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then immediately corrects themselves with 'Oh wait, oh never mind, cancel the whole thing.' This constitutes a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or modification of the original task. Therefore, the final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order. Oh wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "08667f0889b34256e3c11ec667fe9374fc55c5e3d8378ba664afb3a6229ac647", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/ambiguous_halt/halt_010/60dd400e2742f25f/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "215f529929c11985dace545822b987e34c5256ad88b4e07d8726e039cce1e90e", "seconds": 13.48, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there any indication that the task should be resumed. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "1c4bd27_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c23a79719407f18738c0ae65565b4c8e193c3343b8bf4516194f87a1729b7452", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/1c4bd27_3/audio_natural/filler/original/fcb98c494a9a79b1/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "3388213c2bef6fec66adf0357fb6d1e3e8caa8631d68c0bcde532c2159983861", "seconds": 9.8, "spoken_utterances": [ "Initiate returns via uh... FedEx for everything in my hmmm... last four amazon order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Initiate returns via FedEx for everything in my last four Amazon order.' The word 'four' is pronounced clearly and matches the reference. There are no hesitations, self-corrections, or changes in the goal. The final intent is to execute the original instruction as heard, with no ambiguity or cancellation. The use of 'four' instead of '4' is a natural paraphrase and does not change the meaning. Therefore, the task is executed as intended.", "heard_transcript": "Initiate returns via FedEx for everything in my last four Amazon order." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_transaction_relation", "proposed_instruction": "How much money have I received from—no, I meant sent to others on venmo this month so far?", "long_pause_offsets": [], "source_sha256": "03ac185971da4a327733f6588b8a1eceedf811599db26984a063bf3344c006b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'received from' to 'sent to' within the same utterance, preserving the recipient ('others'), app ('Venmo'), and time frame ('this month so far'). The final intended action is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/false_start/false_start_01_transaction_relation/066fde155163fc61/speech.wav", "wer": 0.1, "wav_sha256": "110c29c1e290a19a23516e72bb90e9014d5c4b37c74f439cdfe10b638494f999", "seconds": 6.84, "spoken_utterances": [ "How much money have I received from—uh... no, I meant sent to others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received from' but immediately self-corrects with 'no, I meant sent to others'. This is a clear false start where the final intent matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How much money have I received from uh, no, I meant sent to others on Venmo this month so far?" }, "candidate_id": "wav_0e81ed5d52d0de93", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'received' and immediately corrects to 'sent' with a clear, spontaneous 'no, I meant' phrase. The prosody is consistent, the correction is integrated smoothly into the sentence, and the overall delivery sounds like a genuine moment of recall. The pause before the correction is brief and plausible, fitting a natural thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How much money have I sent to others on cashapp—I should say venmo this month so far?", "long_pause_offsets": [], "source_sha256": "d3085c491537b3ce44f80d651663e532f2456b4a6c04e67fb2405b7bc8b3d5ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the action, recipient scope, time frame, and intent. The final app is Venmo, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/false_start/false_start_02_app/590d841806704ace/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "8db61f04bd91ba3f9dbbeda7961599e6ec827723eaa6afeda47c3252c293d784", "seconds": 8.12, "spoken_utterances": [ "How much money have I sent to others on cashapp—uh... I should say venmo ... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash' but immediately self-corrects to 'venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('cash') is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "How much money have I sent to others on cash uh uh I should say venmo this month so far?" }, "candidate_id": "wav_89266392d290516d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear self-correction from 'cash' to 'Venmo', but the correction is slightly less fluid. The speaker hesitates with 'uh uh' before the correction, which is natural, but the phrase 'I should say Venmo' is a bit more formal and less conversational than a spontaneous 'no, I meant'. The prosody is otherwise natural, but the correction feels slightly more rehearsed compared to audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much money have I sent to others on venmo last month—sorry, that should be this month so far?", "long_pause_offsets": [], "source_sha256": "c671830c0a9d8d4da10714d90d6bf2bbb75afdda4e760a8526aedc0c072f7445", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month so far' within the same utterance, preserving the app, recipient scope, and intent to query sent amounts." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/false_start/false_start_03_time_reference/046b62a3fcb56513/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "c028e5d9b5afb0c3fcdfba32f0a83f213e783089d70ead20ad8da5d44509b692", "seconds": 10.0, "spoken_utterances": [ "How much money have I sent to others on um, let's see... venmo last month—um... sorry, that should be ... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month so far'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The corrected instruction matches the original reference exactly in action, target, and time scope.", "heard_transcript": "How much money have I sent to others on um, let's see, Venmo last month um, sorry, that should be this month so far?" }, "candidate_id": "wav_1ee4faf97d1eab60", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the most noticeable and less natural self-correction. The speaker starts with 'last month' and corrects to 'this month so far', but the correction is delivered with a longer pause and a more abrupt shift in intonation. The phrase 'sorry, that should be' sounds more like a formal acknowledgment than a spontaneous thought correction. The overall delivery feels slightly more rehearsed or less fluid compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How much money have I sent to others on this month so far? Oh, and one more thing: do this on Venmo.", "long_pause_offsets": [ 59 ], "source_sha256": "bb73b4fde097cc43c26e50de2943fd022c33fefcff0c9fb82698ed7c80922999", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app 'Venmo' to the initial question, which already includes the correct action, recipient scope, and time frame. The final goal combines both utterances without altering the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/addition/addition_02/17edd243d141dda7/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "64d394d29a434a2395144e057923c4e0f42fbad54c21d8ac71eb476fbd460edb", "seconds": 11.12, "spoken_utterances": [ "How much money have I sent to others on um... this month so far? ", "Oh, and one more thing: do this on ... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is a near-verbatim repetition of the original instruction with a minor hesitation ('um'). The speaker then adds 'do this on Venmo', which clarifies the app context. Since the original instruction already specified Venmo, this addition is redundant but does not contradict or alter the core task. The final intent is to execute the original request with the app context confirmed, satisfying the addition type criterion.", "heard_transcript": "How much money have I sent to others on um this month so far? Oh, and one more thing, do this on Venmo." }, "candidate_id": "wav_c191c25fcb7c4235", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The 'um' filler is brief and contextually appropriate, suggesting spontaneous speech. The addition ('Oh, and one more thing...') is delivered with a believable shift in intonation and a slight pause, followed by a coherent, additive clarification. The flow feels like a real-time thought process, with the second part clearly building on the first." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How much money have I others on venmo this month so far? Oh, I forgot to mention, I want to know how much money I've sent to others.", "long_pause_offsets": [ 57 ], "source_sha256": "acfa19d0746015cb73b625680f68a61acddc53f897c3b1591b04958e7581219f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing verb 'sent' to clarify the intended action, preserving the app, recipient scope, and time frame. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/addition/addition_01/07c767d8182fb8ab/speech.wav", "wer": 0.0, "wav_sha256": "94d49a0ebe5311238c8bbb307b6fbacd2ca8bdcffa9f37642b9d8e70055370f8", "seconds": 13.36, "spoken_utterances": [ "How much money have I others on venmo ... this month so far? ", "Oh, I forgot to mention, I want to know how much money I've ... sent to others." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance is incomplete and grammatically incorrect ('How much money have I others...'), but the speaker immediately corrects and adds the missing information ('I want to know how much money I've sent to others'). This constitutes an addition, where the final goal combines the original intent with the added clarification. The final audible intent matches the original instruction exactly, so the criterion is satisfied.", "heard_transcript": "How much money have I others on Venmo this month so far? Oh, I forgot to mention, I want to know how much money I've sent to others." }, "candidate_id": "wav_685b74fc863e1498", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is clear, but the addition ('Oh, I forgot to mention...') is delivered with a slightly longer, more hesitant pause and a less fluid transition. The phrasing 'I want to know how much money I've sent to others' feels more like a rephrased, self-corrected version rather than a natural additive thought. The overall delivery is less seamless than audio_02, with a slightly more robotic or rehearsed quality in the second part." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How much money have I sent to others on cashapp this month so far? Wait, the app should be venmo, not cashapp.", "long_pause_offsets": [ 67 ], "source_sha256": "09134c3a32857370cc1de42d632af7f81e3744a11b4241f7ff35ec2fae9a62bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Cashapp to Venmo, preserving the action, recipient scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/revision/revision_02_app/24e15dee027b1c62/speech.wav", "wer": 0.28, "wav_sha256": "c1f8f28639bdeac2e7ff7dbeed7e492b998321248f6f531bea343cb540a64d1b", "seconds": 11.88, "spoken_utterances": [ "How much money have I sent to others on uh... cashapp this month so far? ", "Wait, the app should be let's see... venmo, not cashapp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Cash App' but immediately corrects it to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction exactly, and the control intent is to execute the corrected task.", "heard_transcript": "How much money have I sent to others on the Cash App this month so far? Wait, the app should be uh, Venmo, not Cash App." }, "candidate_id": "wav_4b65683c2bb98446", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question. After a brief pause and self-correction ('Wait, the app should be...'), they smoothly revise the app name from 'Cash App' to 'Venmo' with a plausible hesitation and intonation shift. The correction is integrated into the same utterance, maintaining a continuous and realistic flow without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much money have I sent to others on venmo last month so far? Make that this month, not last month.", "long_pause_offsets": [ 65 ], "source_sha256": "44f4e0cf9bf2758a7420e8bc13f813a8c60b2497ae3354ba1822d01819313add", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app (Venmo), action (how much money sent), and recipient scope (to others)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/revision/revision_03_time_reference/d7e29896f9c27cd4/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "db1907b316bb347d183999aebd88966bcf3deeb9e77fe88e71e56e2a64f2a86d", "seconds": 12.72, "spoken_utterances": [ "How much money have I sent to others on um... venmo ah... last month so far? ", "Make that ... this month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, with only a minor grammatical variation ('Make that this month' vs. 'this month') that does not alter the meaning.", "heard_transcript": "How much money have I sent to others on um Venmo uh last month so far? Make that this month, not last month." }, "candidate_id": "wav_87bcb6ba54652920", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a natural-sounding question but makes a revision after a longer pause. The correction ('Make that this month, not last month') is delivered as a separate sentence, which is less fluid than a self-repair within the original utterance. While the hesitation and correction are realistic, the break between the initial request and the revision feels slightly more disjointed compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_transaction_relation", "proposed_instruction": "How much money have I received from others on venmo this month so far? Sorry, I meant sent to others, not received from.", "long_pause_offsets": [ 71 ], "source_sha256": "270aaa867de254046d7dbbee3eb217bd8bb689f5b025e388641fae6bcf6f554a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the action from 'received from' to 'sent to', preserving the app, recipient scope, and time frame." }, "audio": { "status": "pending", "spoken_utterances": [ "How much money have I received from others on venmo hmmm... this month so far? ", "Sorry, I meant uh... sent to others, not received from." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "6f89e7651ff5f9eb98e8e3badb62ecd3c1251910314bd11d757c49a73c20bc36", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/cancellation/cancel_025/bee83857604eacfd/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "feb6ee6f25f0ec1ec1155888838d2c385881f39b150cfafb08154d8eeb1ebf25", "seconds": 10.32, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?", "Hold on, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, forget that request. Don't do it,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far? Hold on, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "4154fd0539e53da0305565773e44b8dcf2fe2fb28885320ec91688730234b545", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/ambiguous_halt/halt_006/08fd55429975b668/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "67516f72acabe00be9bb3895ee5ed825401b8bd2d79467d7ec8b5bde736af870", "seconds": 8.52, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the phrase 'Just a moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the agent is expected to hold the task until the user is ready to proceed.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far? Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "973b92b1e04e9339def0cffb67d2b7871a831481a817425b4a1568ff5b4e535d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_1/audio_natural/filler/original/01c50ce8b29cb37d/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "b9dd3bad3ca0135e419716005533b8ebf8cd6a8aed6deea574db62742cf40eb7", "seconds": 5.44, "spoken_utterances": [ "How much money have I sent to others on uh... Venmo ... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'on uh Venmo', where 'uh' is a filler word. The rest of the sentence is clear and matches the original instruction. The final intent is to execute the query as stated, with no change in action, target, or value. The filler does not alter the task goal.", "heard_transcript": "How much money have I sent to others on uh Venmo this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_transaction_relation", "proposed_instruction": "How much money have I sent to—no, I mean received from others on venmo this month so far?", "long_pause_offsets": [], "source_sha256": "72cf356fe718045655bef010bebf0d269a110527227bee1711dd85c95a08f77e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent to' but immediately corrects to 'received from' using 'no, I mean'. The final action is 'received from', and all other components (money, app, time frame) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/false_start/false_start_01_transaction_relation/bfdb7b47b02a3e1a/speech.wav", "wer": 0.0, "wav_sha256": "8d93e8bfc64abfb1423ca9f4327756e85673bc73e15eb14835b13f50d2003bb3", "seconds": 11.0, "spoken_utterances": [ "How much money have I sent to—uh... no, I mean um... received from others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'received'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How much money have I sent to uh, no, I mean um, received from others on Venmo this month so far?" }, "candidate_id": "wav_e9f1a829d448bfeb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'How much money have I sent to...' but immediately self-corrects with 'No, I mean...' and smoothly transitions to the correct request. The intonation and pacing are consistent with spontaneous speech, and the correction feels like a genuine moment of realization rather than a rehearsed edit. The delivery is clear, with appropriate pauses and filler words ('uh', 'um') that enhance realism." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How much money have I received from others on cashapp—I should say venmo this month so far?", "long_pause_offsets": [], "source_sha256": "4e93ea0996435f5f70683abb633217af89ee58579f8e230c73ea34d96bac9bba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the rest of the query including the app, recipient scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/false_start/false_start_02_app/79c85a93881ee3c8/speech.wav", "wer": 0.1, "wav_sha256": "b0b4af4154b49e9b5ba12974b61bb27ce1c805ac19b4169ba2b29b93755a81c2", "seconds": 7.48, "spoken_utterances": [ "How much money have I received from others on cashapp—um... I should say uh... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' after a brief hesitation. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "How much money have I received from others on Cash App this month so far? I should say uh Venmo this month so far?" }, "candidate_id": "wav_972ee687a719e949", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start and self-correction, but the correction is slightly less fluid than audio_02. The speaker starts with 'Cash App' and then corrects to 'Venmo'. The 'uh' filler before the correction is natural, but the overall pacing feels a bit more hesitant and less confident than audio_02. The prosody is still realistic, but the flow between the initial error and the correction is slightly less seamless." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much money have I received from others on venmo last month—scratch that, make it this month so far?", "long_pause_offsets": [], "source_sha256": "8762aa5dd785634037d64119152c4026bf859140168f9ac7c5399426e351db85", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month so far'. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/false_start/false_start_03_time_reference/07fb9ccc1d6d2b3e/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "0596713520c30623d0dc28e5ada0397c224fda067bbf40df6f49f9b3e074ed71", "seconds": 8.52, "spoken_utterances": [ "How much money have I received from others on um... venmo last month—uh... scratch that, make it this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects with 'scratch that, make it this month so far'. This is a clear false start where the final value replaces the initial one. The corrected instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "How much money have I received from others on um Venmo last month uh scratch that, make it this month so far?" }, "candidate_id": "wav_cd60a7342af86daa", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start and correction, but the delivery is less natural. The speaker starts with 'How much money have I received from others on Venmo last month' and then says 'scratch that, make it this month so far'. The phrase 'scratch that' is a bit abrupt and less conversational than a typical self-correction. The correction is clear, but the overall prosody feels slightly more robotic, and the transition lacks the smoothness of the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How much money have I received on venmo this month so far? Oh, I forgot to mention, only include money received from others.", "long_pause_offsets": [ 59 ], "source_sha256": "4cdea874ea06ea0acb2d889405d8c15de28cb836a9dd8103d13f1f751815fe5f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that narrows the scope to money received from others, which is consistent with the original instruction's intent and does not contradict it." }, "audio": { "status": "pending", "spoken_utterances": [ "How much money have I received on venmo um... this month so far? ", "Oh, I forgot to mention, only include money received uh... from others." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "How much money have I received from others on this month so far? I should also mention, this is specifically on Venmo.", "long_pause_offsets": [ 65 ], "source_sha256": "e30ca6b8cc7421c441ecb8738a392a12b74d940b5e0b09bc9cf59b1825d08234", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Venmo' to the initial question, which was otherwise complete. The final goal combines the original intent with the added specificity." }, "audio": { "status": "pending", "spoken_utterances": [ "How much money have I received from others on ... this month so far? ", "I should also mention, this is specifically on uh... Venmo." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "21abae1_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How much money have I received from others on cashapp this month so far? Let me correct the app: use 'venmo', not 'cashapp'.", "long_pause_offsets": [ 73 ], "source_sha256": "40e4bd010fe094d04810ee2ba6150d95e6c163a66c2a20e6c1cee28b5b4224f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'cashapp' to 'venmo', preserving the rest of the query including the action, recipient scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/revision/revision_02_app/edf1c23d76b97d08/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "40f1a4471e12532f10ae8a01365f1e927ebfa503a9ab635a4cdbfe88106f9d6e", "seconds": 13.52, "spoken_utterances": [ "How much money have I received from others on cashapp ... this month so far? ", "Let me correct the app: use 'let me think... venmo', not 'cashapp'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task for Cash App but then corrects it to Venmo after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the amount received on Venmo, which matches the original reference instruction. All other constraints (e.g., 'this month so far') are preserved.", "heard_transcript": "How much money have I received from others on Cash App this month so far? Let me correct the app, use, let me think, Venmo, not Cash App." }, "candidate_id": "wav_d111967d23b33225", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. The speaker begins with a clear request, then hesitates with 'let me think' before self-correcting the app name from 'cash app' to 'venmo'. The correction is integrated smoothly into the utterance with a natural pause and filler, and the prosody reflects genuine recollection rather than a rehearsed or robotic delivery. The rhythm and intonation are consistent with a person formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_01_transaction_relation", "proposed_instruction": "How much money have I sent to others on venmo this month so far? Wait, the transaction relation should be 'received from', not 'sent to'.", "long_pause_offsets": [ 65 ], "source_sha256": "6663bce580f3d5507cb56dfa2a30a04dccfd0c6e0e4743e2fd244c03c607d599", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the transaction relation from 'sent to' to 'received from', preserving the app, time frame, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/revision/revision_01_transaction_relation/3e3f7af1d027095e/speech.wav", "wer": 0.0, "wav_sha256": "0b7ce2c010364826acb5ff88d3d3d6a642c3ddff0b7a2cf8978335353930e268", "seconds": 13.12, "spoken_utterances": [ "How much money have I sent to others on venmo uh... this month so far? ", "Wait, the transaction relation should be '... received from', not 'sent to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about money sent to others, but then corrects the transaction relation to 'received from' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "How much money have I sent to others on Venmo uh this month so far? Wait, the transaction relation should be received from, not sent to." }, "candidate_id": "wav_023cf6b2df6fe663", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a plausible revision with a self-correction from 'sent to' to 'received from'. The speaker's tone and pacing are natural, and the correction is delivered as a spontaneous realization. However, the initial request is slightly less fluent than audio_03, with a longer pause and a more abrupt shift in phrasing. The correction is clear but less seamlessly integrated than in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much money have I received from others on venmo last week so far? Change the time reference to 'this month', not 'last week'.", "long_pause_offsets": [ 70 ], "source_sha256": "da5f0d954bd416774d85f2c8dbb58a80093de12d384942f170668f068de30b68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last week' to 'this month', preserving the rest of the query including the app, action, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/revision/revision_03_time_reference/a47643783d329eae/speech.wav", "wer": 0.04, "wav_sha256": "c9716bedf4630ebd4fbd35e6dc1029280713a14b3ffc24d903e6470cb4b0b007", "seconds": 13.56, "spoken_utterances": [ "How much money have I received from others on ... venmo last week so far? ", "Change the time reference to 'um... this month', not 'last week'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'last week' but then corrects the time reference to 'this month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed.", "heard_transcript": "How much money have I received from others on Venmo last week so far? Change the time reference to um this month, not last week." }, "candidate_id": "wav_94030385f76b5812", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but feels less natural due to the phrasing of the correction. The speaker says 'change the time reference to um this month not last week', which sounds more like a direct instruction to a system than a spontaneous thought. The 'um' filler is awkwardly placed, and the correction is delivered in a more declarative, less conversational tone. The overall prosody is flatter and less expressive compared to the other two, making it sound less like a genuine human request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "9fe44cfaa2a19046190e6ce6fd591f17ce73c86688fdc1b8c003cf41ed0d2656", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/cancellation/cancel_018/ace2e00bd4bec21a/speech.wav", "wer": 0.0, "wav_sha256": "143d486e6af138dfb7ebc80721bdd8d368c7af09ec364ca73a42e9f9b7a3955b", "seconds": 12.88, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?", "Hang on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question but then clearly revokes the entire task with 'Hang on, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the original request. The final interjection is a complete cancellation command, not a pause or ambiguous halt.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far? Hang on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "7316d0e2d169ad6f7f714c177f408c3ac4cfdb87f33de44f8faac4d778dedb4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/ambiguous_halt/halt_033/72eca3ee8822509f/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "7282b3e77c0bcb52c006adb202ad8550430de1f3fcc1cf9d8202cc7e6ccbd7bf", "seconds": 11.64, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on, do something,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary stop rather than a cancellation or execution. The initial request is not revoked, and the speaker does not resume or correct it, so the final intent is to pause.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far? Hold on, do something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "54cbd3b2db0b1f77c3c647967b7ab38ff6498a71abf1ecb5baceed709cf1940e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_2/audio_natural/filler/original/35ae898b7e7a346a/speech.wav", "wer": 0.0, "wav_sha256": "f81fc62e780d2d43c27d75e34fb93f7d6af902a8b216e1945ac2ac10653d87fc", "seconds": 7.44, "spoken_utterances": [ "How much money have I received from others on ... Venmo uh... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent remains to ask for the amount received from others on Venmo this month. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How much money have I received from others on Venmo uh this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How much money have I sent to or received from others on cashapp—I should say venmo this month so far?", "long_pause_offsets": [], "source_sha256": "3c48c4d005089474fde64c81edb2c17a6094990a3ed2aef0fa54a03257e52e5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the full scope of the original request including the app, action, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/false_start/false_start_02_app/6895fd81e02718ff/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "d415fb78f70a8e67424679e8d3839b6ba8a19c7b21972248c0d303ba66203d65", "seconds": 12.88, "spoken_utterances": [ "How much money have I ... sent to or received from others on cashapp—hmmm... I should say venmo let's see... this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is executed with the correction applied. The rest of the instruction remains unchanged.", "heard_transcript": "How much money have I sent to or received from others on Cash App, uh, I should say Venmo, let's see, this month so far?" }, "candidate_id": "wav_bc4862140713270f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('How much money have I...') and immediately self-corrects from 'CashApp' to 'Venmo' with a plausible 'I should say' repair. The intonation is conversational, and the final phrase 'this month so far' is delivered smoothly without further interruption, suggesting a single, continuous thought process." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much money have I sent to or received from others on venmo last month—no, that should be this month so far?", "long_pause_offsets": [], "source_sha256": "be3b5089953bfd270155916090bb6830194dfe4019a8a5491200f7a62feeebb6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month so far'. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/false_start/false_start_03_time_reference/9d03db381ff67bad/speech.wav", "wer": 0.12, "wav_sha256": "5df343f241d15461b1d77c4392d47bfde87bcc70bd482d8aeee69576a8d3d0d7", "seconds": 13.76, "spoken_utterances": [ "How much money have I hmmm... sent to or received from others on ... venmo last month—uh... no, that should be this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month so far'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('last month') is discarded. The final audible task matches the original instruction exactly.", "heard_transcript": "How much money have I sent to or received from others on Venmo last month? No, that should be this month so far." }, "candidate_id": "wav_60c923c3f6035505", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a more complex false start with a throat clear ('uhem'), a self-correction from 'last month' to 'this month so far', and a final 'uh' hesitation. While the repair is realistic, the multiple pauses and vocal fillers make the flow slightly less fluid than audio_01, though still credible as spontaneous speech." } ] }, "is_default": false }, { "variant_id": "false_start_01_transaction_relation", "proposed_instruction": "How much money have I sent from or received to—sorry, I mean sent to or received from others on venmo this month so far?", "long_pause_offsets": [], "source_sha256": "8d21269bc810b8615505a1b4c19859775e424d4bdaf354e665ef8558ec19b92d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects the prepositional phrase 'sent from or received to' to 'sent to or received from', which matches the original instruction. The action, app, and time frame remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/false_start/false_start_01_transaction_relation/1fcb7b311caf6da0/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "4ea399db16a4be70d96dd95aed06a6124d67f3b804ee3fdea1b31abca7a53348", "seconds": 12.88, "spoken_utterances": [ "How much money have I sent from or received to—ah... sorry, I mean sent to or received from others on let's see... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent from or received to a' which is a false start. They immediately self-correct to 'sent to or received from others' and continue with the rest of the original instruction. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much money have I sent from or received to a, sorry, I mean sent to or received from others on let's see, Venmo this month so far?" }, "candidate_id": "wav_fb57e2de13884c5e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a slightly awkward phrasing ('sent from or received to'), then corrects to 'sent to or received from others'. The correction is grammatically sound but feels more like a rehearsed edit than a natural slip. The 'let's see' pause is placed awkwardly mid-sentence, and the overall delivery is more robotic and less conversational than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How much money have I sent to or received from others on this month so far? I should also mention, this is on Venmo.", "long_pause_offsets": [ 76 ], "source_sha256": "44234c2a9c1ce3c5c81b029807cd578cc192d83224db9bf30526a5f0f0999518", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the app as Venmo, which was omitted in the initial utterance. The core question about money sent/received on Venmo this month remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/addition/addition_02/cd0ec0df7e37e94e/speech.wav", "wer": 0.04, "wav_sha256": "20e761c652f45c858129f72be33e81eb5949c6916336daf05bb45b16cbb5d42b", "seconds": 12.48, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... this month so far? ", "I should also mention, this is on Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and completed. The speaker adds the detail 'this is on Venmo' after a pause, which clarifies the platform. This is a valid addition that narrows the scope of the original request without changing its core intent. The final goal matches the original instruction with an added constraint, satisfying the addition type criterion.", "heard_transcript": "How much money have I um sent to or received from others on this month so far? I should also mention, this is on Venmo." }, "candidate_id": "wav_24346d54ae797a38", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How much money have I others on venmo this month so far? Oh, I forgot to mention, include both money sent to and received from others.", "long_pause_offsets": [ 57 ], "source_sha256": "cf60040ce7bf4e2eb1103655ff8fc337afadcabc1992f85f5a8f5eeaf5a37b4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'include both money sent to and received from others', which completes the original instruction's intent. The final goal matches the original exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "How much money have I others on uh... venmo this month so far? ", "Oh, I forgot to mention, include both money ... sent to and received from others." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_transaction_relation", "proposed_instruction": "How much money have I sent to or received by others on venmo this month so far? Wait, it should be 'sent to or received from', not 'sent to or received by'.", "long_pause_offsets": [ 80 ], "source_sha256": "d9c67968d8e3eb6fa64969cff54d90cf0870b382d0b499252d6fb9a79f1c6682", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the preposition from 'by' to 'from', which is the only difference. The rest of the query, including the app (Venmo), time frame (this month so far), and action (sent to or received from) remains consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/revision/revision_01_transaction_relation/634b4db091ddb2af/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "76e7ad92051d863451fccc2b4a7be1b9ef7ac9c292f09dd34e4cfeff5fb22624", "seconds": 16.12, "spoken_utterances": [ "How much money have I um... sent to or received by others on venmo uh... this month so far? ", "Wait, it should be 'sent to or received from', not 'sent to or received by'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received by' but immediately corrects it to 'received from' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "How much money have I um sent to or received by others on Venmo uh this month so far? Wait, it should be sent to or received from, not sent to or received by." }, "candidate_id": "wav_8790ab93560701a5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a plausible self-correction. The initial phrasing 'sent to or received by' is slightly awkward, leading to a believable hesitation and revision to 'sent to or received from'. The correction is smoothly integrated with a natural 'wait' and clear intonation shift, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much money have I sent to or received from others on venmo last month so far? Make that 'this month', not 'last month'.", "long_pause_offsets": [ 82 ], "source_sha256": "1b8a69270ce53bf83af71418a6415c6dd5606b032d0320fa1ebe17bb6a74c7c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app, action, and scope of money sent/received." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/revision/revision_03_time_reference/0248e574cbd49282/speech.wav", "wer": 0.04, "wav_sha256": "373500242bee41b7a09c54cb062f55636160edadbed19edecbc9667735c9c24a", "seconds": 12.92, "spoken_utterances": [ "How much money have I uh... sent to or received from others on venmo last month so far? ", "Make that '... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "How much money have I uh sent to or received from others on Venmo last month so far? Make that this month, not last month." }, "candidate_id": "wav_27dad2174f9c00cb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request and then corrects 'last month' to 'this month'. The correction is delivered with a natural 'make that' and a clear shift in intonation. The overall flow is smooth, and the pause before the correction feels like a genuine moment of recall. The only minor issue is a slightly abrupt transition to the correction, but it remains highly realistic." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How much money have I sent to or received from others on cashapp this month so far? Change the app to 'venmo', not 'cashapp'.", "long_pause_offsets": [ 84 ], "source_sha256": "6481dfdb8b85e09e20413870c049078d1a1e37f19e9e042c7e5e09e571e9c3f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'cashapp' to 'venmo', preserving the rest of the query including the action, direction, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/revision/revision_02_app/4014bd3d45ea5ae0/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "ac7a820471b37fc9f4fa9c960382d34e4094aae6a413580c599788c450045a41", "seconds": 17.84, "spoken_utterances": [ "How much money have I uh, let's see... sent to or received from others on um... cashapp this month so far? ", "Change the app to 'hmmm... venmo', not 'cashapp'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Cash App' but immediately corrects it to 'Venmo' with the explicit phrase 'Change the app to Venmo, not Cash App.' This constitutes a revision where the final audible intent is to query the Venmo app, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "How much money have I uh let's see sent to or received from others on um Cash App this month so far? Change the app to Venmo, not Cash App." }, "candidate_id": "wav_d6df160c8133d238", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a natural hesitation ('uh') and a filler ('let's see'), which adds realism. However, the correction from 'Cash App' to 'Venmo' is delivered with a flat, robotic intonation and lacks the natural prosodic shift expected in a spontaneous correction. The overall delivery feels less fluid and more like a rehearsed or read-aloud task, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "3f8bd089df016a19eb50420c5620e8f7e39df554f48608d18df35c70af5622fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/cancellation/cancel_001/4dc57d4104336130/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "dcdd41a6ae8f706d94e88934e835e3be672271340f55baf7c791c52474ba123e", "seconds": 13.36, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately self-corrects with 'Oh wait, cancel that,' which is a clear and final cancellation of the entire task. The audio ends with this cancellation, so no execution is expected. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far? Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "bd17ed2cb49c844644f8fb71f6264f12c7465be5226e448a79c917890bd07a8c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/ambiguous_halt/halt_002/d983b3d27993e954/speech.wav", "wer": 0.05, "wav_sha256": "fcd1ebc9dc4d4fba0ceeb159b5c2e2ab3d7c400fc37687af6721105c8447f6c6", "seconds": 12.64, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then audibly says 'Hold on,' which is a clear request to pause or wait. There is no cancellation or instruction to resume. The final interjection is a pause command, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "21abae1_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "17377c100a87dcd38fe55793df4fa7b3090be93a86c4594d3ae379bb149264a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/21abae1_3/audio_natural/filler/original/ed8e5634c86f86db/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "161521122ed0c9c21b5185746415e41334eb347830720634a6a2a0a94e8d60c8", "seconds": 9.36, "spoken_utterances": [ "How much money have I um... sent to or received from others on ... venmo this month so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um' as a hesitation filler but completes the original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task is executed as intended.", "heard_transcript": "How much money have I um sent to or received from others on Venmo this month so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I received—no, make that sent this month, have in total?", "long_pause_offsets": [], "source_sha256": "bcce32c5aa049d4d7cd4028018fb7ba3416227b3c5f45d4e1e14626d82a6222e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'received' but immediately corrects it to 'sent' within the same utterance. The final intended direction of the transactions is 'sent', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/false_start/false_start_03_transaction_direction/e44e851755944f29/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "0a9f2a1edb47e45e5da5c1afb0f45bb6aae5d9c134b7fd5e9cedc80e28c88e26", "seconds": 11.12, "spoken_utterances": [ "How many likes did uh... all Venmo transactions, let me think... I received—um... no, make that sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I received' but immediately self-corrects to 'make that sent' after a pause. According to the 'false_start' type criterion, the final corrected value ('sent') should be applied, and the original incorrect value ('received') should be discarded. The final intent correctly reflects the original instruction with the corrected verb. All other constraints (Venmo transactions, this month, total likes) are preserved.", "heard_transcript": "How many likes did uh all Venmo transactions, let me think. I received um, no, make that sent this month, have in total?" }, "candidate_id": "wav_16bdab9cb1a9d48e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. The speaker begins with 'likes' and 'Venmo transactions', then hesitates with 'uh' and 'let me think', before correcting 'received' to 'sent'. The correction is integrated smoothly into the sentence with a natural 'no, make that' phrase, and the prosody includes believable pauses and filler words that suggest genuine thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How many likes did all CashApp—I should say Venmo transactions, I sent this month, have in total?", "long_pause_offsets": [], "source_sha256": "f8c4b32bd724af3cec65e2d13f439c046c1d47d6293c5fb888d43aca22bc5fe6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the rest of the query including the action (counting likes), the app (Venmo), and the time scope (this month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/false_start/false_start_02_app/a7ed39047dc4a4b0/speech.wav", "wer": 0.3, "wav_sha256": "458be1b0c49a3d3b9c824d2a084e1792987407513dda25180e18662697f4445c", "seconds": 9.4, "spoken_utterances": [ "How many likes did all CashApp—uh... I should say Venmo transactions, hmmm... I sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'I should say Venmo transactions'. This is a false start where the initial value is corrected. The final audible intent matches the original instruction with the correct app name. The control intent is to execute the corrected task.", "heard_transcript": "How many likes did all Cash App, uh, I should say Venmo transactions, I sent this month, have in total?" }, "candidate_id": "wav_643eb9a77086f9f7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'cash app' corrected to 'Venmo transactions'. The speaker uses a natural 'huh' after the correction, indicating a moment of reflection. The prosody is conversational, and the correction is integrated into the sentence without awkwardness. However, the initial error ('cash app') is less plausible than 'received/sent', making the correction slightly less natural than in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent last month—sorry, that should be this month, have in total?", "long_pause_offsets": [], "source_sha256": "058de81cdf7f633758788aa97537d983814545b93613ed942955c9c9424b4f7e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month' within the same utterance. The final time reference is 'this month', matching the original instruction. All other elements, including the action, subject, and app, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/false_start/false_start_04_time_reference/1e7763424cad4df9/speech.wav", "wer": 0.0, "wav_sha256": "7a5a7df5c7ac96b82e55b3602c76df550ae63fca912fabe6789e50f38b673e80", "seconds": 7.68, "spoken_utterances": [ "How many likes did all Venmo transactions, uh... I sent last month—um... sorry, that should be ... this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value ('this month') should be applied, and the original incorrect value ('last month') should be discarded. The final intent matches the original instruction with the corrected time frame.", "heard_transcript": "How many likes did all Venmo transactions, uh, I sent last month, um, sorry, that should be this month, have in total?" }, "candidate_id": "wav_34f061b5d0ec377e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker first says 'last month' and then corrects to 'this month'. The correction is made with 'sorry, that should be', which is a common and natural phrase. However, the initial request is slightly more complex and the correction is less integrated into the sentence structure, making the flow feel a bit more disjointed compared to audio_03 and audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_01_engagement_metric", "proposed_instruction": "How many comments—sorry, I meant likes did all Venmo transactions, I sent this month, have in total?", "long_pause_offsets": [], "source_sha256": "870ed3046849f605ca4001fb8f7c9bfb8425ae8e91ad5a60ee6f953c416be17e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'comments' to 'likes' within the same utterance, preserving the rest of the query including the app (Venmo), the scope (transactions I sent), the time frame (this month), and the aggregation (total)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/false_start/false_start_01_engagement_metric/7e937fbd4729f6ce/speech.wav", "wer": 0.0, "wav_sha256": "678ebe531465297f11505de09f9a2b25860cabfc9b2c723b905433d22f7ffa8c", "seconds": 9.6, "spoken_utterances": [ "How many comments—uh... sorry, I meant ... likes did all Venmo transactions, I sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'comments' but immediately self-corrects to 'likes' with 'sorry, I meant likes'. According to the 'false_start' type criterion, the final corrected value ('likes') should be applied, and the original incorrect value ('comments') should be discarded. The rest of the instruction, including the target ('Venmo transactions I sent this month') and the action ('how many... in total'), remains unchanged. The final audible intent matches the corrected instruction, so the task should be executed as requested.", "heard_transcript": "How many comments uh, sorry, I meant likes did all Venmo transactions I sent this month have in total?" }, "candidate_id": "wav_7b33fdf5813f6231", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start where 'comments' is corrected to 'likes'. While the correction is grammatically correct, the phrasing 'sorry, I meant' is less natural in this context compared to 'no, make that' or 'sorry, that should be'. The prosody is somewhat flat, and the correction feels more like a simple substitution rather than a spontaneous thought process, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How many likes did all Venmo transactions, this month, have in total? I didn't mention this, but only count the ones I sent.", "long_pause_offsets": [ 70 ], "source_sha256": "c779f33cfbb5ff03fbdb84d7ade8b89f66838c852d62e94cd4d0f04d4e5e6816", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that only counts transactions I sent, which is consistent with the original intent of focusing on the user's own transactions. The app (Venmo), time frame (this month), and metric (total likes) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/addition/addition_02/c950c5161da81de2/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "51641c6be5135fb875099d566847cf428069bb63e130558119ac57e60dc8cea1", "seconds": 12.04, "spoken_utterances": [ "How many likes did um... all Venmo transactions, this month, have in total? ", "I didn't mention this, but only count the ones ... I sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is a slightly garbled version of the reference, but the final addition 'only count the ones I sent' correctly narrows the scope to transactions sent by the speaker. The addition is spoken after a pause and modifies the original request, satisfying the 'addition' type criterion. The final intent matches the reference with the added constraint.", "heard_transcript": "How many likes did um all Venmo transactions this month have in total? I didn't mention this, but only count the ones I sent." }, "candidate_id": "wav_e92c86078c2b7385", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um') and a plausible self-correction ('I didn't mention this, but...'). The addition of the restriction 'only count the ones I sent' is smoothly integrated with a believable pause and change of thought, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "How many likes did all Venmo transactions, I sent, have in total? To narrow that down, only count the ones from this month.", "long_pause_offsets": [ 66 ], "source_sha256": "e6fb15b9b35eea4688cb9267894c2090563b28cd421deae1d64131fe6f59b827", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'this month' to the initial broad request, refining the scope without changing the action or other components." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/addition/addition_03/34e51a59804925d8/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "2b742792fff47fc5abf8728a614998da4f149391547797322a67cea5b16ef8af", "seconds": 13.28, "spoken_utterances": [ "How many likes did um... all Venmo transactions, uh... I sent, have in total? ", "To narrow that down, only count the ones from let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How many likes did all Venmo transactions, I sent this month, have in total?'. The speaker then adds a clarifying constraint: 'To narrow that down, only count the ones from this month.' This addition refines the scope of the initial request to only include transactions from 'this month'. The final intent combines the original request with the added constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "How many likes did um all Venmo transactions uh I sent have in total? To narrow that down, only count the ones from let's see, this month." }, "candidate_id": "wav_85d0fe24d2b8ae75", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request and adds a restriction ('only count the ones from...') with a natural pause and filler ('let's see'). The flow is mostly smooth, though the initial phrase 'how many likes did... all Venmo transactions...' is slightly awkward, and the addition feels a bit tacked on rather than fully integrated." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "How many likes did Venmo transactions, I sent this month, have in total? Oh, I forgot to mention, include all Venmo transactions.", "long_pause_offsets": [ 73 ], "source_sha256": "5bc7b0ed493553dd45590af575e7c2058d3eeb4af010231ed422b6b71793c5c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing scope 'all Venmo transactions', which clarifies the initial request. The initial utterance already specified 'Venmo transactions' and 'this month', so the addition completes the intended scope without contradicting or removing any prior condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/addition/addition_01/0d0f7045aa95f097/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "7828e539713700128d31a1ea5a63f1ed00d05e3f0adcfc48f7295119927585bc", "seconds": 14.96, "spoken_utterances": [ "How many likes did Venmo transactions, hmmm... I sent this month, have in total? ", "Oh, I forgot to mention, include um, let me think... all Venmo transactions." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'Venmo transactions I sent this month' and then adds 'all Venmo transactions' after a pause. The addition clarifies that the scope should include all Venmo transactions, not just those sent by the speaker. The final intent combines the original request with the added scope, resulting in a broader but coherent task that matches the reference instruction. The addition is clear and does not contradict the original request.", "heard_transcript": "How many likes did Venmo transactions, hmm, I sent this month, have in total? Oh, I forgot to mention, include um, let me think, all Venmo transactions." }, "candidate_id": "wav_a1af1a7f1522398d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The delivery is less realistic due to unnatural phrasing ('how many likes did Venmo transactions... I sent this month have in total') and a jarring, repetitive structure. The speaker adds a restriction ('include... let me think... all Venmo transactions') after already having mentioned 'this month', creating a confusing and illogical sequence. The prosody is inconsistent, and the overall flow feels disjointed and less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent last month, have in total? Make that this month, not last month.", "long_pause_offsets": [ 77 ], "source_sha256": "eb0e6b456516ea884f12292953a28234fb3f5e2512aae00f5f85e528503306af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app (Venmo), the action (counting likes), and the scope (transactions I sent)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/revision/revision_04_time_reference/79218930e4585eb1/speech.wav", "wer": 0.04, "wav_sha256": "a950154c7b6637bc3ded845874e5b8ff172d6e9bfa2c9ea256ed9aa03d56f963", "seconds": 12.44, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I sent last month, have in total? ", "Make that um, let me think... this month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a self-correction phrase ('let me think'). According to the revision type criterion, the correction spoken after a pause (or self-correction) overrides the earlier instruction. The final audible intent is to ask about 'this month', which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "How many likes did all Venmo transactions, I sent last month, have in total? Make that, um, let me think, this month, not last month." }, "candidate_id": "wav_2ef414a477028280", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects 'last month' to 'this month' with a plausible 'um' and 'let me think' filler. The correction is smoothly integrated into the sentence, maintaining a consistent rhythm and intonation. The prosody and pacing are highly realistic, resembling a spontaneous thought process without robotic or rehearsed qualities." } ] }, "is_default": true }, { "variant_id": "revision_01_engagement_metric", "proposed_instruction": "How many comments did all Venmo transactions, I sent this month, have in total? Wait, I meant likes, not comments.", "long_pause_offsets": [ 80 ], "source_sha256": "be692b9647e01d5c6935a77156f2f034cae3e3d08e606b5fdfa426b0647bd590", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the metric from 'comments' to 'likes', preserving the app (Venmo), scope ('all transactions I sent this month'), and the request for a total count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/revision/revision_01_engagement_metric/1add7f5f5c043a07/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "6bf2a376d43c5badd7e77ff85dad0096c0a4796bfed45de6e4d171576f84e92d", "seconds": 12.8, "spoken_utterances": [ "How many comments did all Venmo transactions, let me think... I sent this month, have in total? ", "Wait, I meant um... likes, not comments." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'comments' but immediately self-corrects to 'likes' after a pause and the phrase 'Wait, I meant'. This constitutes a revision where the final intent is to inquire about 'likes'. The rest of the instruction, including the app (Venmo), the action (transactions I sent), and the time frame (this month), remains unchanged. The revision is clear and the final control intent is to execute the corrected request.", "heard_transcript": "How many comments did all Venmo transactions, let me think, I sent this month have in total? Wait, I meant um, likes not comments." }, "candidate_id": "wav_1c2d29acb66dba72", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'let me think' before correcting 'comments' to 'likes'. The correction is natural, with a slight pause and 'wait, I meant' phrasing that reflects a genuine change of mind. The flow is slightly less smooth than audio_01 due to the longer filler and more abrupt correction, but still highly realistic and credible." } ] }, "is_default": false }, { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I received this month, have in total? I misspoke about the direction: it should be sent, not received.", "long_pause_offsets": [ 81 ], "source_sha256": "bbfe4be844d7e37ed993000b3e56f4a4ad5162ab0c973fda6bac1dc9b0d6ff73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the transaction direction from 'received' to 'sent', which is the only mismatched value. The app (Venmo), time frame (this month), and the action (counting likes) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/revision/revision_03_transaction_direction/f9f87b8f167fc7d1/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "adc5d04ee84168139e60c8db85284377491b5b694fb47fffcb6537f4ca38348b", "seconds": 14.88, "spoken_utterances": [ "How many likes did ... all Venmo transactions, um... I received this month, have in total? ", "I misspoke about the direction: it should be hmmm... sent, not received." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I received this month' but immediately corrects themselves with 'I misspoke about the direction, it should be sent, not received.' This constitutes a revision where the final intent is the corrected version. The original instruction matches the final audible intent, so the task is executed as revised.", "heard_transcript": "How many likes did all Venmo transactions, um, I received this month have in total? I misspoke about the direction, it should be sent, not received." }, "candidate_id": "wav_0feabcd5d1148f9d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request, then corrects 'received' to 'sent' with a self-aware 'I misspoke' and a thoughtful 'hmm'. The correction is believable, but the pacing is slightly slower and the intonation more deliberate, making it sound a bit more rehearsed. The filler and pause are natural, but the overall delivery feels slightly less spontaneous than the top two." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How many likes did all PayPal transactions, I sent this month, have in total? Let me correct the app: use Venmo, not PayPal.", "long_pause_offsets": [ 78 ], "source_sha256": "d5528fa4c74829e48876a6bc447ca3045a6e01031fce69bfe9ca04c21f66a33c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the rest of the query including the scope (all transactions sent this month) and the requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/revision/revision_02_app/0b0b1e049dd62133/speech.wav", "wer": 0.0, "wav_sha256": "1f4055e77b4b079eedbd23f4d3af3da6903317d112238065b448edb549e44b6e", "seconds": 11.8, "spoken_utterances": [ "How many likes did ... all PayPal transactions, I sent this month, have in total? ", "Let me correct the app: use uh... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal' but immediately self-corrects to 'Venmo' after a pause, which aligns with the revision type. The final audible intent is to query Venmo transactions, matching the original instruction. The correction is clear and applied to the earlier request, preserving all other conditions.", "heard_transcript": "How many likes did all PayPal transactions I sent this month have in total? Let me correct that. Use uh Venmo, not PayPal." }, "candidate_id": "wav_72340ec1ea42710a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request, then corrects 'PayPal' to 'Venmo' with 'let me correct the app use'. The phrasing is slightly awkward and less conversational than the others. The correction feels more like a mechanical fix than a natural thought process, and the prosody is flatter. The delivery is still intelligible and realistic, but the phrasing and pacing reduce its spontaneity compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "1ad842b53eaabaed84ae53b3829a1dcc9544c0bea46637d4a8f1b812fa2cb0e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/cancellation/cancel_013/416c6c31e1c48d85/speech.wav", "wer": 0.12, "wav_sha256": "059df336f3059e891799835317c7919887815703795abe89015869217b9b5f76", "seconds": 13.36, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?", "Wait, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a question, then includes a clear self-correction with 'Wait, I changed my mind. Please cancel the task.' This constitutes a final, explicit cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. Therefore, the final control intent is cancellation, matching the expected behavior for the 'cancellation' type.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total? Wait, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "49e2773055700b9025208efe72746379eb24ae919a60b3d47a0586722cbc5833", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/ambiguous_halt/halt_039/4b7a9fa4e15c316d/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "6d64189d92bf4185e3653b80dc8bcdbaa0ce2e0d3de9be0c4a02a0f331354a5f", "seconds": 12.44, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then clearly interjects with 'Wait, give me a second first,' which is a standard expression indicating a need to pause or take a moment. This is a final interjection that does not cancel the task or request a different action, but rather asks for a temporary halt. The audio ends with this pause request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2eb41f1b031e9f64d9a531bd9cc276177029c540f8da87a01e7ae3481c5626c9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_1/audio_natural/filler/original/8d0c2bbee44f0797/speech.wav", "wer": 0.1875, "wav_sha256": "76e1961aa4b278db543eafd0fa84978a10dbd36a4753b27328ee550197258b92", "seconds": 8.04, "spoken_utterances": [ "How many likes did um... all Venmo transactions, hmmm... I sent this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "How many likes did um all Venmo transactions, I sent this month, have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I received last year—sorry, that should be this year, have in total?", "long_pause_offsets": [], "source_sha256": "383ee606573c66e7f47fb06870b0d506a874fa2876010df50e8263622138b2ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year'. The rest of the query, including the app (Venmo), the subject (transactions I received), and the action (total likes), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/false_start/false_start_04_time_reference/ccdd47174f51a024/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4859febc9043eeb7d8a77c665f2513e942e49674098b551cd9206e64f745f390", "seconds": 8.24, "spoken_utterances": [ "How many likes did all Venmo transactions, I received last year—hmmm... sorry, that should be ... this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('last year') is discarded. The final instruction matches the original reference instruction in both action ('how many likes') and target ('all Venmo transactions I received this year').", "heard_transcript": "How many likes did all Venmo transactions I received last year have? Sorry, that should be this year have in total." }, "candidate_id": "wav_3312309351bddd45", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage self-correction. The initial request is clear, followed by a believable 'uh' hesitation, a self-aware 'sorry' that explicitly references the error ('that should be this year'), and a smooth, unforced correction. The prosody includes a slight pitch drop on 'uh' and a return to the original tone after the correction, mimicking genuine thought process. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How many likes did all CashApp—I should say Venmo transactions, I received this year, have in total?", "long_pause_offsets": [], "source_sha256": "f0ee95c6ef9ac484a87b67d56ff775b095abf8d61a8b8191380ce8994e3d4aac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the rest of the query including the scope (all transactions), condition (I received this year), and requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/false_start/false_start_02_app/3de8fccee13c9275/speech.wav", "wer": 0.2, "wav_sha256": "c5000363ba815722a88c2b500be6d2d6cf360e26425df51dd7707501a9ca195a", "seconds": 8.68, "spoken_utterances": [ "How many likes did all uh... CashApp—hmmm... I should say Venmo transactions, I received this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash app' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('cash app') is discarded. The rest of the instruction, including the action ('likes'), the target ('Venmo transactions'), and the condition ('I received this year'), remains unchanged and is preserved. Therefore, the final audible intent matches the original instruction after correction.", "heard_transcript": "How many likes did all uh, cash app, I should say Venmo transactions I received this year have in total?" }, "candidate_id": "wav_e9b6d8372fc02418", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start ('Ola Cash App') immediately corrected to 'Venmo'. The correction is immediate and the speaker continues smoothly. The 'uh' pause before 'I should say' is a natural filler for a self-correction. The overall delivery is fluent and the correction is integrated into the sentence without awkwardness, making it highly realistic." } ] }, "is_default": false }, { "variant_id": "false_start_01_engagement_metric", "proposed_instruction": "How many comments—sorry, I meant likes did all Venmo transactions, I received this year, have in total?", "long_pause_offsets": [], "source_sha256": "ed4f0a4c598ba58cd87509bbbfcb027db232c84b1727adf385afee78c1a03ccc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'comments' to 'likes' within the same utterance, preserving the rest of the query including the app (Venmo), the scope (all transactions I received this year), and the action (total count)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/false_start/false_start_01_engagement_metric/839e2dd132bbc036/speech.wav", "wer": 0.15, "wav_sha256": "f15b21c62d1ba7cb659214c296601eff306cb10aa10bc8a9145bf9bb0871eafa", "seconds": 8.8, "spoken_utterances": [ "How many comments—um... sorry, I meant likes did all Venmo transactions, uh... I received this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'comments' but immediately self-corrects to 'likes'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is ignored. The corrected request matches the original instruction in action ('likes'), target ('Venmo transactions'), and condition ('I received this year'). No other constraints are changed or lost.", "heard_transcript": "How many comments, um, sorry, I meant likes, did all Venmo transactions, uh, I received this year, have in total?" }, "candidate_id": "wav_e02e5bc0790ee06b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a false start ('comments'), self-corrects to 'likes', and then continues. However, the correction is less integrated; the phrase 'some' before the correction sounds like a filler that doesn't clearly connect to the intended word. The 'uh' after 'transactions' is a bit long and the flow feels slightly more hesitant than the top two, making it less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I sent—no, I meant received this year, have in total?", "long_pause_offsets": [], "source_sha256": "a610e1ced6d4a63d78a681ca4383ef68e19f3b63af0d4f402e8a603b778e2ec3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'sent' to 'received' within the same utterance, preserving the app (Venmo), the time frame (this year), and the action (counting likes). The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/false_start/false_start_03_transaction_direction/7279b0030ebac68c/speech.wav", "wer": 0.0, "wav_sha256": "f91271270f5c274a65129cc349276038b614235f636ae68a0627d79eb2912bab", "seconds": 10.0, "spoken_utterances": [ "How many likes did ... all Venmo transactions, I um... sent—uh... no, I meant received this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'received' with 'no, I meant received'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original mistaken value is discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "How many likes did all Venmo transactions I um sent uh no, I meant received this year have in total?" }, "candidate_id": "wav_799c98b50b0128ea", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker makes a false start ('sent') and corrects to 'received'. However, the correction is abrupt and the phrase 'I am' inserted between the false start and correction sounds unnatural and out of place. The prosody is flat, and the overall delivery lacks the natural flow of the others, sounding more like a rehearsed or robotic correction than spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many likes did Venmo transactions, I received this year, have in total? Oh, I forgot to mention, include all Venmo transactions.", "long_pause_offsets": [ 76 ], "source_sha256": "7c6dca1308e3efcdd4ba82639eb5171795cc859dbffb6130db5b61da9c098d5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the phrase 'include all Venmo transactions', which clarifies that the scope should be broadened to all transactions, not just those previously implied. However, the original instruction already specified 'all Venmo transactions, I received this year', so the addition is redundant but not contradictory. The final goal remains the same." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/addition/addition_01/bb8a20f350262750/speech.wav", "wer": 0.08, "wav_sha256": "3886718cf3f333438c6002bb8fdccdedd0488b9aaa98a8b563686b79468755f6", "seconds": 13.96, "spoken_utterances": [ "How many likes did Venmo transactions, ... I received this year, have in total? ", "Oh, I forgot to mention, include uh, let's see... all Venmo transactions." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many likes did Venmo transactions I received this year have in total?'. After a pause and an interjection, the speaker adds 'include uh let's see, all Venmo transactions.' This addition clarifies and broadens the scope to include all Venmo transactions, which is consistent with the original intent of counting likes for Venmo transactions received this year. The addition does not contradict or remove any prior conditions and is a natural expansion of the initial request. Therefore, the final intent is to execute the combined task.", "heard_transcript": "How many likes did Venmo transactions I received this year have in total? Oh, I forgot to mention, include uh let's see, all Venmo transactions." }, "candidate_id": "wav_3d272b9feaa07cfc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, then adds a detail ('Oh, I forgot to mention...') with a believable hesitation ('uh, let's see') before smoothly incorporating the missing information. The prosody is consistent, and the addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "How many likes did all Venmo transactions, I received, have in total? Oh, and one more thing: just for this year.", "long_pause_offsets": [ 70 ], "source_sha256": "acd249e067c22b3f2ae3a56436419dae22b031a2ca96c84d5e67b580e4966f9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time restriction 'this year' to the initial request, which was missing in the original instruction. The final goal combines the original request with the added temporal constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/addition/addition_03/6a4f74ca0629418a/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "c94f16c5d1f4a559a40e416d11737402ea6eaafaf8d95627846b9f56bd304894", "seconds": 12.72, "spoken_utterances": [ "How many likes did um... all Venmo transactions, ... I received, have in total? ", "Oh, and one more thing: just for ... this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'all Venmo transactions I received' without a time frame, then adds 'just for this year' as a clarifying constraint. This matches the 'addition' type, where the final goal combines the original request with the added temporal condition. The final intent is a narrowed version of the original, and the addition is clearly spoken after a pause, satisfying the criterion.", "heard_transcript": "How many likes did um all Venmo transactions I received have in total? Oh, and one more thing, just for this year." }, "candidate_id": "wav_815974d130b2be3e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a plausible addition, but the transition is slightly less smooth. The speaker starts with a request, then adds 'Oh, and one more thing' before specifying the time frame. The delivery is natural, but the addition feels a bit more abrupt compared to audio_03, and the initial request is slightly less fluent with the 'um'." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "How many likes did all Venmo transactions, this year, have in total? I left out a detail: only count the ones I received.", "long_pause_offsets": [ 69 ], "source_sha256": "d877b2c35a463c5ae60ae300a88a2414520562a8d71ca64f864146f12cc3e1a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that only transactions received by the speaker should be counted, which refines the original broad request. The app (Venmo), time frame (this year), and metric (total likes) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/addition/addition_02/c067a9418e9897db/speech.wav", "wer": 0.0, "wav_sha256": "415fce13865dd1ea837c7433837dbb5a43664e89477fd7c7eacbbc89e0a3e410", "seconds": 12.04, "spoken_utterances": [ "How many likes did ... all Venmo transactions, this year, have in total? ", "I left out a detail: only count the ones uh... I received." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is a broad request for the total likes of all Venmo transactions this year. The speaker then adds a clarifying constraint, 'Only count the ones uh I received,' which narrows the scope to only incoming transactions. This is a valid addition that refines the original request without contradicting it. The final intent is a combined, more specific version of the original task, matching the addition type criterion.", "heard_transcript": "How many likes did all Venmo transactions this year have in total? I left out a detail. Only count the ones uh I received." }, "candidate_id": "wav_df763e1cccc43c3b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio sounds the least realistic. The speaker starts with a clear request, then says 'I left out a detail' before correcting the scope. However, the correction ('only count the ones uh...') is delivered with a flat, robotic intonation and lacks the natural hesitation or prosodic variation of a genuine self-correction. The overall delivery feels more like a rehearsed or read instruction than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I received last year, have in total? Make that this year, not last year.", "long_pause_offsets": [ 80 ], "source_sha256": "885b71bb96d118ac281c9301e2e7f567b3689ec0568464654612d9a006a26487", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last year' to 'this year', preserving the rest of the query including the app (Venmo), the subject (transactions I received), and the requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/revision/revision_04_time_reference/56342bf9b672ac03/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "8c8ba0423550e7d1578f6187f7d79db3716b83b7c04430e33ee2d09ca71982de", "seconds": 12.88, "spoken_utterances": [ "How many likes did uh... all Venmo transactions, ... I received last year, have in total? ", "Make that um, let me think... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects to 'this year' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to ask about 'this year', which matches the original reference instruction. All other constraints (likes, Venmo transactions, received, total) are preserved.", "heard_transcript": "How many likes did uh all Venmo transactions I received last year have in total? Make that, um, let me think, this year, not last year." }, "candidate_id": "wav_6fdf125ac4e64b7b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then naturally self-corrects the time frame from 'last year' to 'this year' with a plausible 'um' and 'let me think' filler. The correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody. The intonation and rhythm are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I sent this year, have in total? I misspoke about the direction: it should be received, not sent.", "long_pause_offsets": [ 76 ], "source_sha256": "8563c78cbc640c1306663d67fadfb77094b2d42309bb0eed01094c8ca4d3ea12", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the transaction direction from 'sent' to 'received', which matches the original instruction. The app (Venmo), time frame (this year), and metric (total likes) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/revision/revision_03_transaction_direction/42a015aa043e1de6/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "6c0e011a196534d803047957fecf650eb1b86aa55bc5448033d8b4ba3126048a", "seconds": 13.0, "spoken_utterances": [ "How many likes did ah... all Venmo transactions, um... I sent this year, have in total? ", "I misspoke about the direction: it should be received, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'I sent this year' but immediately corrects themselves with 'I misspoke about the direction, it should be received, not sent.' This is a clear revision where the final intent is to request the total likes for Venmo transactions received this year. The correction is spoken after a pause and directly addresses the error in the initial request. The final audible intent matches the original instruction, so the task should be executed as revised.", "heard_transcript": "How many likes did uh all Venmo transactions um I sent this year have in total? I misspoke about the direction, it should be received, not sent." }, "candidate_id": "wav_eb5caa6886101446", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, includes a natural pause and filler ('um'), and then self-corrects the transaction direction from 'sent' to 'received'. The correction is delivered with a slight shift in intonation and a brief pause, sounding like a genuine moment of realization. The prosody remains consistent and the delivery is fluent." } ] }, "is_default": false }, { "variant_id": "revision_01_engagement_metric", "proposed_instruction": "How many comments did all Venmo transactions, I received this year, have in total? Wait, the metric should be likes, not comments.", "long_pause_offsets": [ 83 ], "source_sha256": "2e123971ec8348a9a4addbff74924e2a049de7d7766adaad0fdc0cd6f576a49e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the metric from comments to likes, preserving the app (Venmo), scope (all transactions received this year), and the total count request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/revision/revision_01_engagement_metric/57c624c8327a5f4e/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "1ab1f61105331ee1832e7b0f0dd969e6a5c7eaedbead9f9b3380e73672101548", "seconds": 13.72, "spoken_utterances": [ "How many comments did um... all Venmo transactions, ... I received this year, have in total? ", "Wait, the metric should be ... likes, not comments." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'comments' but immediately corrects themselves to 'likes' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to ask for the total number of 'likes', which matches the original reference instruction. The correction is clear and unambiguous, and the rest of the context (Venmo transactions, 'I received this year') remains unchanged.", "heard_transcript": "How many comments did um all Venmo transactions I received this year have in total? Wait, the metric should be likes, not comments." }, "candidate_id": "wav_1d65f9b897e5ab29", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request, then corrects the metric from 'comments' to 'likes'. The correction is made after a short pause and a 'wait' interjection, which is a plausible way to signal a change of mind. However, the correction is slightly abrupt and the intonation shift feels a bit less fluid than in audio_01 and audio_03, making it slightly less natural." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How many likes did all PayPal transactions, I received this year, have in total? Change the app to Venmo, not PayPal.", "long_pause_offsets": [ 81 ], "source_sha256": "c53a888b9e8793a207354a5068b1f3d60e6358ca47238795d694212013e19ba6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, preserving the rest of the query including the action (counting likes), the scope (all transactions received this year), and the object (likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/revision/revision_02_app/2479c08ae35b3b43/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "6fc22ae9cc76e75d01f9545cabe5082a36b6acf26ea646f988cc0e48995971fc", "seconds": 9.76, "spoken_utterances": [ "How many likes did ah... all PayPal transactions, ... I received this year, have in total? ", "Change the app to Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects it to 'Venmo' with 'Change the app to Venmo, not PayPal.' This is a clear revision where the final intent is to query Venmo transactions. The correction is spoken after a short pause and directly modifies the app from PayPal to Venmo, satisfying the revision type criterion. The rest of the instruction remains unchanged and is executed as intended.", "heard_transcript": "How many likes did all PayPal transactions I received this year have in total? Change the app to Venmo, not PayPal." }, "candidate_id": "wav_609d3f2960031338", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request, then corrects the app from 'PayPal' to 'Venmo'. The correction is made after a pause and a 'change' interjection, which is acceptable. However, the delivery feels slightly more robotic and less conversational compared to the others, with less natural intonation and a slightly stiffer rhythm. The correction is also less integrated into the flow of the sentence." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "ea23cd321dbf92a213eb3b85255f477639bf7cfc457c91aa2274a6561a0f3617", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/cancellation/cancel_029/d7bea74086de87dc/speech.wav", "wer": 0.05, "wav_sha256": "2a8ca893d52853cff4e0b237a4ff83ae37757e28f3d1996292735a0c9fa2321c", "seconds": 10.64, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?", "Hold on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task and then clearly and unambiguously cancels it with 'Hold on, I'm withdrawing that request.' This matches the expected behavior for a 'cancellation' type, where the final intent is to revoke the task. No execution of the original request is required.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total? Hold on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "e22f4ddda9947eac2465a788726e82d1055b843fa269a9054cbea47546839cd5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/ambiguous_halt/halt_039/b6338bd8f9333a6a/speech.wav", "wer": 0.05, "wav_sha256": "f4ea8f631dba30787f9201d961bf6e99164a9faf370925b320189f07fdd39705", "seconds": 9.64, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, give me a second first,' which is a clear request to pause or hold. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the verdict is pass.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c64bee225adaa125097d3ca322ddf15c64a4adce603751170a58a81792e32c51", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_2/audio_natural/filler/original/d483edc0f3c8c650/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "9acc0c627aed3a597ea85d13449f8244347c2cae96ef7b131304739b2d1f3bc0", "seconds": 6.12, "spoken_utterances": [ "How many likes did ... all Venmo transactions, ... I received this year, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction as intended, with no changes to the action, target, or value. The 'filler' type allows for natural pauses or hesitations, but the final goal is clear and matches the original instruction.", "heard_transcript": "How many likes did all Venmo transactions I received this year have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How many likes did all CashApp—sorry, I meant Venmo transactions, I sent or received this month, have in total?", "long_pause_offsets": [], "source_sha256": "8e373dd13bf6f48f1bb8d8e624fb9901574dd5bc6e369315962fde7f7493d1b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance. The rest of the instruction, including the action (counting likes), the scope (all transactions sent or received), and the time frame (this month), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/false_start/false_start_02_app/6f86730ee34c0db8/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "6773dc3244dab4b870aa43531ba73a88d7a3b9001156d199191aa956e0bd4203", "seconds": 9.48, "spoken_utterances": [ "How many likes did all CashApp—um... sorry, I meant ... Venmo transactions, I sent or received this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'sorry, I meant'. According to the 'false_start' type criterion, the final corrected value ('Venmo') should be applied, and the original incorrect value ('Cash App') should be discarded. The rest of the instruction, including the scope ('transactions I sent or received this month') and the action ('how many likes... have in total'), remains unchanged. The final audible intent matches the corrected instruction.", "heard_transcript": "How many likes did all Cash App um, sorry, I meant Venmo transactions I sent or received this month have in total?" }, "candidate_id": "wav_34898d611f329976", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a clear, spontaneous false start with 'Cash App' followed by a natural self-correction to 'Venmo transactions'. The speaker's intonation and pacing are consistent with someone recalling the correct term mid-sentence. The correction is immediate and integrated smoothly into the flow, with no awkward pauses or unnatural phrasing. The overall delivery sounds like a genuine, unscripted moment of retrieval." } ] }, "is_default": true }, { "variant_id": "false_start_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent or received last month—wait, make that this month, have in total?", "long_pause_offsets": [], "source_sha256": "68bb80dfab36647181edc9a7b1383df067c7ce5c2dcd887bd14a5aa49c13e049", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' within the same utterance. The final time reference is 'this month', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/false_start/false_start_04_time_reference/eef0ca31a80320ae/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "b26afdbde13ab087bfefbb146090f27c2668847ce96119313f58695f388fd7f7", "seconds": 11.76, "spoken_utterances": [ "How many likes did all um... Venmo transactions, I uh... sent or received last month—um... wait, make that this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'Wait, make that this month'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is discarded. The final audible task matches the reference instruction after the correction.", "heard_transcript": "How many likes did all um Venmo transactions I uh sent or received last month um. Wait, make that this month, have in total?" }, "candidate_id": "wav_d94f30d6abc45a44", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a false start ('Cash App') and a revision ('last month' to 'this month'), both of which are plausible. However, the speaker's delivery is slightly more hesitant and fragmented, with multiple fillers ('um', 'uh') and a less fluid transition between the initial and revised phrases. The correction feels a bit more labored than in audio_01, reducing its naturalness despite being semantically correct." } ] }, "is_default": false }, { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I sent—no, that should be sent or received this month, have in total?", "long_pause_offsets": [], "source_sha256": "02aa6b293238735482372e57780bbfce0dc414d75bc616f8e5529ce92c1f7447", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'I sent' to 'sent or received', which matches the original instruction. The action, app, time frame, and aggregation goal remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/false_start/false_start_03_transaction_direction/cbe8eb4e246e6089/speech.wav", "wer": 0.0, "wav_sha256": "5b517967e0a4d1d1d59c1bb9bcd180f4ccbf4b44c52b643ed679f540d24d248a", "seconds": 14.72, "spoken_utterances": [ "How many likes did all let me think... Venmo transactions, I sent—uh... no, that should be um... sent or received this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'likes' but immediately corrects to 'Venmo transactions'. The final, corrected instruction matches the original reference. The false start is properly handled by the 'false_start' type criterion, which requires executing the self-corrected version.", "heard_transcript": "How many likes did all let me think Venmo transactions I sent uh no that should be um sent or received this month have in total?" }, "candidate_id": "wav_3a1e7acc63bc9b8b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio includes a false start ('Cash App'), a filler ('let me think'), and a self-correction ('no, that should be...'). While the structure is realistic, the correction is less immediate and more hesitant than in the others. The phrase 'no, that should be' sounds slightly more rehearsed or unnatural compared to a spontaneous 'sorry, I meant'. The overall pacing is slower and more disjointed, making it feel less like a real-time thought process." } ] }, "is_default": false }, { "variant_id": "false_start_01_engagement_metric", "proposed_instruction": "How many comments—no, I mean likes did all Venmo transactions, I sent or received this month, have in total?", "long_pause_offsets": [], "source_sha256": "911b699c67aa3ab045821d153d867b74557cf0bd38516459abe232167809535e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'comments' to 'likes' within the same utterance. The rest of the query, including the app (Venmo), scope (transactions sent or received), and time frame (this month), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/false_start/false_start_01_engagement_metric/0db20d4cfcee5356/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "6236111babab90cb84112f200b31958a1e4b1702cf064c9e8c558c7878584776", "seconds": 11.0, "spoken_utterances": [ "How many comments—um... no, I mean likes did all hmmm... Venmo transactions, I sent or received ... this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'comments' but immediately self-corrects to 'likes'. The final utterance matches the original instruction after the correction. The 'false_start' type requires applying the immediate self-correction, which is done here. The rest of the instruction, including the scope of Venmo transactions and the time frame 'this month', remains unchanged and correctly interpreted.", "heard_transcript": "How many comments from no, I mean, likes did all Venmo transactions I sent or received this month have in total?" }, "candidate_id": "wav_a3dc1453a9a90b7e", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip begins with a false start ('comments'), then corrects to 'likes'. However, the correction is abrupt and lacks the natural flow of the others. The speaker says 'no, I mean' followed by a long pause and a restart, which sounds more like a forced edit than a genuine thought process. The delivery is less continuous and more fragmented, with an unnatural rhythm that suggests a synthetic or poorly recorded origin." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many likes did all Venmo transactions, I this month, have in total? Oh, I forgot to say, include both transactions I sent and received.", "long_pause_offsets": [ 72 ], "source_sha256": "126a05f2e90c691de70d483ee29f8db93acf2ba97c7b844d85e998b258bf38b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'include both transactions I sent and received', which clarifies the scope of transactions to include. The original instruction already specified 'all Venmo transactions' and 'this month', so the final goal is fully specified and consistent with the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/addition/addition_01/0c8eb979721d7b4c/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "3a65828148e9a3ce7b19d3bcfe8dfd1428daa7576489814fae71e008dbbf243f", "seconds": 17.28, "spoken_utterances": [ "How many uh... likes did all um... Venmo transactions, I this month, have in total? ", "Oh, I forgot to say, include both transactions I uh, let me think... sent and received." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is partially garbled but the core intent to count likes on Venmo transactions is clear. The speaker then adds the missing constraint 'include both transactions I sent and received' after a pause and self-correction, which completes the original instruction. The addition clarifies the scope without changing the action or target, so the final goal matches the reference.", "heard_transcript": "How many uh likes did all um Venmo transactions I this month have in total? Oh, I forgot to say, include both transactions I uh let me think, sent and received." }, "candidate_id": "wav_fcd7019524eefa43", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('How many uh likes did all um...') and then adds a clarifying detail ('Oh I forgot to say...') with a believable 'let me think' pause. The addition is smoothly integrated into the same utterance, maintaining a consistent rhythm and intonation. The prosody reflects genuine recall and spontaneous correction, making it sound like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many likes did all Venmo transactions, I sent or received, have in total? I left out a detail: only count transactions from this month.", "long_pause_offsets": [ 78 ], "source_sha256": "b74a64531c889c84511151b7bb5ba406a06a70afb7225e2e2884b170895944f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'this month' to the initial request, refining the scope without changing the action, app, or other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/addition/addition_02/cd097a00096e0718/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "53ba9f7bcdf36cbae71fe4f81c55d484fed191c4d26c0d58587271c38ad8d22e", "seconds": 16.48, "spoken_utterances": [ "How many likes did all hmmm... Venmo transactions, I ... sent or received, have in total? ", "I left out a detail: only count transactions from um, let me think... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How many likes did all Venmo transactions, I sent or received this month, have in total?'. The speaker first asks a similar question without the 'this month' constraint, then adds the constraint 'only count transactions from this month' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added condition. The final audible intent is to count likes for Venmo transactions sent or received this month, which aligns with the reference instruction. No errors or contradictions remain.", "heard_transcript": "How many likes did uh Venmo transactions I sent or received have in total? I left out a detail, only count transactions from um, let me think, this month." }, "candidate_id": "wav_464c3239d0b01e9a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar hesitant tone but the addition ('I left out a detail...') is delivered as a separate, slightly abrupt sentence. The 'let me think' pause is longer and feels more like a reset than a natural continuation. While still realistic, the flow is less seamless than audio_01, with a more noticeable break between the initial request and the revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_time_reference", "proposed_instruction": "How many likes did all Venmo transactions, I sent or received last month, have in total? That should be 'this month', not 'last month'.", "long_pause_offsets": [ 89 ], "source_sha256": "c4a673908b66b3604c0e30d308f824d8dbe4ebfaf78ecff5a00ee6b80614544f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this month', preserving the rest of the query including the app (Venmo), transaction scope (sent or received), and the requested metric (total likes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/revision/revision_04_time_reference/2ad55c56bf20f562/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "7c7e03baf04d1d6077e8d0e06b47d63d6cc4c68e09d21abf0a632d3ad91a46f1", "seconds": 15.84, "spoken_utterances": [ "How many um... likes did all uh... Venmo transactions, I sent or received last month, have in total? ", "That should be 'um, let me think... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'last month' but immediately self-corrects to 'this month' after a pause and repair. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the total likes for Venmo transactions from 'this month', which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "How many um likes did all uh Venmo transactions I sent or received last month have in total? That should be um let me think, this month, not last month." }, "candidate_id": "wav_1694bdc29eeffac2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage revision process. They begin with a hesitant 'um' and 'uh', then correct 'last month' to 'this month' with a plausible 'let me think' filler and a self-correcting 'that should be'. The prosody is continuous and the intonation reflects genuine thought and correction, making the revision sound spontaneous and realistic." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "How many likes did all PayPal transactions, I sent or received this month, have in total? Let me correct the app: use Venmo, not PayPal.", "long_pause_offsets": [ 90 ], "source_sha256": "f402d1db34d184be0a1d036f5d654a582c3e4b77e743ece0e403031655f71c08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from PayPal to Venmo, while retaining the rest of the query including the action (counting likes), the scope (all transactions sent or received), and the time frame (this month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/revision/revision_02_app/161d80a5fc24ad02/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "7f8b899216b1264902126f8b9652dd4686a39648e8fa135d7811b048ee2b6b61", "seconds": 14.12, "spoken_utterances": [ "How many ... likes did all ... PayPal transactions, I let's see... sent or received this month, have in total? ", "Let me correct the app: use Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal' but immediately corrects it to 'Venmo' in a revision. The final intent is the corrected version, which matches the original instruction. The revision type requires applying the correction spoken after a pause to the earlier instruction, which is satisfied here.", "heard_transcript": "How many likes did all PayPal transactions I sent or received this month have in total? Let me correct the app, use Venmo, not PayPal." }, "candidate_id": "wav_82dcf82ed0e88fb0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then self-corrects the app from 'PayPal' to 'Venmo'. The correction is delivered with a natural 'let me correct the app' phrase and a pause before the correction, which sounds like a genuine mid-sentence realization. The prosody is consistent and the correction is integrated smoothly into the flow." } ] }, "is_default": false }, { "variant_id": "revision_01_engagement_metric", "proposed_instruction": "How many comments did all Venmo transactions, I sent or received this month, have in total? Wait, I meant likes, not comments.", "long_pause_offsets": [ 92 ], "source_sha256": "0113fa635284be9da28114a2f1a1c571d8a9c9e7edd804aa9b192fe85d5f4126", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the metric from 'comments' to 'likes', preserving the app (Venmo), scope (transactions sent or received), and time frame (this month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/revision/revision_01_engagement_metric/8a014e3382e932a7/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "c0dfcde53d3d021f639bb9ebb96855bd3b8b692c76a8a65707e74e9a8bd4af8b", "seconds": 14.92, "spoken_utterances": [ "How many um... comments did all Venmo transactions, I um, let me think... sent or received uh... this month, have in total? ", "Wait, I meant likes, not comments." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'comments' but immediately self-corrects to 'likes' after a pause and the interjection 'Wait, I meant likes, not comments.' This constitutes a revision where the final intent is to inquire about 'likes'. The rest of the instruction, including the scope of 'all Venmo transactions I sent or received this month', remains unchanged and matches the original reference. Therefore, the final audible intent aligns with the revision type criterion.", "heard_transcript": "How many um comments did all Venmo transactions I am let me think sent or received uh this month? Wait, I meant likes, not comments." }, "candidate_id": "wav_abb0e10f6d4da088", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start by saying 'comments' and then self-corrects to 'likes'. The correction is immediate and the intonation shift is natural. However, the initial hesitation ('um') and the correction are somewhat abrupt, and the overall delivery feels slightly rushed compared to the others, reducing its realism." } ] }, "is_default": false }, { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "How many likes did all Venmo transactions, I sent this month, have in total? Change 'sent or received' to just 'sent'.", "long_pause_offsets": [ 77 ], "source_sha256": "3531a1e56fe20b85d06bca0be1ad3119c5700a13d8f730897a19cc29e85e4156", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision narrows the scope from 'sent or received' to only 'sent', which changes the set of transactions considered." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/revision/revision_03_transaction_direction/fe64837dee90392c/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "cc084b07be98f7cae82444f2bbcd440a9e031eb5abac7637577f8380bc879d76", "seconds": 12.96, "spoken_utterances": [ "How many um... likes did all Venmo transactions, I sent this month, have in total? ", "Change '... sent or received' to just 'sent'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent or received' but then immediately corrects to 'just sent' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to count likes for transactions that were sent, which matches the revised instruction. All other constraints remain unchanged.", "heard_transcript": "How many um likes did all Venmo transactions I sent this month have in total? Change sent or received to just sent." }, "candidate_id": "wav_5424b34fa59d2df9", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker begins with a hesitant 'um' and then makes a revision from 'sent or received' to 'just sent'. However, the revision is delivered as a separate, declarative sentence rather than being integrated into the original request. This makes the correction sound more like a post-hoc instruction than a spontaneous self-correction, reducing its naturalness." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "2906abcc9354e74d106ae123bce2a33a93fa4c78222153d8d8cebc7a5d997aec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/cancellation/cancel_029/6fe7a1be4c71d05d/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "622e82a6016d9092ffcb8ca214c98949b4dd685af4acf2e8cd64a7171f3e5e86", "seconds": 14.64, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?", "Wait, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins a request but then clearly states 'Wait, I'm withdrawing that request' as the final interjection, which constitutes a cancellation of the entire task. The initial request is irrelevant once the cancellation is audibly confirmed.", "heard_transcript": "How many uh likes did all, let me think, Venmo transactions I um sent or received this month have in total? Wait, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "cde8ea0d977fd77cd4f046b9d60123b4c95a2df667ce9cf3aa4280dfb2cd63f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/ambiguous_halt/halt_008/dff53c7153202209/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "fbc3b58d0071283730bd910f4942d7b92daf57585ba03db77420ed9052eeb4a9", "seconds": 13.24, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then says 'let me think' and 'wait a second' as final interjections. These are audible requests to hold or pause, with no indication of cancellation or resumption. The final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How many uh likes did all, let me think, Venmo transactions I um sent or received this month have in total? Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "23cf851_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0b997508f1f0ae92b45080ac2ad357a9c5dd182dfc788280e88da8df733857dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/23cf851_3/audio_natural/filler/original/f1c904514d4aa08a/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ecfac9d84834fd4905d18c5dc726386063100a79f58a2a281ead41e5b46e3246", "seconds": 10.28, "spoken_utterances": [ "How many uh... likes did all let me think... Venmo transactions, I um... sent or received this month, have in total?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let me think', 'um') and pauses but does not change the core request. The final audible intent matches the original instruction. The filler type criterion allows ignoring hesitations and backchannels, so the task is executed as intended.", "heard_transcript": "How many uh likes did all let me think Venmo transactions I um sent or received this month have in total?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail inbox—scratch that, make it inbox and outbox before the last calendar month as read.", "long_pause_offsets": [], "source_sha256": "193aed7be3e5cbdc22abba6d7dcec927815c9ddd9b928dd7aca35239bfdc0add", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'inbox and outbox'. The final scope includes both folders, matching the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/false_start/false_start_02_email_folder_scope/fc937eb9cda3e138/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "fe8559e79778a21156aececd60f684bd969610fc9907d773d5383b86c731521c", "seconds": 10.08, "spoken_utterances": [ "Mark everything in my Gmail inbox—hmmm... scratch that, make it ... inbox and outbox before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Mark everything in my Gmail inbox' but immediately self-corrects with 'Scratch that. Make it inbox and outbox before the last calendar month as read.' This is a false_start where the final instruction replaces the initial one. The final instruction matches the original reference in both action ('mark as read') and scope ('inbox and outbox before the last calendar month'). No other constraints differ. Therefore, the final intent is fully aligned with the reference.", "heard_transcript": "Mark everything in my Gmail inbox. Hmm. Scratch that. Make it inbox and outbox before the last calendar month as read." }, "candidate_id": "wav_9228e9abdf8c4fdd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with a thoughtful 'hmm' before self-correcting 'scratch that' and smoothly revising the task to include 'outbox' and 'last calendar month as read'. The correction is immediate and the prosody reflects genuine reconsideration, with a natural pause and intonation shift. The flow is continuous and the repair feels spontaneous." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the last week—sorry, that should be before the last calendar month as read.", "long_pause_offsets": [], "source_sha256": "0de4f93d5c343ddfe82361d9aee8286cdce59574578dba96f9b446a0429abc1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last week' but immediately corrects it to 'last calendar month'. The final time filter matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/false_start/false_start_03_time_filter/b8ba6f057b7ebfca/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "face8dda8bbd0906309c13d71ad7b093067f8939a7c61eb767e197c3eff00113", "seconds": 16.32, "spoken_utterances": [ "Mark everything in let's see... my Gmail inbox and outbox um, let's see... before the last week—hmmm... sorry, that should be before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'before the last week' but immediately self-corrects to 'before the last calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Mark everything in my Gmail inbox and outbox, um, let's see, before the last week. Sorry, that should be before the last calendar month as read." }, "candidate_id": "wav_f18ab80d7d09caa7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but includes multiple fillers ('let's see', 'um', 'hmm') and hesitations, which are realistic but make the delivery feel slightly less fluent. The self-correction from 'last week' to 'last calendar month' is natural, with a pause and a slight restart, but the overall pacing is more fragmented than audio_01. The repair is believable but the initial hesitation and repetition reduce the smoothness of the utterance." } ] }, "is_default": false }, { "variant_id": "false_start_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the last calendar month as unread—no, that should be read.", "long_pause_offsets": [], "source_sha256": "892d81066f780858c16cba50880c244af11a2631dba2eb74c57bf24f23d68438", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately corrects it to 'read'. The final instruction is to mark the specified items as read, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/false_start/false_start_04_email_state/b7fc4082556417c1/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2894378a76816edbbf860e69ae4457cb51314946d0e82d73abacf097ea8be9ef", "seconds": 9.8, "spoken_utterances": [ "Mark everything in my Gmail inbox and outbox before the last calendar month as unread—ah... no, that should be uh... read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately self-corrects to 'read'. According to the 'false_start' type criterion, the final corrected value should be applied. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the last calendar month as unread. No, that should be uh read." }, "candidate_id": "wav_890f9bbd00b3158f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker delivers a clear initial request but then makes a self-correction from 'unread' to 'read'. However, the correction is less natural due to the awkward phrasing 'no, that should be uh read' and the insertion of 'uh' mid-correction, which sounds slightly forced. The prosody lacks the smoothness of the other two, and the repair feels more like a mechanical edit than a spontaneous thought process, reducing its realism." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Mark everything in my Outlook—I should say Gmail inbox and outbox before the last calendar month as read.", "long_pause_offsets": [], "source_sha256": "631a485bd2e0138e5edbfdd88294d301eca005d1aa11530c6cb6521dd583dbc4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the rest of the instruction including the target folders (inbox and outbox), time condition (before the last calendar month), and action (mark as read)." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in my Outlook—ah... I should say ... Gmail inbox and outbox before the last calendar month as read." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as read. Oh, I forgot to mention, mark everything before the last calendar month as read.", "long_pause_offsets": [ 54 ], "source_sha256": "bc9b6dca7f8e366c5b32bc1f4709ab6b9b2cd644d9cd6b8b76b92bede366b029", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'before the last calendar month' to the initial instruction, which already specified the app (Gmail), locations (inbox and outbox), and action (mark as read). The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/addition/addition_01/d39c2e92da92e1db/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "1fedc40944873724900fd06b5638bef39dd4b49e358622a9f4b5a2c8e5b5fe0f", "seconds": 15.44, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and um... outbox as read. ", "Oh, I forgot to mention, mark everything let's see... before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to mark everything in the inbox and outbox as read. After a pause and self-correction, the speaker added the constraint 'before the last calendar month'. The final instruction combines the original request with the added temporal condition. The final intent matches the original instruction with an additional constraint, satisfying the 'addition' type criterion.", "heard_transcript": "Mark everything in my Gmail inbox and um outbox as read. Oh, I forgot to mention, mark everything, let's see, before the last calendar month as read." }, "candidate_id": "wav_abcb84f6c932262f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly natural, multi-stage thought process. It begins with a clear initial request, includes a believable self-correction ('Oh, I forgot to mention'), and smoothly adds a new temporal condition ('before the last calendar month as read'). The speaker uses natural fillers ('um', 'let's see') and pauses that align with real-time formulation. The prosody is conversational and the transitions between ideas feel spontaneous, not rehearsed." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Mark everything in my Gmail before the last calendar month as read. One more thing: do this for both my inbox and outbox.", "long_pause_offsets": [ 68 ], "source_sha256": "6512cde680ed3ad1d460771f2b01e157396d9181b35b040311faf8e7c20f5793", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'outbox' requirement, preserving the original 'inbox' and 'before the last calendar month' conditions. The final goal now includes both folders." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/addition/addition_02/7ca25b6e316aec97/speech.wav", "wer": 0.0, "wav_sha256": "e2afbb76acc904056271064cd35c88b09ae85851da4a433a81894f9a5c690d55", "seconds": 16.08, "spoken_utterances": [ "Mark everything in ... my Gmail ... before the last calendar month as read. ", "One more thing: do this for both my uh, let's see... inbox and outbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to mark all emails before the last calendar month as read in both inbox and outbox. The speaker initially says 'inbox and outbox', then adds 'do this for both my uh, let's see, inbox and outbox' as an addition. The final intent combines the original instruction with the clarification that both folders should be included, which is consistent with the original instruction. The addition does not change the core task but reinforces the scope. Therefore, the final intent matches the original instruction and the addition, and the task should be executed.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the last calendar month as read. One more thing, do this for both my uh, let's see, inbox and outbox." }, "candidate_id": "wav_7188fd29cd02336b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a natural-sounding request with a follow-up addition. However, the initial request is slightly less fluent ('Mark everything in my Gmail before the last calendar month as read') and the addition ('One more thing...') is delivered with a more abrupt, almost robotic transition. The speaker's intonation is flatter, and the overall delivery feels more like a rehearsed or read instruction compared to the more organic flow of audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before last week as read. I misspoke about the time filter: it should be before the last calendar month, not before last week.", "long_pause_offsets": [ 71 ], "source_sha256": "27394dd482f0cb0a936090599c6c27b1dc9f2c6de8ed3b6a490d5e4e933222a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time filter from 'last week' to 'last calendar month', while retaining the app (Gmail), locations (inbox and outbox), and action (mark as read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/revision/revision_03_time_filter/90d67470b0cea631/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "39a27a92f75004954d4dd39ddcd001f0b9fd42e970e30abded1e582d17e93f3d", "seconds": 16.04, "spoken_utterances": [ "Mark everything in uh, let's see... my Gmail inbox and um... outbox before last week as read. ", "I misspoke about the time filter: it should be before the last calendar month, not before last week." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'before last week' but immediately corrects themselves with 'I misspoke about the time filter. It should be before the last calendar month, not before last week.' This is a clear revision where the final intent is the corrected version. The original instruction is overridden by the revision, and the final goal matches the reference instruction exactly.", "heard_transcript": "Mark everything in uh let's see my Gmail inbox and um outbox before last week as read. I misspoke about the time filter. It should be before the last calendar month, not before last week." }, "candidate_id": "wav_6fcbb664e07c8fbb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio presents a highly realistic revision. The speaker begins with a natural, slightly hesitant formulation ('mark everything in uh let's see my Gmail inbox and um outbox before last week as read'), including a filler ('uh') and a self-correction ('let's see'). After a brief pause, they clearly state the error ('I misspoke about the time filter') and provide a coherent, grammatically correct revision ('it should be before the last calendar month, not before last week'). The prosody is natural, with appropriate stress and intonation, and the correction flows seamlessly from the initial mistake, mimicking how a real person would spontaneously correct themselves." } ] }, "is_default": true }, { "variant_id": "revision_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail inbox before the last calendar month as read. Make that inbox and outbox, not just inbox.", "long_pause_offsets": [ 74 ], "source_sha256": "2d990629e98228a34a10b876b90fb6ce5e8671b3923aa969c08f7909cb41995e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the scope from only inbox to include both inbox and outbox, while preserving the app, action, time condition, and target state (read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/revision/revision_02_email_folder_scope/5e0497ab12409cf0/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "dfe8092bf67270e4660cfe83ed1d1bccdd133d9aff600f8884241855910641cd", "seconds": 11.96, "spoken_utterances": [ "Mark everything in ... my Gmail ... inbox before the last calendar month as read. ", "Make that inbox and outbox, not just inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'inbox and outbox' after a short pause, which is a revision. The final instruction matches the original reference exactly, including the scope of both inboxes and the date condition. The revision is clear and applied to the earlier instruction.", "heard_transcript": "Mark everything in my Gmail inbox before the last calendar month as read. Make that inbox and outbox, not just inbox." }, "candidate_id": "wav_08fc1fa5e4c43c32", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is a revision but less realistic than audio_01. The speaker starts with a clear, fluent request ('mark everything in my Gmail inbox before the last calendar month as read'), which is followed by a pause and a revision ('make that inbox and outbox, not just inbox'). While the correction is grammatically sound and the pause is plausible, the initial request is too smooth and lacks the hesitations or fillers that would make it sound like a spontaneous utterance. The revision itself is abrupt and lacks the natural self-awareness or hesitation present in audio_01, making it sound more like a rehearsed or edited statement than a genuine correction." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Mark everything in my Outlook inbox and outbox before the last calendar month as read. Wait, the app should be Gmail, not Outlook.", "long_pause_offsets": [ 87 ], "source_sha256": "66415fbcc877083daca760d0e576cbe8bb4635ed9b8e4fe30702fd46970f6ef0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, preserving the action, scope (inbox and outbox), and time condition (before the last calendar month)." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in ... my Outlook inbox and outbox ... before the last calendar month as read. ", "Wait, the app should be ... Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the last calendar month as unread. Change the state to read, not unread.", "long_pause_offsets": [ 87 ], "source_sha256": "4d64289b9e11397e1daf58d3da6c19820fd22506c5477e236fc58215487b6185", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly reverses the state from unread to read, contradicting the initial instruction's state." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in ... my Gmail inbox and outbox ... before the last calendar month as unread. ", "Change the state to read, not unread." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e68d62ccf751d4a484e56ee00b52b96c4e9ce4fb7f217e324dc0212822cb62ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/cancellation/cancel_009/3c04a28d2108cf6c/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "2a7736d3e8bb9ce00e955182f72c98331756546f1b18a9c66167cdc3a28572ca", "seconds": 12.2, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task and then immediately follows with 'Oh wait, cancel that altogether,' which is a clear and unambiguous cancellation of the entire prior request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "88c51a1334e96e51084b8564158765bba813334cbc8888db33aabe4e8dfbd1f9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/ambiguous_halt/halt_027/250186e6c3fe0a44/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "252a192d98bd0a4f1e94e1cb991feadb253ef4471d5975b3d2f3e9dbb5900529", "seconds": 12.4, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, I need a moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the verdict is pass.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1f57d5a49606b97f0dc379c81ea900135b17ccc8300d8def668b384639c7b4d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_1/audio_natural/filler/original/372e4ff5758b7235/speech.wav", "wer": 0.0625, "wav_sha256": "109a016d61eb346afa923088f9f987a231d691ef849b75b10695706fd12a6c97", "seconds": 7.92, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and uh... outbox before the last calendar month as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core task. The final intent matches the original instruction, with no correction or cancellation. The task is executed as requested.", "heard_transcript": "Mark everything in my Gmail inbox and uh outbox before the last calendar month as read." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox in the current calendar month as read—sorry, I meant unread.", "long_pause_offsets": [], "source_sha256": "73b501ba9b6eb62ecd6e6c760377f3de95e6e07506c11c463d7063b33d159d17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'read' but immediately corrects to 'unread' within the same utterance. The final intended action is to mark items as unread, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/false_start/false_start_04_email_state/f6b2d9ec625d6573/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "8e4a97925e83d7e1ae154c7e0cb0e3484b8333d0d1da432f12484a7a5d536699", "seconds": 11.6, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and outbox in the current calendar month as read—uh... sorry, I meant um... unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as read' but immediately self-corrects to 'unread'. According to the false_start type criterion, the final corrected value should be applied. The correction is clear and immediate, and the rest of the instruction remains unchanged. Therefore, the final intent is to mark items as unread, which matches the original instruction after correction.", "heard_transcript": "Mark everything in my um Gmail inbox and outbox in the current calendar month as read. Uh, sorry, I meant um unread." }, "candidate_id": "wav_030f1c48351d46c5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start with 'is read uh sorry I meant um unread'. The self-correction is immediate, the filler 'uh' and 'um' are placed at natural points of hesitation, and the intonation reflects genuine recollection. The prosody is consistent and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Mark everything in my Outlook—I should say Gmail inbox and outbox in the current calendar month as unread.", "long_pause_offsets": [], "source_sha256": "a40c20759fa4f56a31fe4897e4a34163e4c1bcd2e38c0897cee46e234b49cb13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the rest of the instruction including the target folders (inbox and outbox), time frame (current calendar month), and action (mark as unread)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/false_start/false_start_01_app/2f42f56ffc3e353a/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "c5dc65219f4336be8438f9ce87e3057ca8cfc80f29a3bdd4d5ca4e3d578d8c1a", "seconds": 13.0, "spoken_utterances": [ "Mark everything in my Outlook—hmmm... I should say ... Gmail inbox and outbox in the um... current calendar month as unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail' before continuing with the rest of the instruction. This is a false start, and the final corrected version matches the original instruction exactly. The control intent is to execute the corrected task.", "heard_transcript": "Mark everything in my Outlook, hmm, I should say, Gmail inbox and outbox in the current calendar month as unread." }, "candidate_id": "wav_50f2a55288b62850", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The false start 'Outlook hmm I should say Gmail inbox' is credible, with a natural pause and filler. The correction is clear and the speaker continues smoothly. However, the initial 'Outlook' is not a direct self-correction of a word but a replacement of the entire email service, making the repair slightly less fluid than in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox last calendar month—no, that should be in the current calendar month as unread.", "long_pause_offsets": [], "source_sha256": "fce379507d1fd42b259b5e5875e0d737e388c62000f1f739c25de87f46b3f787", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'current calendar month'. The final intended time filter is 'current calendar month', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/false_start/false_start_03_time_filter/039dc022db3682e2/speech.wav", "wer": 0.125, "wav_sha256": "304740bc8757186fa1f68edba06787d6633aef2d08077defe838ec0d1339f4d4", "seconds": 12.04, "spoken_utterances": [ "Mark everything in my uh... Gmail inbox and outbox ... last calendar month—um... no, that should be in the current calendar month as unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately self-corrects to 'current calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Mark everything in my uh Gmail inbox and outbox last calendar month. No, that should be in the current calendar month as unread." }, "candidate_id": "wav_782e55f45d77e542", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker says 'last calendar month' and then corrects to 'current calendar month'. The revision is clear, but the correction is delivered with a slightly abrupt shift in intonation and the 'no' feels a bit too definitive for a natural hesitation. The filler 'uh' and 'um' are used, but the overall flow is less smooth than in the top two." } ] }, "is_default": false }, { "variant_id": "false_start_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail sent and drafts—scratch that, make it inbox and outbox in the current calendar month as unread.", "long_pause_offsets": [], "source_sha256": "a55c68c063e0504f107ddd3e511b3ad195209cc1e364f3f10a5e43abaa1020fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'sent and drafts' but immediately corrects to 'inbox and outbox'. The final target folders are correctly specified, and all other conditions (Gmail, current calendar month, action of marking as unread) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/false_start/false_start_02_email_folder_scope/9707ee7130ed60a9/speech.wav", "wer": 0.24, "wav_sha256": "686eb0f60089f50aa2b638683f105b8dd5ac97eec58549959a1094e54a5ca02b", "seconds": 14.44, "spoken_utterances": [ "Mark everything in my Gmail ah... sent and drafts—um... scratch that, make it inbox and outbox in the uh... current calendar month as unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent and drafts' but immediately self-corrects with 'scratch that, make it inbox and outbox'. This is a false_start type, where the final corrected values override the initial mistaken ones. The final intent matches the original instruction exactly, including the target folders (inbox and outbox) and the condition (current calendar month).", "heard_transcript": "Mark everything in my Gmail uh sent and drafts um scratch that, make it inbox and outbox in the uh current calendar month as unread." }, "candidate_id": "wav_ac0671fba4bacb8e", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker says 'Gmail uh sent and drafts um scratch that make it inbox and outbox'. The 'scratch that' is a common conversational phrase, but the correction is delivered with a flat, almost robotic intonation. The pause and filler are present, but the overall delivery lacks the natural prosody and emotional nuance of the others, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as unread. Oh, I forgot to mention, this should apply only to the current calendar month.", "long_pause_offsets": [ 56 ], "source_sha256": "bc66294daf1fdeaf9d74cfaf67ae8b27a9e3a1236a635cbf5dab4e0c879c0d6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time constraint that narrows the scope of the initial instruction. The core action and targets remain unchanged, and the new condition is a restriction, not a contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/addition/addition_01/53b6b7fbefd666eb/speech.wav", "wer": 0.08, "wav_sha256": "4ae3c779ef592120363cc0f981cf45a51f5d63b7da0e1fb4474d128bb0c161a3", "seconds": 13.88, "spoken_utterances": [ "Mark everything in my Gmail inbox and hmmm... outbox as unread. ", "Oh, I forgot to mention, this should apply only to the ... current calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to mark all emails in the inbox and outbox as unread. After a pause, they add a constraint that this should only apply to the current calendar month. Since the addition is a clarifying constraint and not a contradiction, the final intent is the original task with the added restriction. The type is 'addition', and the final goal matches the original instruction with the added condition, so it passes.", "heard_transcript": "Mark everything in my Gmail inbox and outbox as unread. Oh, I forgot to mention, this should apply only to the current calendar month." }, "candidate_id": "wav_90a211a9299633cc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and filler ('hmm') before correcting 'inbox' to 'outbox'. The self-repair is smooth and integrated into the flow. The addition of the calendar month restriction is a believable, contextually relevant follow-up with a natural-sounding 'oh, I forgot to mention'. The prosody is consistent, and the speaker's rhythm reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Mark everything in my Gmail inbox in the current calendar month as unread. Oh, and one more thing: also include the outbox.", "long_pause_offsets": [ 75 ], "source_sha256": "11aec2ce92eaa10d7dc80a5cf2e0e9e32116adf2ea25dff67533cc879dddbd37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the outbox to the initial request for the inbox, combining both folders under the same time and action constraints." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/addition/addition_02/3bfee518c1e96509/speech.wav", "wer": 0.0, "wav_sha256": "760a58503c5f9cfc42032f8ec80c686579a10d996577e7b9acecbe2d50d70a28", "seconds": 16.56, "spoken_utterances": [ "Mark everything in my uh... Gmail inbox in the ... current calendar month as unread. ", "Oh, and one more thing: also include the uh, let me think... outbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests marking Gmail inbox items as unread for the current month. After a pause and an interjection, they add 'also include the outbox' as a new constraint. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final audible intent is complete and matches the reference instruction.", "heard_transcript": "Mark everything in my uh Gmail inbox in the current calendar month as unread. Oh, and one more thing, also include the uh, let me think, outbox." }, "candidate_id": "wav_10425b865628d266", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is fluent and clear, but the follow-up 'oh, and one more thing' is slightly abrupt. The speaker then hesitates with 'uh, let me think' and attempts to add 'inbox', but the correction is cut off mid-word. The interruption and incomplete phrasing make the addition feel less natural and more like a recording glitch or an uncompleted thought, reducing the overall realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox in the current calendar month as read. That should be unread, not read.", "long_pause_offsets": [ 84 ], "source_sha256": "d775e9d7c54ea61987d8fc1011166f9107a97b15f2d3a07f93f7df43757470b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'read' to 'unread', preserving all other conditions including the app, locations, and time scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/revision/revision_04_email_state/484b31948c9ce6a0/speech.wav", "wer": 0.0, "wav_sha256": "4c45d1e52947cfd4c99193546abe2c86fe95022bb458b6a821655ed4e7c0c9ee", "seconds": 16.6, "spoken_utterances": [ "Mark everything in my um, let's see... Gmail inbox and ... outbox in the let me think... current calendar month as read. ", "That should be unread, not read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as read' but immediately corrects themselves with 'That should be unread, not read.' This constitutes a revision where the final audible intent is to mark items as unread. The correction is clear and directly modifies the initial incorrect value, satisfying the revision type criterion.", "heard_transcript": "Mark everything in my um, let's see, Gmail inbox and outbox in the current calendar month as read. That should be unread, not read." }, "candidate_id": "wav_dc6b5ea67c8f6c5d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'let me see', 'let me think') that are well-placed and reflect genuine hesitation. The revision ('That should be unread, not read') is a clear, spontaneous correction with a believable change-of-mind intonation and a direct reference back to the initial request. The prosody is continuous and conversational, with a natural rhythm and pacing throughout the entire utterance." } ] }, "is_default": true }, { "variant_id": "revision_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox last calendar month as unread. Make that in the current calendar month, not last calendar month.", "long_pause_offsets": [ 76 ], "source_sha256": "fd1f5ddc3e806bbe42e54f2e99b7897d6e02b7c04fb6f25a7f5ab2fd8a8afb0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time filter from 'last calendar month' to 'current calendar month', preserving the action, app, and scope of 'inbox and outbox'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/revision/revision_03_time_filter/a8748a3f7d1fe84d/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "147e6f7083d9f7cc105e7a50572de411537a3c837ffd121e8f5f8743b6b22404", "seconds": 15.76, "spoken_utterances": [ "Mark everything in my uh... Gmail inbox and outbox um... last calendar month as unread. ", "Make that in the current calendar month, not last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'current calendar month, not last calendar month'. This is a clear revision where the final intent overrides the initial incorrect value. The rest of the instruction, including the target (Gmail inbox and outbox) and the action (mark as unread), remains consistent. The correction is spoken after a brief pause and is unambiguous, satisfying the revision type criterion.", "heard_transcript": "Mark everything in my uh Gmail inbox and outbox um last calendar month as unread. Make that in the current calendar month, not last calendar month." }, "candidate_id": "wav_391f0c672de6d211", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker uses fillers ('uh', 'um') and a self-correction ('Make that in the current calendar month, not last calendar month'), which is a plausible revision. However, the correction is delivered with a slightly flatter intonation and less of a natural 'change-of-mind' cadence compared to audio_02. The overall delivery is still realistic but feels slightly more rehearsed or less emotionally spontaneous in the correction phase." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Mark everything in my Outlook inbox and outbox in the current calendar month as unread. Wait, use Gmail instead of Outlook.", "long_pause_offsets": [ 88 ], "source_sha256": "7fc56b05f69763157e27eb5dc2a799a5f01439ea8f83c0d4955bd4439fda7062", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the email app from Outlook to Gmail, while retaining the action, scope (inbox and outbox), and time condition (current calendar month)." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in my um... Outlook inbox and ah... outbox in the um... current calendar month as unread. ", "Wait, use Gmail instead of Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail sent and drafts in the current calendar month as unread. Change the folders to inbox and outbox, not sent and drafts.", "long_pause_offsets": [ 85 ], "source_sha256": "547e3db6c2c581e134d55df417e1ef2dd74135f9c6ba7747d5c1056871732879", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the folders from 'sent and drafts' to 'inbox and outbox', preserving the action, time scope, and target state (unread)." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in my ... Gmail sent and drafts in the uh... current calendar month as unread. ", "Change the folders to hmmm... inbox and outbox, not sent and drafts." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "5ccb15ed12c1f9cdba1a3fccd8643e926838a2bc4a5606484222610dfe28a8d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/cancellation/cancel_026/31c7173adc6e6847/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "ca1f135ba0600582b8e325b43c11aaa46a3727e3d475843a8cef7d6e5a0275f2", "seconds": 14.68, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread.", "Oh, wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to mark emails as unread, then immediately says 'forget what I just asked, cancel it,' which is a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread. Oh wait, forget what I just asked, cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "e94fcbc518d3430802af55499806f7a5ad7db3bb07798bbf8b174ec1831324e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/ambiguous_halt/halt_007/8b2f908dac872aa9/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "65e74458d62ff42aa1d0e01c0476439d54aa112e664d63c09d01607ccd83fe54", "seconds": 13.12, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full instruction and then says 'One moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original instruction is fully stated before the pause.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5603b16745a2524a2d9d75ba1fbf752db3932e375344a63733bb049d0b8ad770", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_2/audio_natural/filler/original/27024453b0168fcf/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "941d827d7f1dd13e6b57f5dad1520141d4a990cca8322653eebdc6be5bfdc9c6", "seconds": 9.64, "spoken_utterances": [ "Mark everything in my um... Gmail inbox and ah... outbox in the ... current calendar month as unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'uh') but does not change the core task. The final intent matches the original instruction, with no self-correction or modification. The task is to mark items as unread, which is preserved. The use of 'inbox and outbox' is consistent with the original request.", "heard_transcript": "Mark everything in my um Gmail inbox and uh outbox in the current calendar month as unread." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the current fiscal year—sorry, that should be before the current calendar year as read.", "long_pause_offsets": [], "source_sha256": "36938f00401532d839f65d5dff0e15c819e65bf72ef371087d9fc160de6dcbcd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'fiscal year' but immediately corrects it to 'calendar year'. The rest of the instruction, including the action, scope (inbox and outbox), and target condition (before the year), remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/false_start/false_start_03_time_filter/04d830f3fb508296/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "d15a512ff77ca93f99af45672d208783f56a8f3de434b4142b87020561aa49d4", "seconds": 12.84, "spoken_utterances": [ "Mark everything in my Gmail inbox and outbox hmmm... before the current fiscal year—um... sorry, that should be before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fiscal year' but immediately self-corrects to 'calendar year' and maintains the rest of the instruction. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the current fiscal year. Um, sorry, that should be before the current calendar year as read." }, "candidate_id": "wav_f03934c2008a43b1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker initially says 'fiscal year' and then smoothly corrects to 'calendar year' with a believable 'sorry' and a clear, continuous intonation shift. The pause and filler 'um' are placed naturally, and the overall prosody sounds like a spontaneous, real-time thought process without any robotic or rehearsed qualities." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Mark everything in my Outlook—I should say Gmail inbox and outbox before the current calendar year as read.", "long_pause_offsets": [], "source_sha256": "b74701ea674ab67de766d790c32ca83f7c5095866249f63093a36f80a9ae186f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the rest of the instruction including the target folders (inbox and outbox), the time condition (before the current calendar year), and the action (mark as read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/false_start/false_start_01_app/86e180931bc5df1c/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "862d4583fffa2e1cc941ae0691910bf929013a7e40a0d43ff1e136f53ce79048", "seconds": 12.2, "spoken_utterances": [ "Mark everything in my um... Outlook—uh... I should say Gmail inbox and outbox hmmm... before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value ('Gmail') should be applied, and the original incorrect value ('Outlook') should be discarded. The rest of the instruction, including 'inbox and outbox', 'before the current calendar year', and 'as read', remains unchanged and is executed as intended.", "heard_transcript": "Mark everything in my um Outlook uh I should say Gmail inbox and outbox before the current calendar year as read." }, "candidate_id": "wav_5fcfd19aa71f3386", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'Outlook' corrected to 'Gmail', but the correction is less fluid than audio_02. The speaker hesitates with 'um' and 'uh' and the correction feels slightly more abrupt. The prosody is less continuous, and the overall delivery sounds a bit more hesitant and less like a natural thought process, though still realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the current calendar year as unread—no, that should be read.", "long_pause_offsets": [], "source_sha256": "2d634c05490a95e63f07977d32c93910273f85f79b37a37ce295d0dbec0d6e65", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately corrects it to 'read'. The final intended action is to mark the specified emails as read, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/false_start/false_start_04_email_state/c055e8d88ad81def/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "bbb1a7cb37fc38563dc31c1770a1a01e11655b82f2cd185bf843f81de75f3f42", "seconds": 11.08, "spoken_utterances": [ "Mark everything in my Gmail inbox and outbox hmmm... before the current calendar year as unread—uh... no, that should be read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately self-corrects to 'read'. According to the 'false_start' type criterion, the final corrected value should be applied. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the current calendar year as unread. Uh, no, that should be read." }, "candidate_id": "wav_1735f8a16a2f88b2", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker initially says 'unread' and then corrects to 'read'. The correction is clear, but the phrasing 'no, that should be read' sounds slightly more like a mechanical correction than a spontaneous thought. The prosody is somewhat flat, and the delivery lacks the natural flow of the top two, making it less realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail inbox—scratch that, make it inbox and outbox before the current calendar year as read.", "long_pause_offsets": [], "source_sha256": "57180b9055b9566e102df39ef7a5a0a00dec69207b7704ea278c2c2ef0f88f29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'inbox and outbox'. The final scope includes both folders, matching the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/false_start/false_start_02_email_folder_scope/73a2c8db3dabccfe/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "329b98c39ac5aa8189de7554b516be5b7e692dc0892de08006eefba78c06136a", "seconds": 11.72, "spoken_utterances": [ "Mark everything in my Gmail hmmm... inbox—uh... scratch that, make it inbox and outbox before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately self-corrects with 'scratch that, make it inbox and outbox'. This is a false start where the final corrected value is 'inbox and outbox'. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final intent is to mark both inbox and outbox as read.", "heard_transcript": "Mark everything in my Gmail inbox, scratch that, make it inbox and outbox before the current calendar year as read." }, "candidate_id": "wav_576c5554e89e6edb", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. The speaker says 'inboces' and then corrects to 'inbox and outbox' with 'scratch that'. The correction is abrupt and the phrase 'scratch that' is more of a meta-commentary than a natural self-correction. The prosody is disjointed, and the overall flow feels less like a genuine thought process and more like a rehearsed or artificially constructed example." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Mark everything in my Gmail inbox and outbox as read. Oh, I forgot to mention, only mark emails from before the current calendar year.", "long_pause_offsets": [ 54 ], "source_sha256": "03b85a336c7b1c65a33fa20b2369d7e7ca4f9c0d752198bfa890891d8a9c68ae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing temporal restriction that narrows the initial broad request. The final goal combines the original scope (inbox and outbox) with the new condition (emails before the current calendar year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/addition/addition_01/0a7d5f154a3f09eb/speech.wav", "wer": 0.1, "wav_sha256": "906fcc4aa294f15b726a930d567f1e18e1c86abea568677bbe391bc7b5104e7d", "seconds": 20.08, "spoken_utterances": [ "Mark everything in um, let me think... my Gmail inbox and uh... outbox as read. ", "Oh, I forgot to mention, only mark emails from um... before the current calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the base instruction to mark everything in the inbox and outbox as read. After a pause and an interjection, they add a new constraint: 'only mark emails from before the current calendar year.' This is a clear addition that narrows the scope of the original task. The final intent combines the original action with the new condition, matching the reference instruction. The addition is spoken after a longer pause, satisfying the 'addition' type criterion.", "heard_transcript": "Mark everything in um, let me think, my Gmail inbox and outbox as read. Oh, I forgot to mention, only mark emails from um, before the current calendar year." }, "candidate_id": "wav_17d4abb5b52f576d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Mark everything in my Gmail before the current calendar year as read. One more detail: do this for both my inbox and outbox.", "long_pause_offsets": [ 70 ], "source_sha256": "900c0896c473815fac43c51eeefb516d79f6a53655cb934da81d053c2a18d803", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'outbox' scope to the initial instruction, which only mentioned 'Gmail'. The action, condition, and target app remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Mark everything in my Gmail uh... before the current calendar year as read. ", "One more detail: do this for both my ... inbox and outbox." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_filter", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the end of last year as read. I misspoke about the time filter: it should be before the current calendar year, not before the end of last year.", "long_pause_offsets": [ 82 ], "source_sha256": "70ad2060c304488e25d10af46018905500707f560010f4a2746c16704ee2fdfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time filter from 'before the end of last year' to 'before the current calendar year', while retaining the app (Gmail), locations (inbox and outbox), and action (mark as read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/revision/revision_03_time_filter/ddf716836ac42b0b/speech.wav", "wer": 0.05, "wav_sha256": "47d5cde0b5e3b2da9dba6f1075fd72943c7edd14a800c78839e8793801c200e2", "seconds": 21.52, "spoken_utterances": [ "Mark everything in um... my Gmail inbox and outbox ... before the end of last year as read. ", "I misspoke about the time filter: it should be let's see... before the current calendar year, not before the end of last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives an incorrect time filter ('before the end of last year') but immediately self-corrects with a longer pause and revision ('before the current calendar year'). According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to mark items before the current calendar year as read, which matches the original instruction. The correction is clear and unambiguous, and the rest of the task (targeting Gmail inbox and outbox, marking as read) remains unchanged.", "heard_transcript": "Mark everything in um my Gmail inbox and outbox before the end of last year as read. I misspoke about the time filter. It should be, let's see, before the current calendar year, not before the end of last year." }, "candidate_id": "wav_126a73ecbbc9bb12", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker first makes a clear request, then self-corrects with a believable hesitation ('I misspoke about the time filter'), followed by a smooth, self-repairing explanation ('It should be...') that directly references and corrects the earlier value. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_04_email_state", "proposed_instruction": "Mark everything in my Gmail inbox and outbox before the current calendar year as unread. Change the state to read, not unread.", "long_pause_offsets": [ 89 ], "source_sha256": "b7b833a15c90e7d7d0031dba8758ddf1288cc40de1fd84c71c73756ef89cf396", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly reverses the action from 'unread' to 'read', correcting the initial instruction. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/revision/revision_04_email_state/42f70cc663d2bfa5/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "93d1f0872b68f386937c579641c5c06a666ee3a7120c9158c79764369758c344", "seconds": 16.56, "spoken_utterances": [ "Mark everything in ... my Gmail inbox and outbox ... before the current calendar year as unread. ", "Change the state to um... read, not unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as unread' but immediately corrects themselves to 'read, not unread' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to mark the emails as read, which matches the original instruction. Therefore, the task should be executed as revised.", "heard_transcript": "Mark everything in my Gmail inbox and outbox before the current calendar year as unread. Change the state to um, read, not unread." }, "candidate_id": "wav_7004b0a13a8fc8a5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision where the speaker first states a task, then corrects the final state from 'unread' to 'read'. The correction is delivered with a natural hesitation ('change the state to...') and a self-repair that flows smoothly within the same utterance. The prosody and pacing are consistent, and the correction feels spontaneous, though the initial request is slightly less detailed than audio_03." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Mark everything in my Outlook inbox and outbox before the current calendar year as read. Wait, use Gmail instead of Outlook.", "long_pause_offsets": [ 89 ], "source_sha256": "b320ba59f0d11b04b415cbbc8ba989a694bfea3e961e6b8d1cb202a0ff7efa4e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email app from Outlook to Gmail, while retaining the action, scope (inbox and outbox), and time condition (before the current calendar year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/revision/revision_01_app/fd18faa7457479a1/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "c90d59e2b053417044df6ddec439ddf673834fb2e9674b8d8da0dd9ace8519f1", "seconds": 14.56, "spoken_utterances": [ "Mark everything in um... my Outlook inbox and outbox ... before the current calendar year as read. ", "Wait, use uh... Gmail instead of Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly.", "heard_transcript": "Mark everything in um my Outlook inbox and outbox before the current calendar year as read. Wait, use uh Gmail instead of Outlook." }, "candidate_id": "wav_18806472c79a6712", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker initially requests a task for Outlook, then corrects it to Gmail. The correction is introduced with a natural 'Wait' and a self-repair ('use uh... Gmail instead of Outlook'). However, the initial request is very brief and lacks detail, and the correction feels slightly abrupt, making the overall flow less complex and less convincing than the top two." } ] }, "is_default": false }, { "variant_id": "revision_02_email_folder_scope", "proposed_instruction": "Mark everything in my Gmail sent and drafts before the current calendar year as read. Make that inbox and outbox, not sent and drafts.", "long_pause_offsets": [ 86 ], "source_sha256": "46f08fbde1a858f78a1a27692d7c66da69dfd98e3009f4e3179514010998d379", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial folders from 'sent and drafts' to 'inbox and outbox', preserving the app, action, date condition, and target state (read)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/revision/revision_02_email_folder_scope/3a54554c3d50937f/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "2505e04f99701c45e2fd4f772398225d5d1170b1641d81e14c0f7af1f7007622", "seconds": 13.04, "spoken_utterances": [ "Mark everything in my Gmail uh... sent and drafts um... before the current calendar year as read. ", "Make that inbox and outbox, not sent and drafts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent and drafts' but immediately corrects to 'inbox and outbox' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Mark everything in my Gmail uh sent and drafts um before the current calendar year as read. Make that inbox and outbox, not sent and drafts." }, "candidate_id": "wav_ac772d6ab1eac0c5", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip shows a revision where the speaker first mentions 'sent and drafts', then corrects it to 'inbox and outbox'. However, the correction is delivered as a separate, slightly disjointed sentence ('Make that inbox and outbox...') rather than a smooth self-repair within the same utterance. The prosody is also flatter and less natural, and the correction feels more like a post-hoc addition than a spontaneous recollection, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "54faded14461597bddbe612c8549f8db3930c201f6e8f0513e46dd158733df03", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/cancellation/cancel_009/989a996d807601a3/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "8cb118a09f9a3bd71b60ba4b4def9a9b7de399399e5821f2306882d19b8526b4", "seconds": 13.96, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly states 'cancel that altogether' as the final interjection, which constitutes a direct cancellation of the entire task. The cancellation is unambiguous and not followed by any resumption or modification of the original request.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "9973d4853a155ef7a2791d617adde81f7aac2b40abbcfc704ff54dfedd1fbb3f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/ambiguous_halt/halt_024/e6dd716031da44ba/speech.wav", "wer": 0.04, "wav_sha256": "64e08172e16314d9fae19427f9e8bb7b730efaa8eaf966ed83ca985c1efa5866", "seconds": 13.08, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler phrase 'um, let's see' and a final interjection 'Give me a moment to think' which clearly indicates a request to pause. The original instruction is not canceled or revised, and the speaker does not execute or correct the task. The final audible intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "277d81d_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8fc2f3ef4a99da50e62fc19a1576e35244aca23beace8eb58604ff4764424715", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/277d81d_3/audio_natural/filler/original/f028080affb76b51/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "a19591b65cfd7fbefb67b20a99e297cdd8f3fc0c24885a60a233acbf8a97a4fc", "seconds": 9.32, "spoken_utterances": [ "Mark everything in um, let's see... my Gmail inbox and outbox uh... before the current calendar year as read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see', 'uh') and pauses but does not change the core task. The final intent matches the original instruction exactly, with no self-correction or modification. The filler type allows ignoring hesitations and backchannels, so the task is executed as stated.", "heard_transcript": "Mark everything in um, let's see, my Gmail inbox and outbox uh before the current calendar year as read." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_recommendation_order", "proposed_instruction": "Name the artist least recommended—no, I meant most recommended to me on Spotify.", "long_pause_offsets": [], "source_sha256": "6d9f84a485d6197033927e5561f48978c9dfbaa370e2e5bfb36f16e826198b1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects to 'most recommended'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/false_start/false_start_01_recommendation_order/1de6bc11fe5e2d2a/speech.wav", "wer": 0.0625, "wav_sha256": "30924511d230fa159c7e58c6a69fcc24cdff6455d7ae9fcbb5ad3420a0e71bd2", "seconds": 9.72, "spoken_utterances": [ "Name the artist least recommended—um... no, I meant uh... most recommended to me on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately self-corrects to 'most recommended' with a pause and 'no, I meant'. According to the 'false_start' type criterion, the final corrected value should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Name the artist least recommended um, no, I meant uh most recommended to me on Spotify." }, "candidate_id": "wav_0ad733b42165d01b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start and self-correction. The initial 'least recommended' is followed by a clear 'no, I meant' with a believable pause and intonation shift, leading into the corrected 'most recommended to me on Spotify'. The pacing, filler 'um', and the pragmatic flow of the correction feel spontaneous and unscripted, closely mimicking how a person would actually rephrase a thought mid-sentence." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Name the artist most recommended to me on Apple Music—I should say Spotify.", "long_pause_offsets": [], "source_sha256": "5238fcc48fcb9216b306467ff9a561fd64c3a196d1c7981786dacc1f3e6a419c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, scope, and intent of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/false_start/false_start_02_app/5b78db227c495fe5/speech.wav", "wer": 0.0, "wav_sha256": "48a1327fd4636b17f5e70b4834cd34456758f8fe57f268061d9ab90309099e58", "seconds": 7.44, "spoken_utterances": [ "Name the artist ... most recommended to me on Apple Music—uh... I should say ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "Name the artist most recommended to me on Apple Music uh, I should say Spotify." }, "candidate_id": "wav_b87242c8b5c2dec9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also performs a false start, correcting 'Apple Music' to 'Spotify'. However, the correction is less fluid; the initial 'uh' after 'Apple Music' is followed by a slightly longer, more hesitant pause before the correction. The phrase 'I should say Spotify' sounds a bit more like a rehearsed correction than a spontaneous thought, and the overall delivery is slightly less natural in its rhythm compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Name the artist on Spotify. Oh, I forgot to say, it should be the artist most recommended to me.", "long_pause_offsets": [ 28 ], "source_sha256": "71fc8bdc7aebbbae9c75c188374b9771229dfbbf05d97a79e32663b1e6a52bd9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'most recommended to me' to the initial broad request 'Name the artist on Spotify', refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/addition/addition_01/7689c921b62e771c/speech.wav", "wer": 0.0, "wav_sha256": "9d6860cfe32cd4ec13bf3d356e4a7695a813a2e43d671b7473780455b613fbff", "seconds": 11.76, "spoken_utterances": [ "Name the artist on let me think... Spotify. ", "Oh, I forgot to say, it should be the artist ... most recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Name the artist on Spotify' and then adds the constraint 'most recommended to me' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added information. The final intent is clear and matches the original instruction.", "heard_transcript": "Name the artist on, let me think, Spotify. Oh, I forgot to say it should be the artist most recommended to me." }, "candidate_id": "wav_f8c16c8297cbb068", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') and then clearly adds a crucial detail ('it should be the artist most recommended to me') as a post-hoc clarification. The addition is smoothly integrated into the sentence, with a believable shift in intonation and pacing that suggests spontaneous thought. The prosody is consistent, and the speaker does not sound rehearsed or robotic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Name the artist most recommended to me. I forgot to add, do this on Spotify.", "long_pause_offsets": [ 40 ], "source_sha256": "97fa17fded0a047c30aaa4b26a1c6cd9be997884ec05da03f7eb1788d717a5cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify' to the initial request, preserving the action and the recommendation scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/addition/addition_02/410f56f9ae6b4fe6/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f42c1de5535ccdc2abd8c8cfc19889f2ec739d2d94584e3b0cced0bf4e2810d1", "seconds": 13.32, "spoken_utterances": [ "Name the artist uh, let me think... most recommended to me. ", "I forgot to add, do this uh, let's see... on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially begins the task 'Name the artist most recommended to me' and then adds the missing context 'on Spotify' after a pause and self-correction. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and complete, matching the original instruction exactly.", "heard_transcript": "Name the artist, uh, let me think, most recommended to me. I forgot to add, do this, uh, let's see, on Spotify." }, "candidate_id": "wav_9193421f686b53fc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail ('do this uh let's see on Spotify'), the phrasing is less coherent and the flow is disrupted. The addition feels more like a disjointed afterthought rather than a natural continuation. The 'uh' and 'let's see' are placed awkwardly, and the final phrase 'on Spotify' is tacked on without a clear grammatical connection to the prior sentence, making the overall delivery less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_recommendation_order", "proposed_instruction": "Name the artist least recommended to me on Spotify. Wait, the artist should be most recommended to me, not least recommended.", "long_pause_offsets": [ 52 ], "source_sha256": "1c17bfad75e7da5c5801d6eb7e3512f4d3b90fb56808c6336a84aaa83567992d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial instruction by replacing 'least recommended' with 'most recommended', while retaining the app and the scope of 'artist recommended to me'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/revision/revision_01_recommendation_order/94ae1dc2112ab8fd/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ec7c77a891c7a594c86bfc4e144d3b389be162052f6cbf7a60bbd3d4b79bfac6", "seconds": 12.76, "spoken_utterances": [ "Name the artist ... least recommended to me on um... Spotify. ", "Wait, the artist should be ... most recommended to me, not least recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects themselves with 'Wait, the artist should be most recommended to me, not least recommended.' This is a clear revision where the final intent is the corrected version. The original instruction was 'Name the artist most recommended to me on Spotify,' and the final audible intent matches this exactly. Therefore, the revision criterion is satisfied.", "heard_transcript": "Name the artist least recommended to me on um Spotify. Wait, the artist should be most recommended to me, not least recommended." }, "candidate_id": "wav_6405e3580afade86", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural revision pattern: the speaker starts with 'least recommended', then self-corrects to 'most recommended' with a clear, spontaneous 'wait' and a smooth, grammatically correct revision. The prosody is conversational, with a believable hesitation and intonation shift during the correction. The pacing and filler usage ('um') are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Name the artist most recommended to me on Apple Music. Change Spotify to Apple Music.", "long_pause_offsets": [ 55 ], "source_sha256": "ecbdd484abca067c66a66bdfad291f377f1d8019fc1c9241c5350127d201c53f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Spotify to Apple Music, aligning with the final spoken instruction. The action and other conditions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/revision/revision_02_app/66c200a09d1fb602/speech.wav", "wer": 0.0, "wav_sha256": "ce3e470a079d18ff8c61991dd192453bab4215e2db82e2c5041f55b38e968feb", "seconds": 11.04, "spoken_utterances": [ "Name the artist most recommended to me on um... Apple Music. ", "Change uh... Spotify to Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify' but immediately corrects it to 'Apple Music' after a pause and a self-correction ('Change uh, Spotify to Apple Music'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to request the artist most recommended on Apple Music, which matches the original instruction's structure but with the correct target platform. No other constraints are altered.", "heard_transcript": "Name the artist most recommended to me on um, Apple Music. Change uh, Spotify to Apple Music." }, "candidate_id": "wav_7f1e724923d85c35", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision but with less natural flow. The speaker begins with 'Apple Music', then corrects to 'Spotify to Apple Music'. However, the correction is abrupt and lacks a clear hesitation or self-correction marker like 'wait' or 'actually'. The prosody is flatter, and the transition between the two parts feels more like a disjointed edit than a spontaneous thought process, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "29b7955c72285f66c5452915b1dda3505c9e93274861b609c67c91ef9ed458db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/cancellation/cancel_018/9abd378e83b48fe2/speech.wav", "wer": 0.0, "wav_sha256": "c7104f8fdbb371bce324c5e68eb88994949f25b3e5efb4a3873147b6ba8324e1", "seconds": 11.56, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify.", "Oh, wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to name the artist, but immediately follows with 'Oh wait, I'd rather not do that after all. Please cancel it.' This constitutes a clear and final cancellation of the original request. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "Name the artist most recommended to me on Spotify. Oh wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "df3f7ea240160850c294f650335e65ff5d6a77d884b80293d55c9528d9b98cdf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/ambiguous_halt/halt_028/6a34cec30a019983/speech.wav", "wer": 0.0, "wav_sha256": "aca4de2b1f42d8188f0e8a4cf29f8b3cee3ffb33f386142a35f8640a1f3099a5", "seconds": 10.0, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. No cancellation or execution is requested, and the original task is not revoked.", "heard_transcript": "Name the artist most recommended to me on Spotify. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8ed19a907434b9f76abbe74b8932e01662f36e6a0ac5c4e1c878b35a98c09f04", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_1/audio_natural/filler/original/d113222461630072/speech.wav", "wer": 0.0, "wav_sha256": "f0dd52b3c0d91ffd5b20e04341c2a84e420ee227e56b50d4698963c43226d848", "seconds": 5.12, "spoken_utterances": [ "Name the artist ... most recommended to me on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or repair. The speaker delivers the full request without interruption, so the final intent matches the original instruction exactly. As this is a 'filler' type, any initial hesitation or pause within the instruction is ignored, and the final goal is to execute the task as stated.", "heard_transcript": "Name the artist most recommended to me on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Name the artist least recommended to me on Apple Music—sorry, I meant Spotify.", "long_pause_offsets": [], "source_sha256": "043a14bbb5f587d33436dd967e0ef0a9eb0b6af69051a93e1df6bb6d63a97fe2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, direction, and target entity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/false_start/false_start_02_app/d5600b107d75ea33/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8006f136a5cdc8122a49e1672f7c945cffa805fe823cc239090dd1ef22c996d3", "seconds": 11.48, "spoken_utterances": [ "Name the artist um, let's see... least recommended to me on Apple Music—ah... sorry, I meant ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Name the artist, um, let's see, least recommended to me on Apple Music, uh, sorry, I meant Spotify." }, "candidate_id": "wav_3abac10807ec9012", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural, spontaneous flow. The initial hesitation ('um', 'let's see') and the self-correction ('sorry I meant Spotify') are delivered with authentic prosody and pacing. The correction is immediate and integrated smoothly into the ongoing thought process, making it sound like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_01_recommendation_order", "proposed_instruction": "Name the artist most recommended—no, I mean least recommended to me on Spotify.", "long_pause_offsets": [], "source_sha256": "4292b36346d66844bc646c18fc232e340b56cf0daeddff32bfe0137bdeb644e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'most recommended' to 'least recommended' within the same utterance, preserving the app (Spotify) and the user perspective ('to me'). The final intended action is to identify the least recommended artist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/false_start/false_start_01_recommendation_order/1af61db7075eeb00/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "3b557f3569332ba98830d5bd06c3c69139199764b9d0834a99b774848e01bba1", "seconds": 9.2, "spoken_utterances": [ "Name the uh... artist most recommended—ah... no, I mean hmmm... least recommended to me on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most recommended' but immediately self-corrects to 'least recommended'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Name the uh artist most recommended uh no, I mean least recommended to me on Spotify." }, "candidate_id": "wav_bf59cd509146615b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While this clip also contains a false start and self-correction, the delivery is less fluid. The initial hesitation ('uh') and the correction ('no I mean') are present, but the overall pacing feels slightly more stilted. The phrase 'me's recommended to me' is awkwardly phrased and delivered with a less natural intonation, making the correction feel more like a forced edit than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Name the artist to me on Spotify. Oh, I forgot to mention, it should be the artist least recommended to me.", "long_pause_offsets": [ 34 ], "source_sha256": "0ed74da8d48bba1f2f47ec478dd47da39f8af1962ec2c7988a0184ca984b6e49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'least recommended to me', which refines the initial vague request. The app (Spotify) and action (name the artist) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/addition/addition_01/a4bf53eed120f8e4/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5bc606c07193cbee73b34921bcfee163290fa5000ecf438c9870243096255d33", "seconds": 14.84, "spoken_utterances": [ "Name the let's see... artist to me ... on Spotify. ", "Oh, I forgot to mention, it should be the artist uh, let's see... least recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Name the let's see artist to me on Spotify', which is incomplete and unclear. After a pause, they correct and add the missing constraint 'it should be the artist uh let's see least recommended to me'. This addition completes the original instruction by specifying the correct artist type. The final intent matches the original instruction, and the addition is applied correctly.", "heard_transcript": "Name the let's see artist to me on Spotify. Oh, I forgot to mention, it should be the artist uh let's see least recommended to me." }, "candidate_id": "wav_e1de749aaf860e0b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Name the artist least recommended to me. I left out a detail: do this on Spotify.", "long_pause_offsets": [ 41 ], "source_sha256": "1a3e99cc4a6a4a5cf6bfe627cfcd7c557b8e7da6c36a086ade1ea016fcb64035", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify' to the initial request, preserving the action and the recommendation scope." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist um, let's see... least recommended to me. ", "I left out a detail: do this um... on Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Name the artist least recommended to me on Apple Music. Let me correct the app: the artist should be least recommended on Spotify, not Apple Music.", "long_pause_offsets": [ 56 ], "source_sha256": "1c358bb0052878859b2a873bc88cc098dd9decf8c40278e0ac9c11d8f50e3b2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the action, the user reference, and the recommendation direction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/revision/revision_02_app/a9cdccc51ccef5e8/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "1847a14b6387eed39d3ec72cf93ebef9445b43b12fb9d74a821a54b0bc80d20f", "seconds": 13.84, "spoken_utterances": [ "Name the ... artist least recommended to me on Apple Music. ", "Let me correct the app: the artist should be least recommended on hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "Name the artist least recommended to me on Apple Music. Let me correct the app. The artist should be least recommended on Spotify, not Apple Music." }, "candidate_id": "wav_5489329460ab7df1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural prosody and a believable self-correction. The pause before 'Let me correct the app' is plausible, and the revision from 'Apple Music' to 'Spotify' is smoothly integrated with a natural change-of-mind intonation and reference back to the earlier statement. The pacing and rhythm are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_01_recommendation_order", "proposed_instruction": "Name the artist most recommended to me on Spotify. Wait, the artist should be least recommended, not most recommended.", "long_pause_offsets": [ 51 ], "source_sha256": "e09d2dffd8e929b428a82bd53f754e6a02400547653d7a152686cc8162d9e0ab", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recommendation direction from 'most' to 'least', preserving the app (Spotify) and the scope of 'artist recommended to me'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/revision/revision_01_recommendation_order/f970cd12b9a227dc/speech.wav", "wer": 0.05, "wav_sha256": "883518d79dff77d69a48d915ee1599f09bcd04292f64728d0dce881fc90b9fa9", "seconds": 12.04, "spoken_utterances": [ "Name the artist um... most recommended to me ... on Spotify. ", "Wait, the artist should be ... least recommended, not most recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most recommended' but immediately corrects themselves with 'Wait, the artist should be least recommended, not most recommended.' This constitutes a revision where the final intent is the corrected version. The original instruction was 'least recommended,' so the final audible intent matches the reference.", "heard_transcript": "Name the artist um most recommended to me on Spotify. Wait, the artist should be least recommended, not most recommended." }, "candidate_id": "wav_0ca4046778fc3f45", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then self-corrects from 'most recommended' to 'least recommended'. While the correction is natural, the initial phrase 'Name the artist um' includes a filler and a slightly awkward pause that makes the start less fluid. The correction is immediate and plausible, but the overall delivery feels slightly less cohesive than audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "6e367a622305cdbfa3b7382a29fd2594fbca0943cbc6a813a9077909bc1154df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/cancellation/cancel_003/b2a04900cbfa552f/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1bc0ec145fc0e47efed92259ba5063dbb010c0277f48e5a9d92ac41bbda69b4a", "seconds": 11.68, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify.", "Wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the original task but immediately corrects themselves with 'Wait, please cancel that,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify. Wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "ead5ac39a51b22d339013d24e69791233ec9a88edc06b6f96e627e1587888271", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/ambiguous_halt/halt_032/4891670b48c7c390/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "ecc04da179438d556114a2e1098d01f4b15225336d7930ab6ce6198c37c31b31", "seconds": 11.64, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Wait, before you do that,' which is a clear request to pause the action. The final interjection is a pause command, not a cancellation or a new instruction. The audio ends with this request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify. Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "71342c7f979cf1e41456b312ff93ad728c4a4d9af10d9b4b2e4310b126e24974", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_2/audio_natural/filler/original/27968835c2499b93/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "743eb9205c8e29ffcc0fef39a897d20628bd1febf0621efb1e36c1dbea2053f6", "seconds": 7.56, "spoken_utterances": [ "Name the artist um... least recommended to me uh, let's see... on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not change the core request. The final intent remains to name the artist least recommended on Spotify, matching the original instruction. The fillers do not constitute a pause or cancellation, so the task should be executed.", "heard_transcript": "Name the artist um least recommended to me uh let's see on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "287e338_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_recommendation_order", "proposed_instruction": "Name the artist least recommended—no, that should be most recommended to me on Spotify.", "long_pause_offsets": [], "source_sha256": "6d368664f10f93ccdac337a3f7630560887d13e37dd414068e63886ae6a59078", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately corrects it to 'most recommended'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/false_start/false_start_01_recommendation_order/18708be9eb9ae1dd/speech.wav", "wer": 0.0, "wav_sha256": "08c26762093e96b9b6547c42088d48807fdfbbe9324342ec0fe00b7fe969b3b3", "seconds": 11.48, "spoken_utterances": [ "Name the artist um, let's see... least recommended—uh... no, that should be most recommended to me on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least recommended' but immediately self-corrects with 'no, that should be most recommended'. This is a clear false start where the final, corrected value is the intended goal. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Name the artist um let's see least recommended uh no, that should be most recommended to me on Spotify." }, "candidate_id": "wav_be7af67a54344757", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Name the artist most recommended to me on Apple Music—I should say Spotify.", "long_pause_offsets": [], "source_sha256": "5238fcc48fcb9216b306467ff9a561fd64c3a196d1c7981786dacc1f3e6a419c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, scope, and intent of the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist most recommended to me on um... Apple Music—uh... I should say um... Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Name the artist on Spotify. I didn't mention this, but it should be the artist most recommended to me.", "long_pause_offsets": [ 28 ], "source_sha256": "d1e815ce0acbacb147df5013e177d7deb660a3cad2031b784b5571565f7e40b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'most recommended to me' to the initial broad request 'Name the artist on Spotify', refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/addition/addition_01/8bf4fe6bb38edf13/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "b114f50bd3111ea5e92d3d440d096ad2827a91914df4532a687f8ea87cf267fa", "seconds": 11.6, "spoken_utterances": [ "Name the artist on um... Spotify. ", "I didn't mention this, but it should be the artist uh, let's see... most recommended to me." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Name the artist on Spotify' is broad. The speaker then adds the constraint 'most recommended to me' after a pause, refining the request. The final intent combines the original request with the added detail, matching the reference instruction. The addition is clear and not contradicted.", "heard_transcript": "Name the artist on um Spotify. I didn't mention this but it should be the artist uh let's see, most recommended to me." }, "candidate_id": "wav_110bc411b429b529", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Name the artist most recommended to me. Oh, and one more thing: do this on Spotify.", "long_pause_offsets": [ 40 ], "source_sha256": "0a37cf5142a3f70b111b79ac4b5a4fbeccc05ac4de6efde55723737049a01ad8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing platform requirement, preserving the original action and scope." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist uh... most recommended to me. ", "Oh, and one more thing: do this let's see... on Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_recommendation_order", "proposed_instruction": "Name the artist most listened to me on Spotify. Wait, the artist should be most recommended to me, not most listened.", "long_pause_offsets": [ 48 ], "source_sha256": "a298b33aad55223d21ee04a6c0c54d998ee910806155b73cba76a7cb6a0e29c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'most listened to' to 'most recommended to', preserving the app and the recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/revision/revision_01_recommendation_order/a3eb0c154d4d3493/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "26a72a9ea0012e053391a88d5efd14ae12db99402e638a29ebd679ceaf1b1665", "seconds": 13.76, "spoken_utterances": [ "Name the artist most listened to me on hmmm... Spotify. ", "Wait, the artist should be um... most recommended to me, not most listened." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most listened to me' but immediately corrects themselves with 'Wait, the artist should be um most recommended to me, not most listened.' This is a clear revision where the final intent is to find the most recommended artist, not the most listened-to artist. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Name the artist most listened to me on Spotify. Wait, the artist should be um most recommended to me, not most listened." }, "candidate_id": "wav_2719e7d5503c6c13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly realistic revision pattern. The speaker begins with a plausible initial request ('Name the artist most listened to me on Spotify'), then hesitates with 'Hmm' and a pause, followed by a self-correction ('Wait, the artist should be...') that clearly references and revises the earlier phrasing ('most listened' to 'most recommended to me'). The intonation and pacing of the correction sound spontaneous and natural, with a believable shift in thought. The flow from initial statement to correction is seamless and typical of real human speech." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Name the artist most recommended to me on Apple Music. Change Spotify to Apple Music.", "long_pause_offsets": [ 55 ], "source_sha256": "ecbdd484abca067c66a66bdfad291f377f1d8019fc1c9241c5350127d201c53f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Spotify to Apple Music, aligning with the final spoken instruction. The action and other conditions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/287e338_3/audio_natural/revision/revision_02_app/7a1a305726360a6d/speech.wav", "wer": 0.0, "wav_sha256": "c94fb5944409535db6e446911315f5faf5489065c0e367c61cb356fc838c5882", "seconds": 12.52, "spoken_utterances": [ "Name the artist ... most recommended to me on let me think... Apple Music. ", "Change ... Spotify to Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify' but then corrects it to 'Apple Music' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to name the artist most recommended on Apple Music, which matches the revised goal. The original instruction is overridden by the correction.", "heard_transcript": "Name the artist most recommended to me on, let me think, Apple Music. Change Spotify to Apple Music." }, "candidate_id": "wav_826cf2f74d9548de", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is less realistic due to its unnatural structure. It starts with a request ('Name the artist most recommended to me on...'), then includes a filler ('Let me think') and a correction ('change Spotify to Apple Music'). However, the correction is delivered as a separate, abrupt command rather than a self-repair within the same utterance. The transition from the initial request to the correction feels like a disjointed edit or a robotic insertion, lacking the natural flow and intonation of a genuine speaker reconsidering their words mid-sentence." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "287e338_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_020", "hide_variant_original": true, "source_sha256": "e107f46008c9fa2e71f52f049242522e5a2d3fc6a66d0c944cb33db5cecb1e91", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist uh... most recommended to me on ... Spotify.", "Oh, wait, never mind, cancel that request." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "287e338_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "f6d370885e6adcd90564d2bc6edae5c031b9c478372297c6ed0f30be24b866da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist uh... most recommended to me on ... Spotify.", "Wait, give me a second first." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "287e338_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e4f2bc04e2715f22c5603d9309c0d8efa2ad3c247b593c888e3a378b21db5be7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Name the artist uh... most recommended to me on ... Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "2c544f9_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $250 on venmo to Caitlyn—sorry, I meant Catherine.", "long_pause_offsets": [], "source_sha256": "16b44c72224b2f224a55e521a83b3b543f54f36466e2a6200c805f3f898f35db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caitlyn' but immediately corrects to 'Catherine' with 'I meant Catherine'. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/false_start/false_start_03_person_name/de9dafedec603577/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "31ce76f16856ef58eb4277dbda0c349512ec966e178229114231cb4722d8cb05", "seconds": 9.12, "spoken_utterances": [ "Send um... two hundred fifty dollars on venmo to Caitlyn—uh... sorry, I meant Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caitlyn' but immediately self-corrects to 'Catherine' after a pause and apology. According to the 'false_start' type criterion, the final corrected target should be applied, and the original incorrect value discarded. The amount and app remain consistent. The final intent matches the original instruction.", "heard_transcript": "Send um two hundred and fifty dollars on Venmo to Caitlyn. Oh, sorry, I meant Catherine." }, "candidate_id": "wav_567b69d2e129f3a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker initially says 'Caitlyn' and then clearly and smoothly corrects to 'Catherine' with a believable 'sorry, I meant' phrase. The pause before the correction and the slight hesitation in 'um' at the start sound spontaneous. The overall prosody is conversational and the flow from initial error to correction feels authentic." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Send $250 on CashApp—I should say venmo to Catherine.", "long_pause_offsets": [], "source_sha256": "23b016c459c1c7121490360e29a3ea04e7e3af1f972b0f81f66ee296d75445dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions CashApp but immediately corrects to Venmo. The amount and recipient remain unchanged, and the final intended app is Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/false_start/false_start_02_app/3f757245dc925ef3/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "ddeea7a04a485a963e9717c06e3a1a87fe93085ca5f0a4eec9bfa5902dd08af6", "seconds": 12.2, "spoken_utterances": [ "Send uh, let's see... two hundred fifty dollars on CashApp—ah... I should say venmo to let's see... Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash app' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('cash app') is discarded. The amount ($250) and recipient (Catherine) remain unchanged and correctly heard. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Send a, let's see, two hundred and fifty dollars on cash app, uh, I should say Venmo to, let's see, Catherine." }, "candidate_id": "wav_59e04096a100dc21", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with a clear self-correction from 'Cash App' to 'Venmo'. The speaker's hesitation ('uh', 'let's see') and the natural-sounding correction are realistic. However, the correction is less integrated than in audio_03, as the speaker restarts the phrase rather than smoothly incorporating the correction. The prosody is slightly less fluid, making it feel a bit more like a rehearsed correction than a spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $200—make that $250 on venmo to Catherine.", "long_pause_offsets": [], "source_sha256": "d21b86f8c50fa134228765ee5768396edbbb275ac9c1964195436a45e6c041de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces two hundred with two hundred fifty; the recipient, app, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/false_start/false_start_01_amount/007d28a898eb83f4/speech.wav", "wer": 0.0, "wav_sha256": "fd4f8ea522b8529a39bc84025a50888cbd88cb80b4d62b6b4e3f96eb1e34b9fd", "seconds": 7.6, "spoken_utterances": [ "Send two hundred dollars—uh... make that two hundred fifty dollars on um... venmo to Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'two hundred dollars' but immediately self-corrects to 'make that two hundred fifty dollars'. According to the false_start type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The corrected amount matches the reference instruction exactly. All other constraints (app: Venmo, recipient: Catherine) are preserved and correct.", "heard_transcript": "Send two hundred dollar uh make that two hundred fifty dollars on um Venmo to Catherine." }, "candidate_id": "wav_1b2da959c69c8f54", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a revision where the speaker changes '200' to '250'. While the 'make that' construction is natural, the correction is abrupt and lacks the hesitation or filler that would make it sound spontaneous. The speaker's voice is very steady and the correction feels more like a read-aloud edit than a genuine moment of reconsideration. The prosody is flat and the delivery lacks the natural rhythm of someone formulating a request in real time." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Send $250 on to Catherine. Oh, and one more thing: use venmo to send the money.", "long_pause_offsets": [ 27 ], "source_sha256": "13bb1d60023b950a37108bb6bae33b1e0aa70c5e364e80d52fa12f1c256ead98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the payment method as Venmo, which is consistent with the original instruction. The amount and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/addition/addition_02/f9a38c7fd9ba4fa2/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "5251d0768ea1f738fa05bb5aec06bd8f4deff6ec421debfcc0b34e0ebe1b8f52", "seconds": 11.84, "spoken_utterances": [ "Send uh, let me think... two hundred fifty dollars on to um... Catherine. ", "Oh, and one more thing: use venmo to send the money." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction to send $250 on Venmo to Catherine is completed by the addition of 'use Venmo to send the money' after a pause, which clarifies the method. The final intent combines the initial request with the added detail, matching the original instruction exactly. No contradictions or omissions remain.", "heard_transcript": "Send uh, let me think, two hundred and fifty dollars on to um Catherine. Oh, and one more thing, use Venmo to send the money." }, "candidate_id": "wav_f0d1b01ae0bedb55", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a clear intent, uses a believable 'let me think' filler, and smoothly adds a new detail ('Use Venmo') after the initial request. The prosody is consistent, with natural pauses and intonation that suggest spontaneous thought. The addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "Send $250 on venmo to. I left out a detail: send it to Catherine.", "long_pause_offsets": [ 23 ], "source_sha256": "c627debe944bf5f9fe8748aea2bc46fd847513920b377676a016a08d462a6b2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing recipient 'Catherine' after an incomplete initial utterance. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/addition/addition_03/7e89c8dc538fa7d1/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "707c62be94afb711425e05c0e9bd6fdf289912360b7b5d539cf9dc14a60353f4", "seconds": 11.12, "spoken_utterances": [ "Send ... two hundred fifty dollars on venmo to. ", "I left out a detail: send it to um... Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send $250 on Venmo, then corrects by adding the recipient 'Catherine' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and matches the original instruction.", "heard_transcript": "Send two hundred and fifty dollars on Venmo to. I left out a detail. Send it to um, Catherine." }, "candidate_id": "wav_0437d3a9259c83f7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision pattern with a believable 'I left out a detail' self-correction. The speaker's tone and pacing are natural, and the pause before the correction feels like a moment of recollection. However, the initial request is slightly less fluid than audio_01, and the correction is delivered with a slightly more abrupt transition, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Send on venmo to Catherine. Oh, I forgot to mention, send $250.", "long_pause_offsets": [ 28 ], "source_sha256": "51a8741bd6aaf06e7b33ea9818138f7e6f1a22391a4f55303ab2bd822b2fdfed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing monetary amount, completing the initial instruction. The recipient and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/addition/addition_01/0bd9c7329ba0ffba/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "723ef43bf5e2c55bcccd065d4567f3ffba11d697c999a97623013114aa8a6f6e", "seconds": 12.2, "spoken_utterances": [ "Send on venmo to ... Catherine. ", "Oh, I forgot to mention, send let me think... two hundred fifty dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Send on Venmo to Catherine' is followed by a longer pause and an addition of the amount 'two hundred fifty dollars'. The addition completes the original request by specifying the missing monetary value. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Send on Venmo to Catherine. Oh, I forgot to mention, send, let me think, two hundred fifty dollars." }, "candidate_id": "wav_7bc4994c78944629", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. It begins with a clipped, incomplete request ('Send on Venmo to...'), followed by a filler ('Oh, I forgot to mention') and then a hesitant 'let me think' before stating the amount. The sequence feels disjointed, and the speaker's voice seems to restart mid-thought rather than flow naturally. The prosody is less continuous, and the overall impression is of a rehearsed or poorly connected set of phrases." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_person_name", "proposed_instruction": "Send $250 on venmo to Caleb. Make that Catherine, not Caleb.", "long_pause_offsets": [ 29 ], "source_sha256": "198c0180f149a88e4e9655706d427466ad99ef136eebb85b16daf27f93e1d240", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from Caleb to Catherine, while retaining the amount, app, and action. The original instruction's recipient is replaced as intended." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/revision/revision_03_person_name/03b78c964f79488f/speech.wav", "wer": 0.0, "wav_sha256": "0b4c207cc8148460577fcf44f73a6bf879b09992e4a625f9205897ced87009ff", "seconds": 12.48, "spoken_utterances": [ "Send um... two hundred fifty dollars on venmo to uh, let me think... Caleb. ", "Make that Catherine, not Caleb." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Caleb' but immediately corrects to 'Catherine' with 'Make that Catherine, not Caleb.' This is a clear revision where the final target is Catherine. The amount and app remain unchanged. The revision type criterion is satisfied.", "heard_transcript": "Send um two hundred and fifty dollars on Venmo to uh let me think Caleb. Make that Catherine, not Caleb." }, "candidate_id": "wav_8225011549c01bac", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear intent, uses a believable 'let me think' pause to recall the recipient, and then smoothly self-corrects from 'Caleb' to 'Catherine' with a natural intonation shift. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Send $250 on PayPal to Catherine. Wait, send it on Venmo, not PayPal.", "long_pause_offsets": [ 34 ], "source_sha256": "b21949bbc3be6c3a25f8c4fa66b0ca1bb96a30511f80de6ae25bc587361f23a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the amount and recipient. The amount 'two hundred fifty dollars' matches the original '$250'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/revision/revision_02_app/fd7276187f2dbf07/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "3d74a2d68307574396562933198e1b7387e073d18c3af05f18784c40fdf7c12e", "seconds": 12.04, "spoken_utterances": [ "Send um... two hundred fifty dollars on PayPal to ... Catherine. ", "Wait, send it on ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with the correct amount, recipient, and app. Therefore, the task is executed as intended.", "heard_transcript": "Send um two hundred and fifty dollars on PayPal to Catherine. Wait, send it on Venmo, not PayPal." }, "candidate_id": "wav_000f55110f7a86be", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic but slightly less so than audio_01. The speaker starts with a clear request, then hesitates with 'Wait' before correcting the platform from 'PayPal' to 'Venmo'. The correction is plausible, but the 'Wait' is abrupt and the shift in intonation feels a bit more like a script than a spontaneous thought. The prosody is otherwise natural, but the correction lacks the smoothness of audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_amount", "proposed_instruction": "Send $200 on venmo to Catherine. Sorry, I meant $250, not $200.", "long_pause_offsets": [ 33 ], "source_sha256": "c9ec92681b14c88d37d8793b3ae4969bc8baeba314f38eb5a3cb5c1481bd5916", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from two hundred to two hundred fifty dollars, while retaining the recipient, app, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/revision/revision_01_amount/85482f6eed924250/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "192b7ce241876ad9068a3498c430ec515847f99ff8d2287719611bf05660396e", "seconds": 13.16, "spoken_utterances": [ "Send uh... two hundred dollars on um... venmo to Catherine. ", "Sorry, I meant hmmm... two hundred fifty dollars, not two hundred dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says $200, then immediately corrects to $250 with 'not two hundred dollars'. This is a clear revision where the final value is $250, matching the original instruction. The correction is spoken after a short pause and is not accompanied by any cancellation or pause request. The final intent is to send $250, which aligns with the expected control intent for a revision.", "heard_transcript": "Send uh two hundred dollars on um Venmo to Catherine. Sorry, I meant two hundred fifty dollars, not two hundred dollars." }, "candidate_id": "wav_30536e07183ef200", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. The speaker starts with a clear request, then after a pause, corrects the amount from '$200' to '$250'. However, the correction is delivered with a flat, rehearsed tone and lacks the natural prosody of the other clips. The 'Sorry, I meant' phrase sounds like a formulaic insertion rather than a spontaneous realization, and the overall delivery feels more like a read-aloud than a genuine spoken request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "1f8e60be03eb1b2c63ded40b150aaddd62d693ae0941c0d0b08c39de78cb4435", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/cancellation/cancel_034/4dc062c558a96447/speech.wav", "wer": 0.15, "wav_sha256": "5a266b1ac360b1811462db9f644dad1329cacc996c2278d4165110fdf39a637e", "seconds": 13.12, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the task but then clearly states 'Hold on, let's drop the whole task,' which is a direct cancellation of the entire request. The final interjection is a complete revocation, satisfying the cancellation type criterion.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "286591a3e23bdd04366a002169dae7605c9c639046b609ee4425111727ebe6a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/ambiguous_halt/halt_024/dc81311856722a93/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "dd580bfe2dd668ed85b05e081103a3cf956072d28159e3a375af07b7f2ca6b61", "seconds": 12.08, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Give me a moment to think,' which is a clear request to pause. The task is not canceled, and the speaker does not provide a correction or new instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a3683cf762e5c4ea63e9009bacd24c675e591571af89758fd1e774fc12b5ae86", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_1/audio_natural/filler/original/bc8b982d41857f71/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "11f41f1a093d497a49a6320ceb68c785e3ff68d7076fa39acf57aeb8aafac507", "seconds": 8.24, "spoken_utterances": [ "Send um... two hundred fifty dollars on let's see... venmo to um... Catherine." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'let's see' but ultimately delivers the full original instruction with correct values, target, and action. The hesitations are fillers that do not alter the final goal. The final intent matches the original instruction exactly.", "heard_transcript": "Send um two hundred fifty dollars on let's see, Venmo to um Catherine." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $427 on venmo to Amanda—I meant Anita.", "long_pause_offsets": [], "source_sha256": "cf7d81375eea8017291b5a64245078c5c81401d8a5bc36e408d44ea106f399f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amanda' but immediately corrects to 'Anita' with 'I meant Anita.' The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/false_start/false_start_03_person_name/7478412a2a8a81c0/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "20d737744177da861970365704c3e211384cc499ac45036d45a33565f44be2a5", "seconds": 7.12, "spoken_utterances": [ "Send four hundred twenty-seven dollars on uh... venmo to Amanda—um... I meant Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amanda' but immediately self-corrects to 'Anita' with 'I meant Anita.' This is a clear false start where the final target is corrected. The amount and app remain unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send four twenty seven dollars on uh Venmo to Amanda. Um, I meant Anita." }, "candidate_id": "wav_0e1ee54154b727f4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with a clear self-correction. The initial request is slightly hesitant ('send 427 dollars on uh Venmo to Amanda'), followed by a believable pause and a self-repair ('I meant Anita'). The correction is smoothly integrated into the same utterance with natural intonation and pacing, making it sound like a spontaneous recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Send $427 on PayPal—sorry, make that venmo to Anita.", "long_pause_offsets": [], "source_sha256": "91fec512690b738b370319344ba29d03fb8164361e8f7a81a740799ff625e983", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions PayPal but immediately corrects to Venmo. The amount and recipient remain unchanged. The final intended app is Venmo, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/false_start/false_start_02_app/26a9dacd4ea11b6b/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "6ec70fd30340739a34e12d6f11759f2a855b538320a512d1872571c55b49a769", "seconds": 9.88, "spoken_utterances": [ "Send uh... four hundred twenty-seven dollars on PayPal—um... sorry, make that uh... venmo to Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' with 'sorry, make that uh Venmo'. This is a false start where the final value is the corrected one. The amount and recipient remain unchanged. The final intent matches the original instruction.", "heard_transcript": "Send uh four hundred twenty seven dollars on PayPal, um, sorry, make that uh Venmo to Anita." }, "candidate_id": "wav_8ed742af0b66b466", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a false start where the speaker initially says 'PayPal' and then corrects to 'Venmo'. The correction is immediate and the prosody is natural, but the initial error ('PayPal') is less plausible than a name mix-up, and the correction feels slightly abrupt. The overall delivery is fluent but less nuanced than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $430—no, that should be $427 on venmo to Anita.", "long_pause_offsets": [], "source_sha256": "a78a1909ba6778757b53b7d112d834c25c8e49d40279ea622b9bb3326e45baca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces four hundred thirty with four hundred twenty-seven; the recipient, app, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/false_start/false_start_01_amount/e10b5ac131d32c1b/speech.wav", "wer": 0.1, "wav_sha256": "8d6a09110594b56eec5c5c056a0b899ee74587e6f856ad05b55ccf473bfe8c88", "seconds": 9.16, "spoken_utterances": [ "Send uh... four hundred thirty dollars—um... no, that should be four hundred twenty-seven dollars on ... venmo to Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says $430, then immediately corrects to $427. According to the 'false_start' type, the final corrected value should be executed. The amount is corrected, and all other conditions (Venmo, Anita) remain consistent. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Send a four hundred thirty dollars on Venmo to Anita. No, that should be four hundred twenty seven dollars on Venmo to Anita." }, "candidate_id": "wav_bcc9d10bea33fbb1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with '430 dollars', then corrects to '427 dollars'. While the correction is present, the initial number is incorrect and the correction is delivered with a flat, robotic intonation. The pause before the correction is long and unnatural, and the overall delivery lacks the spontaneous flow of the other two clips. The prosody is less varied, making it sound less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Send $427 on venmo to. I left out a detail: send it to Anita.", "long_pause_offsets": [ 23 ], "source_sha256": "fc59e8698b1fc3c6f543da679925c72368dfaa26194ba9164fbce21a3689e412", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing recipient 'Anita' after an incomplete initial utterance. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/addition/addition_03/b57577df505ba05c/speech.wav", "wer": 0.2916666666666667, "wav_sha256": "3b4f7f90b83dd9c7a77f826107d952815876250a83d009c68b4899041575d8bc", "seconds": 13.88, "spoken_utterances": [ "Send let's see... four hundred twenty-seven dollars on um, let me think... venmo to. ", "I left out a detail: send it to ... Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the amount and app, then pauses to correct the recipient by adding 'send it to Anita' after a longer pause. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final intent is clear and matches the reference.", "heard_transcript": "Send, let's see, four twenty seven dollars on um, let me think, Venmo to, I left out a detail, send it to Anita." }, "candidate_id": "wav_cb0b5feeecd53ebc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural-sounding addition with a clear, spontaneous self-correction. The initial request 'send $427 on Venmo to' is followed by a believable 'let me think' pause, then a smooth, pragmatic revision where the speaker explicitly acknowledges the omission ('I left out a detail') and adds the recipient 'Anita'. The prosody is consistent, the repair is integrated into the flow, and the intonation reflects a genuine moment of recall." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send $427 on to Anita. One more thing: send it on venmo.", "long_pause_offsets": [ 23 ], "source_sha256": "0f754a53a238dd41dbf43f06e2b02f14def53a95176bb1ddb1af305230b4fd9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app 'Venmo' to the initial instruction, which already specified the amount and recipient. The final goal combines both utterances without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/addition/addition_02/92893fd1297e84ca/speech.wav", "wer": 0.15, "wav_sha256": "64beeef841ef08681e0808f5a64bc80ee13d7b12f4844f3c5ddbd35b1efd1bdd", "seconds": 12.68, "spoken_utterances": [ "Send uh... four hundred twenty-seven dollars on to let me think... Anita. ", "One more thing: send it on ... venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'four twenty seven dollars' and then corrects to 'send it on Venmo' after a pause. The addition of 'send it on Venmo' clarifies the platform, combining with the original instruction to form the complete task. The final intent matches the original instruction with the addition of the platform, satisfying the 'addition' type criterion.", "heard_transcript": "Send a four twenty seven dollars on to let me think. Anita. One more thing. Send it on Venmo." }, "candidate_id": "wav_89e9d078ce491abc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable addition with a 'let me think' filler and a subsequent request to 'send it on Venmo'. The delivery is natural, but the addition of 'Anita' is less smoothly integrated than in audio_02. The speaker says 'send it on Venmo' after the pause, which slightly disrupts the flow compared to audio_02's more cohesive revision. The prosody is still credible, but the connection between the initial and added information is less seamless." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Send on venmo to Anita. Oh, I forgot to mention, send $427.", "long_pause_offsets": [ 24 ], "source_sha256": "c84db99f0201207a16957e6c3570fd28f6d3d4c8c6109c59c20c67deb0107766", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing monetary amount, which was omitted in the initial utterance. The recipient and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/addition/addition_01/28cfd3c0c98b8fe9/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "24039b0ca2551070fdb393908f77cf58cbea2c9479db0c62694d31f85c9d69dd", "seconds": 11.36, "spoken_utterances": [ "Send on venmo to let's see... Anita. ", "Oh, I forgot to mention, send hmmm... four hundred twenty-seven dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, missing the amount. The speaker then adds the amount after a pause, fulfilling the addition type criterion. The final audible goal matches the original instruction exactly.", "heard_transcript": "Send on Venmo to let's see, Anita. Oh, I forgot to mention, send four hundred twenty-seven dollars." }, "candidate_id": "wav_be19a82b1961f640", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with 'send on Venmo to let's see Anita', which sounds like a rehearsed or slightly awkward phrasing. The addition 'Oh I forgot to mention send $427' is delivered with a natural-sounding filler, but the initial part of the utterance lacks the spontaneity of the others. The prosody is somewhat flat, and the overall flow feels less like a natural thought process and more like a constructed sentence, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Send $427 on PayPal to Anita. Wait, send it on Venmo, not PayPal.", "long_pause_offsets": [ 30 ], "source_sha256": "b689f3e69c56c51f42df0952183efbe6776f91449f3a7b7f01c3bea62eb0640a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the amount and recipient. The monetary value is expressed in words but matches the reference amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/revision/revision_02_app/de698c84cf17d8ad/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "fc7b5be2b8650bd395f05397863fb4811446f528f504c83af4bd79998619c8ed", "seconds": 11.88, "spoken_utterances": [ "Send hmmm... four hundred twenty-seven dollars on um... PayPal to Anita. ", "Wait, send it on ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send the money on Venmo, which matches the original reference instruction. All other constraints (amount, recipient) are preserved.", "heard_transcript": "Send four twenty seven dollars on um PayPal to Anita. Wait, send it on Venmo, not PayPal." }, "candidate_id": "wav_7a5b0fe2cd044a2a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and spontaneous revision. The speaker begins with a plausible false start ('Send... um...'), then corrects the payment platform from PayPal to Venmo mid-sentence with a believable 'Wait, send it on Venmo not PayPal.' The hesitation and self-correction are integrated smoothly into the flow, with natural fillers and a clear, pragmatic change-of-mind. The prosody and pacing reflect genuine formulation and correction." } ] }, "is_default": true }, { "variant_id": "revision_01_amount", "proposed_instruction": "Send $450 on venmo to Anita. Sorry, I meant $427, not $450.", "long_pause_offsets": [ 29 ], "source_sha256": "8358bfcf68a7c9e7321409eee7861d95413a724f9993bef6101e0cc332265cc2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from four hundred fifty to four hundred twenty-seven, while retaining the recipient, app, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/revision/revision_01_amount/792ae47c3cfdccd0/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "08506f44b6cecd49f5cf57769acd4a3519b8fe93073dfd7e5cee5b9180aeeabc", "seconds": 14.44, "spoken_utterances": [ "Send um... four hundred fifty dollars on venmo to ... Anita. ", "Sorry, I meant ... four hundred twenty-seven dollars, not four hundred fifty dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states $450, then immediately corrects to $427 after a pause and self-correction. According to the revision type criterion, the final corrected value overrides the initial incorrect one. The recipient (Anita) and app (Venmo) remain unchanged, and the correction is clear and unambiguous. Therefore, the final intent matches the original reference instruction.", "heard_transcript": "Send um, four hundred fifty dollars on Venmo to Anita. Sorry, I meant four hundred twenty-seven dollars, not four hundred fifty dollars." }, "candidate_id": "wav_2979905587efb7ca", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision with a self-repair ('Sorry, I meant $427 not $450'). The initial request is slightly hesitant ('Send um...'), and the correction is immediate and natural. However, the repair is less integrated than in audio_03; the speaker stops, says 'sorry,' then restarts the number. While realistic, the flow is slightly more segmented compared to the seamless mid-sentence correction in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_person_name", "proposed_instruction": "Send $427 on venmo to Amanda. Change Anita to Amanda instead.", "long_pause_offsets": [ 30 ], "source_sha256": "e8563227b68be8215a699f5d357829e3fe29db03f416292be7b56913dc7c4382", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recipient from Anita to Amanda, aligning with the first utterance. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/revision/revision_03_person_name/42606217c5549096/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "6cd27915d9432dd74a06824d800d66e50e4885fefa64d7cba1ecae239f6d19cd", "seconds": 11.8, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on venmo to ... Amanda. ", "Change Anita to ... Amanda instead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send four twenty seven dollars on Venmo to Anita' and then corrects the recipient from 'Anita' to 'Amanda' with 'Change Anita to Amanda instead.' This is a revision where the final intent is to send the money to Amanda. The amount and app remain unchanged, and the correction is clear and final. The revision type requires applying the correction to the earlier instruction, which is done here.", "heard_transcript": "Send four twenty seven dollars on Venmo to Anita. Change Anita to Amanda instead." }, "candidate_id": "wav_23bedbb043d43d9a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It presents a revision but in an unnatural, almost robotic way. The speaker first makes a complete request ('Send $427 on Venmo to Amanda'), then immediately follows with a separate, declarative sentence ('Change Anita to Amanda instead'). This lacks the natural hesitation, fillers, or mid-sentence correction seen in the others. The delivery is flat and the revision feels like a post-hoc addition rather than a spontaneous correction, making it sound less like a real-time thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "7aff6152b0605fd6d559236cd6f415f83382e29e21b00b7d6d8974fb9fc8b436", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/cancellation/cancel_012/b7c4d303833804b2/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b37efe7029b5c6ec056cb8d4526094fb68993ac4bd137a33a1d3539e36bd1990", "seconds": 10.92, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita.", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests sending $427 to Anita on Venmo, but immediately corrects themselves with 'Wait, actually, I've changed my mind. Don't do that,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly revokes the original instruction, satisfying the cancellation type criterion.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita. Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "87c7164aece390e5396d2115f561490a5f7cd8685ed17e60bd471b036518caf8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/ambiguous_halt/halt_040/e5aa4ae59416ceab/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "984b03957f1c753c7e46d9449174cdf4f1e71cd87bce6a38338a478ff740d6c4", "seconds": 9.64, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction with a final interjection 'Hang on, just a minute' that clearly requests a pause. This matches the 'ambiguous_halt' type criterion, which requires an audible request to hold/wait/pause with no permanent cancellation or permission to resume. The initial request is fully audible and complete, and the final interjection is a clear pause command.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9806321d2f483a1055ce3ab871a72ee4eeb4aea2923bcf5c18b0f0e38e59f4c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_2/audio_natural/filler/original/62695ae60c34455e/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "887e5efd689190b4fe6e6718b899bbe33bb0ad21ac0ca99aba2dcfd2dd99f0b4", "seconds": 6.04, "spoken_utterances": [ "Send ... four hundred twenty-seven dollars on um... venmo to let's see... Anita." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'let's see' but ultimately completes the original instruction without changing the action, amount, app, or recipient. The hesitation and backchanneling are typical filler behavior and do not alter the final task goal. The final intent matches the original instruction.", "heard_transcript": "Send four twenty seven dollars on um Venmo to let's see, Anita." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_person_name", "proposed_instruction": "Send $234 on venmo to Jill—sorry, I meant Jack.", "long_pause_offsets": [], "source_sha256": "4767b2717c08f7c299b2d3a78ff48a1807d813d582a2779f0967a96bc5618a4e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately corrects to 'Jack' with 'I meant Jack'. The amount and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/false_start/false_start_03_person_name/ad70a045dfb8d78b/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "cd27e174223ea5ad3fa31706cc9e9a83f95815bd51507184bd8771cad141e7da", "seconds": 10.04, "spoken_utterances": [ "Send two hundred thirty-four dollars on um... venmo to Jill—uh... sorry, I meant um... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately self-corrects to 'Jack' after a pause and 'sorry I meant'. This is a clear false start where the final value replaces the mistaken one. The amount and app remain unchanged. The final intent matches the original instruction.", "heard_transcript": "Send two hundred thirty four dollars on um Venmo to Jill uh sorry I meant um Jack." }, "candidate_id": "wav_91663b0ddb625f8a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'uh' and 'um' fillers, a believable self-correction from 'Jilla' to 'Jack', and maintains a consistent, conversational pace. The repair is integrated smoothly within the utterance, and the prosody reflects genuine hesitation and recollection." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Send $234 on CashApp—I should say venmo to Jack.", "long_pause_offsets": [], "source_sha256": "158c2e212e98b1fc2253c9eba47c2c9d5244d35760d9a36c1fbf3369276412cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions CashApp but immediately corrects to Venmo. The amount and recipient remain unchanged, and the correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/false_start/false_start_02_app/90f07e1ed2dfd005/speech.wav", "wer": 0.125, "wav_sha256": "771b50004b27f0a5c9b9b535f5c33980971e97915276cf8377a387ad6be2c7c6", "seconds": 9.48, "spoken_utterances": [ "Send ... two hundred thirty-four dollars on CashApp—uh... I should say venmo to um... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash app' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('cash app') is discarded. The amount and recipient remain consistent. Therefore, the final audible intent matches the original instruction after correction.", "heard_transcript": "Send two thirty four dollars on cash app of, I should say Venmo to um Jack." }, "candidate_id": "wav_452525e8eb2dc960", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While this clip also contains a false start and correction, the initial mention of 'Cash App' is followed by a more abrupt and less fluid correction to 'Venmo'. The transition feels slightly less organic, and the speaker's intonation during the correction lacks the subtle pitch shift seen in audio_01, making the repair feel a bit more mechanical." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $250—no, that should be $234 on venmo to Jack.", "long_pause_offsets": [], "source_sha256": "abe537fa87cad615234b8c3dc9969a6290191a86170efbedee122a3cdb7dcd05", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces two hundred fifty with two hundred thirty-four; the recipient, app, and action remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Send two hundred fifty dollars—uh... no, that should be ... two hundred thirty-four dollars on venmo to um... Jack." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Send $234 on venmo to. One more thing: send it to Jack.", "long_pause_offsets": [ 23 ], "source_sha256": "dc527209bdf7b6bf060e502a77109e1f73297a4dfbcd658ad818f690837374c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The first utterance is incomplete, lacking a recipient. The followup adds the missing recipient, Jack, without changing any other parameters. The amount and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/addition/addition_03/c0c7addb4f2d3cee/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "68b3b7cf6d71a8034c63ad86adbd18115e25fd529bbc917981d38cb9f1714de9", "seconds": 11.56, "spoken_utterances": [ "Send let's see... two hundred thirty-four dollars on ... venmo to. ", "One more thing: send it to hmmm... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Send two hundred thirty four dollars on Venmo to' is followed by a pause and the addition 'send it to, hmm, Jack.' The final target 'Jack' is clearly stated after the initial amount and app, satisfying the addition type criterion. The final goal combines the original instruction with the added target, matching the reference.", "heard_transcript": "Send, let's see, two hundred thirty four dollars on Venmo to, one more thing, send it to, hmm, Jack." }, "candidate_id": "wav_13ec3aae73d73cbf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident instruction ('Send let's see two hundred thirty four dollars on Venmo to...'), then pauses thoughtfully ('One more thing') before adding the recipient ('Jack'). The addition is natural, with a slight hesitation and filler ('hmm') that sounds spontaneous. The prosody is consistent, and the flow from initial command to addition feels like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Send on venmo to Jack. Oh, I forgot to mention, send $234.", "long_pause_offsets": [ 23 ], "source_sha256": "ea78cb104d32176ae885ae7aa88207d83ea663a4cff1185952424f459ea91864", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing monetary amount, which is equivalent to $234. The recipient and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/addition/addition_01/4397c535f765723a/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "dff03c3ff99e056d066538946e141c856121989959c46dfaf981c6540ef47d7d", "seconds": 10.64, "spoken_utterances": [ "Send on venmo to ... Jack. ", "Oh, I forgot to mention, send um... two hundred thirty-four dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Sent on Venmo to Jack' is incomplete, but the speaker immediately adds the missing amount 'two hundred thirty four dollars' after a pause. The addition completes the original request with the correct value. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Sent on Venmo to Jack. Oh, I forgot to mention, send um, two hundred thirty four dollars." }, "candidate_id": "wav_686afd83289ae5e0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a complete but slightly hesitant request ('Sent on Venmo to Jack'), then adds the amount after a pause. The addition ('send um two hundred thirty four dollars') is less fluid than audio_01, with a more abrupt transition and a filler ('um') that sounds slightly forced. The prosody is less continuous, and the overall delivery feels more like a correction than a natural thought process, making it less realistic." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Send $234 on to Jack. I left out a detail: use venmo to send the money.", "long_pause_offsets": [ 22 ], "source_sha256": "8adf2bd3baa2c91522ad095d4eeddf16716fcab37bcc6dff60e25f1ce37bc5ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app requirement; the amount and recipient remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Send ... two hundred thirty-four dollars on to Jack. ", "I left out a detail: use uh... venmo to send the money." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_person_name", "proposed_instruction": "Send $234 on venmo to Jill. Wait, the recipient should be Jack, not Jill.", "long_pause_offsets": [ 28 ], "source_sha256": "33f75de332034e2f319be21d0fc8c4cccd1ad4050e150c2c99dc3b4a65a34f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from Jill to Jack, while retaining the amount and app. The original instruction specified Jack as the recipient, and the final spoken instruction confirms this." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/revision/revision_03_person_name/7934192acbf06618/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "0183ef842caefea151a2710d01de55267da37aa2d250d962280e15bc1e3df7c4", "seconds": 12.76, "spoken_utterances": [ "Send ... two hundred thirty-four dollars on venmo to um... Jill. ", "Wait, the recipient should be um, let's see... Jack, not Jill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Jill' but immediately corrects to 'Jack' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final audible intent is to send $234 to Jack, matching the original instruction. All other constraints (amount, app) are preserved.", "heard_transcript": "Send two hundred thirty four dollars on Venmo to um Jill. Wait, the recipient should be um let's see Jack, not Jill." }, "candidate_id": "wav_cf955f0e686c49bc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic revision pattern: the speaker begins with a complete request, then hesitates with 'um' before correcting the recipient from 'Jill' to 'Jack'. The self-correction is natural, with a plausible change-of-mind intonation and a brief pause ('wait') before the revision. The prosody is conversational, with appropriate stress and pacing, and the correction flows seamlessly within the utterance, making it sound like a spontaneous real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Send $234 on PayPal to Jack. Use Venmo instead of PayPal.", "long_pause_offsets": [ 29 ], "source_sha256": "3c0e35a595262e1e97d24b07d46beae0f52395372f4f57e08f86ee7902ad7e43", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the amount and recipient. The monetary value is expressed in words but matches the original amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/revision/revision_02_app/f59fff57819705de/speech.wav", "wer": 0.0, "wav_sha256": "472ae8c111647ca6c056de5576131a258b3f823d04722248d7064089006e9887", "seconds": 9.44, "spoken_utterances": [ "Send ... two hundred thirty-four dollars on PayPal to Jack. ", "Use ... Venmo instead of PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send $234 on Venmo to Jack, which matches the original reference instruction. All constraints are preserved and the correction is clear.", "heard_transcript": "Send two thirty four dollars on PayPal to Jack. Use Venmo instead of PayPal." }, "candidate_id": "wav_d3a303a08b8ded14", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is less realistic because it sounds like a two-part command rather than a single, natural utterance. The speaker first says 'Send $234 on PayPal to Jack' and then immediately corrects to 'Use Venmo instead of PayPal'. The delivery is flat and lacks the natural hesitation or prosodic cues of a real-time correction. The abrupt transition between the two statements feels more like a scripted instruction than a spontaneous revision, reducing its realism compared to audio_02." } ] }, "is_default": false }, { "variant_id": "revision_01_amount", "proposed_instruction": "Send $250 on venmo to Jack. Sorry, I meant $234, not $250.", "long_pause_offsets": [ 28 ], "source_sha256": "6b082676840a3b0d32dde6532566720618a320db6ef31ab02a1ec75662d179cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from two hundred fifty to two hundred thirty-four, while retaining the recipient, app, and action." }, "audio": { "status": "pending", "spoken_utterances": [ "Send two hundred fifty dollars on venmo to uh, let me think... Jack. ", "Sorry, I meant uh... two hundred thirty-four dollars, not two hundred fifty dollars." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "e3e7a3cefa0b4af7e9bf3350c8d832956534458abb2ea3673c9e062dd593a13d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/cancellation/cancel_019/115801f389aaad37/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "20460c0903365e2b9c6e399f5a39cba1028868c3acddb38e3e96aeafa5db86ec", "seconds": 11.32, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack.", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hang on, never mind, don't do that.' This constitutes a cancellation of the entire task as per the type-specific criterion. The final control intent is a cancellation, which matches the expected behavior for this type.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on Venmo to Jack. Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "b5d75ae1c989c89c870229f741133d99329a27b9ca657a741ec15888d0c66124", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/ambiguous_halt/halt_003/6958f2549ba16596/speech.wav", "wer": 0.25, "wav_sha256": "410d000060dcb1a46947a19f5663de94be23d119b9dfd014787876cd43401473", "seconds": 9.96, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack.", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then says 'Hang on' as a final interjection, which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the final audible control is a pause.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on venmo to Jack. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "2c544f9_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4732f3943148fac16b1075901de870e11adee0dcaedcbb3a36874b6bcb8c7957", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/2c544f9_3/audio_natural/filler/original/b94e5a7fa496dde9/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "e55ddd80fb6cf8f519592d36017f792d062aa9c6c05c71832005f2e643306b8c", "seconds": 7.08, "spoken_utterances": [ "Send um, let me think... two hundred thirty four dollars on venmo to ... Jack." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'um, let me think' but does not change the core task. The final audible instruction matches the original request in action, amount, app, and recipient. The hesitation is a common backchannel and does not alter the goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Send um, let me think, two hundred thirty four dollars on Venmo to Jack." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 10 minutes—no, that should be 5 minutes.", "long_pause_offsets": [], "source_sha256": "2f11905e0678ff5be05a4843d37db6992ecad33719cdc3dc310289d2f84c42dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces ten minutes with five minutes; the action, alarm type, and day scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/false_start/false_start_03_duration/3864642d70a3af4c/speech.wav", "wer": 0.0, "wav_sha256": "1e666351a11a981c5e4bf43b403eadc87d267776bd3868b09a3174db66311d4f", "seconds": 9.48, "spoken_utterances": [ "Set my um... weekday wake up alarm snooze to ten minutes—uh... no, that should be ... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'ten minutes' but immediately self-corrects to 'five minutes'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Set my um weekday wake up alarm snooze to ten minutes uh no, that should be five minutes." }, "candidate_id": "wav_1818603b9080e8e7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with a clear request, then after a brief pause and filler ('uh'), corrects the value from 'ten minutes' to 'five minutes' with a believable change-of-mind intonation. The prosody is consistent, and the correction is integrated smoothly into the utterance without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "false_start_01_day_category", "proposed_instruction": "Set my weekend—sorry, I meant weekday wake up alarm snooze to 5 minutes.", "long_pause_offsets": [], "source_sha256": "ee89b01af70689009774cdf2d00a735a72e57061abdc627074d1093060c1c4f7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'weekend' to 'weekday' within the same utterance, preserving the app function (alarm snooze), the value (five minutes), and the target (wake up alarm)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/false_start/false_start_01_day_category/dd9cd201ff8d84ac/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "90fde563c720179a4f14791c2b5321c0b859b7cf35b72588a0120d4a4ae279ce", "seconds": 7.44, "spoken_utterances": [ "Set my weekend—um... sorry, I meant uh... weekday wake up alarm snooze to um... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekend' but immediately self-corrects to 'weekday'. The final audible intent matches the original instruction exactly, with the only difference being the natural hesitation and repair. The false_start type requires the immediate correction to be applied, which it is. All other constraints (wake up alarm, snooze, 5 minutes) are preserved and correctly stated.", "heard_transcript": "Set my weekend um, sorry, I meant uh weekday wake up alarm snooze to um five minutes." }, "candidate_id": "wav_5a9193ac9c2eca95", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible false start with 'weekend' corrected to 'weekday', followed by a pause and a rephrased request. The correction is natural, and the speaker uses fillers ('um') appropriately. However, the initial error is more significant (a different day type), and the second part of the sentence is slightly less fluent than audio_02, making it slightly less realistic overall." } ] }, "is_default": false }, { "variant_id": "false_start_02_alarm_description", "proposed_instruction": "Set my weekday morning alarm—or rather, wake up alarm snooze to 5 minutes.", "long_pause_offsets": [], "source_sha256": "2c5ac033d462056c771fd34cfed4fa1de98b122d1ac670312c44a6034896026f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately corrects it to 'wake up alarm'. The final goal specifies the correct alarm type ('wake up alarm') and retains the snooze duration and weekday condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/false_start/false_start_02_alarm_description/c0bde880b7252a4f/speech.wav", "wer": 0.0625, "wav_sha256": "ecbf60259f5ad9ac471ba6c7b3ae7d23702c2807ef554b0cc7a2e7a512ba649f", "seconds": 8.96, "spoken_utterances": [ "Set my ... weekday morning alarm—uh... or rather, hmmm... wake up alarm snooze to five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately self-corrects to 'wake up alarm snooze' after a pause and hesitation. This constitutes a false start where the final corrected value replaces the mistaken one. The final instruction matches the original intent of setting the snooze to five minutes, with the only difference being the phrasing of the alarm type, which is functionally equivalent in this context.", "heard_transcript": "Set my weekday morning alarm uh, or rather, hmm, wake up alarm snooze to five minutes." }, "candidate_id": "wav_653fc5a6dc64e855", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'weekday morning alarm' corrected to 'wake up alarm snooze'. However, the correction is less smooth and the speaker's intonation sounds more hesitant and less confident. The pause and filler ('uh') are longer and less integrated, and the final phrase 'to five minutes' is delivered flatly, reducing the overall naturalness compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Set my weekday wake up alarm snooze to. I forgot to add, set the snooze to 5 minutes.", "long_pause_offsets": [ 40 ], "source_sha256": "61cbf40105f050bf6026f24b4eeddf0bb0d6cd7601434df68d991e4201f23876", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing snooze duration, preserving the original alarm type and schedule. The initial incomplete utterance is completed by the second." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/addition/addition_02/dbad033bac8f2260/speech.wav", "wer": 0.1, "wav_sha256": "1b9384f5903b45c6044175fa4b042eef5bb24461d1ff52677a2f84b92e2ab11c", "seconds": 11.28, "spoken_utterances": [ "Set my uh... weekday wake up alarm snooze to. ", "I forgot to add, set the snooze to hmmm... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to set the snooze, then adds the specific value 'five minutes' after a pause and self-correction. This matches the addition type criterion: the original instruction is combined with the added information to form the final goal, which aligns with the reference instruction.", "heard_transcript": "Set my uh weekday wake up alarm snooze to. I forgot to add, set the snooze to five minutes." }, "candidate_id": "wav_fd11f3e0a4f2b61d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Set my wake up alarm snooze to 5 minutes. Oh, right, this is just for my weekday wake-up alarm.", "long_pause_offsets": [ 42 ], "source_sha256": "a997cb8af4625187ba8a7bce94c8b14929ebc1bc75ed416d266fa662bc0d801b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial alarm type, specifying it applies only to the weekday wake-up alarm. The snooze duration remains five minutes, matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Set my wake up alarm snooze to ... five minutes. ", "Oh, right, this is just for my ... weekday wake-up alarm." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_day_category", "proposed_instruction": "Set my weekend wake up alarm snooze to 5 minutes. Sorry, that should be my weekday alarm, not my weekend alarm.", "long_pause_offsets": [ 50 ], "source_sha256": "ec5b778950d456218c3ce4be3c426801505d9087f45a1264f8fe9c2f2133f878", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm category from 'weekend' to 'weekday', while retaining the snooze duration of five minutes. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/revision/revision_01_day_category/06e6a78df7e89029/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "fb720c7d1d4f64fc290b1be167ffb447e19df3007ba75f5091b275ac13854b29", "seconds": 14.72, "spoken_utterances": [ "Set my uh... weekend wake up alarm snooze to um... five minutes. ", "Sorry, that should be my weekday alarm, not my uh, let's see... weekend alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekend alarm' but immediately corrects it to 'weekday alarm' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original instruction exactly.", "heard_transcript": "Set my uh weekend wake up alarm snooze to um five minutes. Sorry, that should be my weekday alarm, not my uh let's see, weekend alarm." }, "candidate_id": "wav_e1ab1ae2bc95bcd8", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then self-corrects from 'weekend' to 'weekday' with a believable hesitation ('uh', 'let's see') and a smooth, pragmatic repair. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a forced edit. The pacing and intonation suggest a real person formulating and adjusting a command." } ] }, "is_default": true }, { "variant_id": "revision_03_alarm_description", "proposed_instruction": "Set my weekday bedtime alarm snooze to 5 minutes. I meant my wake up alarm, not my bedtime alarm.", "long_pause_offsets": [ 50 ], "source_sha256": "ce3be62a0658f14b572479db778c7bc03355c5125d8e745ec1aec0b991df9261", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm type from 'bedtime' to 'wake up', while retaining the 'weekday' scope and 'five minutes' snooze duration." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/revision/revision_03_alarm_description/4cfbb7ec27ad2a12/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "55454fe673e7feabd80a4fc6d5459199d204beb37cbd2b69fd80750f1d907cff", "seconds": 14.12, "spoken_utterances": [ "Set my uh... weekday bedtime alarm snooze to ah... five minutes. ", "I meant my wake up alarm, not my uh, let's see... bedtime alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bedtime alarm' but immediately corrects themselves to 'wake up alarm' after a pause, which is the final and intended instruction. The value 'five minutes' is consistent. This matches the revision type criterion: the correction spoken after a pause overrides the earlier incorrect target.", "heard_transcript": "Set my uh weekday bedtime alarm snooze to uh five minutes. I meant my wake up alarm, not my uh let's see, bedtime alarm." }, "candidate_id": "wav_24eae7ab7571c069", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision from 'bedtime alarm' to 'wake-up alarm'. The speaker uses a natural 'uh' and 'let's see' filler, and the correction is integrated into the same utterance. However, the initial request is slightly less fluent than audio_03, and the correction feels a bit more abrupt. The prosody is still credible, but the flow is less smooth than audio_03." } ] }, "is_default": false }, { "variant_id": "revision_02_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 10 minutes. Set the snooze to 5 minutes instead of 10 minutes.", "long_pause_offsets": [ 51 ], "source_sha256": "f0b54af668888b2f192c89074064c27ae4ffa7dc0e65400a7c1127e811f8924c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the snooze duration from ten minutes to five minutes, while retaining the original context of the weekday wake up alarm." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/revision/revision_02_duration/46c6c12032c504f9/speech.wav", "wer": 0.12, "wav_sha256": "cae19208346ce4717eafca72f199e05854be477b8af08c2976d19d6629009282", "seconds": 16.28, "spoken_utterances": [ "Set my let me think... weekday wake up alarm snooze to ... ten minutes. ", "Set the snooze to let's see... five minutes instead of ten minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'snooze to ten minutes' but immediately corrects to 'five minutes instead of ten minutes' after a brief pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Let me think. Weekday wake up alarm snooze to ten minutes. Set the snooze to let's see, five minutes instead of ten minutes." }, "candidate_id": "wav_cdee67dd35f5c538", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip has the least realistic delivery. It contains two separate, disjointed requests ('set my...snooze to 10 minutes' and 'set the snooze to...5 minutes instead of 10 minutes') with no clear connection or natural transition. The speaker uses fillers ('let me think', 'let's see'), but the overall structure feels like a concatenation of two different attempts rather than a single, coherent thought process. The prosody is less consistent, and the second part sounds like a pasted-on addition rather than a spontaneous revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_022", "hide_variant_original": true, "source_sha256": "aff8d50bd84b59495104b565c3c4fd4cb9bd364afbde19dbe14cd67451f23a75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/cancellation/cancel_022/a7dd8bce01fc961d/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "f930058348e44c4d1422f61376a817810998ecc3f98437bb84d6c8f786e83692", "seconds": 11.2, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes.", "Wait, actually, never mind. Cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'wait, actually never mind, cancel the task.' This constitutes a final, unambiguous cancellation of the entire task. The initial partial execution is irrelevant once the cancellation is stated.", "heard_transcript": "set my uh weekday wake up alarm snooze to um five minutes. wait, actually never mind, cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_035", "hide_variant_original": true, "source_sha256": "d17c0a053b4e202632fde18999a4bfcd9dc2611d7edbd773843b127f9c119e50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/ambiguous_halt/halt_035/4d33c046e15ae900/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "aa51d19d6ea710bc8d54ee756122a550c7880fcdba0e5a7747d5ab1ee8928e12", "seconds": 10.36, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes.", "Hold on before you do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to set the alarm snooze to five minutes but then immediately says 'Hold on, before you do that,' which is a clear request to pause or wait. This constitutes an ambiguous_halt as the final interjection is a pause command with no cancellation or resumption permission. The original instruction is not executed, but the intent to pause is audible and final.", "heard_transcript": "Set my uh weekday wake up alarm snooze to um five minutes. Hold on, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0c59011e4a92a30237ea037656a2d741784844083a2b4ec7ac3bc86ce5d471de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_1/audio_natural/filler/original/34df81ee479e04d0/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "2e2bf97aa23ca3b2b0ee0b1d87e7f82ba3fcd2184464d2294a30693042b28bbf", "seconds": 6.24, "spoken_utterances": [ "Set my ah... weekday wake up alarm snooze to um... five minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core instruction. The final audible intent matches the original instruction with no correction or modification. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Set my uh weekday wake up alarm snooze to um five minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_day_category", "proposed_instruction": "Set my weekday—I meant weekend wake up alarm snooze to 15 minutes.", "long_pause_offsets": [], "source_sha256": "77c198e84411988185360fbf09303db5c5965bdd61986447271d854a0c9c9fe9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'weekday' to 'weekend' within the same utterance, preserving the action, snooze duration, and alarm type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/false_start/false_start_01_day_category/39f121d3bb6eb98e/speech.wav", "wer": 0.0, "wav_sha256": "39c73e679b548438325f442c91de61bf2a8c1c18516372ac232a8b86b9b9c099", "seconds": 7.76, "spoken_utterances": [ "Set my ... weekday—uh... I meant weekend wake up alarm snooze to fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday' but immediately self-corrects to 'weekend'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Set my weekday uh, I meant weekend wake up alarm snooze to fifteen minutes." }, "candidate_id": "wav_afc4557d208d0734", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start. The speaker begins with 'weekday', hesitates with 'uh', then self-corrects with 'I meant weekend' before continuing the full request. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and correction. The pacing and intonation are consistent with someone thinking aloud while formulating a command." } ] }, "is_default": true }, { "variant_id": "false_start_03_duration", "proposed_instruction": "Set my weekend wake up alarm snooze to 10 minutes—sorry, that should be 15 minutes.", "long_pause_offsets": [], "source_sha256": "4c2455eedecc36427040aca49e418fae1391675013e259095ed3f4b0c568314a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ten minutes' to 'fifteen minutes' within the same utterance, preserving the app, alarm type, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/false_start/false_start_03_duration/0a723f6084bd1e6b/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "6a344fc8b35670f50e2b3592666866f8d97e8c070a694297a20ec488eb534431", "seconds": 9.88, "spoken_utterances": [ "Set my weekend wake up alarm snooze to um... ten minutes—uh... sorry, that should be um... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ten minutes' but immediately self-corrects to 'fifteen minutes'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Set my weekend wake up alarm snooze to um, ten minutes uh, sorry, that should be um, fifteen minutes." }, "candidate_id": "wav_a4e52206c6fec7a2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a believable change of mind. The speaker starts with 'ten minutes', then hesitates with 'uh', and corrects to 'fifteen minutes'. The pause and filler are placed naturally, and the correction is delivered with a slight shift in intonation indicating reconsideration. However, the correction is slightly less integrated than in audio_02, as it feels like a separate thought rather than a seamless self-correction within the same utterance." } ] }, "is_default": false }, { "variant_id": "false_start_02_alarm_description", "proposed_instruction": "Set my weekend morning alarm—make that wake up alarm snooze to 15 minutes.", "long_pause_offsets": [], "source_sha256": "391fbb2648d823efa3791b7064adb2474b6ab9b84abb145b206357560c0ddabc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'morning alarm' but immediately corrects it to 'wake up alarm' within the same utterance. The final target is the wake up alarm, and the snooze duration is correctly specified as fifteen minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/false_start/false_start_02_alarm_description/e378b6c17d29cf08/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "aee5479fbd61799b72ee3eee4a11912c697eccd19d7ea34857279473489d6609", "seconds": 7.76, "spoken_utterances": [ "Set my weekend morning alarm—uh... make that wake up alarm snooze to ... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekend morning alarm' but immediately self-corrects to 'wake up alarm'. According to the 'false_start' type criterion, the final corrected instruction should be executed. The corrected target 'wake up alarm' and the value '15 minutes' are both clearly stated and preserved. The initial incorrect phrase 'morning alarm' is discarded as a false start.", "heard_transcript": "Set my weekend morning alarm uh make that wake up alarm snooze to 15 minutes." }, "candidate_id": "wav_7184c097b948b0cf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start but sounds less realistic due to its abrupt and unnatural delivery. The speaker says 'weekend morning alarm' and then immediately corrects to 'wake up alarm' without a natural pause or hesitation. The correction feels like a mechanical replacement rather than a spontaneous self-correction. The prosody is flat and lacks the natural variation of someone thinking aloud, making it sound more like a rehearsed or read instruction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Set my weekend wake up alarm snooze to. I left out a detail: set the snooze to 15 minutes.", "long_pause_offsets": [ 40 ], "source_sha256": "93c8d604c917552f0daa665091c2458ad9f0f684ba63155dae330e016c8e08ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing snooze duration value, which is consistent with the original instruction's context and target. The initial incomplete utterance is completed by the addition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/addition/addition_02/6945501082dadda6/speech.wav", "wer": 0.2, "wav_sha256": "0a6dbc09b2b91749b58c91729eea6dffa979dab3c05a36eba3f3ccf443703d1a", "seconds": 11.68, "spoken_utterances": [ "Set my um... weekend wake up alarm snooze to. ", "I left out a detail: set the snooze to ... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Set my um weekend wake up alarm snooze to' and then corrects by adding 'Set the snooze to 15 minutes.' This is a clear addition that completes the original instruction. The final audible intent matches the original instruction exactly, with the addition of the specific snooze duration. The 'um' and the self-correction are typical of natural speech and do not alter the final goal.", "heard_transcript": "Set my um weekend wake up alarm snooze to. I left out a detail. Set the snooze to 15 minutes." }, "candidate_id": "wav_dcaa739e972a865b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant 'um' and a clear false start ('Set my um...'), followed by a natural self-correction ('I left out a detail') and a smooth, additive continuation ('Set the snooze to 15 minutes'). The prosody is conversational, with a believable pause and filler before the correction, and the final instruction is delivered with consistent intonation and rhythm, making the entire sequence sound like a spontaneous, real-time thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Set my wake up alarm snooze to 15 minutes. Oh, right, this is just for my weekend wake-up alarm.", "long_pause_offsets": [ 43 ], "source_sha256": "078615807dc4be1275a46e82e9e3aa20ac1b86f5d2b3ab2d0ce60402a26542e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint specifying that the alarm is the weekend wake-up alarm, while preserving the initial snooze duration of fifteen minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/addition/addition_01/f6f49907ae8248ff/speech.wav", "wer": 0.0, "wav_sha256": "21f961d4fc54278ff752cc1673116c456ffc8d2dc45d479479f6bec2b6ba6913", "seconds": 10.76, "spoken_utterances": [ "Set my wake up alarm snooze to um... fifteen minutes. ", "Oh, right, this is just for my ... weekend wake-up alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a general instruction to set the wake-up alarm snooze to 15 minutes, then adds a clarifying constraint specifying that this applies only to the 'weekend wake up alarm'. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final intent is clear and executable, matching the reference instruction exactly.", "heard_transcript": "Set my wake up alarm snooze to um fifteen minutes. Oh right, this is just for my weekend wake up alarm." }, "candidate_id": "wav_2f5a9762b37022d6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear, slightly hesitant 'um' and delivers the core request fluently. However, the revision ('Oh right, this is just for my weekend wake-up alarm') is delivered as a separate, abrupt sentence with a noticeable pause and a shift in intonation that sounds more like a post-hoc clarification than a natural thought process. The flow feels less continuous than audio_01, and the correction lacks the immediate, integrated feel of a self-repair, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_day_category", "proposed_instruction": "Set my weekday wake up alarm snooze to 15 minutes. Wait, set the alarm for the weekend, not the weekday.", "long_pause_offsets": [ 51 ], "source_sha256": "b398f9bf2d9f194a156ffa54a093663df847031bf96b16a2000da3779013077a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm category from 'weekday' to 'weekend' while retaining the snooze duration of fifteen minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/revision/revision_01_day_category/684218ce6d6b5837/speech.wav", "wer": 0.12, "wav_sha256": "543ffb825db095cf39bd2bc3e96d438ed21762b50e8b1e6173edd35a4450d719", "seconds": 12.2, "spoken_utterances": [ "Set my um, let me think... weekday wake up alarm snooze to fifteen minutes. ", "Wait, set the alarm for the hmmm... weekend, not the weekday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the target as 'weekday' but immediately corrects it to 'weekend' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The value 'fifteen minutes' remains consistent. The final intent matches the original instruction, so the task passes.", "heard_transcript": "Set my um, let me think, weekday wake up alarm snooze to fifteen minutes. Wait, set the alarm for the weekend, not the weekday." }, "candidate_id": "wav_fc538e738113fec0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a plausible 'let me think' filler, formulates a clear initial request, then pauses and self-corrects with a believable change-of-mind intonation ('Wait, set the alarm for the weekend, not the weekday'). The prosody is consistent, the correction is smoothly integrated, and the pacing reflects genuine thought processing without robotic or disjointed transitions." } ] }, "is_default": true }, { "variant_id": "revision_03_duration", "proposed_instruction": "Set my weekend wake up alarm snooze to 10 minutes. Change the snooze to 15 minutes instead of 10 minutes.", "long_pause_offsets": [ 51 ], "source_sha256": "b5cc63e4e6188fd74e12d6966ec1832767aff0c9f1b0ec4440417cd494b6f380", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the snooze duration from ten to fifteen minutes, preserving the alarm type and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/revision/revision_03_duration/bca9ed0de9c54f8a/speech.wav", "wer": 0.08, "wav_sha256": "5c384230442d8b9760c3480ace3c4c95ef222b081f9486dcf63c90fa651fcd70", "seconds": 13.16, "spoken_utterances": [ "Set my uh, let me think... weekend wake up alarm snooze to um... ten minutes. ", "Change the snooze to ... fifteen minutes instead of ten minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ten minutes' but then corrects to 'fifteen minutes' after a pause, which matches the original instruction. The revision type requires applying the correction spoken after a longer pause, and the final value is 'fifteen minutes'. All other conditions (weekend wake up alarm snooze) are preserved.", "heard_transcript": "Set my uh let me think weekend wake up alarm snooze to um ten minutes. Change the snooze to fifteen minutes instead of ten minutes." }, "candidate_id": "wav_edada4ea2bc8afc8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a natural revision with a 'let me think' start and a clear initial request. The speaker then adds a follow-up correction ('Change the snooze to 15 minutes instead of 10 minutes') that sounds like a remembered detail or adjustment. The prosody is generally natural, though the second part of the utterance feels slightly more additive than a spontaneous self-correction, making it less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_02_alarm_description", "proposed_instruction": "Set my weekend morning alarm snooze to 15 minutes. Make that the wake up alarm, not the morning alarm.", "long_pause_offsets": [ 51 ], "source_sha256": "23f4aed57b807c903ca43d3c8be0d965db4f9684f36f45d6ea84b82d5aecf381", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm type from 'morning alarm' to 'wake up alarm' while preserving the snooze duration and weekend scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/revision/revision_02_alarm_description/4688de5d28c0ca0a/speech.wav", "wer": 0.0, "wav_sha256": "37fa292af76692cc296e3ee69fdd0f319460d90ca1c9055413289ce1a460877a", "seconds": 11.24, "spoken_utterances": [ "Set my um... weekend morning alarm snooze to fifteen minutes. ", "Make that the uh... wake up alarm, not the morning alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the alarm as 'weekend morning alarm' but immediately corrects it to 'wake up alarm' after a pause. This constitutes a revision where the final target alarm is correctly specified. The snooze duration of 'fifteen minutes' is preserved and correctly applied to the revised alarm. All other constraints match the original instruction.", "heard_transcript": "Set my um weekend morning alarm snooze to fifteen minutes. Make that the uh wake up alarm, not the morning alarm." }, "candidate_id": "wav_f730fff29689be21", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a clear revision structure but feels less natural due to a slightly abrupt transition. The speaker starts with 'set my um' and then corrects 'Make that the uh wake-up alarm, not the morning alarm.' The 'uh' fillers are present, but the correction is delivered with a flatter intonation and less of a change-of-mind inflection compared to audio_02. The overall delivery is more robotic and less like a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "28cf7b256b8c7469c1e9b7734b92d8d651fa53ab4db1ae66afa930e7cb8c8a56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/cancellation/cancel_008/c857733e7406ce14/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "955ad8203517d22e5be2cf9c4cab770c2d3833a638edf5ebb7c17ed04cdfde88", "seconds": 11.44, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes.", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first attempts to set the alarm snooze to 15 minutes, then immediately follows with 'Hang on, I want to cancel that task.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, not to execute the original instruction. The cancellation is explicit and complete, satisfying the 'cancellation' type criterion.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes. Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "2a0265d6d7706eae02849e2e3def86d58a0c9ea07dad95dd8d64fa3e2f917d67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/ambiguous_halt/halt_019/61b9f12208bb473a/speech.wav", "wer": 0.1, "wav_sha256": "f39914d94686fd6c45f4234af5b6db6ae62b60e603815241b1f22f40114e177d", "seconds": 10.52, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The user begins the task but then explicitly requests a pause with 'Can you give me a moment?'. This is a clear, final interjection to hold the process, fitting the 'ambiguous_halt' type. The initial request is not canceled, but execution is paused. The final control intent is to pause, which matches the type's criterion.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6c205f559ac9e4dc752942274c3a54fc8fbfb42e9bc1318fb44fe0e10b65b75b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_2/audio_natural/filler/original/f1cad63ed0929daf/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "cdbded738128870043ebbff624a25f149b0254fa60380ec540037b91da3e4ab0", "seconds": 6.76, "spoken_utterances": [ "Set my um, let's see... weekend wake up alarm snooze to um... fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') and pauses but does not change the core request. The final audible instruction matches the original task intent, with 'fifteen minutes' confirming the value. The filler type allows ignoring hesitations and backchannels, so the task is executed as intended.", "heard_transcript": "Set my um, let's see, weekend wake up alarm snooze to um, fifteen minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_alarm_description", "proposed_instruction": "Set my weekday morning alarm—sorry, I meant wake up alarm snooze to 20 minutes.", "long_pause_offsets": [], "source_sha256": "4a8548fdfec8f9f5d2a0bdc8ed74cc32ceeb27c9fe3eb70f0b73c86c42340052", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately corrects to 'wake up alarm'. The snooze duration is specified as 'twenty minutes' in both the original and spoken instruction. The final goal matches the original instruction in action, target, and value." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/false_start/false_start_02_alarm_description/501bb83533b2b365/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "8f3c19f4d9b186d52978d7d9f13c9dd466a61a617c0e871d0ab1a1f23d8faed6", "seconds": 8.64, "spoken_utterances": [ "Set my weekday morning alarm—uh... sorry, I meant ah... wake up alarm snooze to twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekday morning alarm' but immediately self-corrects to 'wake up alarm'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistaken value should be discarded. The final instruction matches the original reference in intent and value, with only a minor paraphrase ('weekday morning alarm' → 'wake up alarm') that does not change the core task.", "heard_transcript": "Set my weekday morning alarm uh, sorry, I meant, uh, wake up alarm snooze to twenty minutes." }, "candidate_id": "wav_fd18c03b15071a7e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'weekday morning alarm', hesitates with 'uh', then clearly self-corrects to 'wake up alarm' with a smooth, conversational 'sorry, I meant'. The correction is integrated into the sentence without awkwardness, and the final instruction is delivered with consistent prosody and a believable, slightly rushed but coherent flow." } ] }, "is_default": true }, { "variant_id": "false_start_01_day_category", "proposed_instruction": "Set my weekend—no, that should be weekday wake up alarm snooze to 20 minutes.", "long_pause_offsets": [], "source_sha256": "bc7e3ff9be72ffc285eec98c9a606102f80aaa4e3356575912dc2f4664719b17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'weekend' to 'weekday' and '20 minutes' to 'twenty minutes', preserving the action, target alarm, and snooze duration." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/false_start/false_start_01_day_category/40941fc63d34476d/speech.wav", "wer": 0.0, "wav_sha256": "253991737f49dc77dc622f1926cf2c8d68cd9b6e3ea7aadbf83cdccb9512d3a7", "seconds": 9.88, "spoken_utterances": [ "Set my weekend—um... no, that should be ... weekday wake up alarm snooze to uh, let's see... twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'weekend' but immediately self-corrects to 'weekday'. The final value '20 minutes' matches the original instruction. The correction is immediate and unambiguous, satisfying the 'false_start' criterion where the self-corrected version is the final intent.", "heard_transcript": "Set my weekend um uh no, that should be weekday wake up alarm snooze to uh let's see, 20 minutes." }, "candidate_id": "wav_f2894a4eb423ba64", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'weekend' followed by a self-correction to 'weekday'. The repair is natural, and the speaker continues with 'let's see' before stating the value, which adds to the impression of spontaneous thought. However, the initial word 'weekend' is more of a misstatement than a misnomer, and the correction feels slightly more abrupt than in audio_03, making it less fluid overall." } ] }, "is_default": false }, { "variant_id": "false_start_03_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 15 minutes—scratch that, make it 20 minutes.", "long_pause_offsets": [], "source_sha256": "cd60d555cfb21ec76b59dd0cc909ceea1d8d1b33b366451a0944c3d1e3960b7f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says fifteen minutes but immediately corrects to twenty minutes. The final value is twenty minutes, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/false_start/false_start_03_duration/5b6ba7a8a6ac16c3/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "99e61fbc0270488301e88008d256a0c89ab826cee969297f24e46356da1322d9", "seconds": 9.32, "spoken_utterances": [ "Set my uh... weekday wake up alarm snooze to fifteen minutes—um... scratch that, make it ... twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen minutes' but immediately corrects with 'scratch that, make it twenty minutes.' According to the 'false_start' type criterion, the final corrected value ('twenty minutes') should be executed, and the initial mistaken value ('fifteen minutes') should be discarded. The final intent matches the original instruction exactly, so the task is correctly executed.", "heard_transcript": "Set my uh weekday wake up alarm snooze to fifteen minutes. Um, scratch that, make it twenty minutes." }, "candidate_id": "wav_fe17f68db56a7d8e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision rather than a self-correction. The speaker first states '15 minutes', then says 'scratch that' and changes to '20 minutes'. While this is a plausible task revision, the phrase 'scratch that' is a bit more formal and less conversational than the natural 'sorry, I meant' in audio_03. The delivery also has a slightly robotic cadence, and the correction feels more like a command than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Set my wake up alarm snooze to 20 minutes. Oh, right, this is just for my weekday wake-up alarm.", "long_pause_offsets": [ 43 ], "source_sha256": "2ec03544c4397844808e2f3ac9cb979d2ffed3956d16ef550b597c1a5165891d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup restricts the initial broad alarm to only the weekday wake-up alarm, preserving the snooze duration of twenty minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/addition/addition_01/624a8c5eacbdd0d8/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "95c04c8ac5ee254554d9b2a1bb71c555228a7522db4762f865d1d1e7fd67e279", "seconds": 10.8, "spoken_utterances": [ "Set my wake up alarm snooze to um... twenty minutes. ", "Oh, right, this is just for my uh... weekday wake-up alarm." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to set the wake up alarm snooze to 20 minutes. The speaker then adds a clarifying constraint, specifying that this should apply only to the 'weekday wake up alarm'. This addition narrows the scope of the original instruction without contradicting it. The final intent is a combination of the original request and the added constraint, matching the expected behavior for an 'addition' type. The value '20 minutes' is correctly heard and preserved.", "heard_transcript": "Set my wake up alarm snooze to um twenty minutes. Oh right, this is just for my uh weekday wake up alarm." }, "candidate_id": "wav_9c3e7b8b3e2b8690", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um') and a plausible self-correction ('Oh right, this is just for my...'). The addition of 'weekday wake-up alarm' is smoothly integrated as a clarifying detail after the initial command, maintaining a consistent rhythm and intonation. The prosody is fluid, and the pause before the addition feels like genuine recollection rather than a scripted edit." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Set my weekday wake up alarm snooze to. I forgot to add, set the snooze to 20 minutes.", "long_pause_offsets": [ 40 ], "source_sha256": "1c859a796cd1921d2e3bf8f32acec23f5deabd3de854c549105f1560ccb9ecdc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing snooze duration, replacing the incomplete initial utterance. The original instruction specifies the alarm type and context (weekday wake up), and the spoken instruction completes the value with 'twenty minutes', which is equivalent to '20 minutes'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/addition/addition_02/f3935b13b529606e/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "e157dfecd751c9ddf889ae6e434fdbe5d0ac448f35332770b2bee6d0eac8b4d9", "seconds": 11.8, "spoken_utterances": [ "Set my uh, let me think... weekday wake up alarm snooze to. ", "I forgot to add, set the snooze to uh... twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the instruction, then adds the missing value 'twenty minutes' after a pause and self-correction. The final audible instruction matches the original reference instruction exactly, with the addition of the value 'twenty minutes' completing the task. The 'addition' type requires combining the initial instruction with the added information, which results in the correct final goal.", "heard_transcript": "Set my uh let me think. Weekday wake up alarm snooze to. I forgot to add, set the snooze to uh twenty minutes." }, "candidate_id": "wav_85083710f1cb59ac", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a filler ('let me think') and a self-correction ('I forgot to add'), the delivery is less fluid. The phrase 'set this snooze to uh twenty minutes' is awkwardly inserted after the initial request, creating a disjointed flow. The prosody is slightly stilted, and the correction feels more like a mechanical addition than a spontaneous thought process, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_duration", "proposed_instruction": "Set my weekday wake up alarm snooze to 15 minutes. Make that 20 minutes, not 15 minutes.", "long_pause_offsets": [ 51 ], "source_sha256": "94a76e4f16debf14760c2011f3872a3b447f15590a0b3655ab23720c7f7f3f86", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the snooze duration from fifteen to twenty minutes, while retaining the alarm type (weekday wake up) and action (set snooze)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/revision/revision_03_duration/04779ae2dfe4e0b8/speech.wav", "wer": 0.0, "wav_sha256": "f5b015fece016ddaac73636fde1794b79c59e8568305016c2d00821d3e4cd184", "seconds": 11.0, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to ... fifteen minutes. ", "Make that twenty minutes, not fifteen minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'fifteen minutes' but immediately corrects with 'Make that twenty minutes, not fifteen minutes.' This constitutes a revision where the final audible intent is to set the snooze to twenty minutes, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Set my weekday wake up alarm snooze to fifteen minutes. Make that twenty minutes, not fifteen minutes." }, "candidate_id": "wav_577c64505e7102d6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural revision flow. The speaker first makes a clear request, then immediately corrects the value from '15 minutes' to '20 minutes' with a smooth, self-correcting intonation. The phrasing 'Make that 20 minutes, not 15 minutes' is a highly plausible, spontaneous way to revise a number, and the delivery is continuous and fluent without awkward pauses or fillers." } ] }, "is_default": true }, { "variant_id": "revision_01_day_category", "proposed_instruction": "Set my weekend wake up alarm snooze to 20 minutes. Wait, the alarm should be set for weekday, not weekend.", "long_pause_offsets": [ 51 ], "source_sha256": "32b56dcbf2fffe63b9d7cd7750acb1bad138bd63b45c4864d07feac72e2d539a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the alarm category from 'weekend' to 'weekday' while retaining the snooze duration of twenty minutes." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/revision/revision_01_day_category/4386297d0c07b0dd/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "56d0a429e59f778958fa76316b4a21e10fddff32b8272ee50e350bdc433c4971", "seconds": 11.2, "spoken_utterances": [ "Set my ... weekend wake up alarm snooze to twenty minutes. ", "Wait, the alarm should be set for uh... weekday, not weekend." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives an incorrect instruction ('weekend') but immediately self-corrects with 'Wait, the alarm should be set for uh weekday, not weekend.' This constitutes a revision where the final audible intent is to set the alarm for the weekday, matching the original instruction. The correction is clear and applied to the earlier instruction, satisfying the revision type criterion.", "heard_transcript": "Set my weekend wake up alarm snooze to twenty minutes. Wait, the alarm should be set for uh weekday, not weekend." }, "candidate_id": "wav_34c98a159023759b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a believable revision where the speaker initially says 'weekend wake-up alarm' and then corrects to 'weekday'. The self-correction is natural, and the pause before the correction is plausible. However, the correction is slightly less fluid than audio_03, with a small hesitation and a slightly more abrupt shift in phrasing, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_alarm_description", "proposed_instruction": "Set my weekday morning alarm snooze to 20 minutes. Change the alarm description to 'wake up alarm' instead of 'morning alarm'.", "long_pause_offsets": [ 51 ], "source_sha256": "16c5c68130ab46f83cd072cbd0034a7c0d385e6903c3b876f51f2a2fc2a90edc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the alarm description from 'morning alarm' to 'wake up alarm', which is a different value than the original instruction's 'wake up alarm'. The original instruction specified 'wake up alarm' as the description, but the spoken instruction starts with 'weekday morning alarm', which is inconsistent with the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/revision/revision_02_alarm_description/e3953ad1cf618329/speech.wav", "wer": 0.0, "wav_sha256": "7ebbdd431c43102d331e509ab5a0700842963bf430628c53d5210efac39fe921", "seconds": 12.6, "spoken_utterances": [ "Set my weekday morning alarm snooze to ... twenty minutes. ", "Change the alarm description to 'um... wake up alarm' instead of 'morning alarm'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a correct instruction to set the snooze to 20 minutes, followed by a revision to change the alarm description from 'morning alarm' to 'wake up alarm'. The revision is spoken after a pause and clearly modifies the initial instruction. The final intent matches the reference instruction with the correction applied. All constraints are preserved and the revision is properly interpreted.", "heard_transcript": "Set my weekday morning alarm snooze to twenty minutes. Change the alarm description to um, wake up alarm instead of morning alarm." }, "candidate_id": "wav_b233c943b8d8c5af", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains two distinct requests rather than a single revision. The speaker first sets a weekday alarm, then immediately makes a separate request to change the description. The second request is introduced with a filler 'um', which is natural, but the lack of a direct connection between the two requests makes the flow feel less like a single, coherent thought process. The revision is less integrated, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "64390c0bfa798cc329b02e77ec85c7ed9254f7a91b4901e8f476f749bc730944", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/cancellation/cancel_009/acc675c296c30ca3/speech.wav", "wer": 0.125, "wav_sha256": "77dcc9f71a6d59c6fc674a1810c898fb20dba2f1da16e7c6944731feee990b78", "seconds": 10.28, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes.", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially attempts to set the snooze duration but immediately corrects themselves with 'Oh wait, cancel that altogether,' which is a clear and final cancellation of the original task. The cancellation is not ambiguous and applies to the entire task. The audio ends with this cancellation, satisfying the cancellation type criterion.", "heard_transcript": "Set my weekday wake up alarm snooze to um twenty minutes. Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_029", "hide_variant_original": true, "source_sha256": "23735584050652d0dae4a670af930f7dff77d48bf08dd55c3e86ff1c68fb7fa3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/ambiguous_halt/halt_029/6be435cb28c1afb5/speech.wav", "wer": 0.0, "wav_sha256": "c8293f580fd8df1466b5de87961725c057996a9ba76f0bd5546b97af1ab2f2d0", "seconds": 9.48, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes.", "Wait, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, actually...', which is a clear interjection indicating the speaker is halting the current task to revise or reconsider. This matches the 'ambiguous_halt' type, where the final intent is to pause, not execute or cancel. The initial instruction is not completed, and no new instruction is given, so the task is paused.", "heard_transcript": "Set my weekday wake up alarm snooze to um, twenty minutes. Wait, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "31dc501_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ebe9674d76f0746fa8bdeab37f03266167f758d078a6bb5df41578d507a38c75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/31dc501_3/audio_natural/filler/original/0f6aab46a191d28a/speech.wav", "wer": 0.0, "wav_sha256": "1f4780932a139801399404e8cb335c4008f13157b925bbc1da01ddafb58d54a8", "seconds": 6.04, "spoken_utterances": [ "Set my ... weekday wake up alarm snooze to um... twenty minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'um' but immediately corrects and completes the original instruction with the correct value 'twenty minutes'. The final audible intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Set my weekday wake up alarm snooze to um, twenty minutes." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_note_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Gmail—sorry, I should have said Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "c87283ea1cbc335a5cdab3f4e0abd5657a64ea69e8e1908706b488153151751f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately corrects to 'Simple Note'. The final intended source of information is Simple Note, which matches the original instruction. All other conditions, including the action, destination app, and context, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/false_start/false_start_02_note_app/1abf125d08b6e0d8/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "f08b5e917a6315ac4ee26058db113e9964b0f374252e47740716e7a2b5e41486", "seconds": 22.12, "spoken_utterances": [ "I went on a few trips each with some of my friends. My Gmail—ah... sorry, I should have said Simple Note has information on who owes whom what um, let's see... from each trip. I have already created Splitwise groups for the trips. uh... Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately self-corrects to 'Simple Note'. The final instruction is to 'record the expenses accordingly in the respective groups', which matches the original instruction. The correction is immediate and does not alter any other conditions. Therefore, the false_start type is satisfied with a pass.", "heard_transcript": "I went on a few chips each with some of my friends. My Gmail uh sorry I should have said Simple Note has information on who owes whom what um let's see. From each trip, I have already created Splitwise groups for the trips. Uh record the expenses accordingly in the respective groups." }, "candidate_id": "wav_72cc4d62bb0d5145", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and credible false start. The speaker begins with 'Gmail', then self-corrects to 'Simple Note' with a clear, spontaneous 'sorry I should've said' phrase. The correction is integrated smoothly into the flow, and the 'let's see' filler after the correction sounds like genuine recollection. The prosody is consistent, and the pacing feels like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my family—I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "7730299e89346fe0079103cb73b30160120222c95df5996a29c9b6701331331c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance. The final intended group is 'friends', which aligns with the original instruction. All other conditions—recording in Splitwise groups, using Simple Note data, and the context of trips—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/false_start/false_start_01_companion_group/81fb780395fd56cf/speech.wav", "wer": 0.0625, "wav_sha256": "11cabaf407a17a447c8aadd6f6b13e94914b0fd6b7682e8e72b60d3092e930e5", "seconds": 20.04, "spoken_utterances": [ "I went on um... a few trips each with some of my family—ah... I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. um... Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. According to the 'false_start' type criterion, the final intent is determined by the immediate correction, so the task is to record expenses in the respective groups. The rest of the instruction is consistent with the original, and no other changes or cancellations are made.", "heard_transcript": "I went on um a few trips each with some of my family uh I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Um record the expenses accordingly in the respective groups." }, "candidate_id": "wav_69eaaeb6d7c6137c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a false start where the speaker says 'family' and immediately corrects to 'friends'. The correction is natural, but the 'I meant' phrasing is slightly less spontaneous than 'sorry I should've said'. The pause before the correction is brief, and the overall delivery is fluent, but the self-correction feels a bit more rehearsed than in audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo—make that Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "b381b73da5406c0687a76e6935c7fea7a1db5d895feb28de78f5e00bcad9aec6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately corrects to 'Splitwise' in the same utterance. The final intended app is Splitwise, and all other conditions—recording expenses in respective groups, based on Simple Note data—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/false_start/false_start_03_expense_app/b604beb0f8ab2afc/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "b231e9797248cea546d419d08613b7bb58c7294a205811e09bed625add3b266f", "seconds": 19.04, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what from each trip. I have already created Venmo—ah... make that Splitwise groups for the trips. Record the expenses accordingly um... in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise'. According to the 'false_start' type criterion, the final corrected value ('Splitwise') should be applied, and the original mistaken value ('Venmo') should be discarded. The rest of the instruction remains consistent. The final intent is to record expenses in the respective Splitwise groups, which matches the original instruction after the correction.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. I have already created Venmo, uh, make that Splitwise groups for the trips. Record the expenses accordingly, um, in the respective groups." }, "candidate_id": "wav_09749a5a7e297933", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start where the speaker says 'Venmo' and corrects to 'Splitwise'. However, the correction is abrupt and lacks a natural self-repair phrase like 'sorry' or 'I meant'. The transition from 'Venmo' to 'Splitwise' is jarring, and the 'uh' after 'Splitwise' sounds like a filler rather than a genuine hesitation. The delivery feels less like a spontaneous thought and more like a scripted correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. One more detail: it's about who owes whom what.", "long_pause_offsets": [ 215 ], "source_sha256": "0637b2f07a99fa002c84dd63568ca1d17867b1d332fb692bce3cf301f531e2dc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail about the content of the Simple Note, specifying that it contains 'who owes whom what'. This clarifies the information source for the expense recording task, without changing the action, app, or groups. The original instruction already specified recording expenses in the groups, so the addition completes the missing condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_03/d6d2961b35296b08/speech.wav", "wer": 0.0196078431372549, "wav_sha256": "b4c8ecc4926a7e63f483b3774ddda0777e7138930600b7bdc75851f07d871f16", "seconds": 23.64, "spoken_utterances": [ "I went on ... a few trips each with some of my friends. My Simple Note has information on um, let me think... from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "One more detail: it's about ... who owes whom what." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and completed. The speaker adds the detail 'who owes whom what' after a pause, which is an addition to the original request. The final intent combines the original task with the added detail, matching the expected behavior for an 'addition' type. No contradictions or cancellations are present.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on um, let me think. From each trip, I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. One more detail, it's about who owes whom what." }, "candidate_id": "wav_8a5ed8fa03eb87be", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear statement, then uses a believable 'let me think' filler pause before continuing. The addition of 'one more detail' followed by the clarification about 'who owes whom what' is a pragmatic and plausible revision. The intonation and pacing are consistent with someone recalling information in real time, and the flow feels unscripted and conversational." } ] }, "is_default": true }, { "variant_id": "addition_04", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. And just to be specific, this applies to expenses from each trip.", "long_pause_offsets": [ 219 ], "source_sha256": "a91c86f51335f0f91436f4ce04d8f9510911d2598e5122ab001402e8d2aa99c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds specificity that the instruction applies to expenses from each trip, which is consistent with the original instruction's context and does not contradict or remove any prior requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_04/360c27ff0d4992a6/speech.wav", "wer": 0.07547169811320754, "wav_sha256": "405c8d33a008522d0354d3cb8db8c7510c55c7f8bfa0bb387e492f626edc2a55", "seconds": 21.04, "spoken_utterances": [ "I went on um... a few trips each with some of my friends. My Simple Note has information on ... who owes whom what. I have already created Splitwise groups for the trips. Record the expenses accordingly uh... in the respective groups. ", "And just to be specific, this applies to expenses from each trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed with the addition of a clarifying detail ('this applies to expenses from each trip') after a pause. The addition narrows the scope of the initial request, which is permitted under the 'addition' type. The final intent is clear and matches the combined goal of the original and added information.", "heard_transcript": "I went on um a few trips each with some of my friends. My Simple Note has information on who owes whom what. I have already created Splitwise groups for the trips. Record the expenses accordingly uh in the respective groups. Just to be specific, this applies to expenses from each trip." }, "candidate_id": "wav_fee73939301dc8b8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a natural, conversational rhythm with a 'um' at the start and a 'uh' during the sentence, which are typical of spontaneous speech. The addition at the end ('just to be specific...') is a believable and smoothly integrated clarification. The speaker's intonation varies naturally, and the overall delivery sounds like someone formulating a request while speaking, with a minor correction that fits the context." } ] }, "is_default": false }, { "variant_id": "addition_05", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip.. Record the expenses accordingly in the respective groups. Oh, right, I have already created Splitwise groups for the trips.", "long_pause_offsets": [ 180 ], "source_sha256": "63f6bef06d379baed52745a161d77ecf77325c5d7e8f11be67a8da6116e780c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the information that the Splitwise groups have already been created, which is a supporting detail but does not change the core task of recording expenses in the groups. The main action and target remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_05/ef4e60c7cbf5f6f2/speech.wav", "wer": 0.020833333333333332, "wav_sha256": "5a3bf4041476be2cd9012e4b104e0d25f38bbdcb75f0412df4c5c39459c29e93", "seconds": 20.2, "spoken_utterances": [ "I went on a few trips each with some of my friends. ... My Simple Note has information on who owes whom what from each trip.. Record the expenses accordingly uh... in the respective groups. ", "Oh, right, let me think... I have already created Splitwise groups for the trips." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to record expenses from Simple Note into respective groups. The speaker initially says 'Record the expenses accordingly uh in the respective groups' and then adds 'I have already created Splitwise groups for the trips' after a pause and self-correction. This addition clarifies a pre-existing condition (that the groups are already made) but does not change the core task of recording expenses. The final intent combines the original task with the added context, which is consistent with the 'addition' type. The final control intent is to execute the task as augmented by the addition.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. Record the expenses accordingly uh in the respective groups. Oh right, let me think. I have already created Splitwise groups for the trips." }, "candidate_id": "wav_bec9a2db272a720b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The delivery is mostly natural, with a 'uh' and a 'let me think' pause that suggest genuine hesitation. However, the self-correction ('Oh, right, let me think') feels slightly less fluid than in audio_02, as it interrupts the flow of the sentence. The speaker also seems to restart the sentence after the pause, which is less seamless than a true self-repair. Still, the prosody and fillers are realistic." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "I went on a few trips each with some of my friends. has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. I didn't mention this, but the information is in My Simple Note.", "long_pause_offsets": [ 219 ], "source_sha256": "11198049cf746228fdc3be091cdcd540e0ae83b446c530c2fddcd42158787e1c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing source of information (My Simple Note) without changing any other requirement. The core task—recording expenses in the pre-existing Splitwise groups—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_02/bbbfe41dd280287b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "59a1799a948973c816c98d9e84defd7fcfcdcf5d89fe3f1aee2e12f2c762a127", "seconds": 22.52, "spoken_utterances": [ "I went on a few trips each with some of my friends. has information on ... who owes whom what ... from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly um... in the respective groups. ", "I didn't mention this, but the information is in My Simple Note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and completed. The speaker adds a clarifying detail ('the information is in my Simple Note') after a longer pause, which is an addition. The final goal combines the original task with the new information, matching the expected behavior for an 'addition' type. The control intent is to execute the combined task.", "heard_transcript": "I went on a few trips each with some of my friends. Has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly um in the respective groups. I didn't mention this, but the information is in my Simple Note." }, "candidate_id": "wav_9cb1e0fa8a98465b", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has a relatively smooth delivery with a 'um' filler, but the addition at the end ('I didn't mention this...') is abrupt and lacks the natural intonation of a spontaneous thought. The transition into the final sentence feels more like a post-hoc addition than a genuine memory recall. The overall pace is slightly rushed, and the delivery lacks the organic hesitation seen in higher-ranked clips." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "I went on. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, it was a few trips each with some of my friends.", "long_pause_offsets": [ 193 ], "source_sha256": "be86069f3b6950a357dcbdcd235ac61ee54ba839c4b9653043b5d7cb04d978c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds a clarifying detail about the trips being with friends, but this does not change the core task of recording expenses in the existing Splitwise groups using the Simple Note data. The original instruction already included the key elements: using Simple Note for expense details and recording them in the pre-created Splitwise groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/addition/addition_01/017b45da52815b59/speech.wav", "wer": 0.057692307692307696, "wav_sha256": "22854a5556a39e1700fa451a6d5d11ffe552e74712b76a19b8c9849c979c6687", "seconds": 19.6, "spoken_utterances": [ "I went on. My Simple Note has information on um... who owes whom what uh... from each trip. um... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Oh, I forgot to mention, it was a few trips each with some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and the speaker adds a clarifying detail ('it was a few trips each with some of my friends') after a longer pause. This addition provides more context but does not contradict or remove any existing constraints. The final goal is the original task combined with the added detail, which is consistent with the 'addition' type. The control intent is to execute the combined task.", "heard_transcript": "I went on, my Simple Note has information on um, who owes whom what uh from each trip, um, I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention it was a few trips each with some of my friends." }, "candidate_id": "wav_b79aaef997f90ace", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip has the least realistic delivery. The speaker starts with a clear statement, then adds 'Oh, I forgot to mention it was a few trips...' as an afterthought. However, the phrasing 'it was a few trips' is awkward and sounds like a forced insertion rather than a natural recollection. The intonation is flat, and the flow feels disjointed, as if the speaker is trying to correct a mistake but does so in a way that disrupts the coherence of the utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my colleagues. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be 'friends', not 'colleagues'.", "long_pause_offsets": [ 237 ], "source_sha256": "50a91212fc3bc9112460ef9f44053ed816dcf1030e8fbbf30ffbc52637c07851", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up revision corrects the companion group from 'colleagues' to 'friends', while retaining all other requirements: recording expenses in the respective Splitwise groups based on information in Simple Note. The final goal is consistent with the original instruction except for the corrected group identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/revision/revision_01_companion_group/588a8a45e06c364f/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "84a5ebb5d5b26b590f1e7b2232838b6fc91cdfc04afa15921810aab235740e9b", "seconds": 21.44, "spoken_utterances": [ "I went on um... a few trips each with some of my colleagues. My Simple Note has information on ... who owes whom what hmmm... from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the companion group should be 'friends', not 'colleagues'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'colleagues' but immediately corrects it to 'friends' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to record expenses in the groups, with the companion group being 'friends'. The original instruction is revised to match the final audible intent, so the task is executed as revised.", "heard_transcript": "I went on um a few trips each with some of my colleagues. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be friends, not colleagues." }, "candidate_id": "wav_eb59a1c59f64ab91", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a plausible hesitation ('um'), uses a realistic self-correction ('Wait, the companion groups should be friends, not colleagues') that sounds like a genuine moment of recollection, and maintains a consistent, conversational pace. The correction is integrated smoothly into the utterance, and the overall prosody is fluid and human." } ] }, "is_default": true }, { "variant_id": "revision_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. That should be Splitwise, not Venmo, for recording the expenses.", "long_pause_offsets": [ 230 ], "source_sha256": "1e8cc23fa874ef03432a5e8f9a2f428eb557f12b7ee44664c550c595949eede9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Venmo to Splitwise, changing the target platform for recording expenses." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/revision/revision_03_expense_app/664aaed18abab4ef/speech.wav", "wer": 0.10909090909090909, "wav_sha256": "c90fd7a239e37a043396cb59e7dc4722d41b3ae1a1e5a6519b2a32355b838b47", "seconds": 26.84, "spoken_utterances": [ "I went on a few trips each with some of my friends. My Simple Note has information on um... who owes whom what ah... from each trip. hmmm... I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. ", "That should be Splitwise, not Venmo, for recording the expenses." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo groups' but immediately corrects to 'Splitwise' after a pause, indicating a revision. The final instruction is to use Splitwise, which aligns with the original reference instruction. The correction is clear and the final intent is executable.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on um, who owes whom what uh, from each trip. Hmm. I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. That should be Splitwise, not Venmo for recording the expenses." }, "candidate_id": "wav_de605b2faa69e7ca", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is realistic with natural hesitations ('um', 'uh') and a clear revision ('That should be Splitwise, not Venmo'). However, the revision is delivered as a separate sentence rather than being embedded within the original thought, making it slightly less fluid than audio_03. The speaker also uses a filler ('Hmm') that sounds like genuine contemplation, but the overall delivery is a bit more segmented." } ] }, "is_default": false }, { "variant_id": "revision_02_note_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Let me correct the note app: use 'Simple Note', not Gmail.", "long_pause_offsets": [ 228 ], "source_sha256": "807a2f42030fe90a5180806e8571c6c02efe533d75e898b4a8020be73f7ab870", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source of information from 'Gmail' to 'Simple Note', while retaining the action and the target (Splitwise groups). The final goal is consistent with the original instruction except for the corrected app name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/revision/revision_02_note_app/a7566054bac4a9c3/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f833876d97cd3bb65cb3e9f85f7b31c75943f2b18d4364a1e54ad9f6fd06ccd6", "seconds": 23.84, "spoken_utterances": [ "I went on ah... a few trips each with some of my friends. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. um... Record the expenses accordingly uh... in the respective groups. ", "Let me correct the note app: use 'Simple Note', not Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states that Gmail contains the information, but then immediately corrects this to 'Simple Note, not Gmail' after a pause. This is a revision where the final intent clearly specifies Simple Note as the correct source. The rest of the instruction remains consistent with the original, and the correction is applied to the earlier statement. Therefore, the final task intent matches the reference instruction.", "heard_transcript": "I went on uh a few trips each with some of my friends. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Um, record the expenses accordingly uh in the respective groups. Let me correct the note app. Use Simple Note, not Gmail." }, "candidate_id": "wav_0ff83d6cff278440", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision ('Let me correct the note app, use Simple Note, not Gmail'), but the correction is delivered as a separate, almost robotic-sounding sentence. The initial part of the utterance is fluent, but the correction lacks the natural integration seen in the other clips. The prosody of the correction feels more like a post-hoc instruction than a spontaneous self-correction, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "8d73faf236c1028b8ed15ca58912eae76ec970ae5b001a6e0236c0599d86e4b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/cancellation/cancel_026/00d77ece550875a7/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "b6e68cde3cf85258ad7a368432bbb505281888447d9d99054784fe142f829197", "seconds": 22.4, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Wait, forget what I just asked. Cancel it.' This constitutes a final, unambiguous directive to cancel the entire task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "fc0f21bfb3c1fbdcdc0c78d76d0d7af8c85b8595f84ca1a3e53eba2ef1a6aa9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/ambiguous_halt/halt_026/2afd7bc1ab14ba51/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "f8fd75918a345cee57e061492c43bcdcee1a948e30180a84bb7616fb159def72", "seconds": 22.0, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d1e78497926c98f17b315de07809b6261cb7f892eb6c156db19835f1d442d7f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_1/audio_natural/filler/original/d3d68c3c438dfd27/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "7658266d1707dbd51d91b556267fd35d398eb6dee13b8f985b17765ec0ac3180", "seconds": 17.64, "spoken_utterances": [ "I went on a few trips each with some of my friends. um... My Simple Note has information on who owes whom what ... from each trip. uh, let's see... I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('Um', 'Uh', 'let's see') and pauses but does not change the core task. The final instruction is to 'Record the expenses accordingly in the respective groups,' which matches the original intent. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "I went on a few trips each with some of my friends. Um, my Simple Note has information on who owes whom what from each trip. Uh, let's see. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my friends—wait, I meant coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "30adb8dab03143c9b3b48123007310688081db65c51144bd73c6c4179482a7f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'friends' to 'coworkers' within the same utterance. The final intended group is 'coworkers', matching the original instruction. All other conditions—recording in Splitwise groups, using Simple Note data, and the context of multiple trips—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/false_start/false_start_01_companion_group/75979b99ef20f2a7/speech.wav", "wer": 0.125, "wav_sha256": "131ce9797fdc361d3fe2c89e41b64a9e6064de5f300c484928e0b2143c95059f", "seconds": 21.04, "spoken_utterances": [ "I went on a few trips each with some of my um... friends—uh... wait, I meant coworkers. My ... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'coworkers'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction remains unchanged and fully matches the original. The correction is immediate and does not alter any other conditions, so the task is valid and complete.", "heard_transcript": "I went on a few trips each with some of my um, friends uh, wait, I meant coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_2a54a61ac4d2b971", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'friends', hesitates with 'uh', then self-corrects with 'wait, I meant coworkers'—a very credible and common pattern of self-correction. The intonation and pacing are consistent with someone recalling and adjusting their words in real time. The pause and filler are appropriately placed and not overly long, contributing to a realistic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_note_app", "proposed_instruction": "I went on a few trips each with some of my coworkers. My Gmail—sorry, make that Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "6eb5559f0d87f52859c90571f4f1ff440b9be4e65ec3cc27b1a3e06ff25e74c9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately corrects to 'Simple Note'. The final intended source of information is Simple Note, which matches the original instruction. All other conditions, including the action, destination (Splitwise groups), and context (trip-related expenses), remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/false_start/false_start_02_note_app/69014ce3e876d77f/speech.wav", "wer": 0.125, "wav_sha256": "ffe5d7240577cdd875f4e2d9ea57b09cc2c841b18619947b5d9aa9c8b040db68", "seconds": 19.4, "spoken_utterances": [ "I went on a few trips each with some of my um... coworkers. My Gmail—uh... sorry, make that Simple Note has information on who owes whom what from each trip. I have already created ... Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'My Gmail' but immediately self-corrects to 'make that Simple Note'. This is a false start where the mistaken value is corrected, and the final instruction matches the original reference. The control intent is to execute the task as corrected.", "heard_transcript": "I went on a few trips each with some of my um coworkers. My Gmail uh sorry, make that Simple Note has information on who owes whom what from each trip. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_8527888d8379f591", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a false start, with the speaker saying 'Gmail' and then correcting to 'simple note'. However, the correction is slightly less natural—the pause before 'sorry' feels a bit abrupt, and the correction is delivered with a flatter intonation compared to audio_03. The overall flow is still realistic, but the self-correction lacks the same level of spontaneous hesitation and recovery." } ] }, "is_default": false }, { "variant_id": "false_start_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Venmo—no, that should be Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "7f36354c1dcd67d992c078354ebc5ad805e0d91dae229dae78fa022a4af50f47", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately corrects to 'Splitwise' within the same utterance. The final intended app is Splitwise, matching the original instruction. All other conditions, including the source of information (Simple Note), the context (trips with coworkers), and the action (recording expenses), remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/false_start/false_start_03_expense_app/150e56a89daa7e22/speech.wav", "wer": 0.10416666666666667, "wav_sha256": "d7210f0372486f3427fc00496ae9cad9a4b725ffba6e4aeb1c9e153b3abfea0d", "seconds": 19.4, "spoken_utterances": [ "I went on a few trips each with some of my ... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Venmo—um... no, that should be Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise' after a brief hesitation ('um, no'). According to the 'false_start' type criterion, the final corrected value ('Splitwise') should be applied, and the original incorrect value ('Venmo') should be discarded. The final instruction to 'Record the expenses accordingly in the respective groups' remains unchanged and is consistent with the reference. Therefore, the final intent matches the original instruction with only a corrected target application.", "heard_transcript": "I went on a few trips each with some of my coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Venmo um, no, that should be Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_d1927b1a96ee69f7", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic of the three. It is delivered in a flat, robotic tone with no natural pauses or fillers. The speaker says 'Venmo' and then corrects to 'Splitwise' without any hesitation or filler, making the correction sound forced and unnatural. The lack of prosodic variation and the abruptness of the correction make it sound more like a rehearsed or synthesized instruction than a spontaneous one." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I went on a few trips each with. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, it's only some of my coworkers.", "long_pause_offsets": [ 215 ], "source_sha256": "76294d77d9cd5aeb96a2a07b9966450c9e239984c123afd38d26962f567b4945", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the scope to 'some of my coworkers', which is a narrowing of the original instruction's broader implication of all coworkers. The core action—recording expenses in Splitwise groups based on Simple Note data—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/addition/addition_01/48e1055e85e9b01b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "8c2972417ffabe0a0768d1f4d5cd43b3bb1cef39953e285a04f1d56a3105e8dc", "seconds": 23.48, "spoken_utterances": [ "I went on a few trips each with. My let's see... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Oh, I forgot to mention, it's only hmmm... some of my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition 'it's only some of my coworkers' narrows the scope of the initial request. The final intent is to record expenses in the respective groups, but limited to 'some of my coworkers'. This matches the addition type criterion, where the final goal combines the original instruction with the added constraint.", "heard_transcript": "I went on a few trips each with my let's see. Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, it's only some of my coworkers." }, "candidate_id": "wav_0cfba5590bc4f85e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see') and a filler ('um') before continuing. The addition ('Oh, I forgot to mention...') is delivered with a believable, slightly apologetic tone and a natural pause before the new information. The phrasing 'it's only some of my coworkers' flows smoothly as a remembered detail, maintaining a consistent rhythm and prosody throughout." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "I went on a few trips each with some of my coworkers. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. I left out a detail: the note contains who owes whom what from each trip.", "long_pause_offsets": [ 167 ], "source_sha256": "9691fc56d43aaf5fe8699b088eda60b7844c5af0ceeaf9d44a29d3ab961614f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing detail about the source of the expense data (the Simple Note), but does not change the core action (recording expenses in Splitwise groups) or the condition (using the note's content). The original instruction already specified recording expenses accordingly, and the addition clarifies how to do so." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/addition/addition_02/d78e015bbd2ea2bb/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "896d424d16d44a1530a5e3144e5de4a49d8105be29ec283ad582166fe17ea722", "seconds": 20.48, "spoken_utterances": [ "I went on a few trips each with some of my um... coworkers. I have already created ... Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "I left out a detail: the note contains let me think... who owes whom what from each trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to record expenses in Splitwise groups based on a note. The speaker adds a clarifying detail about the content of the note ('who owes whom what from each trip') after a pause and self-correction ('let me think'). This addition specifies the exact information needed from the note to complete the task. The final intent combines the original instruction with the added detail, which is consistent with the 'addition' type. The task is fully specified and executable.", "heard_transcript": "I went on a few trips each with some of my um coworkers. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. I left out a detail, the note contains, let me think, who owes whom what from each trip." }, "candidate_id": "wav_35aeab57321fac66", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a natural hesitation ('um') and maintains a conversational pace. The revision ('I left out a detail...') is introduced with a clear shift in intonation and a brief pause, followed by a plausible self-correction ('the note contains...'). However, the delivery of the correction is slightly more abrupt and less fluid than in audio_01, with a less natural-sounding 'let me think' filler that feels a bit more rehearsed." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the group should be coworkers, not friends.", "long_pause_offsets": [ 234 ], "source_sha256": "87805c210882c5dffb9d5eb4be4c4bbe855a91cca1c338806fb91774a3d8aa44", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group membership from 'friends' to 'coworkers', preserving the action, app, and context of recording expenses in existing groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/revision/revision_01_companion_group/877a13271cf322b8/speech.wav", "wer": 0.09615384615384616, "wav_sha256": "4c79ac5f7c195e19efc5aeee5901c36a7bb9afbab0ce1f41951613647905a268", "seconds": 23.52, "spoken_utterances": [ "I went on a few trips each with some of my ... friends. My Simple Note has information on who owes whom what from each trip. I have already created hmmm... Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the group should be um... coworkers, not friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially refers to 'friends' but immediately corrects themselves to 'coworkers' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to record expenses in the groups, but the group context is revised from 'friends' to 'coworkers'. All other conditions remain unchanged, so the final task is valid and matches the revision type requirement.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I've already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the groups should be um coworkers not friends." }, "candidate_id": "wav_4e5e5b94d518c369", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and pragmatic flow. The speaker begins with a plausible initial statement ('friends'), then after a realistic pause and self-correction ('wait, the group should be...'), corrects the group type to 'coworkers'. The correction is integrated smoothly into the ongoing utterance with a believable change-of-mind intonation and reference back to the earlier request. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "revision_02_note_app", "proposed_instruction": "I went on a few trips each with some of my coworkers. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Let me correct the note app: use Simple Note, not Gmail.", "long_pause_offsets": [ 230 ], "source_sha256": "a439ab096b44adfdeb3e2e9efdf4b25d6bb5474a7f951ccc2ca5818a2722e981", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source of the information from Gmail to Simple Note, while retaining the action and the destination group. The final goal is to record expenses in Splitwise groups based on data from Simple Note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/revision/revision_02_note_app/19f28689506574fd/speech.wav", "wer": 0.1, "wav_sha256": "ba0e8ea9c1dd9abf848b9ef03c87ff6d84b64d69272604346311e2cde9adb23f", "seconds": 24.84, "spoken_utterances": [ "I went on a few trips each with some of my uh... coworkers. My uh, let me think... Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Let me correct the note app: use um, let's see... Simple Note, not Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Gmail as the source of information but immediately corrects themselves to 'Simple Note, not Gmail' after a pause. According to the 'revision' type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to use Simple Note, which matches the original instruction. All other conditions (recording expenses in Splitwise groups) are preserved. Therefore, the final audible intent satisfies the revision criterion.", "heard_transcript": "I went on a few trips each with some of my uh coworkers. My uh let me think. Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Let me correct that. Note app. Use um let's see. Simple Note, not Gmail." }, "candidate_id": "wav_bd8807f8f1735e2c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear self-correction ('let me correct the note app, use um... simple note, not Gmail'), but the correction is delivered as a separate, slightly disjointed clause rather than being fully integrated into the original request. The speaker also uses fillers ('uh', 'let me think') and pauses naturally, but the repair feels more like an afterthought than a seamless part of the thought process, making it slightly less realistic than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my coworkers. My Simple Note has information on who owes whom what from each trip. I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. Change Splitwise to Venmo for recording the expenses.", "long_pause_offsets": [ 232 ], "source_sha256": "d4181b23e9dd4769631930b0921337db3ccb9e2426fe5db6a076b4c420b00ab9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the app from Splitwise to Venmo for recording the expenses, which alters the target application from the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/revision/revision_03_expense_app/7c5a36bc927a0112/speech.wav", "wer": 0.11320754716981132, "wav_sha256": "fa32afbf10e209e7919867ddf7a4a6524d4c637364a936e3e4bab0cffbdcaf44", "seconds": 22.6, "spoken_utterances": [ "I went on a few trips each with some of my coworkers. My ... Simple Note has information on who owes whom what from each trip. I have already created uh, let's see... Venmo groups for the trips. Record the expenses accordingly in the respective groups. ", "Change ... Splitwise to Venmo for recording the expenses." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task using 'Splitwise groups' but then corrects it to 'Venmo groups' and explicitly revises the action to 'Change Splitwise to Venmo for recording the expenses.' This constitutes a clear revision where the final intent is to use Venmo, not Splitwise. All other conditions remain unchanged, satisfying the revision type criterion.", "heard_transcript": "I went on a few trips each with some of my coworkers. My Simple Note has information on who owes whom what from each trip. I have already created uh, let's see, uh, Venmo groups for the trips. Record the expenses accordingly in the respective groups. Change Splitwise to Venmo for recording the expenses." }, "candidate_id": "wav_a71f2ad9328036a9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision ('change splitwise to Venmo'), but the revision is delivered as a separate, abrupt command at the end rather than being integrated into the flow of the original request. The speaker also uses fillers ('uh', 'let's see') and pauses, but the final instruction sounds more like a post-hoc addition than a natural part of the speaker's thought process, reducing its realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "deb711d8345e5deeb65d051c035ae7089f31e1e172cd6e76777a3b19c567b457", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/cancellation/cancel_023/4875dfcccab64b38/speech.wav", "wer": 0.1, "wav_sha256": "721bb33ba1c1320410b7f9af0b1ffb70643e59e99e08d6c113694bded85d2424", "seconds": 24.2, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker explicitly states 'Hang on, never mind about that. Don't carry out the request' after completing the full description of the task. This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "I went on a few trips each with some of my uh, let me think, coworkers. My Simple Note has information on who owes whom what from each trip. I have already created, let me think, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "1349931c9f705a4da641e42abca056bd8245f7666d9903a9ef520fa9c635333c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/ambiguous_halt/halt_033/7df83e870bcdc91b/speech.wav", "wer": 0.11320754716981132, "wav_sha256": "70a9b6c360f6019387005be24b3c435451b6d12d974dbe72d8a58b6aa5fa382c", "seconds": 22.44, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something...', which is a clear request to pause or hold the current task. The speaker does not cancel the task or provide a new instruction. The initial request is fully stated and understood, but the final interjection indicates a need to pause before proceeding. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "I went on a few trips, each with some of my uh, let me think, coworkers. My Simple Note has information on who owes whom what from each trip. I have already created, let me think, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Hold on, there's something..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "feaa852ddcab90acb767768909aedb4b5e5c7aeef0040d3a966a6a6edc73318b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_2/audio_natural/filler/original/c059555148844a80/speech.wav", "wer": 0.12244897959183673, "wav_sha256": "0ab23a3c8db5514ab20b61fd130c299f1209208134b21eb19190e71ec6df2f92", "seconds": 18.52, "spoken_utterances": [ "I went on a few trips each with some of my uh, let me think... coworkers. My Simple Note has information on who owes whom what from each trip. I have already created let me think... Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes two filler phrases ('uh let me think') but does not change the core task. The final instruction is to record expenses in the respective groups, which matches the original instruction. The filler phrases do not alter the goal, so the task should be executed.", "heard_transcript": "I went on a few trips each with some of my uh let me think coworkers. My simple note has information on who owes whom what from each trip. I have already created let me think splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my roommates—I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "0502c2a9b1c067c1b7dae2eb226750ccc2b730d1be28f02efee73f3e37c12e37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'roommates' to 'friends' within the same utterance. The final intended group is 'friends', but the core task—recording expenses in the appropriate Splitwise groups based on Simple Note data—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/false_start/false_start_01_companion_group/5482c1f608a0793d/speech.wav", "wer": 0.0425531914893617, "wav_sha256": "a514c2f9206053e4a93a3ad9927c4ff834c5091a75488206b0b5b5d498f4b61f", "seconds": 20.32, "spoken_utterances": [ "I went on a few trips each with some of my roommates—um... I meant friends. My hmmm... Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'roommates' but immediately self-corrects to 'friends'. The final intent is to record expenses in Splitwise groups as per the original instruction. The correction is a false start, and the final goal matches the original instruction.", "heard_transcript": "I went on a few trips each with some of my roommates, some. I meant friends. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_f1af66694fe73c27", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with 'roommates' but immediately self-corrects to 'friends' with a believable 'I meant' repair, followed by a realistic 'hmm' pause that suggests genuine thought. The prosody is varied and conversational, with appropriate intonation and pacing. The correction is integrated smoothly into the flow, and the overall rhythm feels like a person recalling and refining a task in real time." } ] }, "is_default": true }, { "variant_id": "false_start_02_note_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Gmail—I should say Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "3a74a0021eff7e759812e30670caf289ed2cbac04cbbbca48b28c45cb71128ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately corrects to 'Simple Note'. The final intended source of information is Simple Note, which matches the original instruction. All other conditions, including the action, destination app, and context, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/false_start/false_start_02_note_app/6fc17b8ddaf9e9fa/speech.wav", "wer": 0.0425531914893617, "wav_sha256": "8fe01f8903a53e7a07de2d58051a8913b0afbda900a8752da9ddbea4eaa47a3d", "seconds": 17.64, "spoken_utterances": [ "I went on a few trips each with some of my ... friends. My Gmail—um... I should say Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'My Gmail' but immediately self-corrects to 'I should say Simple Note'. This is a false start, and the final, corrected target is 'Simple Note'. The rest of the instruction remains unchanged and is clear. The final intent matches the original instruction with the correction applied, so the task should be executed.", "heard_transcript": "I went on a few trips each with some of my friends. My Gmail, um, I should say Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_84bc4e913aa27f8f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is very realistic, with a clear false start where the speaker says 'Gmail' before correcting to 'Simple Note'. The 'um' filler and the self-correction are natural and well-timed. The delivery is fluent and the intonation varies appropriately. However, the initial error ('Gmail') is slightly less plausible than 'roommates' in the context of expense tracking, making the correction feel a bit less organic than in audio_02, though still very believable." } ] }, "is_default": false }, { "variant_id": "false_start_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo—scratch that, make it Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "long_pause_offsets": [], "source_sha256": "4c876b7a5d1c67eb65d45970fbd1414b6fc4521eae87634254fbb27dc4c847fc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Venmo' but immediately corrects it to 'Splitwise' within the same utterance. The final intended app is Splitwise, and all other conditions—recording expenses in respective groups based on Simple Note data—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/false_start/false_start_03_expense_app/5dedbdb5f36f6d37/speech.wav", "wer": 0.0625, "wav_sha256": "dff1b1011bab2b01635773f3a11bfcada629c5f3568de3610f5283af1037d2a9", "seconds": 18.64, "spoken_utterances": [ "I went on a few trips each with some of my friends. My ... Simple Note has information on who owes whom what from each trip. I have already created Venmo—uh... scratch that, make it Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('Venmo') is discarded. The rest of the instruction remains unchanged and is executed as intended.", "heard_transcript": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo uh, scratch that, make it Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "candidate_id": "wav_ef006f977eb997e4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a more robotic and less natural flow. The false start 'Venmo' is replaced with 'Scratch that, make it Splitwise', which is a more abrupt and less conversational repair. The speaker does not use a natural filler like 'uh' or 'um' before the correction, and the transition feels more like a mechanical edit than a spontaneous thought process. The prosody is flatter and the pacing is less varied, reducing the sense of realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. One more thing: the Splitwise groups for the trips are already set up.", "long_pause_offsets": [ 179 ], "source_sha256": "ef222cea2cd05945d77906bfdc9c5e27a349b8a76c5a3bec4ae0fe6c378f3d30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds confirmation that the Splitwise groups are already created, which supports but does not alter the core task of recording expenses in those groups. The original instruction already assumed the groups existed, so this addition is consistent and non-contradictory." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/addition/addition_02/2e55e7ee297027af/speech.wav", "wer": 0.02127659574468085, "wav_sha256": "2fd2a358eaaf8c7fbcfd41c13790423c678d31db09d80f52c30fc10c4aba2d4d", "seconds": 19.8, "spoken_utterances": [ "I went on a few trips each with some of my um... friends. My ... Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. ", "One more thing: the ... Splitwise groups for the trips are already set up." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to record expenses in respective groups. The speaker adds the information that the Splitwise groups are already set up, which is an additional constraint that narrows the scope of the task. The final intent is to execute the original task with this added detail, matching the 'addition' type criterion.", "heard_transcript": "I went on a few trips each with some of my um friends. My Simple Note has information on who owes whom what from each trip. Record the expenses accordingly in the respective groups. One more thing, the Splitwise groups for the trips are already set up." }, "candidate_id": "wav_d16c74c6c9c2bfeb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker's delivery is highly natural, with a clear, conversational rhythm and plausible fillers ('um'). The addition of 'One more thing' followed by a revision of the group setup information is a believable pragmatic flow. The intonation and pacing suggest spontaneous thought, and the correction ('the splitwise groups... are already set up') is smoothly integrated without awkwardness. The voice is consistent and the speech is free of distortions or unnatural pauses." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "I went on a few trips each with some of my friends. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, the note contains the details of who owes whom.", "long_pause_offsets": [ 165 ], "source_sha256": "9aba85680f8e25bbfb28bd1dd13887f6b1d5644ad031f765c061205e3f6fde4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing detail about the source of the expense information (the Simple Note), which clarifies how to execute the original instruction. The core action and target remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/addition/addition_01/52897315f20fe4cf/speech.wav", "wer": 0.08888888888888889, "wav_sha256": "0b8f35f0d277be94cadc87a6ef1ca2ff70d79657c06cade5e4173f692be288b6", "seconds": 20.36, "spoken_utterances": [ "I went on a few trips each with some of my uh... friends. I have already created ... Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Oh, I forgot to mention, the note contains the details of hmmm... who owes whom." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the core task of recording expenses in Splitwise groups. After a pause, they add the crucial detail that the note contains the information on who owes whom. This addition clarifies the source of the expense data and narrows the scope of the task. The final intent is to execute the original task with the added constraint of using the note for the 'who owes whom' details. The addition is not a correction or cancellation, but a clarification that enhances the original instruction. The final control intent is to execute the combined task.", "heard_transcript": "I went on a few trips each with some of my uh friends. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Oh, I forgot to mention, the note contains the details of whom who owes whom." }, "candidate_id": "wav_aaabd3de238c4f57", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the content is similar, the delivery is slightly less natural. The speaker uses 'uh' and 'oh' as fillers, but the addition of 'Oh, I forgot to mention' followed by a new clause feels a bit more abrupt and less integrated than in audio_02. The pacing is slightly slower, and the overall prosody is flatter, making it sound a bit more rehearsed or less spontaneous. The correction is present but less fluidly connected to the prior sentence." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_companion_group", "proposed_instruction": "I went on a few trips each with some of my roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion group should be 'friends', not 'roommates'.", "long_pause_offsets": [ 236 ], "source_sha256": "e6a0913279ff58b17f903dbc6473cd89bc4a1531d02bdbcc169aed8aea101beb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the companion group from 'roommates' to 'friends', while retaining all other requirements including the action, app, and context of recording expenses based on Simple Note data." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/revision/revision_01_companion_group/92b4f61d7e408cd1/speech.wav", "wer": 0.0, "wav_sha256": "6d8e053a5acedbb8cd41c6d958ce7961eba6b19a1f1bb3115674207dd1387b2b", "seconds": 23.36, "spoken_utterances": [ "I went on a few trips each with some of my um... roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Wait, the companion group should be '... friends', not 'roommates'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'roommates' but immediately corrects to 'friends' after a pause, which is a revision. The final intent is to execute the task with the corrected target group name. The rest of the instruction remains unchanged and is executable.", "heard_transcript": "I went on a few trips each with some of my um, roommates. My Simple Note has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, the companion groups should be, friends, not roommates." }, "candidate_id": "wav_179069e10bae4bd0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous initial request with a believable self-correction. The initial mention of 'roommates' is followed by a pause and a clear, spontaneous realization that it should be 'friends', with a smooth back-reference to the earlier term. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The pacing and intonation are very human." } ] }, "is_default": true }, { "variant_id": "revision_02_note_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Let me correct the note app: use 'Simple Note', not Gmail.", "long_pause_offsets": [ 228 ], "source_sha256": "25fbaab2728c798007f083a3b6bfaaa7fa53551131d369e7a608dd7983341791", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source of information from 'Gmail' to 'Simple Note', while retaining the action and the target (Splitwise groups). The final goal is consistent with the original instruction except for the corrected app name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/revision/revision_02_note_app/f62bc2c736408fe1/speech.wav", "wer": 0.11320754716981132, "wav_sha256": "a8e243d65da7aea3606f4398ed08134d05e90e135ece9c7370fb23f6c0ac9f12", "seconds": 22.4, "spoken_utterances": [ "I went on a few trips each with some of my uh... friends. My Gmail has information on who owes whom what from each trip. I have already created ... Splitwise groups for the trips. Record the expenses accordingly in the respective groups. ", "Let me correct the note app: use 'Simple Note', not Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states that Gmail contains the information, but immediately corrects this to Simple Note. This is a revision of the source information. The rest of the instruction, including the action (record expenses) and the target (Splitwise groups), remains consistent. The final intent is to execute the task with the corrected source, so the revision is properly applied.", "heard_transcript": "I went on a few trips each with some of my uh friends. My Gmail has information on who owes whom what from each trip. I have already created Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Let me correct, the note app, use Simple Note, not Gmail." }, "candidate_id": "wav_39c188576e5cd548", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a self-correction ('Let me correct the note app, use SimpleNote, not Gmail') that is grammatically and semantically coherent. However, the correction is delivered as a separate sentence rather than being integrated into the flow of the original request. The prosody is slightly flat, and the correction feels more like a post-hoc addition than a natural in-the-moment realization, making it less realistic than audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_expense_app", "proposed_instruction": "I went on a few trips each with some of my friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. Change the expense app to 'Splitwise' instead of 'Venmo'.", "long_pause_offsets": [ 230 ], "source_sha256": "d4f7d49400637cf435e92108158fa6b30b1a7e93510e728da7b94ce7ddd5dbf8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the expense app from Venmo to Splitwise, while retaining the action of recording expenses in the respective groups. The initial description of trips and Simple Note context is preserved as background." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/revision/revision_03_expense_app/e726c2dcb510e1cb/speech.wav", "wer": 0.019230769230769232, "wav_sha256": "93b1d02c6b4b6e9aad06a6dc78e7d8ace322e1ffd45c5a8127515ed3b8430602", "seconds": 24.44, "spoken_utterances": [ "I went on a few trips each with some of my hmmm... friends. My Simple Note has information on who owes whom what from each trip. I have already created ... Venmo groups for the trips. Record the expenses accordingly in the respective groups. ", "Change the expense app to '... Splitwise' instead of 'Venmo'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Venmo groups' but then revises the instruction to 'Splitwise groups' after a pause. The revision is clear and final, changing the target app from Venmo to Splitwise. The rest of the task—recording expenses in the respective groups—remains consistent. The revision type requires applying the correction to the earlier instruction, which is done here. The final intent is fully executable and matches the revised goal.", "heard_transcript": "I went on a few trips each with some of my um friends. My Simple Note has information on who owes whom what from each trip. I have already created Venmo groups for the trips. Record the expenses accordingly in the respective groups. Change the expense app to Splitwise instead of Venmo." }, "candidate_id": "wav_427e1cbecc2394bf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a revision ('Change the expense app to Splitwise instead of Venmo') that is grammatically correct but delivered as a separate, abrupt sentence. The initial request is fluent, but the revision lacks the natural back-reference or hesitation that would make it sound like a spontaneous correction. The prosody is somewhat robotic, and the overall delivery feels less like a real-time thought process and more like a structured command with a later edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "20ec33131e13767765713ee5e8c6309ba44e60759878c9d31b26c6e0d7588744", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/cancellation/cancel_039/7206ac20852a46d0/speech.wav", "wer": 0.057692307692307696, "wav_sha256": "565a08802af3d14e3d7628354fd140c526916d463915ac8c5ac05d4330bfd503", "seconds": 22.6, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Wait, don't go ahead with that. Cancel it.' This constitutes a final, unambiguous directive to cancel the previously stated task. The initial instruction is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "I went on a few trips each with some of my friends. My uh Simple Note has information on who owes whom what from each trip. I have already created um Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "ad7c8fd373a803906e84270cae123dbd10feb7f352ab0202ab0be4fc2e3c73d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/ambiguous_halt/halt_036/8f127384746ad5da/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "9dbd1647ebf6f794f1871a1d72743adedf513441cbbf8f0b4c61c55fc646efb7", "seconds": 22.64, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly says 'Wait a second, before you go ahead,' which is a direct request to pause. The final interjection is a pause command, not a cancellation or a new instruction. The type is 'ambiguous_halt', and the audio ends with a valid pause request, satisfying the criterion.", "heard_transcript": "I went on a few trips each with some of my friends. My uh, Simple Note has information on who owes whom what from each trip. I have already created um, Splitwise groups for the trips. Record the expenses accordingly in the respective groups. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "32616b5_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "29051ac7a1667bf0d4f93a734601321538306c0daf4a8b38af096c94715d72d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/32616b5_3/audio_natural/filler/original/e2dd72751961c680/speech.wav", "wer": 0.06818181818181818, "wav_sha256": "7bc4ceca42f4492ec8a9f467478b47b0bfa324390e377639e130b2121a135c5b", "seconds": 18.04, "spoken_utterances": [ "I went on a few trips each with some of my friends. My uh... Simple Note has information on who owes whom what from each trip. I have already created um... Splitwise groups for the trips. Record the expenses accordingly in the respective groups." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') and pauses but does not change the core request. The final instruction remains to record expenses in the respective groups, matching the original intent. No correction, addition, or cancellation is made. The filler words do not alter the task goal.", "heard_transcript": "I went on a few trips each with some of my friends. My uh, Simple Note has information on who owes whom what from each trip. I have already created um, Splitwise groups for the trips. Record the expenses accordingly in the respective groups." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_subscription_type", "proposed_instruction": "How much have I paid in basic membership—no, I meant prime membership since I made the amazon account?", "long_pause_offsets": [], "source_sha256": "21099084550af770e7b165c8b7a01ca40feb17e5cf4c6ed335585f6a239a5a1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'basic membership' to 'prime membership' within the same utterance, preserving the rest of the query including the app (amazon), the financial scope (how much paid), and the time frame (since I made the account)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/false_start/false_start_01_subscription_type/a480a5557ece9215/speech.wav", "wer": 0.0, "wav_sha256": "315d8f51ec6a4a9dfdc9f6e910d4b6c122b6149b553aac9371a018339c3768c3", "seconds": 8.92, "spoken_utterances": [ "How much have I paid in basic membership—um... no, I meant ... prime membership since I made the amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately self-corrects to 'prime membership'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in basic membership, um, no, I meant prime membership since I made the Amazon account." }, "candidate_id": "wav_efe94806a1d02055", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects 'basic membership' to 'Prime membership' with a natural 'no, I meant' transition. The correction is immediate and integrated smoothly into the sentence, maintaining a consistent pace and intonation. The overall delivery is fluent and sounds like a spontaneous, real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the Apple—sorry, I mean amazon account?", "long_pause_offsets": [], "source_sha256": "423e9237fa88bd51dcef6fa4061a3048c7d6d00c81aa2aa5cd80f4a8951126d2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple' to 'amazon' within the same utterance, preserving the rest of the query including the app (amazon), the service (prime membership), and the time frame ('since I made the account')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/false_start/false_start_02_app/d73fe2638decbcf8/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "4800e61eca3dd88f4b58e77ab91bb97c2cef4a76e15e4f908e53b35aeab342a7", "seconds": 11.32, "spoken_utterances": [ "How much have I paid in um, let me think... prime membership since I made the Apple—uh... sorry, I mean um... amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'apple' but immediately self-corrects to 'amazon' after a pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value ('amazon') should be applied, and the original incorrect value ('apple') should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in um, let me think, prime membership since I made the apple uh, sorry, I mean um, amazon account?" }, "candidate_id": "wav_1eb8bd7cb0b6bb35", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker starts with a similar structure but includes a longer, more drawn-out 'let me think' pause and a more abrupt, less integrated correction from 'Apple' to 'Amazon'. The correction feels slightly more forced, and the overall pacing is less fluid. While still realistic, the self-correction is less seamlessly woven into the utterance compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in since I made the amazon account? Oh, I forgot to mention, it's specifically for prime membership.", "long_pause_offsets": [ 57 ], "source_sha256": "8ea88290d431d37fd0c533911c85a3eb9822f3babac38a0b81286834514aeda9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specificity about 'prime membership', which clarifies the initial broad 'paid in' request. The app (Amazon), the financial context (paid), and the time frame (since I made the account) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/addition/addition_01/23fcc1ea2974e032/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5b7b1c1eb49491538c563fe6a42e525bf89f40f7d34022caac75e4bbf3f1b1ca", "seconds": 12.84, "spoken_utterances": [ "How much have I paid in um... since I made the amazon account? ", "Oh, I forgot to mention, it's specifically for um, let me think... prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('How much have I paid in um since I made the Amazon account?'), but the speaker immediately adds the missing constraint 'Prime membership' after a pause and self-correction ('let me think'). This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is fully specified and matches the reference.", "heard_transcript": "How much have I paid in um since I made the Amazon account? Oh, I forgot to mention, it's specifically for um, let me think, Prime membership." }, "candidate_id": "wav_25cd55da92c5b0ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much have I paid in...') and then adds a clarifying detail ('Oh, I forgot to mention...') about Prime membership. The addition is smoothly integrated with a realistic 'let me think' filler, and the prosody reflects genuine recall rather than a rehearsed or robotic delivery. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in prime membership since I made the account? I should also mention, this is for my Amazon account.", "long_pause_offsets": [ 67 ], "source_sha256": "7eda33b7fa6ec0ecab06fab5d535b6e5768602eeb4f62124f3cfb6bf70fadbb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the account is Amazon's, which is consistent with the original instruction's implied context. No conflicting or new constraints are introduced." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/addition/addition_02/8bb52300c41c80f8/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "40a9414292e0da8a5ca994e5e8d43ac5594d12f20985860e7869a4fd77572f8d", "seconds": 12.84, "spoken_utterances": [ "How much have I paid in um... prime membership since I made the account? ", "I should also mention, this is for my um, let me think... Amazon account." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'How much have I paid in um prime membership since I made the account?'. The speaker then adds 'I should also mention this is for my um let me think, Amazon account.' This addition clarifies the account context, which is consistent with the original instruction's intent. The addition does not contradict or alter the core request. The final intent is a combined request matching the original instruction with added clarification. Therefore, the addition type criterion is satisfied.", "heard_transcript": "How much have I paid in um prime membership since I made the account? I should also mention this is for my um let me think, Amazon account." }, "candidate_id": "wav_dd7996c23810774a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the addition of 'Amazon account' is also natural, the initial phrasing 'How much have I paid in... Prime membership...' is slightly more abrupt and less fluid than audio_01. The 'let me think' pause is placed after 'my', which feels slightly less organic than audio_01's placement. The overall delivery is still realistic but slightly less cohesive in its flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the ebay account? Let me correct the app: look for the payment history in Amazon, not eBay.", "long_pause_offsets": [ 72 ], "source_sha256": "18122217fdd76117f8eaef0dc2ac1d2af63cbede9311c9e89619c63a14c3893f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the original intent of querying the payment history for Prime membership since account creation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/revision/revision_02_app/6c465eb7dd1bd3b6/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "35e0cbc85a9f2b0936bcf048321a42343db4d01e0df43631fd7cdcaa43ff820b", "seconds": 13.24, "spoken_utterances": [ "How much have I paid in uh... prime membership since I made the ebay account? ", "Let me correct the app: look for the payment history in um... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions eBay but immediately corrects the app to Amazon, which is the original target. The revision is clear and final, and the core request (total Prime payment since account creation) remains unchanged. The correction applies to the target app, not the action or value, so the final intent matches the original instruction.", "heard_transcript": "How much have I paid in uh prime membership since I made the eBay account? Let me correct the app, look for the payment history in um Amazon, not eBay." }, "candidate_id": "wav_f00b36799954f820", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much have I paid in uh...') with a plausible filler and a clear, self-corrected revision ('let me correct the app, look for the payment history in um... Amazon not eBay'). The correction is smoothly integrated into the same utterance, with a believable shift in focus and a clear reference back to the initial error. The prosody is consistent, and the pacing reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "revision_01_subscription_type", "proposed_instruction": "How much have I paid in prime video since I made the amazon account? Wait, that should be prime membership, not prime video.", "long_pause_offsets": [ 69 ], "source_sha256": "8ac0f8d3b60750ee4b2b32bb69efc33e49d7d8203bd21cb395a7b1427fa4aff4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the service from 'prime video' to 'prime membership', preserving the action, time frame, and context. The original instruction is fully restored." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/revision/revision_01_subscription_type/9d063b572905c363/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "59a82f22c6aca06d60c9d33bf963cc885bf18f527ce8659f5b30daffec3fc450", "seconds": 12.08, "spoken_utterances": [ "How much have I paid in uh... prime video since I made the amazon account? ", "Wait, that should be ... prime membership, not prime video." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Prime Video' but immediately self-corrects to 'Prime membership' after a pause, which is the final intended target. The revision type requires applying the correction spoken after a longer pause to the earlier instruction, and the final audible intent matches the original instruction exactly.", "heard_transcript": "How much have I paid in uh Prime Video since I made the Amazon account? Wait, that should be Prime membership, not Prime Video." }, "candidate_id": "wav_90f7cf6b21306d60", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar hesitant tone and filler ('How much have I paid in uh...'), but the revision ('Wait, that should be Prime membership not Prime Video') is delivered as a separate, slightly abrupt clause after a short pause. While the correction is semantically clear, the transition feels less fluid than in audio_01, and the 'Wait' introduces a more abrupt cognitive shift. The overall delivery is still realistic but slightly less integrated than the first." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "15854cd31cbd9a5fcc9b2dc7065d7e98a5e0627db9429bfb8da84b07195ef966", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/cancellation/cancel_026/ee5144c702d82247/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "83568747fd9f8c67f9d1a0ee1e4abd113b0f20c4f3894754d854e14862489d77", "seconds": 12.48, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?", "Wait, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question, then pauses and says 'forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How much have I paid in, um, prime membership, uh, let me think, since I made the amazon account? Wait, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "5f0dd743ea0fa4f70ed490c33d30cfc56070af47feea4c774b66a5da6df94843", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/ambiguous_halt/halt_002/349e836d524f47a5/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "a9ac1c53ace03c608486695ac2e8ade2883d971606a65782140a4dec690fb3ae", "seconds": 10.48, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, but the speaker explicitly halts further action. No execution occurs, and the final intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "How much have I paid in prime membership, uh, let me think, since I made the Amazon account? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "365e0a3_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4040c038dcc8e4e5716efa96414ad43d2098dc3567e09400fdaa102c2d4c86a0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_1/audio_natural/filler/original/41700c140bbf3565/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "71500f75ecccf452c46dd6d408ea7caf54574c735549feb9c28220c441e69aca", "seconds": 7.56, "spoken_utterances": [ "How much have I paid in hmmm... prime membership uh, let me think... since I made the amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('mm', 'uh', 'let me think') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "How much have I paid in mm prime membership uh let me think since I made the amazon account?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "365e0a3_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in premium membership since I made the Apple Music—I should say spotify account?", "long_pause_offsets": [], "source_sha256": "a7dd875b386b4e68ede8ba05281829d52f5857655d3a4d7fec0597249532e8bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the app, the financial aspect, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/false_start/false_start_02_app/8a4e81b86449af47/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "b079764d2f1f4d39d2c75519037fc98f61cd8db64dd220a8c493d17209bd4fd0", "seconds": 12.24, "spoken_utterances": [ "How much have I paid in premium membership ... since I made the um... Apple Music—uh... I should say spotify account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify account'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in premium membership since I made the um, Apple Music uh, I should say Spotify account?" }, "candidate_id": "wav_78a52851b8d6b4b6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start with a clear self-correction from 'Apple Music' to 'Spotify'. The speaker's intonation and pacing are consistent with spontaneous speech, including a realistic pause and filler ('um') before the correction. The correction is immediate and smoothly integrated into the sentence, maintaining a natural flow without any robotic or rehearsed quality." } ] }, "is_default": true }, { "variant_id": "false_start_01_subscription_type", "proposed_instruction": "How much have I paid in basic membership—no, I meant premium membership since I made the spotify account?", "long_pause_offsets": [], "source_sha256": "08d25a31d9bd35410f2ca1408a4bad3e56bab2c2c553c95124e16b4faa5ce653", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'basic membership' to 'premium membership' within the same utterance. The final intended membership type is premium, and all other elements of the query (amount paid, time frame, and service) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/false_start/false_start_01_subscription_type/30aeb62173b08946/speech.wav", "wer": 0.1, "wav_sha256": "bf1d46665fbfd991a570505118d38c6af1ba49776e1cb1d5248482f2ff0eefec", "seconds": 9.64, "spoken_utterances": [ "How much have I ... paid in basic membership—um... no, I meant premium membership since I made the spotify account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately self-corrects to 'premium membership' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in basic membership, um, no, I'm in premium membership since I made the Spotify account." }, "candidate_id": "wav_15942f8a86e33a5c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a self-correction from 'basic membership' to 'premium membership', but the correction feels slightly less fluid. The speaker hesitates with 'um' and 'no', but the shift in topic is abrupt and the intonation on 'no' sounds more like a flat negation than a natural self-repair. The overall delivery is slightly more hesitant and less continuous than audio_01, making it less realistic in terms of spontaneous speech flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in since I made the spotify account? Oh, I forgot to mention, it's specifically for premium membership.", "long_pause_offsets": [ 58 ], "source_sha256": "df6932de8c3cb473d8865143050e138c5d052efabcfe89c873756a66a1dddc8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'premium membership' to the initial broad question about payments, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/addition/addition_01/746e5f000371231b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "4868187d2ef6d150d10f565ce7d760aab14cabcac93694843b35e9dbc87ed990", "seconds": 12.44, "spoken_utterances": [ "How much have I let me think... paid in ah... since I made the spotify account? ", "Oh, I forgot to mention, it's specifically for premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about payments since making the Spotify account, then adds the specific constraint 'for premium membership' after a longer pause. This matches the 'addition' type criterion, where the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "How much have I let me think paid in uh since I made the Spotify account. Oh, I forgot to mention it specifically for premium membership." }, "candidate_id": "wav_320066d5154d1309", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in premium membership since I made the account? One more thing: this is about my Spotify account.", "long_pause_offsets": [ 69 ], "source_sha256": "4e0169085b58ab29782448e73ed2c3043e1f33a4d8b6a1d4762c75830d598cb2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the platform context by specifying 'Spotify account', which was implied but not explicitly stated in the original instruction. The core question about total premium payments since account creation remains unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I uh... paid in premium membership since I made the account? ", "One more thing: this is about my ... Spotify account." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in premium membership since I made the apple music account? Wait, the question is about Spotify, not Apple Music.", "long_pause_offsets": [ 81 ], "source_sha256": "b126f8c9e25c3831fca3df47779c98f33f555f5f0dad4c383add0e3e66852a63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the rest of the query including the action, financial scope, and time reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_2/audio_natural/revision/revision_02_app/5c9cdb782df5b92f/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "18ebf693ee44ee3daf38aedc2bb3010113b6f4e9a7fd2eb06436a7fc69ef88d7", "seconds": 16.76, "spoken_utterances": [ "How much have I hmmm... paid in premium membership uh... since I made the apple music account? ", "Wait, the question is about hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects themselves to Spotify after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to inquire about Spotify, which matches the original reference instruction. All other constraints are preserved, and the correction is clear and unambiguous.", "heard_transcript": "How much have I hmm paid in premium membership uh since I made the Apple Music account. Wait, the question is about hmm Spotify, not Apple Music." }, "candidate_id": "wav_268522adc155e6f6", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_subscription_type", "proposed_instruction": "How much have I paid in basic membership since I made the spotify account? Sorry, I meant premium membership, not basic membership.", "long_pause_offsets": [ 75 ], "source_sha256": "84f7c66bec6d8ffc1482493504442fdcea5f2bfba177f5c0c27f28b374acb72a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'basic membership' to 'premium membership', preserving the rest of the query including the time frame and the app context." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I uh... paid in basic membership since I made the spotify account? ", "Sorry, I meant ... premium membership, not basic membership." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_005", "hide_variant_original": true, "source_sha256": "d20b1e29d14aec3416b88d5ceae88e17e92f65c78832b8423afc99b2f906ad50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I um... paid in premium membership ah... since I made the Spotify account?", "Hold on, cancel what I just asked you to do." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "d13372754f389b795eadec5f1b095a0346cfb7493780dec1ce873f721c01c8cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I um... paid in premium membership ah... since I made the Spotify account?", "Wait a second..." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ce079e34d976c4e41bee4ce3d1c6c9ff755900351217dcd2c0343660e31f1d54", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I um... paid in premium membership ah... since I made the Spotify account?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the apple—sorry, make that amazon account?", "long_pause_offsets": [], "source_sha256": "c925f8848a55bd9b0a3a31b78b151e0131bc04776cf3738c4f18e645127b2342", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'apple' to 'amazon' within the same utterance, preserving the rest of the query including the subject (prime membership), the action (how much paid), and the time reference (since I made the account)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/365e0a3_3/audio_natural/false_start/false_start_02_app/992820b8249d4389/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c5ad601c77a750791c9ec661fa38d629dc4e4e928fa45ee2fc3c730b6f3bdb86", "seconds": 10.96, "spoken_utterances": [ "How much have I paid in ah... prime membership since I made the uh... apple—um... sorry, make that amazon account?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'apple' but immediately self-corrects to 'Amazon' with 'sorry, make that Amazon account'. This is a clear false start where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to calculate the total paid for Prime membership since the Amazon account was made.", "heard_transcript": "How much have I paid in uh prime membership since I made the uh apple um sorry, make that Amazon account?" }, "candidate_id": "wav_eb841d51a4a2f0f6", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_01_subscription_type", "proposed_instruction": "How much have I paid in basic membership—wait, I meant prime membership since I made the amazon account?", "long_pause_offsets": [], "source_sha256": "7f6c2bf6f9b4ddc716f49673c5e04387930fb5799f41ae59fbccae67f43bab72", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'basic membership' to 'prime membership' within the same utterance, preserving the rest of the query including the app (amazon), the financial scope (how much paid), and the time frame (since I made the account)." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in basic membership—um... wait, I meant prime membership uh... since I made the amazon account?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "365e0a3_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in since I made the amazon account? Oh, I forgot to mention, it's specifically for prime membership.", "long_pause_offsets": [ 57 ], "source_sha256": "8ea88290d431d37fd0c533911c85a3eb9822f3babac38a0b81286834514aeda9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specificity about 'prime membership', which clarifies the initial broad 'paid in' request. The app (Amazon), the financial context (paid), and the time frame (since I made the account) are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in let me think... since I made the amazon account? ", "Oh, I forgot to mention, it's specifically for ... prime membership." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in prime membership since I made the account? I didn't mention this, but it's for my Amazon account.", "long_pause_offsets": [ 67 ], "source_sha256": "9cbb11c7cb12e40868e1305d52bd420c2cd94684f33460027c1e56569827c5d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the account in question is the Amazon account, which was not explicitly stated in the initial utterance but is the only account referenced in the original instruction. The final goal is identical to the original instruction with the same action, entity, and time frame." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in prime membership ... since I made the account? ", "I didn't mention this, but it's for my uh... Amazon account." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_subscription_type", "proposed_instruction": "How much have I paid in prime video since I made the amazon account? Wait, I meant prime membership, not prime video.", "long_pause_offsets": [ 69 ], "source_sha256": "d599b0d8f426ca2c63eb62aa2411a54ebd66c93b806674375de0f6adefecf014", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the service from 'prime video' to 'prime membership', preserving the action, time frame, and context." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in uh... prime video since I made the amazon account? ", "Wait, I meant um... prime membership, not prime video." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in prime membership since I made the apple account? Let me correct the app: the account is on Amazon, not Apple.", "long_pause_offsets": [ 73 ], "source_sha256": "3e1e54e77ed30a424d0ba1ba936c8a66e7d8fd0a117c77f6c93afb866bdd6f0e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple to Amazon, preserving the original intent of querying the total payment for Prime membership since account creation." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in um, let's see... prime membership since I made the apple account? ", "Let me correct the app: the account is on um... Amazon, not Apple." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_022", "hide_variant_original": true, "source_sha256": "b60065fcc098e84f45b4631ce65b1b1b8af06e87ebc2843b50888ae2b6b3ac21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in uh, let's see... prime membership um... since I made the amazon account?", "Hang on, actually, never mind. Cancel the task." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "b4370fe01473788a7197d8df2d772b4f8513d733668ed7eccb3be08b42f9f10c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in uh, let's see... prime membership um... since I made the amazon account?", "Wait, let me think..." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "365e0a3_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8c50f27e3b89b6f9654463e8a77b500ed46025e4c51579dc209d204b83f7f3b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in uh, let's see... prime membership um... since I made the amazon account?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "37a8675_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_visibility", "proposed_instruction": "Send $91 publicly—sorry, I meant privately on Venmo to the person with this phone number 2192158297.", "long_pause_offsets": [], "source_sha256": "97385442d3b17b255f4b393f4f78567be21d3934396f9ba7795285143a8870ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects to 'privately' within the same utterance. The amount and recipient phone number are correctly stated and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/false_start/false_start_02_visibility/48d7c0df5c099b2e/speech.wav", "wer": 0.0, "wav_sha256": "d4276a256f8006b8af3ef5fb2a23b92e447913bfefb0d70a5c694cef04123769", "seconds": 11.48, "spoken_utterances": [ "Send ninety-one dollars publicly—um... sorry, I meant ... privately on Venmo to the person with this phone number two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately self-corrects to 'privately'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is overridden. The amount ($91) and recipient (phone number 2192158297) remain consistent. The correction is immediate and unambiguous, so the final intent matches the original instruction except for the privacy setting, which is corrected as intended.", "heard_transcript": "Send ninety one dollars publicly, um, sorry, I meant privately on Venmo to the person with this phone number two one nine, two one five, eight two nine seven." }, "candidate_id": "wav_774d05c8e9e651aa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a clear initial request, then self-corrects with a believable 'sorry, I meant' phrase, smoothly transitioning to the correct value. The prosody is consistent, with a natural pause and intonation shift during the correction. The delivery is fluent and sounds like a genuine moment of recall, with no awkwardness or hesitation beyond what's expected in real speech." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Send $91 privately on CashApp—I should say Venmo to the person with this phone number 2192158297.", "long_pause_offsets": [], "source_sha256": "995f8c770563d16c8579032e698745fea244db2e516c4568c10c055cb361d034", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the app from CashApp to Venmo and specifies the correct phone number. The amount and privacy setting are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/false_start/false_start_03_app/12f41c8a4210cb05/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "4c402e1e0f6a1cb34e0a77663206f8be6005f4ce29b0cf11fa364748c8b3244c", "seconds": 13.12, "spoken_utterances": [ "Send ninety-one dollars privately on CashApp—ah... I should say Venmo um... to the person with this phone number uh... two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. The final intent is clear and matches the original instruction after the correction. The amount and recipient phone number are consistent. This is a false start where the correction is immediate and unambiguous.", "heard_transcript": "Send ninety one dollars privately on Cash App. I should say Venmo, um, to the person with this phone number, uh, two one nine, two one five, eight two nine seven." }, "candidate_id": "wav_5799fe6b640b08d2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a credible self-correction with 'I should say Venmo' after initially saying 'Cash App'. The 'uh' and 'um' fillers are placed naturally, and the correction is integrated smoothly into the ongoing sentence. The speaker's voice and prosody remain consistent, and the correction feels like a spontaneous realization rather than a rehearsed edit. It's slightly less fluid than audio_04 due to the initial incorrect app name, but still very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $95—make that $91 privately on Venmo to the person with this phone number 2192158297.", "long_pause_offsets": [], "source_sha256": "1ce931e92571e16bee25e2834ef5e49a70fd6eb5fa11ce46eb837b5d5b5d73f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces ninety-five with ninety-one; the recipient identifier (phone number) and privacy setting remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/false_start/false_start_01_amount/4ca52edf650d8b50/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "41ae1b3b943a12474e63f6c9431934c21bc6ad1c093d18359b09d094cb490e31", "seconds": 14.36, "spoken_utterances": [ "Send ... ninety-five dollars—hmmm... make that ninety-one dollars privately on let me think... Venmo to the person with this phone number two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ninety five dollars' but immediately corrects to 'ninety one dollars'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction, including the app (Venmo), the action (send), the privacy setting (privately), and the recipient identifier (phone number 2192158297), remains consistent and is not altered. Therefore, the final intent matches the original instruction, and the correction is properly applied.", "heard_transcript": "Send ninety five dollars. Make that ninety one dollars privately on Venmo to the person with this phone number two one nine two one five eight two nine seven." }, "candidate_id": "wav_e582f5894086c556", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'Send $95' followed by 'Make that $91'. The correction is immediate and natural. However, the speaker then says 'on let me think Venmo', which introduces a longer, slightly awkward pause and filler that disrupts the flow. The number is stated clearly, but the overall delivery feels a bit more hesitant and less cohesive than the top two." } ] }, "is_default": false }, { "variant_id": "false_start_04_phone_number", "proposed_instruction": "Send $91 privately on Venmo to the person with this phone number 2192158298—scratch that, make it 2192158297.", "long_pause_offsets": [], "source_sha256": "c6e03dcf8b976946233974d630be59848738f9d8312432e43e20121e91f0df22", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker first states the phone number as 2192158298, then immediately corrects it to 2192158297. All other elements—amount, app, privacy, and action—match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/false_start/false_start_04_phone_number/87ca7e62830e115f/speech.wav", "wer": 0.3488372093023256, "wav_sha256": "f8b1106e01b4496e2e8b0a195019f4e9b41d850cd8242a0a66e53c33af8c1465", "seconds": 18.48, "spoken_utterances": [ "Send let me think... ninety-one dollars privately on Venmo um... to the person with this phone number two one nine, two one five, eight two nine eight—uh... scratch that, make it two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misreads the phone number as 2192158298, then immediately self-corrects to 21982158297. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The corrected number matches the reference, and all other constraints (amount, app, privacy) are preserved. Therefore, the final intent matches the original instruction.", "heard_transcript": "Send, let me think. Ninety one dollars privately on Venmo, um, to the person with this phone number, two one nine, two one five, eight two nine eight or eight two. Scratch that, make it two one nine, eight two one five, eight two nine seven." }, "candidate_id": "wav_d30d50d4e814b5f1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has multiple issues that reduce its realism. It starts with 'Send let me think', indicating a significant hesitation. The speaker then gives a number, '2192158298 or 82', which is unclear and possibly a mispronunciation. The correction 'Scratch that, make it 21982158297' is abrupt and lacks a natural intonation shift. The number itself is also mispronounced ('2198' instead of '2192'), and the overall delivery sounds disjointed and less spontaneous than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Send $91 on Venmo to the person with this phone number 2192158297. Oh, and one more thing: send it privately.", "long_pause_offsets": [ 67 ], "source_sha256": "ed2432278eaebd7d4341459fd8dc419ff61fbd1709f30ab65fd6234ec7977e3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing privacy constraint; all other values including amount, app, recipient identifier, and phone number are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/addition/addition_01/f5177d0b5e282ee5/speech.wav", "wer": 0.0, "wav_sha256": "460691986589e5db677fbc58af2dfd2362a82ef552c4648334c877cf6db181db", "seconds": 15.36, "spoken_utterances": [ "Send ... ninety-one dollars on ... Venmo to the person with this phone number let me think... two one nine, two one five, eight two nine seven. ", "Oh, and one more thing: send it privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send $91 on Venmo to the person with the given phone number. After a brief pause and hesitation, they add the constraint 'send it privately'. This is a clear addition to the original instruction, and the final goal combines both the original and the added condition. The addition is not a correction or cancellation, but an extra constraint. Therefore, the final intent matches the expected behavior for an 'addition' type task.", "heard_transcript": "Send ninety one dollars on Venmo to the person with this phone number, let me think, two one nine, two one five, eight two nine seven. Oh, and one more thing, send it privately." }, "candidate_id": "wav_9d8db92f0b575325", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then uses a believable 'let me think' filler pause before providing the phone number. The delivery includes a smooth, pragmatic addition ('Oh, and one more thing') followed by a coherent correction of the initial instruction. The prosody is consistent, and the flow reflects a realistic thought process with a minor correction and addition." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send $91 privately on Venmo to the person with this phone number. I left out a detail: send it to the person with this phone number 2192158297.", "long_pause_offsets": [ 66 ], "source_sha256": "1cb208430b955c0c45ec9fb418d8aa7ce4a40512df0ba0d2bbbf99e5b2e09038", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing phone number, which matches the original. The initial utterance omits the number but otherwise matches the final goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/addition/addition_02/d3d04134d101504a/speech.wav", "wer": 0.05, "wav_sha256": "39fd02b7c8c2ed0ef43e7905b0a7f6ab04a388d21c0d4df083282db2977f65b1", "seconds": 16.28, "spoken_utterances": [ "Send uh... ninety-one dollars privately on Venmo ah... to the person with this phone number. ", "I left out a detail: send it to the person with this phone number two one nine, two one five, eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send $91 privately on Venmo but omits the phone number. After a pause and an interjection, they correct and complete the instruction by adding the full phone number. This matches the 'addition' type criterion: the original instruction is combined with the added information to form the final goal, which matches the reference.", "heard_transcript": "Send uh ninety one dollars privately on Venmo uh to the person with this phone number. I left out a detail. Send it to the person with this phone number two one nine, two one five, eight two nine seven." }, "candidate_id": "wav_363c51b4f8788ee9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but includes a noticeable pause and filler ('uh') before continuing. The second part of the utterance is a self-correction that adds a missing detail. However, the transition between the two parts feels slightly abrupt, and the correction is delivered more as a separate statement than a natural self-repair within the same thought flow. The prosody is less fluid compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Send $91 privately on CashApp to the person with this phone number 2192158297. Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 79 ], "source_sha256": "58b3788d8b9496cfb7a5d7761f812a6fb1c8eaad03c4d1ae20da2ead9a493e7b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the amount, privacy setting, and recipient phone number. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/revision/revision_03_app/4b0140febf88bc60/speech.wav", "wer": 0.38461538461538464, "wav_sha256": "d443d30616135b0cad0175d7f796109b13bab86a12197a799b4d6824f85410d1", "seconds": 19.92, "spoken_utterances": [ "Send ninety-one dollars privately on um, let me think... CashApp to the person with this phone number hmmm... two one nine, two one five, eight two nine seven. ", "Let me correct the app: use um... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, privacy setting, app, and recipient identifier.", "heard_transcript": "Send ninety one dollars privately on um, let me think, Cash App to the person with this phone number, two one nine, two one five, eight two nine seven. Let me correct the app, use um, Venmo not Cash App." }, "candidate_id": "wav_1a3d61719242c274", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear request, then uses a believable 'let me think' filler before correcting the app from 'cash app' to 'Venmo'. The correction is integrated smoothly within the same utterance, with a natural self-repair and no awkward pauses. The prosody and pacing mimic genuine hesitation and recollection, making it sound like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Send $91 publicly on Venmo to the person with this phone number 2192158297. Wait, send it privately, not publicly.", "long_pause_offsets": [ 76 ], "source_sha256": "7d2cba86e64d0f4ca8858b644e8d3e75b58997c42bb962c2e724a925e6f5a90b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the visibility from 'publicly' to 'privately', while retaining the amount, app, recipient phone number, and action. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/revision/revision_02_visibility/054a0354811b20e2/speech.wav", "wer": 0.0625, "wav_sha256": "39ced003e2d6f59856850ec9e84e3a55f5f55f244569529416ebe1296a3eee7b", "seconds": 18.04, "spoken_utterances": [ "Send ninety-one dollars publicly on Venmo um... to the person with this phone number uh... two one nine, two one five, eight two nine seven. ", "Wait, send it ... privately, not publicly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects to 'privately' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send the money privately, which matches the original instruction's intent except for the privacy setting. The amount and recipient remain unchanged.", "heard_transcript": "Send ninety one dollars publicly on Venmo um to the person with this phone number uh two one nine two one five eight two nine seven. Wait, send it privately, not publicly." }, "candidate_id": "wav_a0dc60b6055c72cb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision: the speaker initially says 'publicly' but then corrects to 'privately'. The 'wait' pause and the self-correction are natural and well-timed. However, the initial request is slightly less fluent than audio_02, with a minor stumble ('um') and a less smooth transition into the correction. Still, the revision flow is credible and the prosody remains consistent." } ] }, "is_default": false }, { "variant_id": "revision_04_phone_number", "proposed_instruction": "Send $91 privately on Venmo to the person with this phone number 2192158298. Change the phone number to 2192158297 instead of 2192158298.", "long_pause_offsets": [ 77 ], "source_sha256": "8359352a163a763ebbe21d8be3fb9e1aab2992b775ab2ff968902eb318975e78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the phone number from 2192158298 to 2192158297, while retaining the amount, app, privacy, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/revision/revision_04_phone_number/021838b94dce2637/speech.wav", "wer": 0.019230769230769232, "wav_sha256": "607369af8f40ed954dc19b3e636471da10f3af3835f489c4823931269a924c70", "seconds": 26.48, "spoken_utterances": [ "Send ninety-one dollars privately on ... Venmo to the person with this phone number ... two one nine, two one five, eight two nine eight. ", "Change the phone number to um... two one nine, two one five, eight two nine seven instead of two one nine, two one five, eight two nine eight." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the phone number as 2192158298, then corrects it to 2192158297 after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the amount, privacy setting, app, and corrected phone number.", "heard_transcript": "Send ninety one dollars privately on Venmo to the person with this phone number two one nine two one five eight two nine eight. Change the phone number to um two one nine two one five eight two nine seven instead of two one nine two one five eight two nine eight." }, "candidate_id": "wav_b9d14ff67daf5983", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic due to its robotic and unnatural delivery. The speaker makes a revision from one phone number to another, but the correction is abrupt and lacks any natural hesitation or self-repair. The prosody is flat, the pace is monotonous, and there are no fillers or pauses that would suggest spontaneous thought. The entire utterance sounds like a pre-recorded or synthesized command rather than a human speaking in real time." } ] }, "is_default": false }, { "variant_id": "revision_01_amount", "proposed_instruction": "Send $95 privately on Venmo to the person with this phone number 2192158297. Make that $91, not $95.", "long_pause_offsets": [ 77 ], "source_sha256": "1655e977bbc344ffb9dd0a2ffaa7299ba34202ccbd6be61a4540c18a41f80ccb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the amount from ninety-five to ninety-one, while retaining the recipient phone number, app, and privacy setting." }, "audio": { "status": "pending", "spoken_utterances": [ "Send ... ninety-five dollars privately on Venmo to the person with this phone number uh, let me think... two one nine, two one five, eight two nine seven. ", "Make that uh... ninety-one dollars, not ninety-five dollars." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "27ba739a886c624d977ee97001c5ac5ca3c03b631567481c574e3173340af727", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/cancellation/cancel_018/d5e90d1e78ae2aa3/speech.wav", "wer": 0.1, "wav_sha256": "733022540b37730e211a015fbd56bd08d1ca8263ebc305a7b5a01114c868ef9e", "seconds": 17.72, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven.", "Wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the full transaction request, then clearly revokes it with 'Wait, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the task. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send ninety-one dollars privately on Venmo to the person with this phone number two one nine two one five eight two nine seven. Wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "24c090675eb12bab9c6b579ee27de61d2733b3447a40c08a2dd3be32983123c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/ambiguous_halt/halt_008/b1053211861bcb96/speech.wav", "wer": 0.125, "wav_sha256": "72a5dff15ebf9ec6ab552fa3a3318bdcbe2e6581daf3c65d457e463d42944f7f", "seconds": 15.48, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction with a final interjection 'Wait a second,' which is an audible request to pause or hold. There is no cancellation or permission to resume, and the original task is not executed. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Send ninety one dollars privately on uh Venmo uh let me think to the person with this phone number two one nine two one five eight two nine seven. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e7b2bce7c6e6fb8993ea1813175c9d5fb47220a3508e37e3a165b9b40f52cd57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_1/audio_natural/filler/original/e32fa0e452d7e06a/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "9b5974af3bd6aba289dd072ef5d7dd33c404363f1e8d10fa96924948af5b7f5e", "seconds": 12.44, "spoken_utterances": [ "Send ... ninety-one dollars privately on ah... Venmo uh, let me think... to the person with this phone number two one nine two one five eight two nine seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'uh, let me think' but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification to the amount, app, privacy setting, recipient, or phone number. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Send $91 privately on Venmo, uh, let me think, to the person with this phone number 2192158297." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_app", "proposed_instruction": "Send $100 publicly on CashApp—I should say Venmo to the person with this phone number 5734599766.", "long_pause_offsets": [], "source_sha256": "092246b3e81ce132ca067ae32fa798171956cd076b6ead14720ef4d89c6ef720", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'CashApp' but immediately corrects to 'Venmo'. The amount and recipient phone number are correctly stated in the final version." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/false_start/false_start_03_app/a4157bb11340ca83/speech.wav", "wer": 0.09375, "wav_sha256": "9441f7a265326c5e1f97617b1794c07ecb206c75997fef8ba9e61da09186feaf", "seconds": 13.08, "spoken_utterances": [ "Send one hundred dollars publicly on CashApp—hmmm... I should say Venmo uh, let's see... to the person with this phone number five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. The final intent is clear and matches the original instruction with the correct app. The value and phone number are correctly stated. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Send one hundred dollars publicly on Cash App. Perm. I should say Venmo. Uh, let's see. To the person with this phone number five seven three four five nine nine seven six six." }, "candidate_id": "wav_c28ada8d792cc421", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident request, then self-corrects 'Cash App' to 'Venmo' with a natural hesitation ('uh') and a brief pause ('let's see'), followed by a smooth continuation. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The number is spoken clearly and without hesitation." } ] }, "is_default": true }, { "variant_id": "false_start_02_visibility", "proposed_instruction": "Send $100 privately—no, I mean publicly on Venmo to the person with this phone number 5734599766.", "long_pause_offsets": [], "source_sha256": "1686f7873c5c7c61507d15734862fa7a17149e18f3ba6fe50f37bc5783c5eb79", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'privately' to 'publicly' within the same utterance, preserving the amount, app, recipient identifier, and phone number." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/false_start/false_start_02_visibility/7e5a0a034dd6c9d4/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "a51fdd10b1794f163506591a953f21fe0a6b6d3b4dfdc4f6d36d70cf75ccb789", "seconds": 14.56, "spoken_utterances": [ "Send one hundred dollars um... privately—uh... no, I mean um... publicly on Venmo to the person with this phone number five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'privately' but immediately self-corrects to 'publicly'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference in action, amount, visibility, app, and target. No other constraints are altered.", "heard_transcript": "send one hundred dollars um privately uh no, I mean um publicly on Venmo to the person with this phone number five seven three four five nine nine seven six six" }, "candidate_id": "wav_36ae2c55ff5b96f8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then hesitates and self-corrects 'privately' to 'publicly' with a natural 'no, I mean' structure. The flow is coherent, and the correction is well-integrated. However, the initial hesitation ('um') and the slightly awkward phrasing ('privately uh no I mean') make it slightly less fluid than audio_01, though still very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_phone_number", "proposed_instruction": "Send $100 publicly on Venmo to the person with this phone number 5734599767—scratch that, make it 5734599766.", "long_pause_offsets": [], "source_sha256": "f2de3afd253b3152bb2a14d24d5d3500ab12ed944d29b8b4698a3c725fe95415", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces the final digit of the phone number from seven to six, aligning with the reference number. The amount, app, and public visibility remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/false_start/false_start_04_phone_number/e61f5e5c09ebfb68/speech.wav", "wer": 0.07317073170731707, "wav_sha256": "53a844793a177c9a427e73d914dd185c6a12d8d14b0b42e5901aa98e09167527", "seconds": 18.52, "spoken_utterances": [ "Send um... one hundred dollars publicly on Venmo to the person with this phone number five seven three, four five nine, nine seven six seven—hmmm... scratch that, make it um... five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong phone number (5734599767), then immediately self-corrects with 'Scratch that' and provides the correct number (5734599766). According to the 'false_start' type criterion, the final corrected value should be executed, and the original incorrect value discarded. The corrected number matches the reference exactly, and all other conditions (amount, app, visibility) are preserved.", "heard_transcript": "Send um, one hundred dollars publicly on Venmo to the person with this phone number five seven three four five nine nine seven six seven. Who? Scratch that. Make it um, five seven three four five nine nine seven six six." }, "candidate_id": "wav_9878b93911e0ef2a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but then makes a revision mid-sentence, saying 'who?' and then 'scratch that, make it...'. The revision is plausible, but the 'who?' interjection sounds slightly unnatural and breaks the flow. The correction is delivered with a rushed tone, and the number is changed from 5734599767 to 5734599766, which feels like a last-minute correction rather than a spontaneous recall. The overall delivery is less smooth than the top two." } ] }, "is_default": false }, { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $150—sorry, make that $100 publicly on Venmo to the person with this phone number 5734599766.", "long_pause_offsets": [], "source_sha256": "6db5add2e68b0dd8fb2a548080fc157d27ca917efbd917d8c15874b5547961d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces one hundred fifty with one hundred; the recipient, app, privacy setting, and phone number remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/false_start/false_start_01_amount/d24d84ee637f4a49/speech.wav", "wer": 0.20588235294117646, "wav_sha256": "63fe3fe3d29f521b402dbd34b91f62471a6795b76756b78a8d7449ecd4a132db", "seconds": 16.36, "spoken_utterances": [ "Send um... one hundred fifty dollars—hmmm... sorry, make that one hundred dollars publicly on Venmo um... to the person with this phone number five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'one hundred fifty dollars' but immediately corrects to 'make that a hundred dollars'. This is a clear false start where the final value is corrected. The rest of the instruction, including the action (send), amount ($100), method (Venmo), visibility (publicly), and recipient (via phone number), matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send um, one hundred fifty dollars. Sorry, make that a hundred dollars publicly on Venmo um, to the person with this phone number five seven three four five nine nine seven six six." }, "candidate_id": "wav_66b390b394f27f4a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker begins with a clear request but then makes a revision from '$150' to '$100', saying 'sorry, make that a hundred dollars'. However, the initial amount is stated as 'um one hundred fifty dollars' with a filler and hesitation, and the correction is abrupt. The prosody is inconsistent, and the overall delivery sounds less natural and more like a rehearsed or forced correction. The number is spoken clearly, but the flow is disrupted by the revision and filler words." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Send $100 on Venmo to the person with this phone number 5734599766. Oh, and one more thing: make the payment public.", "long_pause_offsets": [ 68 ], "source_sha256": "e4d2c154aaaae3a37c8f3fa82aaa7e021fb4e226863c9ce9b0151258e9a6ae74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'public' constraint, which was not in the initial utterance but is present in the original instruction. The amount, recipient, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/addition/addition_01/215bc7c0552a9356/speech.wav", "wer": 0.0, "wav_sha256": "2a4e3c650aa400bcc301437aeca0ffa102bae215a4b2635e503bac8d56faa9d6", "seconds": 15.76, "spoken_utterances": [ "Send ... one hundred dollars on Venmo ... to the person with this phone number five seven three, four five nine, nine seven six six. ", "Oh, and one more thing: make the payment public." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to send $100 on Venmo to a person with a given phone number. After a pause and an 'Oh, and one more thing' interjection, the speaker adds the constraint 'make the payment public'. This is a clear addition to the initial request, and the final intent combines both the original and the added condition. The addition is not a correction or cancellation, and the final goal matches the reference instruction with the added public visibility requirement. Therefore, the task is executed with the full intent.", "heard_transcript": "Send $100 on Venmo to the person with this phone number 5734599766. Oh, and one more thing, make the payment public." }, "candidate_id": "wav_5a795051f67463fa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the request with a natural, conversational rhythm and clear prosody. The initial command is fluent, followed by a plausible 'Oh, and one more thing' addition that smoothly introduces the extra instruction. The intonation and pacing suggest spontaneous thought, and the number is spoken clearly without hesitation or distortion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send $100 publicly on Venmo. I should also mention, send it to the person with this phone number 5734599766.", "long_pause_offsets": [ 29 ], "source_sha256": "6a1f8012ef627ae4080816292e06e0324e750f20c0a5c04adaf41da202fe8059", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient identifier using the phone number, preserving the amount, app, and public visibility from the initial instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/addition/addition_02/e0f7ea307dbe057e/speech.wav", "wer": 0.0625, "wav_sha256": "60028c0dabdc545e0587b10a58b2aa130b09315f05cf167d82ab749e2c7a6f20", "seconds": 15.16, "spoken_utterances": [ "Send ah... one hundred dollars publicly on hmmm... Venmo. ", "I should also mention, send it to the person with this phone number five seven three, four five nine, nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Send a uh one hundred dollar publicly on Venmo' is followed by an addition specifying the recipient via phone number. The final audible goal combines the original action with the added detail, matching the reference instruction. The addition is clear and not contradicted, satisfying the addition type criterion.", "heard_transcript": "Send a uh one hundred dollar publicly on Venmo. I should also mention, send it to the person with this phone number five seven three four five nine nine seven six six." }, "candidate_id": "wav_10ef7eeac8156845", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The delivery is less fluid, with noticeable fillers ('uh', 'huh') and a slightly hesitant, fragmented structure. The phrase 'send a uh one hundred dollar publicly on huh Venmo' contains awkward pauses and self-correction that sound less spontaneous and more like a rehearsed or stilted attempt. The addition at the end is grammatically awkward ('I should also mention, send it to...') and lacks the natural flow of the first clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Send $100 publicly on CashApp to the person with this phone number 5734599766. Make that Venmo, not CashApp.", "long_pause_offsets": [ 79 ], "source_sha256": "09b463b3a605718de9d6dd1674f86ff1ceda10f21de7c001e31956cb52f9dd32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup corrects the app from CashApp to Venmo, but the original instruction specifies sending to the person with phone number 5734599766, while the spoken number is five seven three, four five nine, nine seven six six, which is 5734599766. The app is corrected, but the phone number is identical in both forms." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/revision/revision_03_app/9cb9885b73b5074e/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "9a9e4cf1443a6fc913dd9e78e0e3600a05d62d78c1db414eb682cfac4d6bf638", "seconds": 16.4, "spoken_utterances": [ "Send one hundred dollars publicly on um... CashApp to the person with this phone number ... five seven three, four five nine, nine seven six six. ", "Make that Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects it to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the amount, action, visibility, app, and recipient identifier.", "heard_transcript": "Send a hundred dollars publicly on um, Cash App to the person with this phone number, five seven three, four five nine, nine seven six six. Make that Venmo, not Cash App." }, "candidate_id": "wav_397da5ae06d5c131", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('um') before the app name. The correction ('Make that Venmo, not Cash App') is a smooth, self-contained revision with a believable shift in intonation and pacing, indicating a spontaneous realization. The prosody is consistent, and the pause before the correction feels like a genuine moment of reconsideration." } ] }, "is_default": false }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Send $100 privately on Venmo to the person with this phone number 5734599766. Wait, the payment should be publicly, not privately.", "long_pause_offsets": [ 78 ], "source_sha256": "e9e57ce248e4bb835912fda363bf28fda62dd5845ba86fc72bd6a3d3c0a3d402", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the privacy setting from 'privately' to 'publicly', while retaining the amount, app, recipient phone number, and action. The final goal matches the original instruction except for the corrected visibility." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/revision/revision_02_visibility/f2a1b0ae323bf26d/speech.wav", "wer": 0.0, "wav_sha256": "3c624d1072346ab82aba7bb4172258d8d319a48f730b86d366db1c56458ae351", "seconds": 17.44, "spoken_utterances": [ "Send one hundred dollars privately on Venmo ... to the person with this phone number uh... five seven three, four five nine, nine seven six six. ", "Wait, the payment should be um... publicly, not privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested a private payment but immediately corrected it to 'publicly, not privately' after a pause. This is a clear revision of the constraint. The final audible intent matches the original instruction with the corrected privacy setting, so the revision criterion is satisfied.", "heard_transcript": "Send one hundred dollars privately on Venmo to the person with this phone number, uh, five seven three, four five nine, nine seven six six. Wait, the payment should be um, publicly, not privately." }, "candidate_id": "wav_459b01f69a498b6f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear, fluent request, then pauses and revises the privacy setting. The 'Wait' is a natural, pragmatic cue for a change of mind. The correction is delivered with a slight shift in tone and a brief pause, suggesting a moment of recollection. The flow is coherent, though the initial delivery is slightly more rehearsed than audio_01." } ] }, "is_default": true }, { "variant_id": "revision_01_amount", "proposed_instruction": "Send $150 publicly on Venmo to the person with this phone number 5734599766. Sorry, I meant $100, not $150.", "long_pause_offsets": [ 77 ], "source_sha256": "94a106e70240f621ed5aa781c87175fd57d45fbee6d0a254538c929fbed81139", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from one hundred fifty to one hundred dollars, while retaining the app, privacy setting, recipient identifier, and phone number." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/revision/revision_01_amount/a89c47dbfdc58bdb/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "25bd19f24b7bf08e2399d1a88f5bca4de434f4fa0686e44ba582387310989170", "seconds": 19.2, "spoken_utterances": [ "Send one hundred fifty dollars publicly on let's see... Venmo to the person with this phone number uh, let me think... five seven three, four five nine, nine seven six six. ", "Sorry, I meant ... one hundred dollars, not one hundred fifty dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states $150 but immediately corrects to $100 after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier value. The final audible intent matches the original instruction's action, target, and phone number, with only the amount revised from 150 to 100. The correction is clear and unambiguous, satisfying the revision type requirement.", "heard_transcript": "Send one hundred fifty dollars publicly on let's see Venmo to the person with this phone number. Uh, let me think. Five seven three four five nine nine seven six six. Sorry, I meant one hundred dollars, not hundred fifty dollars." }, "candidate_id": "wav_159e04dd6291b319", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a hesitant, filler-laden delivery ('let's see', 'uh', 'let me think'), which is realistic for spontaneous speech. The correction ('Sorry, I meant $100 not $150') is a clear revision with a natural apology and intonation shift. However, the initial request is less fluent, and the number '150' is pronounced with a slight stumble, making the overall delivery feel slightly less confident than the top two." } ] }, "is_default": false }, { "variant_id": "revision_04_phone_number", "proposed_instruction": "Send $100 publicly on Venmo to the person with this phone number 5734599767. Change the phone number to 5734599766 instead of 5734599767.", "long_pause_offsets": [ 77 ], "source_sha256": "bc91cd84d9cd8384c411e15c30182f6149289f6c0a06e3f2ae1c60f4aacdb7e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the phone number from 5734599767 to 5734599766, which differs in the last digit." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/revision/revision_04_phone_number/c80e933cd4c7cec3/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "86a387cd9f26d76f88b5e0e69273b6234ecb191ef74bd199d54ef3b70a72afa0", "seconds": 25.4, "spoken_utterances": [ "Send um... one hundred dollars publicly on ah... Venmo to the person with this phone number five seven three, four five nine, nine seven six seven. ", "Change the phone number to um... five seven three, four five nine, nine seven six six instead of five seven three, four five nine, nine seven six seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the phone number as 5734599767, then immediately revises it to 5734599766. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to send $100 to the person with phone number 5734599766, which matches the original instruction. All other constraints (amount, app, visibility) are preserved.", "heard_transcript": "Send um, one hundred dollars publicly on uh Venmo to the person with this phone number five seven three four five nine nine seven six seven. Change the phone number to um, five seven three four five nine nine seven six six instead of five seven three four five nine nine seven six seven." }, "candidate_id": "wav_c687cac5598969d4", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with multiple fillers ('send um', 'on uh') and a stutter ('a Venmo'), indicating hesitation. The revision is a full sentence that sounds more like a post-hoc addition than a spontaneous correction. The change in phone number is stated as a new instruction rather than a self-repair, and the delivery lacks the natural flow of a real-time correction. The prosody is less consistent, and the overall impression is less like a natural thought process and more like a rehearsed or synthesized request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "653efc5e346080a8f70acd75899ae9a7e2f74b060ea2ae575a09e4e5efef558a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/cancellation/cancel_040/87e5cb07388ed68a/speech.wav", "wer": 0.0, "wav_sha256": "52419912032b766c98fd474b9c9f8937bb76e6877787f28fd34d03eee5e8239c", "seconds": 17.64, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six.", "Wait, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly states 'Wait, I don't want to proceed with that task. Please cancel it.' at the end, which is a direct and unambiguous cancellation of the previously stated task. The cancellation is not conditional or followed by a new instruction, and the original task is fully described before the cancellation is issued. Therefore, the final control intent is to cancel, matching the 'cancellation' type criterion.", "heard_transcript": "Send one hundred dollars publicly on Venmo to the person with this phone number five seven three four five nine nine seven six six. Wait, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "0b07d526b1b44c5ad858e42e4227eb1c8dd53a596dd944d171ee484a41a09c0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/ambiguous_halt/halt_017/8949775e5cbd391f/speech.wav", "wer": 0.0, "wav_sha256": "c0698853f61f63a113c51e56d40f6665cc7bbfe787b7e905ac308719401e4c0f", "seconds": 15.64, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and correct, and no correction or cancellation is made. The final control request is a pause, satisfying the type-specific criterion.", "heard_transcript": "Send one hundred dollars publicly on Venmo to the person with this phone number five seven three four five nine nine seven six six. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c62ae424670dd2061a564468ff9975682d30d7ea820e886fc155f7cc7f87e7d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_2/audio_natural/filler/original/68c616eec1db8e20/speech.wav", "wer": 0.0, "wav_sha256": "5fa07ff1ca42a00a60e2a8be50e1f8ec2f75ff80c5c50cf6a6cb7617997a1a49", "seconds": 12.4, "spoken_utterances": [ "Send ... one hundred dollars publicly on ... Venmo to the person with this phone number ... five seven three four five nine nine seven six six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio matches the original instruction exactly, with no hesitation, correction, or change in intent. The 'filler' type allows for natural speech including pauses or fillers, but the final task goal is clear and fully specified. No issues are present.", "heard_transcript": "Send $100 publicly on Venmo to the person with this phone number 5734599766." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_amount", "proposed_instruction": "Send $75—no, that should be $68 privately on Venmo to the person with this phone number 2114992604.", "long_pause_offsets": [], "source_sha256": "2c738f80e687daa1b6782b2940b8138edc03eb6489e78caa1097b6acf824a922", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces seventy-five with sixty-eight; the recipient, app, privacy setting, and phone number remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/false_start/false_start_01_amount/fb50f108297a2747/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2620d2a6c7ce548d10a610a9f40f3d23e434d003ea060c5fb1f171689b37806c", "seconds": 14.92, "spoken_utterances": [ "Send seventy-five dollars—um... no, that should be hmmm... sixty-eight dollars privately on Venmo to the person with this phone number ... two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'seventy five dollars' but immediately self-corrects to 'sixty eight dollars'. According to the 'false_start' type criterion, the final intent is the corrected value, and the original incorrect value is discarded. All other conditions (private, Venmo, recipient by phone number) are preserved. The final amount and recipient match the original instruction.", "heard_transcript": "Send seventy five dollars dollars um no that should be sixty eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four" }, "candidate_id": "wav_9e4f94cd33bb3b9f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent false-start repair. The speaker begins with 'send seventy-five dollars' and immediately self-corrects with 'no, that should be sixty-eight dollars' in a smooth, conversational tone. The correction is integrated into the same utterance without awkward pauses or stutters, and the prosody reflects genuine hesitation and correction. The delivery is continuous and credible, with no signs of artificiality or forced pacing." } ] }, "is_default": true }, { "variant_id": "false_start_02_visibility", "proposed_instruction": "Send $68 publicly—scratch that, make it privately on Venmo to the person with this phone number 2114992604.", "long_pause_offsets": [], "source_sha256": "0a196dba795cf7e7f90b5564ea40e488a949bdc25c96234e5b6809c32469df15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects it to 'privately'. The amount and recipient phone number are consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/false_start/false_start_02_visibility/bc27460fe536e986/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "7f3417663efd33c72b12e36c600cbecc960ffaa01b20a626d173506e9c28c1aa", "seconds": 14.04, "spoken_utterances": [ "Send ... sixty-eight dollars publicly—uh... scratch that, make it privately on Venmo to the person with this phone number ... two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send sixty-eight dollars publicly' but immediately self-corrects with 'scratch that, make it privately'. This is a clear false start where the final intent is the corrected version. The amount and recipient phone number are consistent with the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send sixty-eight dollars publicly uh, scratch that, make it privately on Venmo to the person with this phone number two one one four nine nine two six zero four." }, "candidate_id": "wav_ac841434f4dff3b5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'send sixty-eight dollars publicly' followed by 'scratch that, make it privately'. The repair is grammatically and semantically coherent, and the speaker's tone shifts naturally from the initial incorrect instruction to the correction. However, the transition is slightly more abrupt than in audio_02, and the 'scratch that' phrase, while plausible, feels a bit more like a script than a spontaneous thought. The prosody is natural, but the flow is less seamless than in the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Send $68 privately on CashApp—I should say Venmo to the person with this phone number 2114992604.", "long_pause_offsets": [], "source_sha256": "cc41efe2b24250d65bda616337d95726e3824e993525751f8f393cf866fff5d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions CashApp but immediately corrects to Venmo. The amount and recipient phone number are consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/false_start/false_start_03_app/03d8f8b383cedf8e/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "4edf7e9cf32c02567d12b7f623f9215bac11c38c940f87fbaef675b4bb261b6b", "seconds": 15.36, "spoken_utterances": [ "Send let me think... sixty-eight dollars privately on CashApp—ah... I should say hmmm... Venmo to the person with this phone number two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final corrected value must be applied. The amount and recipient phone number are unchanged and correctly stated. The final intent matches the original instruction with only the app corrected.", "heard_transcript": "send, let me think, sixty eight dollars privately on Cash App, I should say, Venmo to the person with this phone number, two one one four nine nine two six zero four." }, "candidate_id": "wav_067c0e6f3649712f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip features a filler ('let me think') and a false start ('send... Cash App'), followed by a correction to 'Venmo'. The repair is believable, but the speaker's tone is more hesitant and less confident than in the top two. The correction is delivered with a slight pause and a filler ('I should say'), which makes it feel like a later realization rather than a spontaneous self-correction. The prosody is natural, but the overall delivery is less fluid and slightly more disjointed due to the multiple pauses and filler words." } ] }, "is_default": false }, { "variant_id": "false_start_04_phone_number", "proposed_instruction": "Send $68 privately on Venmo to the person with this phone number 2114992605—I should have said 2114992604.", "long_pause_offsets": [], "source_sha256": "2e7870d19af83a88098485a1ad10d968fdef558fee045a2c34ca2128014b02cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in five, then corrects it to end in four. The final number matches the original instruction, and all other parameters (amount, app, privacy, action) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/false_start/false_start_04_phone_number/89e9c1333eae89c8/speech.wav", "wer": 0.07317073170731707, "wav_sha256": "d8ceabe541180a0aa06c0584328d9808f87957ba8ed3877bcf87f55dc48a3824", "seconds": 19.92, "spoken_utterances": [ "Send uh... sixty-eight dollars privately on um... Venmo to the person with this phone number two one one, four nine nine, two six zero five—uh... I should have said two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong phone number (2114992605) but immediately self-corrects to the correct one (2114992604). According to the 'false_start' type criterion, the final intent is the corrected version, and all other conditions (amount, privacy, app) remain unchanged. The correction is immediate and unambiguous, so the task is to execute the corrected instruction.", "heard_transcript": "Send a sixty-eight dollars privately on um Venmo to the person with this phone number two one one four nine nine two six zero five uh I should have said two one one four nine nine two six zero four" }, "candidate_id": "wav_58b2734b3b66c2d3", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision rather than a false start. The speaker initially says 'send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero five', then corrects the number to 'two one one four nine nine two six zero four'. While the correction is plausible, the initial number is stated with a slight stumble ('two six zero five uh'), and the correction is delivered with a flat, almost robotic tone. The prosody lacks the natural variation seen in the other clips, and the correction feels more like a mechanical update than a spontaneous thought. The overall delivery is less convincing in terms of naturalness and spontaneity." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Send $68 on Venmo to the person with this phone number 2114992604. Oh, I forgot to mention, send it privately.", "long_pause_offsets": [ 67 ], "source_sha256": "ad75f794538fea9dc016f1f497085a337be41b8474955461ac5c0c0a620fe48e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing privacy requirement; all other values including amount, app, recipient identifier, and phone number are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/addition/addition_01/c873b5515dd0715c/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f3a6708189232317277803727a488e0a2a65cb96f10e775c1cab1d9cf6716616", "seconds": 18.2, "spoken_utterances": [ "Send ... sixty-eight dollars on Venmo uh... to the person with this phone number um... two one one, four nine nine, two six zero four. ", "Oh, I forgot to mention, send it privately." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction to send $68 on Venmo to the specified phone number is followed by an addition that explicitly requests the transaction be sent privately. This addition modifies the original instruction by adding a constraint, and the final audible intent is to execute the original task with the added privacy condition. The addition is clear and does not contradict the original request, satisfying the 'addition' type criterion.", "heard_transcript": "Send sixty-eight dollars on Venmo uh to the person with this phone number um two one one four nine nine two six zero four. Oh, I forgot to mention, send it privately." }, "candidate_id": "wav_888e21af667fd9ac", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('uh', 'um') and a believable self-correction ('Oh, I forgot to mention...'). The addition of 'send it privately' is smoothly integrated with a change in intonation and pacing, sounding like a spontaneous afterthought rather than a rehearsed or robotic insertion. The prosody is consistent, and the number is pronounced clearly without distortion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send $68 privately on Venmo to the person with this phone number. I left out a detail: send it to the person with this phone number 2114992604.", "long_pause_offsets": [ 66 ], "source_sha256": "3dba5a5cebc64f3c5dc4711a0091d5f9511e58348ca5ddce38a0f137534abda0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing phone number, which matches the original. The amount and privacy setting are preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/addition/addition_02/4acea3eeef60eea9/speech.wav", "wer": 0.0, "wav_sha256": "873515712ee9d0867f56b9da20657203d53367dd53da77595485f492ba687f5f", "seconds": 20.2, "spoken_utterances": [ "Send um... sixty-eight dollars privately on Venmo ... to the person with this phone number. ", "I left out a detail: send it to the person with this phone number ... two one one, four nine nine, two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, missing the phone number. After a pause and self-correction, the speaker adds the missing phone number. The addition type requires combining the initial instruction with the added information, which results in a complete and coherent task matching the original instruction. The final intent is clear and executable.", "heard_transcript": "Send um sixty eight dollars privately on Venmo to the person with this phone number. I left out a detail. Send it to the person with this phone number two one one four nine nine two six zero four." }, "candidate_id": "wav_86c17689a6117be4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail, the delivery is less fluid. The initial request is interrupted by a long, unnatural pause ('Send um...'), and the phrase 'I left out a detail' sounds slightly rehearsed or formulaic. The rephrased instruction ('Send it to the person with this phone number...') is delivered with a flat, almost robotic intonation, lacking the natural variation of a spontaneous correction. The overall prosody is less dynamic, making it sound less like a real-time thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_amount", "proposed_instruction": "Send $75 privately on Venmo to the person with this phone number 2114992604. Sorry, I meant $68, not $75.", "long_pause_offsets": [ 77 ], "source_sha256": "ec75aae00517c04772ad0ecf5564243022a4d38a49195b711c8c3db13afcfecc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the amount from seventy-five to sixty-eight, while retaining the app, privacy setting, recipient phone number, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/revision/revision_01_amount/e34b58b0c87330ee/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "4df0a41eefcbb2cb7298064ab72587730b21bb4cbfd14dbf04351f144d3bbfe5", "seconds": 23.84, "spoken_utterances": [ "Send seventy-five dollars privately on ah... Venmo let me think... to the person with this phone number uh... two one one, four nine nine, two six zero four. ", "Sorry, I meant sixty-eight dollars, not seventy-five dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states $75 but immediately corrects to $68 after a pause, which is a clear revision. The final intent matches the original instruction exactly, including the app (Venmo), amount ($68), privacy setting (privately), and recipient (by phone number). The revision type criterion is satisfied.", "heard_transcript": "Send seventy five dollars privately on uh Venmo, let me think, to the person with this phone number uh two one one four nine nine two six zero four. Sorry, I meant sixty eight dollars not seventy five dollars." }, "candidate_id": "wav_d056b069d5032577", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, inserts a believable 'let me think' filler, and then self-corrects the amount from $75 to $68 with a smooth, spontaneous-sounding 'sorry, I meant...'. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_04_phone_number", "proposed_instruction": "Send $68 privately on Venmo to the person with this phone number 2114992605. Make that 2114992604 instead of 2114992605.", "long_pause_offsets": [ 77 ], "source_sha256": "d3aa719203ce4ae75fafece3b47cb77e1b5381bd0f015e24136354fcc6794cb9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the phone number from ending in five to ending in four, matching the original instruction. The amount, app, privacy, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/revision/revision_04_phone_number/849e80e5bb818cb0/speech.wav", "wer": 0.08, "wav_sha256": "1d8ec32c1c1a5ffaccac6f06d4fcdedc079cc884506c11bbd3903b55248c05ae", "seconds": 27.28, "spoken_utterances": [ "Send sixty-eight dollars privately on Venmo hmmm... to the person with this phone number ... two one one, four nine nine, two six zero five. ", "Make that um... two one one, four nine nine, two six zero four instead of two one one, four nine nine, two six zero five." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the phone number as 2114992605, then immediately corrects it to 2114992604. This is a revision where the final value replaces the earlier one. The rest of the instruction, including the amount, app, privacy setting, and action, remains consistent. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Send sixty eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero five. Make that um uh two one one four nine nine two six zero four instead of two one one four nine nine two six zero five." }, "candidate_id": "wav_697f0b3320e7587e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic with a clear revision. The speaker first states the correct amount ($68) and a phone number, then pauses with 'make that um...' before correcting the number. The correction is plausible, and the 'um' filler is natural. However, the initial request is slightly less fluid than audio_01, and the revision is a bit more abrupt, making it feel slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_visibility", "proposed_instruction": "Send $68 publicly on Venmo to the person with this phone number 2114992604. Wait, send it privately, not publicly.", "long_pause_offsets": [ 76 ], "source_sha256": "50da7a2c471b5494921e6418ded2ff27d1839a01b8bbc3c3845012d317da0986", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the visibility from 'publicly' to 'privately', while retaining the amount, app, recipient phone number, and action. The original instruction's 'privately' is replaced by the correction, and all other values match." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/revision/revision_02_visibility/43339967943163e7/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "2c133aa3d7e6b5c4eb184d4a92d67761737710817e95d65f314e924b472b89e8", "seconds": 19.52, "spoken_utterances": [ "Send sixty-eight dollars publicly on Venmo ... to the person with this phone number uh, let's see... two one one, four nine nine, two six zero four. ", "Wait, send it privately, not publicly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'publicly' but immediately corrects to 'privately' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to send the money privately, which matches the original instruction's intent except for the privacy setting. The value and recipient remain unchanged, and the correction is clear and unambiguous.", "heard_transcript": "Send sixty eight dollars publicly on Venmo to the person with this phone number, uh, let's see, two one one, four nine nine, two six zero four. Wait, send it privately, not publicly." }, "candidate_id": "wav_a9d8382c5da2878e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip shows a believable revision with a 'wait' correction from 'publicly' to 'privately'. The 'let me see' filler is natural. However, the initial request is delivered in a slightly flat, read-aloud tone, and the correction, while plausible, feels a bit more like a task instruction than a spontaneous thought. The prosody is less varied than in the top two." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Send $68 privately on CashApp to the person with this phone number 2114992604. Change the app to Venmo, not CashApp.", "long_pause_offsets": [ 79 ], "source_sha256": "35bc59a72975ea54b7f5495b7deaa44adadbca507478a84828820ea73e723604", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the amount, privacy setting, and recipient phone number." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/revision/revision_03_app/7f8c99fedb0c1674/speech.wav", "wer": 0.14705882352941177, "wav_sha256": "98db0acc680e5bbe7b0eaea155e0f55cfd5eec85bfa40c793fe89534c72b1839", "seconds": 20.96, "spoken_utterances": [ "Send sixty-eight dollars privately on CashApp um, let's see... to the person with this phone number ... two one one, four nine nine, two six zero four. ", "Change the app to ... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests sending $68 on Cash App, but then corrects the app to Venmo. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent is to send the money on Venmo, which matches the original instruction's action, amount, privacy setting, and recipient. The only change is the app, which is the subject of the revision. Therefore, the final intent is correctly executed.", "heard_transcript": "Send sixty dollars privately on Cash App. Um, let's see. To the person with this phone number, two one one, four nine nine, two six zero four. Change the app to Venmo, not Cash App." }, "candidate_id": "wav_56b1008efc4ebe65", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision but feels less realistic due to its structure. The speaker first makes a request on Cash App, then corrects the app to Venmo. However, the initial request is delivered in a very flat, monotonous tone, and the correction is made as a separate, almost robotic statement ('change the app to...') rather than a natural continuation. The prosody lacks the variation of a real person speaking, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_033", "hide_variant_original": true, "source_sha256": "b763cd235f66dad164df022079ff1b00f54fdc9d3a30cb619b967ea2fa55216c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/cancellation/cancel_033/47b053ffcbf4157e/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "77b2a9c833c5b88ac97365dc3db418d2f2837092309a200223b4460309a23ab5", "seconds": 15.96, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four.", "Hang on, that request is no longer needed. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker stating 'Cancel it' after acknowledging the initial request, which constitutes a direct cancellation of the task. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four. Hang on, that request is no longer needed. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "e5e76b1e922f95a1eb5721fc6043c5047311d876b00874e38cf7050ef859ef3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/ambiguous_halt/halt_037/f8c5536a5863d41f/speech.wav", "wer": 0.1, "wav_sha256": "cddd75c1bddd19053e14783808b302c7bd73e11724c5f3a8a6a5e7846c7fd7c1", "seconds": 14.16, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the final interjection 'Pause for a moment,' which is a clear request to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel. The initial instruction is fully audible and complete, but the final spoken control is a pause request.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "37a8675_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ec0741a1d9cc5a16338098ee7e9ee75a3f80f407f1bf4e8e089056b4012c34bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/37a8675_3/audio_natural/filler/original/d930e7c8fefaa0d2/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "67cbda7c3f9f78d3fa856e8f9a07ed6d48e3a38b9a4eb81fe5ee097178b9b6a9", "seconds": 10.64, "spoken_utterances": [ "Send hmmm... sixty-eight dollars privately on Venmo ah... to the person with this phone number two one one four nine nine two six zero four." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'Send uh' but then clearly states the full correct instruction. The hesitation and filler sound are ignored as per the 'filler' type criterion, which allows for such interruptions as long as the final task meaning matches the original instruction. The amount, action, privacy setting, app, and recipient identifier are all correctly stated and match the reference.", "heard_transcript": "Send sixty-eight dollars privately on Venmo to the person with this phone number two one one four nine nine two six zero four." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_invitation_app", "proposed_instruction": "I got some Venmo—I should say Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fb03a6be220edbafccc9ec7a502d0ee06dfe18d92e984dc33ebe9f42486b77f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Venmo' to 'Splitwise' within the same utterance, preserving the rest of the instruction including the app, message type, condition, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_01_invitation_app/73dcea0d09bf08b0/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "90d288739b2800c1d40be92f9f5b026993bfe9d37faea6211b08d30edae78903", "seconds": 16.28, "spoken_utterances": [ "I got some Venmo—uh... I should say Splitwise group invitations ... over phone text messages yesterday. ... If their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise'. This is a false start, and the final, corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "I got some Venmo uh, I should say Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_5dc224bfb410b532", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers the request with a natural, spontaneous flow. The initial false start 'Venmo uh' is immediately self-corrected to 'Splitwise', which sounds like a genuine moment of recall. The rest of the sentence is spoken at a steady, conversational pace with no awkward pauses or stutters. The prosody is consistent and the correction feels integrated, not forced." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages today—no, I mean yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "a1f988de54a795f55bbc2004f611461fb946908efd5c6b8a4b812ec53c54f532", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'today' to 'yesterday' within the same utterance. The final time reference matches the original instruction, and all other conditions and actions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_03_time_reference/a1d2bc44f0390638/speech.wav", "wer": 0.03225806451612903, "wav_sha256": "8061675acba28bad7f8f780021d20b87d596172d23686ddf33a9641aea94701d", "seconds": 15.16, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages ... today—um... no, I mean yesterday. If their number is in my phone contact book, accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'today' but immediately self-corrects to 'yesterday'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction, including the conditional logic and action, remains unchanged and matches the original instruction. Therefore, the final audible intent is clear and executable.", "heard_transcript": "I got some Splitwise group invitations over phone text messages today. No, I mean yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_8ce6792b2780bc63", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start ('today um') followed by a self-correction ('no, I mean yesterday'). The correction is immediate and the intonation shifts naturally. The speaker maintains a consistent pace and tone after the correction, and the overall delivery sounds like a person recalling a detail in real time." } ] }, "is_default": false }, { "variant_id": "false_start_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone—I meant my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "5e955359a5f8dc10ba9e69cb577c8212a414da66fa05245cf29005b45cf12cc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'phone' to 'phone contact book' within the same utterance, clarifying that the intended reference is the contact book. The condition and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_04_contact_reference/25c816e33f0e5523/speech.wav", "wer": 0.0, "wav_sha256": "6afc4d926a30777bfd3f0b5d85e528a9bd3d88d6321476e03015034c3f085546", "seconds": 15.36, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone text messages uh, let's see... yesterday. If their number is in my phone—um... I meant my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'in my phone' but immediately self-corrects to 'I meant my phone contact book'. According to the 'false_start' type criterion, the final corrected version should be applied. The final intent matches the original instruction with the corrected target ('phone contact book' instead of 'phone').", "heard_transcript": "I got some Splitwise group invitations um over phone text messages uh let's see yesterday. If their number is in my phone um I meant my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_e23b6eb082b662c1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a filler ('um') and a pause, then continues with 'uh let's see', indicating hesitation. The correction 'I meant my phone contact book' is a plausible self-repair. The flow is natural, though the multiple fillers and slight stuttering make it less fluent than audio_01 or audio_04, but still realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_message_channel", "proposed_instruction": "I got some Splitwise group invitations over email—or rather, phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "0a52d1ad2a90b57bb3bfbee94660abd122180afffa4507e5149fc5524de20925", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'email' to 'phone text messages' within the same utterance, and the rest of the instruction, including the conditional logic and action, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_02_message_channel/1171cbc8d18b9948/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "7a769a9e4ac0fae41abacd853f9e0b543b3a320e357525cc0df42baa775d50e5", "seconds": 14.8, "spoken_utterances": [ "I got some Splitwise group invitations over ... email—ah... or rather, phone text messages yesterday. uh... If their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately self-corrects to 'phone text messages'. The final instruction matches the original reference exactly, with the correction applied. The 'false_start' type requires the final intent to be the corrected version, which it is. No other constraints are altered.", "heard_transcript": "I got some Splitwise group invitations over email, uh, or rather phone text messages yesterday. Uh, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_6f8cb5f2da1e5529", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker has a clear false start ('email ah') and a self-correction ('or rather phone text messages'). The correction is natural, but the initial error is more abrupt and the 'ah' filler sounds slightly more forced than in other clips. The overall delivery is still believable, but the correction feels a bit less integrated than in audio_01 or audio_04." } ] }, "is_default": false }, { "variant_id": "false_start_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, decline it—sorry, I mean accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fdc333085859af4678e44fbe7e13f2dcf37a9dfb6e12c9ad02f654a25008533d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately corrects with 'I mean accept it'. The final intended action is to accept invitations when the number is in the contact book, otherwise delete the messages. All other conditions and structure match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_05_matched_contact_action/ba1c1e880d28854b/speech.wav", "wer": 0.0, "wav_sha256": "739702a8a0c20a0a7b20b78bd5e2a3a99b42399cb54cd88d77c7ac93a10a6703", "seconds": 14.96, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages let me think... yesterday. If their number is in my phone contact book, decline it—um... sorry, I mean accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately self-corrects to 'accept it' after a pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value ('accept it') overrides the initial mistaken value ('decline it'), and the rest of the instruction remains unchanged. The final intent matches the original instruction.", "heard_transcript": "I got some Splitwise group invitations over phone text messages, let me think, yesterday. If their number is in my phone contact book, decline it, um, sorry, I mean accept it, otherwise delete those messages." }, "candidate_id": "wav_433fa0bfb62fc0db", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip contains a false start ('let me think') and a self-correction ('sorry I mean accept it'). The correction is clear, but the speaker says 'decline it um' before correcting, which is a bit awkward. The 'um' after 'decline it' sounds like a filler, but the correction feels slightly rushed and less natural than in other clips." } ] }, "is_default": false }, { "variant_id": "false_start_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise archive those messages—scratch that, make it delete those messages.", "long_pause_offsets": [], "source_sha256": "ae7c5bc397a8ace23dd3257c8d8e400c57532271b938f38af7e68377258cefdf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests archiving the messages when the number is not in the contact book, but immediately corrects this to deleting them. The final instruction matches the original requirement to delete non-contact messages." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/false_start/false_start_06_unmatched_contact_action/024ded3aeede3166/speech.wav", "wer": 0.125, "wav_sha256": "2f37a43c4661c0d58c26bba70979db70f31f9dd99d90a86399cd36c0597f0a06", "seconds": 17.88, "spoken_utterances": [ "I got some Splitwise group invitations um, let me think... over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise let's see... archive those messages—um... scratch that, make it delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but immediately self-corrects with 'scratch that, make it delete those messages'. According to the 'false_start' type criterion, the final corrected instruction must be executed. The final intent is to delete the messages, which matches the original instruction's conditional logic (accept if in contacts, otherwise delete). The correction is immediate and unambiguous, so the task is executed as corrected.", "heard_transcript": "I got some Splitwise group invitations um let me think over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise let's see. Archive those messages. Um scratch that, make it delete those messages." }, "candidate_id": "wav_423872041ebbe384", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This speaker has multiple fillers ('um', 'let me think', 'let's see') and a revision ('archive those messages... scratch that, make it delete'). While the revision is a plausible change-of-mind, the multiple fillers and the abrupt 'scratch that' make the delivery feel less smooth and more hesitant than the others. The overall flow is less natural, possibly due to the multiple interruptions." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday.. Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [ 76 ], "source_sha256": "c66e60a6f1eea8e62f925f3af2e30c500637eb7302fd6fd8f64c940b9c5158e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the conditional logic that was missing in the initial utterance. The final goal combines the initial context (Splitwise group invitations via text) with the added condition, preserving the original intent of handling messages based on contact book presence." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/addition/addition_01/f22096089df02c59/speech.wav", "wer": 0.029411764705882353, "wav_sha256": "84033607673a97a5d23f9ff39490de74e51ea42e516c78dbed4e620f06955029", "seconds": 16.4, "spoken_utterances": [ "I got some uh... Splitwise group invitations over phone text messages let's see... yesterday.. ", "Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task and then adds a condition ('if their number is in my phone contact book, accept it, otherwise delete those messages') after a pause. This is a clear addition that refines the original instruction. The final intent is to accept invitations from contacts and delete those from non-contacts, which is a valid execution of the combined request.", "heard_transcript": "I got some uh Splitwise group invitations over phone text messages. Let's see. Yesterday. Oh, I forgot to mention, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_3ecaca228b6b52fc", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_invitation_app", "proposed_instruction": "I got some Venmo group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Sorry, those are Splitwise group invitations, not Venmo invitations.", "long_pause_offsets": [ 160 ], "source_sha256": "ed6fc912c7e94b9444028979860edbfe5634ed40b3a3756a48c42bfd88ca164b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Venmo to Splitwise, which matches the original instruction. The rest of the logic, including the condition based on the phone contact book and the actions (accept or delete), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_03_invitation_app/505bc0fd7d73c072/speech.wav", "wer": 0.1282051282051282, "wav_sha256": "d3aea5b621b643850599c70f428cca646fd296abd1f32393959a8c52b82365f9", "seconds": 23.76, "spoken_utterances": [ "I got some ... Venmo group invitations over phone text messages um, let's see... yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "Sorry, those are Splitwise group invitations, not uh... Venmo invitations." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the app as 'Venmo' but immediately corrects it to 'Splitwise' after a pause. According to the 'revision' type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to handle Splitwise invitations with the same conditional logic (accept if in contacts, delete otherwise). All other conditions from the original instruction are preserved. The correction is clear and unambiguous, satisfying the revision type requirement.", "heard_transcript": "I got some Venmo group invitations over phone text messages, um, let's see, yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Sorry, those are Splitwise group invitations, not uh, Venmo invitations." }, "candidate_id": "wav_bdc355a470268b7a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'I got some Venmo group invitations...' and uses a genuine 'um, let's see' pause before continuing. The self-correction from 'Venmo' to 'Splitwise' is a plausible false start with immediate, smooth repair. The intonation and rhythm are consistent with someone recalling and adjusting a task on the fly, and the final sentence flows naturally without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my Gmail contacts, accept it, otherwise delete those messages. Check my phone contact book for those numbers, not my Gmail contacts.", "long_pause_offsets": [ 160 ], "source_sha256": "c1227cf95e25087c5a4a38811eaccc3146829263af83a5691dfdc600df0445ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the contact source from Gmail to phone, reversing the initial condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_04_contact_reference/52876bb65c541c91/speech.wav", "wer": 0.0, "wav_sha256": "0c0b6c0f962a1edae9501c02f4557f6aaf733eedf5e6211a4c626711059ecca2", "seconds": 21.56, "spoken_utterances": [ "I got some Splitwise group invitations ... over phone text messages let's see... yesterday. If their number is in my Gmail contacts, accept it, um... otherwise delete those messages. ", "Check my phone contact book for those numbers, not my Gmail contacts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the condition as 'Gmail contacts' but immediately corrects it to 'phone contact book' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to use the phone contact book, which matches the original instruction. All other conditions are preserved.", "heard_transcript": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my Gmail contacts, accept it. um, otherwise delete those messages. Check my phone contact book for those numbers, not my Gmail contacts." }, "candidate_id": "wav_a0b7bfeb154f3f06", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear, slightly hesitant delivery and uses 'let's see' as a natural thought transition. The initial error ('Gmail contacts') is followed by a believable self-correction ('check my phone contact book...'). The correction is phrased as a realization rather than a mechanical fix, and the prosody remains consistent throughout, suggesting genuine recollection." } ] }, "is_default": false }, { "variant_id": "revision_01_message_channel", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. To correct what I said, the invitations came over phone text messages, not phone voice messages.", "long_pause_offsets": [ 165 ], "source_sha256": "10347c1b854caf2ad69bf4d1b255e9e5bc5ce3b0aa795d0774ac16f7d9c23296", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type from 'voice messages' to 'text messages', which is the only change. The rest of the logic and conditions remain identical." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_01_message_channel/655e3adbe32166b9/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "b7fec52003e568ba79168d700dfc94559d2eff21e053a5cfce553df74f023421", "seconds": 20.48, "spoken_utterances": [ "I got some um... Splitwise group invitations over phone voice messages ... yesterday. If their number is in my phone contact book, accept it, ... otherwise delete those messages. ", "To correct what I said, the invitations came over phone text messages, not phone voice messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the invitations were received via 'phone voice messages' but immediately corrects this to 'phone text messages' after a pause. According to the 'revision' type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to process invitations received via text messages, which is a revision of the initial medium. All other conditions (accept if in contacts, delete otherwise) remain unchanged and are correctly applied.", "heard_transcript": "I got some um Splitwise group invitations over phone voice messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. To correct what I said, the invitations came over phone text messages, not phone voice messages." }, "candidate_id": "wav_7fc7327866555cad", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a filler 'um' and a slightly hesitant delivery. The initial error ('phone voice messages') is corrected with a clear 'to correct what I said' phrase, which is a pragmatic and believable way to revise a statement. The correction is integrated smoothly, and the overall delivery sounds like a person catching a mistake mid-sentence." } ] }, "is_default": false }, { "variant_id": "revision_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, decline it, otherwise delete those messages. For an invitation from a number in my phone contact book, accept it rather than decline it.", "long_pause_offsets": [ 165 ], "source_sha256": "8065c6d99f3a48355b85687ea9b8658e6d28ecfeeb132f5fad839852208bffc8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the action for contacts from 'decline' to 'accept', reversing the original condition's outcome." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_05_matched_contact_action/1e5120760bd0ded7/speech.wav", "wer": 0.0, "wav_sha256": "56ba36405841449afdbee4da07bc8f856591568b27fc3b3163f155f0690b8106", "seconds": 22.52, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages um... yesterday. If their number is in my phone contact book, let me think... decline it, otherwise delete those messages. ", "For an invitation from a number in my phone contact book, accept it rather than decline it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a rule to 'decline it' if the number is in the contact book, but after a pause, revises this to 'accept it rather than decline it'. According to the 'revision' type criterion, the final instruction after the pause overrides the initial one. The final intent is to accept invitations from contacts, which is a clear revision of the original instruction. All other conditions are preserved, and the revision is unambiguous.", "heard_transcript": "I got some Splitwise group invitations over phone text messages um yesterday. If their number is in my phone contact book, let me think, decline it, otherwise delete those messages. For an invitation from a number in my phone contact book, accept it rather than decline it." }, "candidate_id": "wav_1b12db86fe50b40b", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear, slightly hesitant delivery and uses 'let me think' as a natural pause. However, the revision ('for an invitation from a number... accept it rather than decline it') is a complete rephrasing of the initial instruction, not a simple correction. The shift in phrasing and the lack of a direct reference back to the first sentence make the revision feel less spontaneous and more like a separate thought." } ] }, "is_default": false }, { "variant_id": "revision_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise keep those messages. If the number isn't in my phone contact book, delete those messages; don't keep them.", "long_pause_offsets": [ 162 ], "source_sha256": "63a0e0edd1d00fae2786850daf994046e6e412adc9d538bb8217859fed45a91f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup reverses the action for the 'not in contact book' case: the initial instruction says to 'delete' when not in contacts, but the followup says to 'keep' them, then later contradicts itself by saying to 'delete' them and 'don't keep' them. The final instruction is to delete messages when the number is not in the contact book, which contradicts the initial instruction's 'keep' action for that case." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/revision/revision_06_unmatched_contact_action/cf5ac0e3a1b4a86d/speech.wav", "wer": 0.0, "wav_sha256": "eb78dd814da80d61370a9b6b8f0a057bf829465a9b2c536d85c0dece60805fa2", "seconds": 21.32, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, ... accept it, otherwise keep those messages. ", "If the number isn't in my phone contact book, let me think... delete those messages; don't keep them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'keep those messages' as the action for numbers not in the contact book, but then revises this to 'Delete those messages, don't keep them' after a pause and self-correction. According to the revision type criterion, the final instruction after the correction must be applied. The final intent is to delete messages from non-contacts, which is a revision of the original instruction. All other constraints remain consistent, and the revision is clear and final.", "heard_transcript": "I got some Splitwise group invitations over phone text messages yesterday. If their number is in my phone contact book, accept it, otherwise keep those messages. If the number isn't in my phone contact book, let me think. Delete those messages, don't keep them." }, "candidate_id": "wav_1bfe423c1aadf235", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker starts with a clear, slightly hesitant delivery but makes a significant revision ('delete those messages, don't keep them') that contradicts the initial instruction ('keep those messages'). The correction is abrupt and lacks a natural transition or self-correction phrase. The prosody also becomes more rushed and less natural after the revision, suggesting a less spontaneous delivery." } ] }, "is_default": false }, { "variant_id": "revision_02_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages last week. If their number is in my phone contact book, accept it, otherwise delete those messages. I meant the invitations from yesterday, not last week.", "long_pause_offsets": [ 164 ], "source_sha256": "1b6f91d02a58cb8333a258d93c73e0215921b8d060ebb60b82180a5d4731b1d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the time reference from 'last week' to 'yesterday', while preserving all other conditions including the app, message type, and conditional logic. The correction is explicit and unambiguous." }, "audio": { "status": "pending", "spoken_utterances": [ "I got some Splitwise group invitations uh... over phone text messages last week. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "I meant the invitations from um, let me think... yesterday, not last week." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "85cad0d45022cdb8c9817a72a0b75dfe7ad720f64b858e0b28219c656d83a8c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/cancellation/cancel_031/c0d8e9c12dc52586/speech.wav", "wer": 0.18604651162790697, "wav_sha256": "32821c9e481dec05fadf4ed89ca916a1502fe315399f52355cb402cef49b3920", "seconds": 21.56, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages.", "Hang on, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation command: 'Hang on, there's no need to do that anymore. Cancel it.' This constitutes a final, unambiguous cancellation of the entire task. The initial instruction is irrelevant once the cancellation is stated.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages. Hang on, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "c3c88df45260532752ddd986bb7b9ef12e874393d89d7a562d855faa6c32365b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/ambiguous_halt/halt_008/01b4d6df3529685f/speech.wav", "wer": 0.22857142857142856, "wav_sha256": "e1321f170707a7e779c06fcbb96910537096361dcc98d5f7156cfc8bf58fb451", "seconds": 19.24, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait a second,' which is a clear request to pause or hold. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction is not canceled or revised, and the speaker does not provide a new instruction or cancel the task. Therefore, the final control intent is to pause.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f553ce460479f681d333d37bba3c492dd97070752ac9a470f90cb144fe2ef0fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_1/audio_natural/filler/original/a1a9aba583fc51b9/speech.wav", "wer": 0.25, "wav_sha256": "574d618966f8a412eefed6ba3fc86d981f7885d002bd4dedd3a89020c61d4b20", "seconds": 15.96, "spoken_utterances": [ "I got some uh, let's see... Splitwise group invitations um... over phone text messages yesterday. If their number is in my phone contact book, let's see... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') and a pause but does not change the core task. The final instruction matches the original request. The hesitation and self-repair do not alter the goal, so the 'filler' type criterion is satisfied.", "heard_transcript": "I got some uh let's see Splitwise group invitations um over phone text messages yesterday. If their number is in my phone contact book, let's see, accept it, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages yesterday—no, I mean the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "78aaa23492308a9c3b9d195b1c83b445c278763eb3d57f956c4c3da2dc5aa7e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'yesterday' to 'the day before yesterday' within the same utterance. The final time reference matches the original instruction, and all other conditions (app, message type, recipient condition, and action) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_03_time_reference/3d7fa312b9b152c4/speech.wav", "wer": 0.0, "wav_sha256": "938a02148faaf63418bca7d2701b4f4d3ae8cc570cdc1e79522812587c694d71", "seconds": 15.08, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages yesterday—um... no, I mean ... the day before yesterday. If their number is in my phone contact book, uh... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yesterday' but immediately self-corrects to 'the day before yesterday'. This is a false start, and the final, corrected time reference matches the original instruction. All other conditions (Splitwise group invitations, phone voice messages, accept if number is in contact book, delete otherwise) are preserved. The final intent is clear and executable.", "heard_transcript": "I got some splitwise group invitations over phone voice messages yesterday. Um, no, I mean the day before yesterday. If their number is in my phone contact book, uh, accept it, otherwise delete those messages." }, "candidate_id": "wav_a04167e0574414b1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker initially says 'yesterday', then self-corrects to 'the day before yesterday' with a clear, spontaneous 'no, I mean' transition. The prosody is consistent, and the correction feels like a genuine moment of recall rather than a rehearsed edit. The 'uh' pause before 'accept it' is also well-placed and natural." } ] }, "is_default": true }, { "variant_id": "false_start_02_message_channel", "proposed_instruction": "I got some Splitwise group invitations over text messages—or rather, phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fb8df535024f4918744a6b9e4504a578088d22bbb525cd8194e59130f8fc9647", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'text messages' to 'phone voice messages' within the same utterance, preserving the rest of the instruction including the condition and action. The final goal is unchanged except for the correct message type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_02_message_channel/88836b04c9448130/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "6962f75f679d9bff74161006aac541f618ccac414ab6b839e899e287e562358f", "seconds": 19.36, "spoken_utterances": [ "I got some ... Splitwise group invitations over text messages—uh... or rather, phone voice messages the day before yesterday. If their number is in my phone contact book, hmmm... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'text messages' but immediately self-corrects to 'phone voice messages'. According to the 'false_start' type criterion, the final intent is determined by the immediate correction. The corrected version matches the original instruction's intent, with only a minor paraphrase ('phone voice messages' vs 'phone voice messages') that does not change the meaning. All other conditions are preserved.", "heard_transcript": "I got some Splitwise group invitations over text messages uh or rather phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_820dbc67c4a3f063", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a credible false start with 'text messages' corrected to 'phone voice messages'. The speaker uses 'uh' and 'or rather' to transition, which sounds natural. However, the correction is slightly less fluid than audio_02, and the overall pacing is a bit more hesitant, making it feel slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_01_invitation_app", "proposed_instruction": "I got some Venmo—I should say Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "c8a3d40f0c3a092fd5420c5afde4146e4d9f062536272b1a473863777bf0776b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Venmo' to 'Splitwise' within the same utterance. The rest of the instruction, including the condition about the phone contact book and the actions (accept or delete), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_01_invitation_app/1f035eae0835e703/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "c97d3fd7cc37610e887daf74595ca3c2f03220f203235013dddac6c0f3a4c3c8", "seconds": 18.72, "spoken_utterances": [ "I got some Venmo—uh... I should say ... Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, hmmm... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise'. The final instruction matches the original reference exactly after the correction. The false start is resolved with a clear, immediate correction, so the final intent is to execute the Splitwise task as specified.", "heard_transcript": "I got some Venmo uh I should say Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_2ad7e12234e8816a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio begins with a clear false start ('Venmo' corrected to 'Splitwise'), which is well-executed. The speaker then continues with a natural-sounding 'if their number is in my phone contact book'. The 'um' pause before 'accept it' is believable. However, the initial error is more jarring and the correction feels slightly more forced compared to audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone—sorry, I meant my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "bef93509864ae187d0fc620a4f556704066ef905f941df2aa27ad17ffa44d597", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'my phone' to 'my phone contact book' within the same utterance. The final condition uses the correct reference, and all other elements including the action, condition, and context remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_04_contact_reference/53449639d25acb6f/speech.wav", "wer": 0.15, "wav_sha256": "274d7741e94806e6133c2ebdd799ffc4bf8db72dc8a7093ccb567739e52438b1", "seconds": 22.96, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages um, let me think... the day before yesterday. uh... If their number is in my phone—hmmm... sorry, I meant my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'in my phone' but immediately self-corrects to 'in my phone contact book'. According to the 'false_start' type criterion, the final corrected version should be applied, and the original incorrect value discarded. The final intent matches the original instruction's goal, with only a minor clarification of the condition. No other constraints are altered.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages. Um, let me think. The day before yesterday. Uh, if their number is in my phone. Hmm. Sorry, I meant my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_8589510d1d932191", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a self-correction from 'my phone' to 'my phone contact book', which is a plausible revision. The 'um' and 'let me think' fillers are natural, but the correction is made mid-sentence and feels a bit abrupt. The overall flow is less smooth than the top-ranked clips, and the speaker seems to be formulating the request more slowly." } ] }, "is_default": false }, { "variant_id": "false_start_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, decline it—no, make that accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "8189e0adbf59207af3fc3803d74c0aec2e9d2a880a0cd6a58e8dd9fb6426405e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately corrects to 'accept it' within the same utterance. The final action is to accept invitations when the number is in the contact book, otherwise delete the messages. All other conditions and context match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_05_matched_contact_action/7e3507aa64ce79af/speech.wav", "wer": 0.10256410256410256, "wav_sha256": "ab19ad35f14017905eb7979c9fab88478416ca827005db0532b1e50604edc1db", "seconds": 20.68, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages uh, let me think... the day before yesterday. If their number is in my phone contact book, decline it—uh... no, make that accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately self-corrects with 'no, make that accept it'. According to the 'false_start' type criterion, the final corrected instruction must be applied. The final intent is to accept invitations for contacts in the book and delete for others, matching the original instruction's structure and intent. The correction is immediate and unambiguous, so the task is executed as corrected.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages uh let me think the day before yesterday. If their number is in my phone contact book, decline it uh no, make that accept it. Otherwise delete those messages." }, "candidate_id": "wav_4a52fee01c001f0a", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This audio contains a revision where the speaker first says 'decline it' and then corrects to 'accept it'. While this is a valid pragmatic repair, the correction is made after a pause and sounds slightly less spontaneous. The 'uh' and 'no, make that' transition is natural, but the overall delivery feels more like a rehearsed correction than a genuine moment of reconsideration." } ] }, "is_default": false }, { "variant_id": "false_start_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise archive those messages—scratch that, make it delete those messages.", "long_pause_offsets": [], "source_sha256": "f3b87b8b10999de907a42526ee1e0db19175c490498c22b74757a779eec5a663", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but immediately corrects it to 'delete those messages'. The final instruction matches the original instruction's action for the 'otherwise' case." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/false_start/false_start_06_unmatched_contact_action/19410a228576eb74/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "127ed9e6ac2a3552177260f6981aab53f92e29a8c91a23e43822551a33d2b791", "seconds": 20.24, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages ah... the day before yesterday. If their number is in my phone contact book, accept it, ... otherwise archive those messages—uh... scratch that, make it delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested to 'archive those messages' as the default action for numbers not in the contact book, but immediately self-corrected with 'scratch that, make it delete those messages.' This constitutes a false_start where the final, corrected instruction is to delete the messages. The original instruction's 'accept if in contact book' condition remains unchanged, and the final action is a clear, unambiguous revision of the default action. The final intent matches the corrected instruction.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages uh the day before yesterday. If their number is in my phone contact book, accept it, otherwise archive those messages. Uh, scratch that, make it delete those messages." }, "candidate_id": "wav_8b261fd2fa205a5a", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip has the most complex and least natural flow. It begins with a revision ('archive' corrected to 'delete'), but the 'uh' and 'scratch that' sound more like a post-production edit than a spontaneous correction. The speaker also seems to be struggling with the phrasing, and the overall delivery feels less fluent and more awkward compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I left out a detail: these were the invitations from the day before yesterday.", "long_pause_offsets": [ 155 ], "source_sha256": "79ee9763ad74725b659b329b0b30f6d9c1524e658bff73cd3a41210f60fa7163", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time detail 'from the day before yesterday' to the initial instruction, which was otherwise complete. The action, condition, and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/addition/addition_02/a6429b7cc1b245d6/speech.wav", "wer": 0.06976744186046512, "wav_sha256": "757b53b728f6553c2cb2d0ba4be0dd0a1fe7df5b67b12772f23c59d2bfe4277b", "seconds": 25.0, "spoken_utterances": [ "I got some let's see... Splitwise group invitations hmmm... over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "I left out a detail: these were the invitations from the uh... day before yesterday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the core instruction to accept or delete Splitwise invitations based on contact book presence. After a pause and self-correction ('I left out a detail'), they add the time constraint 'from the day before yesterday'. This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the new temporal condition, matching the expected behavior for an 'addition' type. The control intent is to execute the combined task.", "heard_transcript": "I got some let's see, Splitwise group invitations, hmm, over phone voice messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I left out a detail, these were the invitations from the day before yesterday." }, "candidate_id": "wav_2cdac0284a47fef4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear initial request, uses a plausible 'hmm' pause to transition, and then adds a detail ('these were the invitations from the day before yesterday') in a way that sounds like a spontaneous memory. The prosody is varied and the pacing is conversational, with a natural self-correction from 'accept it' to 'accept it otherwise delete those messages' that fits the context. The addition feels integrated, not tacked on." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday.. Oh, and one more thing: if their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [ 92 ], "source_sha256": "b07bf7cfd933888107f4425acd0ba60b68c94cbf8fdfa58b6cf26abc2bc1f36b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the conditional logic for handling the invitations, which was missing in the initial utterance. The initial utterance correctly identifies the source and timing of the invitations. The final goal combines both parts: receiving invitations on the specified date and applying the contact-based rule to accept or delete them." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/addition/addition_03/1fd0de8bedfc5459/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5a1a8b6e2029ced7195ed790c39422a3c02b592dcfdfdf973b2cfa724f119991", "seconds": 20.28, "spoken_utterances": [ "I got some hmmm... Splitwise group invitations um, let me think... over phone voice messages the day before yesterday.. ", "Oh, and one more thing: if their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'let me think' but continues the original instruction without changing its core meaning. The addition of 'Oh and one more thing' introduces a new conditional clause that is part of the original instruction. The final control intent is to execute the combined task as stated, matching the reference.", "heard_transcript": "I got some um Splitwise group invitations um let me think over phone voice messages the day before yesterday. Oh and one more thing, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_55405de3c03410d6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is realistic but slightly less so than audio_01. The speaker uses 'um' and 'let me think' as fillers, which are natural, but the addition of 'Oh and one more thing' feels a bit more like a formulaic way to add information rather than a truly spontaneous thought. The prosody is good, but the transition to the final clause is slightly abrupt, and the overall delivery is a bit more hesitant and less fluid than audio_01." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "I got some Splitwise group invitations the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Oh, I forgot to mention, the invitations were received over phone voice messages.", "long_pause_offsets": [ 154 ], "source_sha256": "d532089e42e7739f8813fa6aa58fcabc438ef8dd1ef6b107fe87b9b1b15df68f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery method 'over phone voice messages' without changing any other condition or action. The original instruction already specified the timing ('the day before yesterday') and the conditional logic ('if number in contact book, accept; otherwise delete'). The addition clarifies the source of the invitations, completing the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/addition/addition_01/ae581b5b425c43c3/speech.wav", "wer": 0.06976744186046512, "wav_sha256": "9b4d64825bbea87b3c6943c3a64b70b98733cc7ad33b8c806697cd823bd0d750", "seconds": 25.32, "spoken_utterances": [ "I got some Splitwise group invitations the day before yesterday. um, let's see... If their number is in my phone contact book, accept it, uh... otherwise delete those messages. ", "Oh, I forgot to mention, the invitations were received over um... phone voice messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and augmented with the additional detail about the invitation delivery method. The addition is clearly spoken after a pause and does not alter the core conditional logic. The final intent matches the original instruction with an added constraint, satisfying the 'addition' type criterion.", "heard_transcript": "I got some Splitwise group invitations the day before yesterday. Um, let's see. If their number is in my phone contact book, accept it. Uh, otherwise delete those messages. Oh, I forgot to mention, the invitations were received over um, phone voice messages." }, "candidate_id": "wav_e627338c39123502", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with a complete request, then uses 'um' and 'let's see' as fillers, but the final addition ('Oh, I forgot to mention...') is delivered with a flat, almost robotic intonation. The prosody lacks the natural variation of a real person speaking, and the final clause feels tacked on rather than a genuine, integrated thought. The overall pacing is also slightly off, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. I misspoke about the time: it should be the day before yesterday, not yesterday.", "long_pause_offsets": [ 165 ], "source_sha256": "878201181fdaf3b7ffdc53cb20a8e790f600bfe29807570788f2dc3dc64c02d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'yesterday' to 'the day before yesterday', which matches the original instruction. All other conditions and actions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_03_time_reference/c31f4c711f45f71c/speech.wav", "wer": 0.0, "wav_sha256": "49b9d483c03d35495c70adc5a56c6ef9a9ead9ed796bf8f61a6fd4e71090ff2a", "seconds": 24.6, "spoken_utterances": [ "I got some Splitwise group invitations uh, let's see... over phone voice messages yesterday. uh, let me think... If their number is in my phone contact book, accept it, otherwise delete those messages. ", "I misspoke about the time: it should be the day before yesterday, not yesterday." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yesterday' but immediately corrects it to 'the day before yesterday' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original instruction exactly, so the task is executed as revised.", "heard_transcript": "I got some Splitwise group invitations uh let's see over phone voice messages yesterday. uh let me think. if their number is in my phone contact book, accept it, otherwise delete those messages. I misspoke about the time, it should be the day before yesterday, not yesterday." }, "candidate_id": "wav_4252315a2428d0c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and continuous flow. The speaker uses a realistic 'uh' and 'let me think' filler, followed by a self-correction of the time ('day before yesterday, not yesterday') that is integrated smoothly into the sentence. The prosody is conversational, with appropriate pauses and intonation, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_invitation_app", "proposed_instruction": "I got some Venmo group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Wait, the invitation app should be Splitwise, not Venmo.", "long_pause_offsets": [ 176 ], "source_sha256": "5b8adbd241775e6a8daa96027667f12219443f2f57d7bf6f6fabfc2124f30fc4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the app from Venmo to Splitwise, preserving all other conditions including the source (phone voice messages), time (day before yesterday), and the conditional logic (accept if number is in contacts, otherwise delete)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_01_invitation_app/97552c9e08c0e33d/speech.wav", "wer": 0.0, "wav_sha256": "2d6d180217693cdc306526260ce8959aaf85b3caf61f83bd5da120b3052de493", "seconds": 24.48, "spoken_utterances": [ "I got some um, let's see... Venmo group invitations over phone voice messages ... the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "Wait, the invitation app should be ... Splitwise, not Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately corrects it to 'Splitwise' after a pause, which is a revision. The final intent is to apply the same logic but to the correct app, Splitwise. The original instruction was for Splitwise, so the revision corrects the initial error and matches the reference.", "heard_transcript": "I got some um, let's see, Venmo group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Wait, the invitation app should be Splitwise, not Venmo." }, "candidate_id": "wav_feec1c7a130034d1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a natural 'um' and 'let me think', indicating hesitation. The correction of 'Venmo' to 'Splitwise' is a plausible self-repair, and the overall delivery is conversational. However, the initial hesitation is slightly more pronounced and the correction is a bit more abrupt than in audio_01, making it feel slightly less fluid." } ] }, "is_default": false }, { "variant_id": "revision_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone, accept it, otherwise delete those messages. Let me correct the contact reference: use my phone contact book, not my phone.", "long_pause_offsets": [ 167 ], "source_sha256": "6ab9573c276961bbfb1a0336ce1b7dec84729b56061895d6b0ad7319dc5971c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the condition from 'my phone' to 'my phone contact book', which matches the original instruction. The rest of the logic and context remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_04_contact_reference/230fb05e45e30620/speech.wav", "wer": 0.06382978723404255, "wav_sha256": "3bc4d7ab7922a6f440b4aab5532c40e3cb138ab7ed1a61cc71a1fe696410633e", "seconds": 25.52, "spoken_utterances": [ "I got some Splitwise group invitations over phone voice messages uh, let me think... the day before yesterday. If their number is in my phone, let's see... accept it, otherwise delete those messages. ", "Let me correct the contact reference: use ... my phone contact book, not my phone." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'in my phone' but immediately corrects this to 'in my phone contact book' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use the 'phone contact book' as the reference, which matches the original instruction's intent. The only difference is the speaker's initial misstatement, which is corrected. Therefore, the final task intent is clear and matches the original instruction with the corrected target.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages, um, let me think, the day before yesterday. If their number is in my phone, let's see, accepted, otherwise delete those messages. Let me correct the contact reference. Use my phone contact book, not my phone." }, "candidate_id": "wav_cf408efb852d8e48", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip includes a 'let me think' filler and a self-correction of 'my phone' to 'my phone contact book'. The correction is clear and the flow is generally natural. However, the phrase 'let me correct the contact reference' is a slightly unnatural way to introduce a self-correction, making the repair feel a bit more mechanical compared to the more integrated corrections in audio_01 and audio_06." } ] }, "is_default": false }, { "variant_id": "revision_02_message_channel", "proposed_instruction": "I got some Splitwise group invitations over email the day before yesterday. If their number is in my phone contact book, accept it, otherwise delete those messages. Make that over phone voice messages, not email.", "long_pause_offsets": [ 165 ], "source_sha256": "6361b2db0befeb5f62f623005002658840d0ca1b6c4fc422d35cefa04e2232e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the communication channel from 'email' to 'phone voice messages', aligning with the original instruction. All other conditions, including the time frame ('the day before yesterday'), the action logic ('accept if in contacts, otherwise delete'), and the entity ('Splitwise group invitations') are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_02_message_channel/f9735c27b494c5d6/speech.wav", "wer": 0.075, "wav_sha256": "5977964de370a1b7bd5b606375b6eb768a39c717ddd8af82054fa94f822708fa", "seconds": 20.84, "spoken_utterances": [ "I got some Splitwise group invitations over email the day before yesterday. If their number is in my phone contact book, accept it, um... otherwise delete those messages. ", "Make that over uh, let me think... phone voice messages, not email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the invitations were received over 'email' but then corrects themselves after a pause and self-interruption to 'phone voice messages, not email'. This constitutes a revision where the final audible intent matches the original instruction. The correction is clear and applies to the medium of the invitations, preserving all other conditions.", "heard_transcript": "I got some Splitwise group invitations over email the day before yesterday. If their number is in my phone contact book, accept it. Um. Otherwise delete those messages. Make that over. Uh, let me think. Phone voice messages, not email." }, "candidate_id": "wav_907f2526667b9eb8", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a 'um' and corrects 'email' to 'phone voice messages'. The correction is clear, but the initial mention of 'email' is a bit too clean and the correction is introduced with a slightly unnatural 'make that over'. The overall delivery is a bit more hesitant and less fluid than the top-ranked clips." } ] }, "is_default": false }, { "variant_id": "revision_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, otherwise archive those messages. That should be delete those messages, not archive those messages.", "long_pause_offsets": [ 181 ], "source_sha256": "70ba861badbedf735ed64bc5c7df7a17089a72d53a564ddfdbf0a18d79c46d3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects 'archive' to 'delete', changing the action for the 'otherwise' branch. The original instruction specifies 'delete', but the initial utterance uses 'archive', and the correction overrides it." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_06_unmatched_contact_action/5e8a7ce7d86b6777/speech.wav", "wer": 0.024390243902439025, "wav_sha256": "b815bd4e4b1c5c2c8ca46fe51b2dca072a56fa11829d1b938064fbba59420580", "seconds": 22.48, "spoken_utterances": [ "I got some Splitwise group invitations ... over phone voice messages ... the day before yesterday. If their number is in my phone contact book, accept it, let's see... otherwise archive those messages. ", "That should be delete those messages, not archive those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but immediately corrects themselves to 'delete those messages' after a pause, which is a revision. The final intent is to delete the messages if the number is not in the contact book, matching the original instruction. The correction is clear and unambiguous.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, let's see, otherwise archive those messages. That should be delete those messages, not archive those messages." }, "candidate_id": "wav_35dc25ff526b08ee", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip has a self-correction from 'archive' to 'delete'. The correction is clear, but the phrase 'that should be' is a bit more formal and less conversational than the 'it should be' in audio_01. The flow is slightly less natural, and the correction feels a bit more like a post-hoc edit than a spontaneous recollection." } ] }, "is_default": false }, { "variant_id": "revision_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone voice messages the day before yesterday. If their number is in my phone contact book, ignore it, otherwise delete those messages. Change accept it to accept it, not ignore it.", "long_pause_offsets": [ 180 ], "source_sha256": "dc8631da210314861f55c5d1cedf701c1dac88b08f242d71e2069428d3f63b4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the conditional action from 'ignore it' to 'accept it', restoring the original instruction's intended behavior. The condition (number in contact book) and the alternative action (delete messages) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/revision/revision_05_matched_contact_action/01bddfb025412baf/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "97b82b2d48893fc7be8701b5c0b6215767674cec1345bc8d40c50eee0c8706e0", "seconds": 26.56, "spoken_utterances": [ "I got some ... Splitwise group invitations over phone voice messages uh, let me think... the day before yesterday. If their number is in my phone contact book, ignore it, otherwise delete those messages. ", "Change let's see... accept it to accept it, not ignore it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ignore it' but then corrects themselves with 'accept it to accept it, not ignore it' after a pause, which is a revision. The final intent is to accept invitations for numbers in the contact book, not ignore them. This matches the original instruction's intent, so the revision is correctly applied.", "heard_transcript": "I got some Splitwise group invitations over phone voice messages, uh, let me think, the day before yesterday. If their number is in my phone contact book, ignore it, otherwise delete those messages. Change, let's see, accept it to accept it, not ignore it." }, "candidate_id": "wav_3a230f18c26ef111", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "The speaker makes a correction from 'ignore' to 'accept'. The correction is clear, but the phrase 'change, let's see' is a bit awkward and less natural. The overall delivery feels a bit more robotic and less conversational compared to the other clips, with a slightly less smooth flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "883115327695a2fd6a85868507f8118d6f6366764cfb21008fe593f8fc00fbb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/cancellation/cancel_021/5fb8c9a529781de3/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "8eff0b30305c08b01c07670924301ce4f9d11daf1201189384a038aa3b0cd4fc", "seconds": 22.12, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages.", "Oh, wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Oh wait, never mind. I don't need that done anymore,' which is a clear and final cancellation of the entire task. The initial instruction is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages. Oh wait, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "8620cabf17d919cde005b90b5e08235f3e743269aee28d801cd1829e38b2ba88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/ambiguous_halt/halt_019/bde1aa837929286d/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "9e09a1373b3e9a1190a57d1e2327775d761ec41eba7356fc517086a4c6bd6405", "seconds": 20.56, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Can you give me a moment?' is a clear request to pause or wait, with no indication of cancellation or a new instruction. The user's hesitation and backchanneling ('um', 'let's see') are internal processing cues, not a change in the task goal. The final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "df484d4b8146c508ee72c454695ad83a0276cb56e298bcda2aaec196c1240e35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_2/audio_natural/filler/original/1b106929953323d1/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "bb35edc055048afa1e520a5502814cf2d3120cde7a9b973f767ddf3ced66b119", "seconds": 16.56, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone voice messages the day before yesterday. ... If their number is in my phone contact book, accept it, um, let's see... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final instruction remains consistent with the original, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "I got some Splitwise group invitations um over phone voice messages the day before yesterday. If their number is in my phone contact book, accept it, um let's see, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages last month—no, that should be this week. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "fc7d89e007d71b7523c6d232c3e6fa150cd2d10fd94ce47358c99d413051311f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this week' within the same utterance. The rest of the instruction, including the condition and action for handling invitations, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_03_time_reference/87ea9ac6b3f5c3e8/speech.wav", "wer": 0.08571428571428572, "wav_sha256": "8bdc5984f1ef2cb986aace2b8b7fa48e5453376f8f0eb4934b19b9679dc392e3", "seconds": 14.24, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages ... last month—um... no, that should be this week. If their number is in my phone contact book, uh... accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this week'. The final instruction matches the original reference exactly in action and conditions. The false start is corrected, so the final intent is clear and executable.", "heard_transcript": "I got some Splitwise group invitations over phone text messages last month. Um, no, that should be this week. If their number is in my phone contact book, uh, accept it, otherwise delete those messages." }, "candidate_id": "wav_e3e6c0b3411d9357", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural revision pattern: the speaker starts with 'last month', immediately self-corrects to 'no, that should be this week' with a believable hesitation and intonation shift, then continues the sentence smoothly. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause before the correction is also well-placed, suggesting genuine recall." } ] }, "is_default": true }, { "variant_id": "false_start_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone—I mean, my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "b6e9a6e2861aab22b1381900b77e525837dc6882567fd579e84e80fa832f3738", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'my phone' to 'my phone contact book' within the same utterance, clarifying that the condition refers to the contact book. The rest of the logic and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_04_contact_reference/5ab8bb1347c4e1a9/speech.wav", "wer": 0.0625, "wav_sha256": "203e9d87ccff5fa7e337599b8c9b33eac4a40587dc627561b15970a66a6c4873", "seconds": 12.92, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages ... this week. If their number is in my phone—uh... I mean, my phone contact book, accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone' but immediately self-corrects to 'phone contact book'. This is a false start, and the final corrected phrase matches the original instruction. The control intent is to execute the task as corrected.", "heard_transcript": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone uh, I mean my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_0cde20fcfe35a0ba", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'phone uh' followed by a self-repair to 'phone contact book'. The hesitation and correction are natural, and the speaker resumes the sentence without awkwardness. The prosody is mostly consistent, and the repair is integrated smoothly into the flow, making it highly realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_message_channel", "proposed_instruction": "I got some Splitwise group invitations over email—or rather, phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "28ab8a502a64da804aab5f36db40aa3aca4a548b701166526034c11ff65c84f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email' to 'phone text messages' within the same utterance. The final channel is phone text messages, matching the original instruction. All other conditions and actions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_02_message_channel/4b63c2969400067b/speech.wav", "wer": 0.0625, "wav_sha256": "92322fc826a38cb88fda5e53fc9543c9caefdf8ef4efe1ee12fc9e8527fe18da", "seconds": 12.24, "spoken_utterances": [ "I got some Splitwise group invitations over email—um... or rather, phone text messages this week. uh... If their number is in my phone contact book, accept it, otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately self-corrects to 'phone text messages'. This is a false start, and the final, corrected instruction matches the original reference. The control intent is to execute the task as corrected.", "heard_transcript": "I got some Splitwise group invitations over email um, or rather phone text messages this week. Uh, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "candidate_id": "wav_e2a4b2772b1f6aff", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a believable self-correction from 'email' to 'phone text messages'. The speaker uses 'um' and 'uh' appropriately, and the correction is delivered with a natural shift in intonation. The pacing is slightly hesitant but plausible, and the overall delivery sounds like someone formulating a request in real time." } ] }, "is_default": false }, { "variant_id": "false_start_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, decline it—sorry, I meant accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "8205eacd44aa2c72208c07bc180f81592cf204dce9b06bdfee1a9728c277ec30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately corrects with 'I meant accept it'. The final action is accept, which matches the original instruction. The condition (number in contact book) and the alternative action (delete) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_05_matched_contact_action/6552d9bf6469c9b5/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "ddc86d272bfa83a5d9720fd42ac7acfecdc03e4569ba142212d2baca1d99d27e", "seconds": 14.68, "spoken_utterances": [ "I got some um... Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, decline it—ah... sorry, I meant accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'decline it' but immediately self-corrects with 'Sorry, I meant accept it.' According to the 'false_start' type criterion, the final corrected value ('accept it') overrides the initial mistake. The rest of the instruction, including the conditional logic and the 'otherwise delete those messages' clause, remains unchanged and is preserved. The final intent matches the original instruction after the correction.", "heard_transcript": "I got some um, Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, decline it. Sorry, I meant accept it. Otherwise, delete those messages." }, "candidate_id": "wav_25fd594bf6f3edc1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a false start with 'decline it' followed by a self-correction to 'accept it'. The correction is immediate and the intonation shift is realistic. However, the initial hesitation is less pronounced than in other clips, and the correction feels slightly rushed, making it less natural than the top-ranked examples." } ] }, "is_default": false }, { "variant_id": "false_start_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise archive those messages—let me correct that, delete those messages.", "long_pause_offsets": [], "source_sha256": "f6f7b814d23ea25db30fb04114efcdb556a6095671a6b6d2f871e71c3ff2f184", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but immediately corrects to 'delete those messages'. The final instruction matches the original instruction's action for the 'otherwise' case." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_06_unmatched_contact_action/c2d689b26ecf50e6/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "40d1af63b8913ccda626696ff2dc3e93dec16843b4bd1ad8d608d37c633711a3", "seconds": 16.4, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages this week. um... If their number is in my phone contact book, accept it, otherwise archive those messages—uh... let me correct that, delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but immediately self-corrects with 'let me correct that, delete those messages.' According to the 'false_start' type criterion, the final corrected instruction must be executed. The final intent is to delete the messages, which matches the original instruction's intent, with only the action being corrected from 'archive' to 'delete'. The correction is immediate and unambiguous, so the task is executed as corrected.", "heard_transcript": "I got some Splitwise group invitations over phone text messages this week. Um, if their number is in my phone contact book, accept it, otherwise archive those messages. Uh, let me correct that, delete those messages." }, "candidate_id": "wav_9eec361eb19b7a21", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This audio includes a revision where the speaker first says 'archive' and then corrects to 'delete'. While the correction is present, the initial word 'archive' is not a plausible mistake in this context, and the correction feels more like a scripted edit than a spontaneous change. The prosody also becomes slightly strained during the correction." } ] }, "is_default": false }, { "variant_id": "false_start_01_invitation_app", "proposed_instruction": "I got some Venmo—I should have said Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages.", "long_pause_offsets": [], "source_sha256": "0ba168019479d32381191c9d784436b80917d89d93e87640c005c0bcfe204460", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Venmo' to 'Splitwise' within the same utterance, preserving the rest of the instruction including the condition and action for handling the invitations." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/false_start/false_start_01_invitation_app/527333cc14c02298/speech.wav", "wer": 0.0, "wav_sha256": "23b539fcf23fc432104b5886af22019470a09c4f13ccdfa967d38193b0e4d6d1", "seconds": 14.2, "spoken_utterances": [ "I got some Venmo—uh... I should have said Splitwise group invitations over phone text messages this week. ... If their number is in my phone contact book, accept it, let's see... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately self-corrects to 'Splitwise' in a false start. The final, corrected instruction matches the original reference exactly in action, target, and conditional logic. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "I got some Venmo uh, I should've said Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, let's see, otherwise delete those messages." }, "candidate_id": "wav_beb8082bbea32255", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip has multiple issues: it starts with 'Venmo' which is clearly incorrect, followed by a self-correction. However, the correction is delivered with a flat tone and lacks the natural hesitation or intonation shift seen in other clips. The 'let's see' interjection feels out of place and disrupts the flow. The overall delivery sounds rehearsed or unnatural." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "I got some Splitwise group invitations over phone text messages. If their number is in my phone contact book, accept it, otherwise delete those messages. I didn't mention this, but it's specifically for invitations received this week.", "long_pause_offsets": [ 154 ], "source_sha256": "8fa21282734f1a5a08503ebfeda12e7a2520f8e8b44fdb8b8594b425383a39ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this week') that narrows the scope of the initial instruction. The core logic (accept if in contacts, delete otherwise) and the context (Splitwise group invitations via text) remain unchanged. The addition is a clarifying constraint, not a contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/addition/addition_02/9dd120660a2ff927/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "0fe93dc44bf07a34769e611bf4b4a6d5af5b37444a4d238e1c3dd8139219b655", "seconds": 21.56, "spoken_utterances": [ "I got some ah... Splitwise group invitations over phone text messages. uh... If their number is in my phone contact book, ... accept it, otherwise delete those messages. ", "I didn't mention this, but it's specifically for invitations received this week." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition of 'this week' is a narrowing constraint that is clearly stated after a pause. The final intent combines the original task with the new temporal condition, matching the expected behavior for an 'addition' type. The addition is not a correction or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "I got some uh Splitwise group invitations over phone text messages. Uh if their number is in my phone contact book, accept it, otherwise delete those messages. I didn't mention this, but it's specifically for invitations received this week." }, "candidate_id": "wav_6d20a204a7ec8343", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural prosody and a single, plausible addition at the end. The phrase 'I didn't mention this, but it's specifically for invitations received this week' is smoothly integrated as a post-hoc clarification, with appropriate intonation and pacing. The delivery is fluent and sounds like a spontaneous thought being added, fitting the 'addition' augmentation type well." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "I got some Splitwise group invitations this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Oh, and one more thing: these are Splitwise group invitations received over phone text messages.", "long_pause_offsets": [ 139 ], "source_sha256": "873b41dc3387d1d119ef4f36798f65441797848ab10f2f6d962c415b8fd2c6bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery method 'over phone text messages' to the initial invitation description. The core logic (accept if in contacts, delete otherwise) and the entity (Splitwise group invitations) remain unchanged. The addition clarifies the source without contradicting or altering the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/addition/addition_01/78fee24d8dcd29e8/speech.wav", "wer": 0.06976744186046512, "wav_sha256": "c41916386b861e4e8fe30fb36713bcd6a2546e54b05816f1c09536829fee7a93", "seconds": 21.76, "spoken_utterances": [ "I got some Splitwise group invitations um... this week. If their number is in my phone contact book, um, let's see... accept it, otherwise delete those messages. ", "Oh, and one more thing: these are Splitwise group invitations received over uh... phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed. The addition clarifies the medium as 'phone text messages', which is consistent with the original 'over phone text messages'. The final intent combines the initial instruction with the added detail, matching the reference. No contradictions or omissions are present.", "heard_transcript": "I got some Splitwise group invitations um this week. If their number's in my phone contact book, um let's see, accept it, otherwise delete those messages. Oh, and one more thing, these are Splitwise group invitations received over uh phone text messages." }, "candidate_id": "wav_7092279353ec333d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also makes an addition, the delivery is less natural due to multiple fillers ('um', 'uh') and a noticeable hesitation ('let's see') that disrupts the flow. The addition is awkwardly phrased and delivered with a shift in intonation that sounds rehearsed or unnatural. The speaker also repeats the phrase 'Splitwise group invitations' in a way that feels redundant rather than a spontaneous memory, reducing the realism of the addition." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_invitation_app", "proposed_instruction": "I got some Venmo group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Wait, the app should be Splitwise, not Venmo.", "long_pause_offsets": [ 160 ], "source_sha256": "d537218ec93f940153333396f85740afe534978f92108774e6573f57341e5204", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the app from Venmo to Splitwise, which matches the original instruction. All other conditions, including the action logic based on contact book presence, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_01_invitation_app/c5d85d866d716a1d/speech.wav", "wer": 0.09302325581395349, "wav_sha256": "74ed1087eeaa2daf4bd2e7c793d69cc22d3347ccd5ebd24f0aee848759e40691", "seconds": 22.96, "spoken_utterances": [ "I got some Venmo group invitations over phone text messages this week. um... If their number is in my phone contact book, let me think... accept it, otherwise delete those messages. ", "Wait, the app should be um, let me think... Splitwise, not Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Venmo' but immediately corrects to 'Splitwise' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "I got some Venmo group invitations over phone text messages this week. Um, if their number is in my phone contact book, let me think, accept it, otherwise delete those messages. Wait, the app should be, um, let me think, Splitwise, not Venmo." }, "candidate_id": "wav_1574fffe6d36df75", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and complex revision pattern. It begins with a plausible false start ('Venmo group invitations'), self-corrects to 'Splitwise', then pauses with 'let me think' before stating the condition. It later revises the app name again with 'wait, the app should be... let me think', showing a believable moment of recollection. The prosody includes natural hesitations and a genuine-sounding correction process, making it the most realistic." } ] }, "is_default": true }, { "variant_id": "revision_04_contact_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my address book, accept it, otherwise delete those messages. Let me correct the contact reference: use 'my phone contact book', not 'my address book'.", "long_pause_offsets": [ 158 ], "source_sha256": "24b7938c2212ea92d1397ec295a617be6adf16621fd4a763880aa8088c92abfc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'address book' to 'phone contact book', aligning with the original instruction. All other conditions and actions remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_04_contact_reference/9615391c4f6c57e4/speech.wav", "wer": 0.10869565217391304, "wav_sha256": "a782de0af33e0bbbd7686ec3eb6b24a83533160725c875e14047fd8872357b16", "seconds": 21.4, "spoken_utterances": [ "I got some um... Splitwise group invitations over phone text messages this week. If their number is in my address book, uh, let's see... accept it, otherwise delete those messages. ", "Let me correct the contact reference: use 'um... my phone contact book', not 'my address book'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'address book' but immediately corrects it to 'phone contact book' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction with the corrected target ('phone contact book' instead of 'address book'). All other conditions are preserved, so the task is executed as revised.", "heard_transcript": "I got some um Splitwise group invitations over phone text messages this week. If their number is in my address book, uh let's see, accept it, otherwise delete those messages. Let me correct the contact reference, use um, my phone contact book not my address book." }, "candidate_id": "wav_29d567d3c8105d5c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision where the speaker initially says 'address book', then corrects to 'phone contact book'. The 'uh let's see' filler is a natural hesitation. The correction is delivered with a slight shift in intonation, indicating a change of mind. The flow is smooth and the self-correction is integrated into the speech, though less complex than audio_03." } ] }, "is_default": false }, { "variant_id": "revision_05_matched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, ignore them, otherwise delete those messages. That should be 'accept it', not 'ignore them'.", "long_pause_offsets": [ 166 ], "source_sha256": "2a0d727f11dd70174d9ebe699bf5c5fe40c21ab520578629dcacc217a9f1ddc7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'ignore them' to 'accept it', restoring the original action. The condition (number in contact book) and the alternative action (delete) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_05_matched_contact_action/b2ab9379cbac4c8e/speech.wav", "wer": 0.1388888888888889, "wav_sha256": "d6bb6e31d470ebae0b546324c88d13243eae9000dc08e4109a1abd72b9f6db1f", "seconds": 17.56, "spoken_utterances": [ "I got some uh... Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, ignore them, ... otherwise delete those messages. ", "That should be '... accept it', not 'ignore them'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ignore them' but immediately corrects themselves with 'That should be accept it, not ignore them.' This constitutes a revision where the final intent is to accept invitations for contacts in the book. The correction is clear and directly replaces the initial incorrect instruction. All other conditions remain unchanged, satisfying the revision type criterion.", "heard_transcript": "I got some uh Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, ignore them. Otherwise delete those messages. That should be accept it, not ignore them." }, "candidate_id": "wav_9f5740850555a370", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker first says 'ignore them', then corrects to 'accept it'. The 'that should be' phrasing is a common way to self-correct. However, the initial request is delivered in a slightly flat, robotic tone, and the correction is abrupt. The prosody lacks the natural hesitation seen in the top two, making it less realistic." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "I got some Splitwise group invitations over phone text messages today. If their number is in my phone contact book, accept it, otherwise delete those messages. Change the time reference to 'this week' instead of 'today'.", "long_pause_offsets": [ 160 ], "source_sha256": "3ebbf8593fb55aaa526f3212f24a427431b685aedbbee3090812440df15760fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces 'today' with 'this week' in the time reference, preserving all other conditions including the app, message type, recipient logic, and action (accept/delete)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_03_time_reference/bb3fa0268ee8c8f8/speech.wav", "wer": 0.1282051282051282, "wav_sha256": "7b0aa5efccf95027e5aac962ac62c08df79e12cbcbd10787dd9a241a45c94354", "seconds": 18.96, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone text messages uh... today. If their number is in my phone contact book, accept it, otherwise delete those messages. ", "Change the time reference to 'hmmm... this week' instead of 'today'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'today' but then clearly revises the time reference to 'this week' after a pause. The revision is applied to the original instruction, and the final goal matches the reference. The correction is not a cancellation or pause, and the task is fully executable with the revised time frame.", "heard_transcript": "I got some Splitwise group invitations um over phone text messages uh today. If their number is in my phone contact book, accept it, otherwise delete those messages. Change the time reference to hmm this week instead of today." }, "candidate_id": "wav_519dc5d1cb986e4d", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a revision where the speaker changes 'today' to 'this week'. The 'change the time reference to...' is a clear, but slightly unnatural, way to state a correction. The initial request is delivered in a flat, read-aloud manner with little prosodic variation. The revision feels more like a mechanical instruction than a spontaneous correction." } ] }, "is_default": false }, { "variant_id": "revision_06_unmatched_contact_action", "proposed_instruction": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise archive those messages. Use 'delete those messages' instead of 'archive those messages'.", "long_pause_offsets": [ 165 ], "source_sha256": "7ee98cb588cab6d50577efdbd2ebf8f148ba37fc0da92d2de673992386d73730", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the action for non-contact numbers from 'archive' to 'delete', which contradicts the original instruction's 'delete' action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/revision/revision_06_unmatched_contact_action/c9a8aa37d6fd3fc9/speech.wav", "wer": 0.08108108108108109, "wav_sha256": "17baa011f753c58156d926d66d84f3a2b83cc8713d7085e6b1ace026a898cf9f", "seconds": 15.72, "spoken_utterances": [ "I got some Splitwise group invitations um... over phone text messages this week. If their number is in my phone contact book, ... accept it, otherwise archive those messages. ", "Use 'delete those messages' instead of 'archive those messages'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archive those messages' but then immediately corrects it with 'Use delete those messages instead of archive those messages.' This is a clear revision where the final instruction overrides the initial one. The final intent matches the original reference instruction, so the task should be executed as revised.", "heard_transcript": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise archive those messages. Use delete those messages instead of archive those messages." }, "candidate_id": "wav_79c230f46191ccbc", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip is the least realistic. It begins with a clear request, then abruptly states 'use delete those messages instead of archive those messages' as a correction. The delivery is flat, lacks any hesitation or natural flow, and the correction is stated as a separate command rather than being integrated into the speech. The prosody is monotonous and the transition is jarring, suggesting a synthetic or scripted origin." } ] }, "is_default": false }, { "variant_id": "revision_02_message_channel", "proposed_instruction": "I got some Splitwise group invitations over email this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Make that phone text messages, not email.", "long_pause_offsets": [ 150 ], "source_sha256": "ff75fd38ef8ea7d2e1fd3ca54a1c83009c468d85fe113e9eb946f5999d00334b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the communication channel from email to phone text messages, preserving the rest of the logic and conditions." }, "audio": { "status": "pending", "spoken_utterances": [ "I got some Splitwise group invitations uh... over email this week. If their number is in my phone contact book, accept it, ... otherwise delete those messages. ", "Make that ... phone text messages, not email." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "416f857b13a06192c858b1aed8b0eee0f27f0a377b3017cf6301e85a2aa84a0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/cancellation/cancel_034/a4742a150ae71e33/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "3b3c397393e5d0a95afc3e854324307d57d193024ad1882979633cda5a3e2844", "seconds": 19.6, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then concludes with the explicit cancellation phrase 'Hold on, let's drop the whole task.' This clearly revokes the task, satisfying the cancellation type criterion. The final intent is to cancel, not to execute.", "heard_transcript": "I got some Splitwise group invitations over phone text messages this week. If their number is in my phone contact book, accept it, otherwise delete those messages. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "26568bfc60f4af301900fd4c0a68de5ac1b3c71e0000921aa69015d4c6429da0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/ambiguous_halt/halt_007/1433befc45f1effa/speech.wav", "wer": 0.12121212121212122, "wav_sha256": "9d93e961081b4587ed8b04490efc2a9e5c638dad034c6c51391c70f8f4357eaf", "seconds": 18.08, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment,' which is a clear request to pause or wait. The user does not cancel the task or provide a final instruction. The initial part of the audio is a filler ('let me think') followed by the original instruction, but the final control request is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "I got some Splitwise group invitations over phone text messages. Uh, let me think. This week, if their number is in my phone contact book, accept it. Otherwise delete those messages. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3aa1a22_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c60a71174929a04661fdbf593f9bbbaa348db5e2efa9f04953dc8818ab669d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3aa1a22_3/audio_natural/filler/original/3e9ed68dca2b9326/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "13aed50752dfdd73db4ce55ace2a2e6a03cf04fb6224a06dc244aedfbfe4c893", "seconds": 14.96, "spoken_utterances": [ "I got some Splitwise group invitations over phone text messages uh, let me think... this week. If their number is in my phone contact book, ... accept it, ... otherwise delete those messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'let me think' but continues with the original instruction without changing the goal. The final intent matches the original instruction, and the filler does not alter the task meaning. Therefore, the task should be executed as intended.", "heard_transcript": "I got some Splitwise group invitations over phone text messages. Uh, let me think. This week, if their number is in my phone contact book, accept it, otherwise delete those messages." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify playlists that I have saved—sorry, I meant liked.", "long_pause_offsets": [], "source_sha256": "c1044f38a8b3d57e845dca1b5f72458956b58f485fc576c735458d451a436737", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended target is 'liked', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/false_start/false_start_03_interaction_type/6f475b63ef393876/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "9c4b702e4ff04d3e62ae121bccf66681e8a874508aed022496fa0216f82566c7", "seconds": 10.4, "spoken_utterances": [ "Download all the songs from my Spotify playlists uh... that I have saved—um... sorry, I meant ... liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value 'liked' should be applied, and the original incorrect value 'saved' should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Download all the songs from my Spotify playlist, uh, that I have saved, um, sorry, I meant liked." }, "candidate_id": "wav_51274bbd52730bc7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'playlists uh' and then smoothly transitions to 'saved um' before correcting to 'liked'. The false start and repair are integrated within a single, continuous utterance with natural hesitation and intonation. The pacing and prosody are consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Download all the songs from my Apple Music—I should say Spotify playlists that I have liked.", "long_pause_offsets": [], "source_sha256": "4f84ab8a6149eb318e513549f7a22cfe68c4d3930abbfa2c6ba46c1188f2d4b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, scope of playlists, and the liked filter." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/false_start/false_start_01_app/1008b0307e0cc90b/speech.wav", "wer": 0.1, "wav_sha256": "22d0f1d6541508e0961d6a3a239bc5a8afadd47e44113eda0d1c85b9749dd7d8", "seconds": 12.52, "spoken_utterances": [ "Download all the songs from my um... Apple Music—uh... I should say Spotify playlists hmmm... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Download all the songs from my um Apple Music uh I should say Spotify playlists that I have liked." }, "candidate_id": "wav_3443a0508b7a5324", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a self-correction from 'Apple Music' to 'Spotify', but the repair is less fluid. The speaker says 'Apple Music uh' and then corrects to 'Spotify playlists'. The pause and filler 'uh' are placed in a way that feels slightly more hesitant and less integrated than in audio_03. The overall delivery is still realistic, but the correction is more abrupt and less seamlessly embedded in the flow." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify albums—scratch that, make it playlists that I have liked.", "long_pause_offsets": [], "source_sha256": "a518b55cb05fa8761eb6dbc5fb2c2df564ba26d4ee3d51ba33fb28eda5e925a6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately corrects to 'playlists'. The rest of the instruction, including the app (Spotify), the action (download), and the condition (that I have liked), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/false_start/false_start_02_source_collection/5e701ac482b38907/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "4d555de92ad26f94700badd371cccb6988bc7592d14288f3cfaadde9a570edab", "seconds": 9.52, "spoken_utterances": [ "Download all the songs from my Spotify albums—uh... scratch that, make it playlists ... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects with 'scratch that' to 'playlists'. This is a false start where the final corrected value overrides the initial mistake. The final intent matches the original instruction, so the task should be executed as corrected.", "heard_transcript": "Download all the songs from my Spotify albums. Uh, scratch that. Make it playlists that I have liked." }, "candidate_id": "wav_485a34b583e39498", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a false start and repair, but the delivery is the least natural. The speaker says 'albums' and then immediately says 'scratch that' before correcting to 'playlists'. The phrase 'scratch that' is a very abrupt and unnatural way to self-correct in casual speech. The correction feels more like a mechanical edit than a spontaneous thought process, and the overall prosody is flatter and less expressive than the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Download from my Spotify playlists that I have liked. I left out a detail: download all the songs.", "long_pause_offsets": [ 54 ], "source_sha256": "e581854fe0b7d7cac6a7673661d074b7c10e104becfd8c6435a1732cf5d05a66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'all the songs' requirement, completing the initial instruction which lacked the scope of the download. The app, source, and condition (liked playlists) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/addition/addition_02/d234a79b9ebc7e9c/speech.wav", "wer": 0.2, "wav_sha256": "30fb8f3bd43f93250dc8cad6f8241731187fd2516cbacf7c299dae8ed5b45e2d", "seconds": 11.44, "spoken_utterances": [ "Download from my Spotify playlists uh... that I have liked. ", "I left out a detail: download ah... all the songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Download from my Spotify playlist uh that I have liked.' They then add, 'I left out a detail, download uh all the songs.' This is a clear addition that expands the initial request to include 'all the songs.' The final intent combines the original instruction with the added detail, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Download from my Spotify playlist uh that I have liked. I left out a detail, download uh all the songs." }, "candidate_id": "wav_74b2f8b46248cc1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then adds a detail ('download all the songs') after a brief, plausible self-correction ('I left out a detail'). The prosody is consistent, and the addition flows smoothly without awkward phrasing or unnatural pauses. The speaker sounds like they are genuinely remembering a missing part of the task." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Download all the songs from my Spotify playlists. Oh, I forgot to mention, only download songs from playlists that I have liked.", "long_pause_offsets": [ 50 ], "source_sha256": "4c9e5643108cdf29dcb208b5dc4f971b0b35eb29f2bbcd237f1bef146837b05f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that restricts the download to only those playlists the user has liked. This narrows the scope of the initial broad request without contradicting it." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/addition/addition_01/20d37a51a30e86b3/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "6bb94f4a8dfcc9d8c829f5356d49cae1e7a4213daa071e1980326f844424f798", "seconds": 14.92, "spoken_utterances": [ "Download all the songs from my uh... Spotify playlists. ", "Oh, I forgot to mention, only download songs from playlists uh, let me think... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to download all songs from Spotify playlists. After a pause and a self-correction, the speaker added the constraint 'only download songs from playlists that I have liked.' This addition narrows the scope of the original request, and the final audible intent combines both the original and the added condition. The final goal matches the reference instruction, so the task passes.", "heard_transcript": "Download all the songs from my uh Spotify playlists. Oh, I forgot to mention, only download songs from playlists uh let me think that I have liked." }, "candidate_id": "wav_fbedf64e071ca3d0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a complete request, then adds a restriction ('only download songs from playlists...') after a hesitation ('oh, I forgot to mention'). However, the follow-up is less natural: the speaker says 'let me think' after already specifying the playlist, which creates a slight logical inconsistency. The addition feels more like a post-hoc clarification than a spontaneous memory, and the prosody is slightly less fluid than in audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify playlists that I have saved. That should be liked, not saved.", "long_pause_offsets": [ 68 ], "source_sha256": "f3ccf592afe11ea3bf7222b04f25521670c793319d995c2c5163c848798a52cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the condition from 'saved' to 'liked', preserving the app, action, and scope of 'all songs from playlists'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/revision/revision_03_interaction_type/bc4d104cbf957d57/speech.wav", "wer": 0.2, "wav_sha256": "6f1d9a454e89c291fc59925f1dfab4cd207941e264870be9c845307d98f2703e", "seconds": 12.84, "spoken_utterances": [ "Download all the songs from my hmmm... Spotify playlists um... that I have saved. ", "That should be liked, not saved." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is the revision type. The final intent is to download songs from playlists that are liked, matching the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Download all the songs from my um Spotify playlists um that I have saved. That should be liked, not saved." }, "candidate_id": "wav_7b25c8dd02a98d66", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic revision pattern. The speaker begins with a plausible false start ('my um Spotify playlists um'), uses filler words and pauses that sound spontaneous, and then clearly self-corrects 'saved' to 'liked' with a believable change-of-mind intonation. The correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Download all the songs from my Apple Music playlists that I have liked. Wait, use Spotify, not Apple Music.", "long_pause_offsets": [ 72 ], "source_sha256": "33dfd19f817f15b246331f1135879b276bdc182941c234da86f669af920ba05a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, while retaining the action, scope (all playlists), and condition (songs that I have liked)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/revision/revision_01_app/83c20f193860094c/speech.wav", "wer": 0.1, "wav_sha256": "b3b052225916035269b26e105d30ee0e0a781756fcf7eb13f5f8373e8fc6c7e8", "seconds": 11.6, "spoken_utterances": [ "Download all the songs from my um... Apple Music playlists that I have liked. ", "Wait, use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify, not Apple Music' after a pause. This is a revision where the final audible intent is to use Spotify. The original instruction was to download liked songs from a playlist, and the correction only changes the app from Apple Music to Spotify. The final intent matches the original instruction with the corrected target app, satisfying the revision type criterion.", "heard_transcript": "Download all the songs from my um, Apple Music playlist that I have liked. Wait, use Spotify, not Apple Music." }, "candidate_id": "wav_7fc250c085d66010", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a realistic revision with a clear false start ('my um Apple Music playlists') and a believable self-correction ('Wait, use Spotify not Apple Music'). The speaker's tone and pacing suggest genuine hesitation and mid-sentence correction. However, the correction is slightly more abrupt and less integrated than in audio_02, making it feel a bit more like a post-hoc edit than a fully spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify albums that I have liked. Change playlists to albums.", "long_pause_offsets": [ 65 ], "source_sha256": "a5aacf2605081f2db96b9ea91c93b45f254a655001ee81d62aa93e1ec945c72a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'playlists' to 'albums', replacing the source entity while preserving the action, app, and liked condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/revision/revision_02_source_collection/f90e8f95a2a91556/speech.wav", "wer": 0.0, "wav_sha256": "ae31fa9f36eb14e32a2d0e588090dc183b142a9bd5d2a1a7f38d037761b2e393", "seconds": 10.24, "spoken_utterances": [ "Download all the songs from my ... Spotify albums ... that I have liked. ", "Change playlists to albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlists' but immediately corrects it to 'albums' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to download songs from albums, not playlists, which matches the revised goal. All other constraints remain unchanged.", "heard_transcript": "Download all the songs from my Spotify playlists that I have liked. Change playlists to albums." }, "candidate_id": "wav_ae63ee1799291926", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It sounds like a flat, rehearsed instruction with no natural hesitation, filler, or self-correction. The revision ('change playlist to albums') is delivered as a separate, detached command rather than being integrated into the original request. The prosody is monotonous and lacks the natural rhythm and intonation of a person formulating a request in real time." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "1c59f5f7113140e39560a28b06eb5840b5afb9ce7dae44f624e2a551989aae27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/cancellation/cancel_003/22e0cbdefa0e5d5c/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "9b7d875269666bd766a352544482dc18275df171c54ae877bcb7ecd81bae152f", "seconds": 11.04, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked.", "Hold on, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting a task, then immediately follows with 'Hold on, please cancel that,' which is a clear and unambiguous cancellation of the previously stated request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked. Hold on, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "d4787ebec7f894d8df532d0da5378c85582ad14a68d39caee006a4856c6d215c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/ambiguous_halt/halt_040/92a446d6666a4439/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "adcfbbfae220b578a16142d6064c20027f090ebee8d6bb86530981f5afd46949", "seconds": 10.4, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly requests a pause with 'Hang on, just a minute.' This is a final interjection that indicates a temporary halt, not a cancellation or a change in the task goal. The original instruction is not revoked, but the speaker explicitly asks to wait, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "266b76f2aff1cbc6825c601ec436507c7a1902e0994aebf4ec064cca5ac516ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_1/audio_natural/filler/original/9c6489e39e4a14e6/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "871a4945b23c94039348e34607788e26a1776b3613bbe256b88154f2704f9788", "seconds": 6.32, "spoken_utterances": [ "Download all the songs from my ... Spotify playlists um... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent is identical to the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Download all the songs from my Spotify playlists um that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify song library that I have saved—I meant liked.", "long_pause_offsets": [], "source_sha256": "5d2365743b3884153e0c580d455fa1cd94c96ad96164c0c156e3485479a6b79b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended action is to download songs that have been liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/false_start/false_start_03_interaction_type/cef89a6eab7d6767/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "8b164b2a94500d0561590509ee92726a82ffd4d3d0a6994ec625ba7d47e64c2b", "seconds": 7.12, "spoken_utterances": [ "Download all the songs from my um... Spotify song library that I have saved—uh... I meant liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects with 'I meant liked'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is discarded. The final audible task matches the original instruction exactly after the correction.", "heard_transcript": "Download all the songs from my um Spotify song library that I have saved uh I meant liked" }, "candidate_id": "wav_081be59c9cea1e6a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with 'saved' and immediately self-corrects to 'liked' with a clear, audible 'uh' and a slight pitch drop, mimicking genuine hesitation. The prosody is fluid, with a natural rhythm and appropriate pauses. The correction feels integrated into the thought process rather than tacked on, and the overall delivery is continuous and believable." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Download all the songs from my Apple Music—I should say Spotify song library that I have liked.", "long_pause_offsets": [], "source_sha256": "e7588258757362f9cb736182a85b0dc06792102726aa246763df1657b6e8d1b1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, scope ('all the songs'), and condition ('that I have liked')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/false_start/false_start_01_app/8c00bbed157c5d36/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "ad950a5600d7112dae0ce7281f19e9350d5c74da126ad67f6524f60437159d51", "seconds": 11.84, "spoken_utterances": [ "Download all the songs from my Apple Music—um... I should say Spotify song library uh, let me think... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. The final intent is clear and matches the original instruction. The false start is corrected, and the task is fully specified.", "heard_transcript": "Download all the songs from my Apple Music, um, I should say Spotify song library, uh, let me think, that I have liked." }, "candidate_id": "wav_4e68aa71ffd2be03", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a plausible false start and revision. The speaker corrects 'Apple Music' to 'Spotify' and then hesitates with 'uh let me think,' which adds realism. However, the correction is less fluid than in audio_03; the 'let me think' pause feels slightly longer and more deliberate, and the overall flow is less continuous, making it slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify playlist—scratch that, make it song library that I have liked.", "long_pause_offsets": [], "source_sha256": "ed6a0ad83ad4c3a4f70e791ca72df311c24a0920ccad40a9c984f1979dd76cc5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects to 'song library' using 'scratch that, make it'. The final target is the song library, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/false_start/false_start_02_source_collection/ee5fa7e306159288/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "58e1f59fc12f94dba8a3a57c0dd70eeada5ac0859faf79df9a28f69bb852c7df", "seconds": 9.56, "spoken_utterances": [ "Download all the songs from my Spotify playlist—um... scratch that, make it song library ... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify playlist Amor' but immediately self-corrects with 'Scratch that, make it song library that I have liked.' This is a false start where the final, corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "Download all the songs from my Spotify playlist Amor. Scratch that, make it song library that I have liked." }, "candidate_id": "wav_ca80a44626b77fc4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'Scratch that, make it song library,' but the correction is abrupt and lacks the natural hesitation or filler words seen in the others. The transition from 'Amor' to 'song library' is very quick and the intonation doesn't fully convey a moment of reconsideration. The delivery is more robotic and less like a natural thought process, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Download all the songs from my Spotify song library. Oh, I forgot to mention, only download the songs that I have liked.", "long_pause_offsets": [ 53 ], "source_sha256": "467c4c5567706a8e2c3ed5728669a9930ad37e694e33ccac441adac209e0553b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'that I have liked', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/addition/addition_01/123d2501c0820b00/speech.wav", "wer": 0.04, "wav_sha256": "0269c92c3ea85aa0dcdc10b0f0a6a30ced5ae3b5930645572eab2f53f0d8ae24", "seconds": 13.76, "spoken_utterances": [ "Download all the songs from my uh... Spotify song library. ", "Oh, I forgot to mention, only download the songs let's see... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to download all songs from the Spotify library. After a pause and a self-correction, the speaker added the constraint 'only download the songs that I have liked'. This is a clear addition that narrows the scope of the original request. The final intent is to execute the download with the added constraint, which matches the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Download all the songs from my uh Spotify song library. Oh, I forgot to mention, only download the songs, let's see, that I have liked." }, "candidate_id": "wav_5d31de81afa6c98c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, uses a plausible filler ('uh') and a brief pause, then smoothly adds a restriction ('only download the songs... that I have liked') after a moment of hesitation ('let's see'). The self-correction from 'Spotify song library' to 'that I have liked' is integrated naturally within the flow, with a believable change of mind and intonation. The pacing and prosody are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Download all the songs from my Spotify that I have liked. One more thing: download from my Spotify song library.", "long_pause_offsets": [ 58 ], "source_sha256": "2a27aea7e0ca92aec0a086f9ff90790065db85454fcf34b8aa623680ade1f69b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source 'song library' to the initial request, refining the scope without changing the action, recipient, or liked status condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/addition/addition_02/31a81f2b72d389d7/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "8d182c10bf0573c7b8278417ead787901fd381d5543a96377ec0cf04f2268a63", "seconds": 12.52, "spoken_utterances": [ "Download all the songs from my um... Spotify that I have liked. ", "One more thing: download from my ... Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Download all the songs from my um Spotify that I have liked.' After a pause, they add 'One more thing, download from my Spotify song library.' This is an addition that narrows the scope of the original request. The final combined intent is to download songs from the 'Spotify song library' that have been liked, which matches the original instruction. The addition refines the target but does not contradict the original goal.", "heard_transcript": "Download all the songs from my um Spotify that I have liked. One more thing, download from my Spotify song library." }, "candidate_id": "wav_e65a8eb9ec8d8a62", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request and a natural filler ('um'), but the addition ('one more thing...') is delivered with a slightly abrupt, rehearsed tone. The second part of the request ('download from my Spotify song library') feels tacked on rather than a spontaneous thought, and the transition lacks the organic hesitation or intonation shift seen in audio_01. The overall delivery is fluent but less convincingly spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify song library that I have saved. I misspoke about the interaction: it should be 'liked', not 'saved'.", "long_pause_offsets": [ 71 ], "source_sha256": "4ac4d1b6a69f546497eee112d43487073debb76faceec3053b5c34a4580b2f8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'saved' to 'liked', while retaining the app, scope ('all songs from my library'), and action ('download')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/revision/revision_03_interaction_type/c5145394616ed46f/speech.wav", "wer": 0.12, "wav_sha256": "c082914b72b255cb5325605afda711c649ab54468b072163c956c11aacc87c5c", "seconds": 13.88, "spoken_utterances": [ "Download all the songs from my Spotify song library ... that I have saved. ", "I misspoke about the interaction: it should be 'ah... liked', not 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is a revision. The final audible intent is to download songs that are 'liked', matching the original instruction. The revision is clear and the correction is applied to the earlier instruction.", "heard_transcript": "Download all the songs from my Spotify song library that I have saved. I misspoke about the interaction, it should be uh liked, not saved." }, "candidate_id": "wav_a092ea94f1db1db7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a natural, spontaneous-sounding revision. The speaker begins with a clear request, then hesitates with 'I misspoke about the interaction,' followed by a smooth correction from 'saved' to 'liked.' The intonation and pacing are consistent with a real-time self-correction, and the pause before the correction feels like genuine recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Download all the songs from my Apple Music song library that I have liked. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "be60aa54e2ce9394e2a1b5d40386881499621eeeb8a8c7b5e601cca37727ba74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the action, scope (all liked songs), and library context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/revision/revision_01_app/15afc6ef7c1884b5/speech.wav", "wer": 0.08, "wav_sha256": "e52dc88471beae24c544dbe8f43e6d2c3b7d0af83848ec46da11633552c49d49", "seconds": 14.6, "spoken_utterances": [ "Download all the songs from my ... Apple Music song library uh... that I have liked. ", "Wait, the app should be um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests downloading liked songs from Apple Music, but immediately corrects the app to Spotify. This is a revision where the final intent is to execute the original task with the corrected target app. The original instruction's scope (download liked songs) is preserved, and the only change is the app name, which is corrected audibly.", "heard_transcript": "Download all the songs from my Apple Music song library uh that I have liked. Wait, the app should be um Spotify not Apple Music." }, "candidate_id": "wav_4c9184dd6afa6d55", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision from 'Apple Music' to 'Spotify,' but the correction is less fluid. The speaker says 'Wait, the app should be um... Spotify not Apple Music,' with a longer, more awkward pause and filler ('um') that disrupts the flow. The correction feels slightly more forced compared to audio_02, though still plausible as a spontaneous error." } ] }, "is_default": false }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify playlist that I have liked. Make that 'song library', not 'playlist'.", "long_pause_offsets": [ 67 ], "source_sha256": "1ce679fe6859c8f3de3e5b59f71a74b9bbc7d964e84a7a237263748ebf2bacae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'playlist' to 'song library' while preserving the action, app, and condition on liked songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/revision/revision_02_source_collection/beeb70175b5065ce/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "5ca6ed4843338e726c4335194a9f29a4062a1d2f71fab8b6a6a7055a94a77c68", "seconds": 19.0, "spoken_utterances": [ "Download all the songs from my um, let me think... Spotify playlist uh, let me think... that I have liked. ", "Make that 'uh, let's see... song library', not 'playlist'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify playlist' but then corrects it to 'song library' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to download from the 'song library', which matches the original instruction. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Download all the songs from my um, let me think, Spotify playlist, uh let me think, that I have liked. Make that uh, let's see, song library, not playlist." }, "candidate_id": "wav_e7cbe764e41b5e32", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains multiple fillers ('um', 'let me think', 'let's see') and two revisions, making it sound more hesitant and less confident. The speaker changes 'playlist' to 'song library' after a long pause, but the overall delivery feels less natural and more like a rehearsed or overly cautious formulation rather than a spontaneous thought process. The excessive fillers and backtracking reduce its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "91fff0a3b381b5d4bd16f9b0eb7fa078c4d46fc1a02a4667a2d5e33e6d200f4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/cancellation/cancel_024/aa006c54858aa1fc/speech.wav", "wer": 0.0, "wav_sha256": "f6b4228d5e1da5196a522a1e42dfe0d1f2499dcd2aaf2f7b7f7721fa7b8eab29", "seconds": 13.16, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked.", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to download liked songs from Spotify. After a brief pause and the interjection 'Hang on', they explicitly state 'Oh, never mind. Cancel the whole thing.' This constitutes a clear cancellation of the entire task as per the 'cancellation' type criterion. The final audible intent is to cancel, not to execute or pause.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked. Hang on. Oh, never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "2eac5f3457f53b29a3e00331640029f1d7ad5de2748dbce9036899b1f5f52e38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/ambiguous_halt/halt_034/ebf5b95e520c1bf4/speech.wav", "wer": 0.0, "wav_sha256": "7e2d11aba5ec546419d763c21605dd1b923e1a6435a569ddf9bfb70b639c27e9", "seconds": 12.48, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation. No further instruction is given to resume or cancel, so the agent should pause.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2a1b314753a067b7c08a5cd5bbccec865190b76b3bbc56c253ba6989e45d1977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_2/audio_natural/filler/original/38a3d3c0ad67c3ca/speech.wav", "wer": 0.0, "wav_sha256": "c31d79abe73e5644d5cc989331d1850244c8bb54df06b541ec87378164c8afc7", "seconds": 7.76, "spoken_utterances": [ "Download all the songs from my ... Spotify song library ... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchannel, or pause that would indicate a change in intent. The instruction is delivered as a single, complete utterance. Since the type is 'filler', any initial hesitation or pause within the instruction is ignored, and the final task meaning must match the original instruction. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Download all the songs from my Spotify song library that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify playlist collection—or rather, album library that I have liked.", "long_pause_offsets": [], "source_sha256": "e8dadc757b101f469c06e5855f4f467c549ff0c79bfde660ecaa6753f0ffc36e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately corrects to 'album library'. The final target is the album library, and all other conditions (download all songs, Spotify, liked status) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/false_start/false_start_02_source_collection/81d1b432034fb8ec/speech.wav", "wer": 0.1, "wav_sha256": "fda8db22f8be95ec68c6ce6ce73697fb24aedee1d23d3b55c7e7cf08558067e1", "seconds": 10.04, "spoken_utterances": [ "Download all the songs from my let's see... Spotify playlist collection—um... or rather, album library that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately self-corrects to 'album library'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction, with 'album library' replacing 'playlist collection'. All other constraints are preserved.", "heard_transcript": "Download all the songs from my let's see, Spotify playlist collection or rather album library that I have liked." }, "candidate_id": "wav_e021215093068811", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. It begins with a filler ('let's see'), includes a clear self-correction from 'playlist collection' to 'album library', and maintains a consistent, conversational pace. The intonation and rhythm are highly realistic, with the correction sounding spontaneous and integrated into the sentence rather than tacked on." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Download all the songs from my Apple Music—I should say Spotify album library that I have liked.", "long_pause_offsets": [], "source_sha256": "6510b978521f46e792084b69c2d399d463345ee94c21b5ecb9a036aa59842d41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the instruction including the action, scope ('all songs'), and condition ('from my album library that I have liked')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/false_start/false_start_01_app/15323b6829a1dd3a/speech.wav", "wer": 0.0, "wav_sha256": "99d5a83b74e6d7f8f6cd0719d27a296abbebd7703a1d2373b169a11afdf1af2f", "seconds": 11.48, "spoken_utterances": [ "Download all the songs from my uh, let me think... Apple Music—um... I should say Spotify album library that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify' after a pause and a self-correction phrase ('I should say'). According to the 'false_start' type criterion, the final corrected value must be applied, and the original mistaken value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Download all the songs from my uh let me think, Apple Music um, I should say Spotify album library that I have liked." }, "candidate_id": "wav_64986d2a0074043d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a realistic false start with 'uh' and 'let me think', followed by a self-correction from 'Apple Music' to 'Spotify album library'. The pauses and fillers are well-placed, and the correction is natural. However, the initial hesitation is slightly longer and the correction feels a bit more abrupt compared to audio_03, making it slightly less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify album library that I have saved—sorry, I meant liked.", "long_pause_offsets": [], "source_sha256": "610ef20338de0869553771e475ca9b378b3ec7619ebb2fe557dbdaaa20bc2785", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' with 'sorry, I meant liked.' The final intended action is to download songs that have been liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/false_start/false_start_03_interaction_type/0740f3ce1620d438/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "8303dd00e91b0531b3144840af66f6d31f7e858b2c0e823764487ad038a21b2c", "seconds": 11.16, "spoken_utterances": [ "Download all the songs from my Spotify album library ... that I have uh, let's see... saved—ah... sorry, I meant liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'album library' but immediately self-corrects to 'liked' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value 'liked' should be applied, and the original incorrect value 'album library' should be discarded. The final audible intent is to download songs from the 'liked' library, which matches the original instruction. The correction is immediate and clear, so the task is executed as intended.", "heard_transcript": "Download all the songs from my Spotify album library that I have uh let's see, see, sorry, I meant liked." }, "candidate_id": "wav_caeae1ad9f142e43", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a self-correction ('uh let's see... sorry I meant liked'), the delivery is less natural. The initial phrase is spoken fluently, but the correction is preceded by a long, awkward pause and a stutter ('see... see...'). The 'sorry I meant liked' feels more like a rehearsed or robotic insertion than a spontaneous thought, reducing its realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Download all the songs from my Spotify that I have liked. I left out a detail: download from my album library.", "long_pause_offsets": [ 58 ], "source_sha256": "61c81f60bff15e57c4f717f9e346625ace64bade92f8b1f22aab036c20242300", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source 'album library' to the initial request, combining it with the existing condition 'that I have liked'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/addition/addition_02/5877bc1057888398/speech.wav", "wer": 0.08, "wav_sha256": "368eff07136b874c57d95ab0d6f8e41994f695033612bfb2610a605bc91a69fa", "seconds": 13.48, "spoken_utterances": [ "Download all the songs from my um, let's see... Spotify that I have liked. ", "I left out a detail: download from my um... album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Download all the songs from my Spotify that I have liked.' The speaker then adds a clarifying detail, 'Download from my um album library,' which narrows the scope of the original request. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent is to download liked songs specifically from the 'album library' on Spotify, which matches the reference instruction. The speaker's hesitation ('um') and self-correction are typical of natural speech and do not alter the final intent.", "heard_transcript": "Download all the songs from my um let's see Spotify that I have liked. I left out a detail. Download from my um album library." }, "candidate_id": "wav_622bbf4d890600a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, uses a plausible filler ('um') and a self-correction ('I left out a detail') that leads to a smooth, additive revision of the source ('download from my album library'). The prosody is consistent, and the pause before the correction feels like genuine recollection. The delivery is continuous and sounds like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Download all the songs from my Spotify album library. Oh, I forgot to mention, only download songs that I have liked.", "long_pause_offsets": [ 54 ], "source_sha256": "9fa84613e099c7dd721dffb31426a98dd1432b41ac72d56fa129fe5d530a90cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only download songs that I have liked', which refines the initial broad request. The app, action, and source library remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/addition/addition_01/7ed952c7baf9eba8/speech.wav", "wer": 0.0, "wav_sha256": "7f1de637ca6bfc064bf9d4f8b88e59350a125ab7f39e07a5fa550e71091d37bd", "seconds": 14.48, "spoken_utterances": [ "Download all the songs from my ... Spotify album library. ", "Oh, I forgot to mention, only download songs um, let me think... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to download all songs from the album library. After a pause and a self-correction, the user added the constraint 'that I have liked'. This addition narrows the scope of the original request, and the final audible intent is to download only the liked songs. The addition type requires combining the initial instruction with the added information, which results in a goal that matches the reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Download all the songs from my Spotify album library. Oh, I forgot to mention, only download songs, um, let me think, that I have liked." }, "candidate_id": "wav_3d5eb79acdd08b9d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then adds a restriction ('only download songs that I have liked') after a pause and a filler ('um'). However, the initial request ('Spotify album library') is not corrected or revised, and the addition feels slightly tacked on rather than a natural follow-up. The prosody is slightly more hesitant, and the 'let me think' pause seems less integrated into the flow compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_interaction_type", "proposed_instruction": "Download all the songs from my Spotify album library that I have saved. I misspoke about the interaction: it should be liked, not saved.", "long_pause_offsets": [ 72 ], "source_sha256": "7b0182507bcf61c921ad4bd38e320e1c382db0b373839086a819708285dd44e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'saved' to 'liked', while retaining the app, scope (album library), and action (download all songs)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/revision/revision_03_interaction_type/86e52cf0d505e460/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "43d57966d0d819c0c7a3c570af92798de24cb17289957370f7636d913914ae31", "seconds": 15.64, "spoken_utterances": [ "Download all the songs from my uh... Spotify album library um... that I have saved. ", "I misspoke about the interaction: it should be uh... liked, not saved." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, which is a clear revision. The final intent is to download songs that are 'liked', matching the original instruction. The correction is applied to the earlier instruction, and all other conditions remain unchanged.", "heard_transcript": "Download all the songs from my uh Spotify album library um that I have saved. I misspoke about the interaction. It should be uh liked, not saved." }, "candidate_id": "wav_b8329c9b5a04e3fe", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a clear request, then self-corrects 'saved' to 'liked' with a believable, slightly apologetic tone ('I misspoke about the interaction'). The correction is integrated smoothly into the utterance, and the prosody remains consistent and conversational throughout. The pacing and intonation suggest genuine spontaneous thought, making it the most realistic." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Download all the songs from my Spotify playlist collection that I have liked. Make that album library, not playlist collection.", "long_pause_offsets": [ 78 ], "source_sha256": "22b928bf206bfab1465ae27ae8216581a5e9648ec370f7e49853f86441b1b072", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial collection type from 'playlist collection' to 'album library', preserving the app, action, and liked filter." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/revision/revision_02_source_collection/56897357e207243e/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "263dd5b1a5c7704ff74092edc415dc0eb5c90ce851e9329f05f487d5c43fac4c", "seconds": 13.16, "spoken_utterances": [ "Download all the songs from my uh... Spotify playlist collection um... that I have liked. ", "Make that album library, not playlist collection." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately corrects it to 'album library' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction, so the task is considered passed.", "heard_transcript": "Download all the songs from my uh Spotify playlist collection um that I have liked. Make that album library, not playlist collection." }, "candidate_id": "wav_c31ed2e8ddfc621a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision from 'playlist collection' to 'album library'. The speaker's intonation and pacing are natural, and the correction is delivered as a follow-up thought. However, the initial phrase 'download all the songs from my uh Spotify playlist collection' is slightly less fluent than audio_03, with a more noticeable hesitation ('uh') and a slightly awkward phrasing ('playlist collection'). The revision is believable but feels a bit more like a correction than a natural thought process." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Download all the songs from my Apple Music album library that I have liked. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 76 ], "source_sha256": "35be76f545fd00c5f80bc5162f7d5fb83070f893b464c8f4b46f32bdc5f64945", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the action, library type, and liked filter." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/revision/revision_01_app/62e019e55e486dbe/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "3616617472c93a707a4e3ba8a71ce70f4d2864bcf02355aa891c4f8043322241", "seconds": 15.44, "spoken_utterances": [ "Download all the songs from my uh... Apple Music album library um... that I have liked. ", "Wait, the app should be um, let's see... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction exactly, so the task should be executed.", "heard_transcript": "Download all the songs from my uh Apple Music album library um that I've liked. Wait, the app should be um, let's see, Spotify, not Apple Music." }, "candidate_id": "wav_2bff62108111c81e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision from 'Apple Music' to 'Spotify', but the delivery is less natural. The speaker starts with a clear request, then hesitates ('wait, the app should be um...'), followed by a long pause and a hesitant 'let's see'. This extended hesitation and the abrupt shift to the correction sound less spontaneous and more like a forced or rehearsed correction. The prosody is less fluid, and the overall impression is that the speaker is struggling to recall the correct information rather than naturally revising a thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "96ceb9f9a8f3edfeeeb5c9a4a0143a657025081c5c6b45c026e18f92d61082ce", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/cancellation/cancel_027/a91aaf28251cca85/speech.wav", "wer": 0.05, "wav_sha256": "8ff642de8a255d40b1a0799650770ee2252c00f49aaf05f13fd6f5a81f601c71", "seconds": 13.08, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked.", "Wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to download songs, then immediately follows with 'Wait, please disregard my last request,' which is a clear and unambiguous cancellation of the prior instruction. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked. Wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "8b7367b798ed6b08162d6d7eaf280be8d21faed4f0fb63e4036c2c679419c730", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/ambiguous_halt/halt_007/c37ee94b5bb2bf56/speech.wav", "wer": 0.0625, "wav_sha256": "423907938b792f290313dbb3fd77d95f6b613ac8bb0902ebcdb6ed174b5def53", "seconds": 11.12, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request and then clearly says 'One moment,' which is an audible request to hold or wait. This constitutes a final interjection that aligns with the 'ambiguous_halt' type, indicating a temporary pause rather than cancellation or execution. The original instruction is not revoked, and the speaker does not provide a new command, so the intent is to pause.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3ab5b8b_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "67103aabcc5101a98169d78af478c3952728bbcaed609d24d67f90c97f4fa22c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3ab5b8b_3/audio_natural/filler/original/7d267fe5c412c938/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "22594376761915329a92c2864258d66d0578d74e887e8f0c8a8b3ac55c688541", "seconds": 8.0, "spoken_utterances": [ "Download all the songs from my ... Spotify album library um... that I have liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler word ('um') but the final, complete instruction matches the original request. The user does not correct or modify the task, and the intent to download liked songs from the Spotify library is clear and unchanged.", "heard_transcript": "Download all the songs from my Spotify album library um that I have liked." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts—sorry, that should be payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "dffbb5b2d16efc381689ed8fbadc495ccef4a12a70bca88e87e7037977261aeb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'bank accounts' to 'payment cards' within the same utterance. The final target is 'expired payment cards', and the scope 'all my app accounts that have payment cards' remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/false_start/false_start_02_payment_method_type/8e953785046bb222/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "b8c3fe8b37ae8a5a273f9071ef141e25a674ac942fda86032cd2adec9d6a37e7", "seconds": 9.24, "spoken_utterances": [ "Remove expired bank accounts—um... sorry, that should be payment cards uh... from all my app accounts ... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bank accounts' but immediately self-corrects to 'payment cards' with 'sorry, that should be payment cards'. This is a false start where the corrected value replaces the mistaken one. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Remove expired bank accounts. Um, sorry, that should be payment cards. Uh, from all my app accounts that have payment cards." }, "candidate_id": "wav_1d3a16620ba598ae", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then naturally self-corrects from 'bank accounts' to 'payment cards' with a plausible 'um' and 'sorry' filler. The correction is immediate and integrated smoothly into the sentence, followed by a coherent continuation. The prosody is natural, with appropriate pauses and intonation that suggest spontaneous thought, making the false start sound authentic and well-executed." } ] }, "is_default": true }, { "variant_id": "false_start_03_account_scope", "proposed_instruction": "Remove expired payment cards from only my social media accounts—scratch that, make it all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "9c5b5258b2fc1f4a3713224cf26d704d0fba01ed39e8ae67404c9d6f404de984", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially restricts the scope to 'social media accounts' but immediately corrects it to 'all my app accounts that have payment cards', which matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/false_start/false_start_03_account_scope/57fdf82a194f8785/speech.wav", "wer": 0.0, "wav_sha256": "51e5f154749af738de5c49ab4bc84af847cb6970d037eb82aee9c2ff110a0c18", "seconds": 12.56, "spoken_utterances": [ "Remove expired payment cards from only my social media accounts—uh... scratch that, make it ... all my app accounts um... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only my social media accounts' but immediately self-corrects with 'scratch that' and specifies 'all my app accounts'. This is a false start where the final corrected instruction matches the original reference. The correction is immediate and unambiguous, so the final intent is to execute the corrected task.", "heard_transcript": "Remove expired payment cards from only my social media accounts. Uh, scratch that. Make it all my app accounts. Um, that have payment cards." }, "candidate_id": "wav_3913ee9fc8a10a48", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then says 'scratch that' to cancel the initial part, which is a believable pragmatic choice. However, the revision to 'all my app accounts' is less smoothly integrated, with a longer pause and a slightly awkward phrasing. The prosody is less fluid than audio_01, and the correction feels more like a mechanical replacement than a natural thought process, reducing its realism despite being a valid revision." } ] }, "is_default": false }, { "variant_id": "false_start_01_card_status", "proposed_instruction": "Remove active—no, I meant expired payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "6222063573d3ed9a91b76e57ec93133994b1fc05b232974c053e77298034d9f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'active' to 'expired' within the same utterance, preserving the target action, scope of accounts, and the condition of having payment cards." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove active—uh... no, I meant expired payment cards um, let me think... from all my app accounts uh, let's see... that have payment cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Remove payment cards from all my app accounts that have payment cards. Oh, I forgot to mention, only remove expired payment cards.", "long_pause_offsets": [ 71 ], "source_sha256": "436e5045f37c277c933fb38e0f70d4ff1f31d51ceec6473b15a94cf85b3ac0b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the removal to only expired cards, which is consistent with the original instruction's scope and does not contradict it. The final goal is a more specific version of the original." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove payment cards uh, let's see... from all my app accounts let me think... that have payment cards. ", "Oh, I forgot to mention, only remove um... expired payment cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Remove expired from all my app accounts that have payment cards. One more detail: remove expired payment cards.", "long_pause_offsets": [ 65 ], "source_sha256": "27a0924319385083e1a1058b00326a994ee86e38bffdf917ddfc6de41d56baa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing noun 'payment cards' to the initial instruction, clarifying that expired items are specifically expired payment cards. The scope 'all my app accounts that have payment cards' is preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove ... expired from all my app accounts ... that have payment cards. ", "One more detail: remove expired hmmm... payment cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "3f3c139_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_account_scope", "proposed_instruction": "Remove expired payment cards from all my social media accounts that have payment cards. Let me correct the scope: remove expired payment cards from all my app accounts, not social media accounts.", "long_pause_offsets": [ 88 ], "source_sha256": "45b69e6dd62daea75fec68ae4e44a02147cb3aab525793de7d532589c81fb4f7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the scope from 'social media accounts' to 'app accounts', which matches the original instruction. The action, condition, and target type are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/revision/revision_03_account_scope/ac362c3f2b0651f5/speech.wav", "wer": 0.027777777777777776, "wav_sha256": "9647e62218992fe1ee9db7fa8664b963e7bd2cfd3684fef0d93c3bca55abe1f5", "seconds": 17.12, "spoken_utterances": [ "Remove expired payment cards uh... from all my social media accounts ... that have payment cards. ", "Let me correct the scope: remove expired payment cards from let me think... all my app accounts, not social media accounts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests removal from 'social media accounts' but then corrects the scope to 'all my app accounts, not social media accounts' after a pause. This constitutes a revision where the final intent is the corrected version. The final instruction matches the original reference instruction exactly.", "heard_transcript": "remove expired payment cards uh from all my social media accounts that have payment cards. let me correct the scope, remove expired payment cards from, let me think, all my app accounts, not social media accounts." }, "candidate_id": "wav_b7f21f154b412e44", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_card_status", "proposed_instruction": "Remove active payment cards from all my app accounts that have payment cards. Wait, remove expired payment cards, not active ones.", "long_pause_offsets": [ 78 ], "source_sha256": "55d236bfe7d718405424ab2337a0c5c74b176d563b64ca19c58c5bc5a1bdef60", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the card status from 'active' to 'expired', while retaining the scope of all app accounts that have payment cards." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove um, let's see... active payment cards from all my app accounts hmmm... that have payment cards. ", "Wait, remove uh, let's see... expired payment cards, not active ones." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_payment_method_type", "proposed_instruction": "Remove expired gift cards from all my app accounts that have payment cards. Make that expired payment cards, not expired gift cards.", "long_pause_offsets": [ 76 ], "source_sha256": "003f5b3f39f716e036d3d1250ffa64959764ccfe1d0a77de812e9c1918f6f2c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial mistaken term 'gift cards' to 'payment cards', preserving the action, scope, and condition. The final goal matches the original instruction exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove expired gift cards ... from all my app accounts that have payment cards. ", "Make that expired ... payment cards, not expired gift cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "24c28bd694c8235b1a533bc4eed6544fba627bd74447c5cf7094050f1db58c3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/cancellation/cancel_021/cc85a22564eb8607/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "c4e58b76283bb10f7f925cc4f77020274341768ca83a278720a826032892b9f2", "seconds": 13.72, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards.", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and self-correction ('Hold on, never mind'), explicitly states 'I don't need that done anymore,' which is a clear cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards. Hold on, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "f10ab6f38c37e3865c92356fbaabc67e07c7826b88b60b36c61ee7091723706f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/ambiguous_halt/halt_023/136a329e8d2d7186/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "8dce352312d4484d6f46f379c16cfbb5c7209a84fc5341b43bb68cd1b42c9ce3", "seconds": 12.72, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Hold on, I need a second,' which is a clear and audible instruction to pause. This matches the 'ambiguous_halt' type criterion, where the final interjection is a request to wait or pause without canceling the task. The initial request is not executed, but the final intent is to pause, not to cancel or execute. Therefore, the control intent is correctly identified as 'pause' and the verdict is 'pass'.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "85d1c62aa2b95d1218cb3ff18839d29359ff10cb1671a085ac8754981723be25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_1/audio_natural/filler/original/26809c68e07d5393/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "7ea91e7cba661da669fe690733f91a77e9530b551a706242b389aa5713b101ee", "seconds": 8.4, "spoken_utterances": [ "Remove ah... expired payment cards um... from all my app accounts um, let's see... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitations ('uh', 'um') and a self-pause ('let's see'), but the final, complete instruction matches the original reference. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Remove uh expired payment cards um from all my app accounts um let's see that have payment cards." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_account_scope", "proposed_instruction": "Remove expired payment cards from all my email accounts—no, that should be all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "93ef8d961e127a1cc87de19cacc7bbdad05bbd3d929499cb749ba57e7b163b9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email accounts' to 'app accounts' within the same utterance, preserving the action, condition, and scope of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/false_start/false_start_03_account_scope/212d8872b82ed50c/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "5a8f7b77e7700757fc2ac46dabaa657c190f266ad29aeb99c8f3a600738b2d79", "seconds": 12.12, "spoken_utterances": [ "Remove expired payment cards uh, let me think... from all my email accounts—hmmm... no, that should be all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email accounts' but immediately self-corrects to 'app accounts' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Remove expired payment cards, uh, let me think, from all my email accounts. Hmm. No, that should be all my app accounts that have payment cards." }, "candidate_id": "wav_5ca7a71a96245eb0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage thought process. The initial request is made, followed by a believable 'let me think' filler and a pause, then a self-correction from 'email accounts' to 'app accounts'. The correction is smoothly integrated into the ongoing utterance with a natural prosodic shift, and the final sentence is delivered with a confident, continuous flow. The pacing and intonation suggest genuine spontaneous formulation." } ] }, "is_default": true }, { "variant_id": "false_start_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts—sorry, I mean payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "665dc2701ab9494e3ccdccad3457bb72fa4f3d211e74f872caa8e1a78f0932f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'bank accounts' to 'payment cards' within the same utterance, preserving the action, target scope, and condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/false_start/false_start_02_payment_method_type/6e34569b7c53edd3/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "acc0476969803662ad7186dd5d82d01ea4e88dd29f5ac5409c36e117c1cd3c61", "seconds": 11.6, "spoken_utterances": [ "Remove um... expired bank accounts—ah... sorry, I mean payment cards uh... from all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bank accounts' but immediately self-corrects to 'payment cards'. According to the false_start type criterion, the final corrected value should be applied. The rest of the instruction matches the original, so the final intent is clear and executable.", "heard_transcript": "Remove um expired bank accounts uh sorry I mean payment cards uh from all my app accounts that have payment cards." }, "candidate_id": "wav_aea6d311f9e19c4e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also shows a false start with 'bank accounts' corrected to 'payment cards', but the delivery is less fluid. The correction is abrupt, with a short, sharp 'uh' before the correction and another 'uh' after, creating a staccato effect. The initial phrase 'remove um expired bank accounts' is delivered with a flat, hesitant tone, and the overall prosody lacks the natural rhythm of the first speaker, making the correction feel more like a mechanical switch than a thought process." } ] }, "is_default": false }, { "variant_id": "false_start_01_card_status", "proposed_instruction": "Remove inactive—I meant expired payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "0257891dc77bc6621e8117a467aca96ef48a595fa3f7cdea66ecff38b649cb13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'inactive' to 'expired' within the same utterance, preserving the action, target, and scope of the instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove inactive—um... I meant ... expired payment cards from all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Remove payment cards from all my app accounts that have payment cards. Oh, I forgot to mention, only remove expired payment cards.", "long_pause_offsets": [ 71 ], "source_sha256": "436e5045f37c277c933fb38e0f70d4ff1f31d51ceec6473b15a94cf85b3ac0b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the removal to only expired cards, which is consistent with the original instruction's scope and does not contradict it. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/addition/addition_01/856a19192007209f/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "5f66fbae62d3c5f07d54c8f21af08eccaf5c200cbec644afa5349e478fd9c086", "seconds": 13.04, "spoken_utterances": [ "Remove payment cards uh, let's see... from all my app accounts uh... that have payment cards. ", "Oh, I forgot to mention, only remove expired payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a broad instruction to remove payment cards from all accounts that have them. After a pause, they add the constraint 'only remove expired payment cards'. Since this is an addition that narrows the scope of the original instruction, the final intent is to execute the original task with the added constraint. The final audible intent matches the reference instruction, which is to remove expired payment cards from all app accounts that have payment cards.", "heard_transcript": "Remove payment cards, uh, let's see, from all my app accounts, uh, that have payment cards. Oh, I forgot to mention, only remove expired payment cards." }, "candidate_id": "wav_6857e653823d63f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural hesitation ('uh', 'let's see') and a plausible, smoothly integrated revision. The correction ('only remove expired payment cards') is clearly a follow-up thought, not a mechanical insertion, and the prosody reflects a genuine change of mind. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Remove expired from all my app accounts that have payment cards. I left out a detail: these are payment cards.", "long_pause_offsets": [ 65 ], "source_sha256": "0110e84808e729981400088ceff5ecde507aacaa1fba9ebdceb3164e50e0a201", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that 'expired' refers to expired payment cards, which is consistent with the original instruction. The scope and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/addition/addition_02/11731da1c1fddbc2/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "e52ffa08ceb6a795f416744583348eae90ce8b892dbe281cdb852da7a436563c", "seconds": 12.24, "spoken_utterances": [ "Remove um... expired from all my app accounts ... that have payment cards. ", "I left out a detail: these are uh... payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, missing the word 'payment' before 'cards'. The speaker then adds the clarification 'These are uh payment cards', which completes the original request. The addition corrects the ambiguity without changing the scope or intent. The final audible task matches the original instruction.", "heard_transcript": "Remove um expired from all my app accounts that have payment cards. I left out a detail. These are uh payment cards." }, "candidate_id": "wav_ce90d289ccb1f0ec", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then adds a detail ('these are uh payment cards') after a pause. While the addition is plausible, the delivery feels slightly more abrupt and less fluid than audio_02. The filler 'uh' and the phrasing of the addition are less naturally integrated, making the overall flow feel a bit more like a correction than a spontaneous thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_account_scope", "proposed_instruction": "Remove expired payment cards from only my primary app accounts that have payment cards. Let me correct the scope: remove expired payment cards from all my app accounts, not just the primary ones.", "long_pause_offsets": [ 88 ], "source_sha256": "1232aa2056c23651ca82e3369f2df5c7d1e46b405dbcd45593d96555b75f44e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the scope from 'only my primary app accounts' to 'all my app accounts', which matches the original instruction. The action, target entity, and condition remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/revision/revision_03_account_scope/b113100ba1c7f878/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "ef4616f2e15ae2c92aa272d3c61a608a536a224773be5222cc160d3c82730509", "seconds": 17.84, "spoken_utterances": [ "Remove uh... expired payment cards um... from only my primary app accounts ... that have payment cards. ", "Let me correct the scope: remove expired payment cards from all my app accounts, not just the primary ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially restricts the action to 'primary app accounts' but then explicitly revises the scope to 'all my app accounts, not just the primary ones' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "remove expired payment cards from only my primary app accounts that have payment cards. let me correct the scope, remove expired payment cards from all my app accounts, not just the primary ones." }, "candidate_id": "wav_4e33d1ce65bd913c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural hesitation ('uh', 'um') and a smooth, self-corrected revision. The correction ('not just the primary ones') is integrated into the same utterance with a believable shift in intonation and pacing, showing a realistic thought process. The prosody is consistent, and the flow from initial statement to revision feels spontaneous and complete." } ] }, "is_default": true }, { "variant_id": "revision_01_card_status", "proposed_instruction": "Remove active payment cards from all my app accounts that have payment cards. Wait, remove expired payment cards, not active ones.", "long_pause_offsets": [ 78 ], "source_sha256": "55d236bfe7d718405424ab2337a0c5c74b176d563b64ca19c58c5bc5a1bdef60", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the card status from 'active' to 'expired', while retaining the scope of all app accounts that have payment cards." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/revision/revision_01_card_status/266f3794635642d3/speech.wav", "wer": 0.125, "wav_sha256": "c4fca9e82f4aaaa078999b2373d3f1eefeb0d60e727a65903a6e3cb92bf24aef", "seconds": 13.4, "spoken_utterances": [ "Remove um... active payment cards from all my app accounts uh... that have payment cards. ", "Wait, remove um... expired payment cards, not active ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests removing active payment cards, but after a pause and self-correction ('Wait'), clearly revises the instruction to remove expired payment cards instead. The revision is explicit and final, aligning with the original instruction's intent to remove expired cards. The final control intent is to execute the revised task.", "heard_transcript": "Remove um active payment cards from all my app accounts uh that have payment cards. Wait, remove um expired payment cards, not active ones." }, "candidate_id": "wav_e46629d72e751c7d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a request, hesitates with 'um' and 'uh', and then self-corrects from 'active' to 'expired' payment cards. While the correction is natural, the second part of the utterance ('wait, remove um expired...') feels slightly rushed and less integrated than in audio_02. The pause and filler placement is plausible, but the overall delivery is less fluid and the correction is less smoothly embedded in the sentence structure." } ] }, "is_default": false }, { "variant_id": "revision_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts from all my app accounts that have payment cards. Make that expired payment cards, not expired bank accounts.", "long_pause_offsets": [ 79 ], "source_sha256": "ced84e08b79c78f06194054d7d068a826ddddc0badf00d0e51a9f17e68f6cadf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'bank accounts' to 'payment cards', preserving the action, scope, and condition. The original instruction and final spoken goal match in all meaningful requirements." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove ... expired bank accounts let's see... from all my app accounts that have payment cards. ", "Make that expired ... payment cards, not expired bank accounts." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "e6e327cbc3b4e16305e1069f4eabed670e19c8ce6665aa3aca5de766e77e69d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/cancellation/cancel_006/b080d592138544c2/speech.wav", "wer": 0.0, "wav_sha256": "0da80eb6abf980d1936219a1dbfbb32a5ea8b8df07f495cbcaf18562974335ca", "seconds": 13.72, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards.", "Hold on, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hold on, please cancel the entire task,' which is a direct and unambiguous cancellation of the previously stated task. The initial part of the audio is a false start or filler, but the final interjection is a complete and explicit cancellation command. According to the rubric, a cancellation must clearly revoke the task, and this does so without requiring execution of the original request.", "heard_transcript": "remove, let me think, expired payment cards from all my app accounts that have payment cards. Hold on, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "f6e8a00d89a830ce9b7d4d19b5e63b560b6ab7b22c0d945c0f583ab00b2c1e15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/ambiguous_halt/halt_026/d523abed0c3db472/speech.wav", "wer": 0.0, "wav_sha256": "c31103b34839c6d1611e0469e477d36c7cfb460b0e649dba13b0dd87cbc01f63", "seconds": 13.4, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The initial request is not canceled or revised, and the speaker does not execute or complete the task. The final control intent is to pause, which aligns with the type-specific criterion.", "heard_transcript": "Remove, let me think, expired payment cards from all my app accounts that have payment cards. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "eccaf1e4ef9bb43e06c5d95cc6779d5e62d6d09fb52f7eda4d00eeadb67756b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_2/audio_natural/filler/original/f1dde9ca7ecc370e/speech.wav", "wer": 0.0, "wav_sha256": "e2c386b3f4c15664fe8f7c26edb12f58669ffea17b32c16013f8c417a8f38395", "seconds": 8.68, "spoken_utterances": [ "Remove let me think... expired payment cards from all my app accounts ... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial hesitation and filler phrase 'let me think' are internal processing cues and do not alter the final task goal. The speaker completes the original instruction without changing the action, target, or constraints. The final intent matches the reference instruction, so the task should be executed.", "heard_transcript": "Remove, let me think, expired payment cards from all my app accounts that have payment cards." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "3f3c139_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_card_status", "proposed_instruction": "Remove active—no, I meant expired payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "6222063573d3ed9a91b76e57ec93133994b1fc05b232974c053e77298034d9f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'active' to 'expired' within the same utterance, preserving the target action, scope of accounts, and the condition of having payment cards." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/false_start/false_start_01_card_status/3d045b0e2888b7cd/speech.wav", "wer": 0.15, "wav_sha256": "d455c38f6d7b3fb1af6ac231fbd7557619cc8cbf8d0f221d40335d5022adfe5a", "seconds": 9.8, "spoken_utterances": [ "Remove active—hmmm... no, I meant expired payment cards uh... from all my app accounts um... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'active' but immediately self-corrects to 'expired' with 'no, I meant expired payment cards'. This is a clear false start where the final corrected value is 'expired'. The rest of the instruction matches the original. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Remove active, no, I meant expired payment cards, uh, from all my app accounts, um, that have payment cards." }, "candidate_id": "wav_0832c444ec17fe3f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a false start ('active'), immediately self-corrects to 'expired payment cards' with a natural-sounding hesitation and intonation shift. The correction is followed by a smooth continuation of the full request. The pacing, filler use ('uh', 'um'), and prosody are highly consistent with spontaneous speech, and the correction feels genuinely integrated into the flow rather than tacked on." } ] }, "is_default": true }, { "variant_id": "false_start_03_account_scope", "proposed_instruction": "Remove expired payment cards from all my email accounts—wait, I meant all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "61e1ad2af05511f7de03c4698f3aed66d11045bedb0b95b8ad961d42270015d2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email accounts' to 'app accounts' within the same utterance, preserving the action, condition, and scope of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/false_start/false_start_03_account_scope/515f3ab0de783780/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "3f24f8183ff7a6c31370d26992481be524f5d4e91814615ac3c1ba0a6732a576", "seconds": 9.0, "spoken_utterances": [ "Remove expired payment cards uh... from all my email accounts—um... wait, I meant all my app accounts uh... that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email accounts' but immediately self-corrects to 'app accounts'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "Remove expired payment cards uh from all my email accounts. Um, wait, I meant all my app accounts uh that have payment cards." }, "candidate_id": "wav_214279c310db84f9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start ('email accounts') and a self-correction to 'app accounts'. The speaker's intonation and pacing are natural, and the correction is delivered with a believable hesitation. However, the initial phrase 'remove expired payment cards' is slightly rushed, and the correction feels a bit more abrupt than in audio_03, making the overall flow slightly less fluid despite being a strong example of a false start." } ] }, "is_default": false }, { "variant_id": "false_start_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts—sorry, make that payment cards from all my app accounts that have payment cards.", "long_pause_offsets": [], "source_sha256": "eac24c9254c82b168cce46f4c0234dfcce9138c8f66d66155c21d841f8481c4a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'bank accounts' to 'payment cards' within the same utterance. The final target is 'payment cards', matching the original instruction. The scope 'all my app accounts that have payment cards' is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/false_start/false_start_02_payment_method_type/4fae880916daaf61/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "10f69ab89db7876cd34d718b674767b32733f83f3806ae2f517fb92c25562a6e", "seconds": 8.56, "spoken_utterances": [ "Remove expired bank accounts—um... sorry, make that payment cards uh, let's see... from all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bank accounts' but immediately self-corrects to 'payment cards'. The final instruction matches the original reference exactly, with the correction applied as required by the 'false_start' type. The rest of the instruction is consistent and complete.", "heard_transcript": "Remove expired bank accounts, um, sorry, make that payment cards, uh, let's see, from all my app accounts I have payment cards." }, "candidate_id": "wav_355f85e833c008a6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a false start ('bank accounts') and a correction to 'payment cards', but the correction is followed by a new, unrelated phrase ('let's see') that disrupts the flow. The speaker then restarts the request, which sounds more like a rephrasing than a natural self-correction. The prosody is less consistent, and the overall delivery feels less spontaneous and more like a rehearsed or disjointed attempt, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Remove payment cards from all my app accounts that have payment cards. Oh, I forgot to mention, only remove expired payment cards.", "long_pause_offsets": [ 71 ], "source_sha256": "436e5045f37c277c933fb38e0f70d4ff1f31d51ceec6473b15a94cf85b3ac0b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that limits the removal to only expired cards, which is consistent with the original instruction's scope and does not contradict it. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/addition/addition_01/ba14515ecc19b323/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "64cfca7a9078bd6c083e147f876bb3e053d929b4bce0d4c806e0324407790c8a", "seconds": 13.92, "spoken_utterances": [ "Remove payment cards um, let me think... from all my app accounts um... that have payment cards. ", "Oh, I forgot to mention, only remove hmmm... expired payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Remove payment cards from all my app accounts that have payment cards.' After a pause and self-correction, the speaker adds the constraint 'only remove expired payment cards.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Remove payment cards, um, let me think, from all my app accounts, um, that have payment cards. Oh, I forgot to mention, only remove expired payment cards." }, "candidate_id": "wav_1a09e59fbf0084da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with plausible hesitation and self-correction. The initial request is clear, followed by a filler 'um' and a pause, then a revision ('Oh, I forgot to mention...') that smoothly adds a restriction. The prosody is consistent, and the flow reflects a realistic thought process. The correction is integrated naturally into the utterance, making it sound like a spontaneous clarification rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Remove expired from all my app accounts that have payment cards. I left out a detail: remove expired payment cards.", "long_pause_offsets": [ 65 ], "source_sha256": "6c678b05cb8e7109fce2eba219ef690188c789934e903b8ace27e6f20cf221dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing noun 'payment cards' to the initial instruction, completing the intended object of removal. The scope 'all my app accounts that have payment cards' and the condition 'expired' are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/addition/addition_02/4df82c183b18b985/speech.wav", "wer": 0.04, "wav_sha256": "1476c031246fa7461b8b55269e5f8d5046805ad7f502ddbd4520f8545fcd55bc", "seconds": 12.92, "spoken_utterances": [ "Remove expired ... from all my app accounts let's see... that have payment cards. ", "I left out a detail: remove expired let me think... payment cards." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Remove expired from all my app accounts that have payment cards', then corrects themselves by adding 'payment cards' after 'expired'. This is a clear addition that completes the original instruction. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Remove expired from all my app accounts, let's see, that have payment cards. I left out a detail. Remove expired, let me think, payment cards." }, "candidate_id": "wav_be95357c6225dbbd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a self-correction, but the flow is less natural. The speaker starts with 'Remove expired...', then pauses and says 'Let's see, they have payment cards', which sounds like a mumbled aside rather than a coherent thought. The correction 'Remove expired, let me think, payment cards' is awkwardly phrased and lacks the smooth transition seen in audio_02. The prosody is flatter, and the self-interruption feels more like a disjointed edit than a genuine moment of recollection." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "3f3c139_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_account_scope", "proposed_instruction": "Remove expired payment cards from only my primary app accounts that have payment cards. That should be all my app accounts, not only my primary app accounts.", "long_pause_offsets": [ 88 ], "source_sha256": "82b491f089778dcb46878966e2ca9817fb80c6dd0305f37c21e4bff1252ff9f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly reverses the restriction from 'only my primary app accounts' to 'all my app accounts', which contradicts the initial scope and changes the final goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/revision/revision_03_account_scope/fe0aae1e8dc71c7b/speech.wav", "wer": 0.1, "wav_sha256": "9251c404c660ccac1aef6347cbb38ac425aa940b153f9c0f264e40f1db94f5e9", "seconds": 14.92, "spoken_utterances": [ "Remove expired payment cards uh... from only my primary app accounts um... that have payment cards. ", "That should be hmmm... all my app accounts, not only my primary app accounts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially restricts the action to 'only my primary app accounts' but then corrects themselves with 'That should be hmm all my app accounts, not only my primary app accounts.' This constitutes a revision where the final intent is the original instruction. The correction is clear and applies to the earlier instruction, preserving all other conditions. Therefore, the final intent matches the original instruction, satisfying the revision type criterion.", "heard_transcript": "Remove expired payment cards uh from only my primary app accounts um that have payment cards. That should be hmm all my app accounts, not only my primary app accounts." }, "candidate_id": "wav_68d26a16255f72b3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and realistic speech patterns. It includes a false start ('uh') and a self-correction ('That should be hum... all my app accounts not only my primary app accounts'), which is a very common and credible way for a person to spontaneously correct a mistake mid-sentence. The intonation and pacing are consistent with someone thinking aloud, and the correction flows naturally within the same utterance. The pauses and fillers are appropriately placed and not overly long or robotic." } ] }, "is_default": false }, { "variant_id": "revision_02_payment_method_type", "proposed_instruction": "Remove expired bank accounts from all my app accounts that have payment cards. Use payment cards instead of bank accounts.", "long_pause_offsets": [ 79 ], "source_sha256": "0e0787e1999de76f65b75e884fa3e2c8c81e4f9e83a214e88e854b10c7cd5c50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the target from 'bank accounts' to 'payment cards', which reverses the intended action. The original instruction targets expired bank accounts, but the final instruction targets expired payment cards." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/revision/revision_02_payment_method_type/6a7b80c9cbc97a06/speech.wav", "wer": 0.0, "wav_sha256": "0670c1017e3f833c1795ec90f9d35e9fb0dad8e335da927100246873ebcd5422", "seconds": 10.48, "spoken_utterances": [ "Remove expired bank accounts ... from all my app accounts ... that have payment cards. ", "Use payment cards instead of bank accounts." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bank accounts' but immediately corrects to 'payment cards' after a pause, which is a revision. The final instruction matches the original reference exactly, so the revision is correctly applied and the task is executed as intended.", "heard_transcript": "Remove expired bank accounts from all my app accounts that have payment cards. Use payment cards instead of bank accounts." }, "candidate_id": "wav_9b782739514cb632", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision where the speaker changes 'bank accounts' to 'payment cards'. The correction is delivered as a separate sentence, which is a plausible way to revise a request. The speech is fluent and the intonation sounds natural. However, it lacks the spontaneous self-correction within a single utterance seen in audio_03, making it slightly less realistic in terms of natural speech flow." } ] }, "is_default": false }, { "variant_id": "revision_01_card_status", "proposed_instruction": "Remove active payment cards from all my app accounts that have payment cards. Change expired to active.", "long_pause_offsets": [ 78 ], "source_sha256": "e0d105a19ae6e49328534e9b9b6b14c6f46c87082421acc85e3a6560a2c8cd20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the condition from 'expired' to 'active', reversing the intended selection criteria." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/3f3c139_3/audio_natural/revision/revision_01_card_status/a58d33fbbf456f3c/speech.wav", "wer": 0.0, "wav_sha256": "efb7012a36a153567c5d68c01321f957cb932a9ad09962e96192ecf84ce586a8", "seconds": 10.32, "spoken_utterances": [ "Remove let's see... active payment cards ... from all my app accounts ... that have payment cards. ", "Change expired to active." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Remove let's see, active payment cards...' but then corrects the instruction by saying 'Change expired to active.' This is a revision where the speaker replaces the initial incorrect value 'active' with the correct value 'expired' as per the original instruction. The final audible intent is to remove expired payment cards, matching the original instruction. Therefore, the revision is correctly applied and the task should be executed.", "heard_transcript": "Remove let's see, active payment cards from all my app accounts that have payment cards. Change expired to active." }, "candidate_id": "wav_654ce0a87b813a20", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It begins with 'Remove let's see' which sounds like a prompt injection or a non-spontaneous start. The speaker then corrects 'expired' to 'active' in a way that feels more like a read-aloud correction than a natural self-repair. The delivery is flat and lacks the natural prosody and hesitation patterns of genuine spontaneous speech. The correction is abrupt and doesn't flow naturally within the sentence." } ] }, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "3f3c139_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "3317e48bbe778fbd9d4db36a1a3b9266e045896b9cbb3b67da574dbf655d2c49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove um... expired payment cards let's see... from all my app accounts that have payment cards.", "Hold on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "3f3c139_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "6492afdbb458c07383b3639cc9394be947e9f6c50da0ff13a4749dc154707ca4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove um... expired payment cards let's see... from all my app accounts that have payment cards.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "3f3c139_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "da905e818558cfa9cb2022b854dec4f2ee86a27a227784855793fe59ae8b84d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "Remove um... expired payment cards let's see... from all my app accounts that have payment cards." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "425a494_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the least—no, make that most in my Spotify song library?", "long_pause_offsets": [], "source_sha256": "ee135a49efa2d2af8384a0568b8cd680797a25a353a86ed32442261b8cd42733", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately corrects it to 'liked the most'. The final intended action is to determine the genre with the most liked songs, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/false_start/false_start_02_aggregation_order/eea4bfdb27020762/speech.wav", "wer": 0.1, "wav_sha256": "29b2701573f21c6a434c77e2d92650aa767f8fa78eed03cf526c505c6e39b84d", "seconds": 11.16, "spoken_utterances": [ "Songs of which ... genre have I um... liked the least—uh... no, make that most in my Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately self-corrects with 'no make that most'. According to the 'false_start' type criterion, the final corrected value ('most') should be applied, and the original incorrect value ('least') should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Songs of which genre have I um liked the least uh no make that most in my Spotify song library?" }, "candidate_id": "wav_252fc6e11d110f38", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody with a believable self-correction. The speaker starts with 'liked the least' and immediately corrects to 'most' with a clear change in intonation and a smooth transition. The filler 'um' is placed naturally, and the overall rhythm mimics genuine spontaneous speech. The correction is integrated into the same utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Songs of which genre have I liked the most in my Apple Music—sorry, I meant Spotify song library?", "long_pause_offsets": [], "source_sha256": "cd30fabd347b74de0f4d1db02e0791426383109db610d343b3976a357f0f3bfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the intent to find the genre with the most liked songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/false_start/false_start_03_app/e7054c86846d2f91/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "cd9782f4008adad97c3c41d45f3e902dbfb7454391ce271360d1b24529934a5e", "seconds": 10.96, "spoken_utterances": [ "Songs of which um, let's see... genre have I liked the most in my Apple Music—uh... sorry, I meant Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify song library'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Songs of which um let's see genre have I liked the most in my Apple Music uh sorry, I meant Spotify song library?" }, "candidate_id": "wav_1e8e1c72a867c806", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible self-correction from 'Apple Music' to 'Spotify song library' with a natural 'sorry' and a brief pause. The initial hesitation with 'um' and 'let's see' sounds authentic. However, the correction is slightly more abrupt than in audio_02, and the speaker's voice has a slightly more robotic quality, making the correction feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_01_grouping_attribute", "proposed_instruction": "Songs of which artist—or rather, genre have I liked the most in my Spotify song library?", "long_pause_offsets": [], "source_sha256": "8d65117c4bed9500cbcaa7422342c2012ca1df5fccbabfa1ee64604617b91fda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist' but immediately corrects to 'genre' within the same utterance. The final intended target is 'genre', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/false_start/false_start_01_grouping_attribute/a16d25d5cfd73b95/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "f2bccf31e393ce7f2b2b3fce9c54f20d45ea8256f8096eb23a6ff9db8c4e113e", "seconds": 13.16, "spoken_utterances": [ "Songs of which artist—uh... or rather, genre have I um... liked the most in my uh, let's see... Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist' but immediately self-corrects to 'genre' with 'or rather'. This is a clear false start where the final intended target is 'genre'. The rest of the instruction matches the original, so the final intent is to execute the original task with the corrected target.", "heard_transcript": "Songs of which artist uh or rather genre have I um liked the most in my uh let's see Spotify song library?" }, "candidate_id": "wav_c8510ef39841d482", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a false start ('artist or rather genre') and uses multiple fillers ('uh', 'um', 'let's see'). While this shows natural hesitation, the self-correction is less fluid and the overall delivery feels slightly more rehearsed or hesitant than the others. The multiple fillers and the way the correction is phrased ('or rather') make it sound less like a spontaneous thought process and more like a pre-planned edit." } ] }, "is_default": false }, { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify playlist—scratch that, make it song library?", "long_pause_offsets": [], "source_sha256": "db0b8e9ec3acbef1b8c20583134127c1a62cee9edb7eb849cc22e071495ad593", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects it to 'song library' within the same utterance. The final intended target is the song library, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/false_start/false_start_04_source_collection/d13bb3e28d858844/speech.wav", "wer": 0.05, "wav_sha256": "32c334e69701bbf8733c41bc3c19bcf3fb3e0e328fde7600cdae1d27038996dd", "seconds": 10.64, "spoken_utterances": [ "Songs of which genre have I liked the most in my Spotify ... playlist—hmmm... scratch that, make it song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects with 'Scratch that. Make it song library.' This is a clear false start where the final value 'song library' replaces the initial 'playlist'. The rest of the instruction remains unchanged. The final intent matches the original instruction, so the task is executed as corrected.", "heard_transcript": "Songs of which genre have I liked the most in my Spotify playlist? Hmm. Scratch that. Make it song library." }, "candidate_id": "wav_88e0ee819409b50e", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has a very flat, robotic delivery with little natural prosody. The false start ('playlist') is followed by a clear 'scratch that' and a correction, but the transition is abrupt and the intonation lacks the natural rise and fall of genuine speech. The pauses are too long and the overall rhythm is unnatural, making it sound like a synthesized or poorly recorded prompt rather than a spontaneous utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Songs have I liked the most in my Spotify song library? Oh, I forgot to mention, I want to know by genre.", "long_pause_offsets": [ 56 ], "source_sha256": "592c80e4051b0360d13a4b554be55e9b5521a4dda49945fa27de3b63bfb8c97c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'by genre' to the initial question, clarifying the desired grouping for the comparison. The core action (identifying the most-liked genre) and context (Spotify song library) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/addition/addition_01/9ef43d142aee5130/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "cc579b25df0bbbaf79fd659fe831cddf978de516522e342bf1acd4dcb455ac2a", "seconds": 14.84, "spoken_utterances": [ "Songs have I uh, let's see... liked the most in my Spotify song library? ", "Oh, I forgot to mention, I want to know by uh... genre." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about songs liked the most, then adds the constraint 'by genre' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "Songs have I uh let's see liked the most in my Spotify song library. Oh, I forgot to mention, I want to know by uh genre." }, "candidate_id": "wav_5e3d7dc24aef560d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see') and a self-correction ('songs have I...'), which is followed by a plausible addition of a missing detail ('Oh, I forgot to mention...') and a smooth continuation. The prosody is conversational, with appropriate pauses and filler words that suggest spontaneous thought. The addition feels integrated into the flow rather than tacked on." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Songs of which genre have I in my Spotify song library? I left out a detail: find the genre I've liked the most.", "long_pause_offsets": [ 56 ], "source_sha256": "da54def86a5fa75db2cf3cf1a6d3146d0a2db2b4fb913bfe22172a91a4e2edb0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'liked the most' to the initial incomplete question, completing the intended goal. The initial utterance is grammatically incorrect but the final goal is unambiguously reconstructed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/addition/addition_02/6178f6a40bb54e4d/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "73e12d4be75ba94a690ddc6bc923e47a79b8c881f15282aca4ca5d479ec8b614", "seconds": 13.6, "spoken_utterances": [ "Songs of which ... genre have I in my ... Spotify song library? ", "I left out a detail: find the genre I've um... liked the most." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance is incomplete ('Songs of which genre have I in my Spotify song library?'), but the speaker immediately self-corrects and adds the missing detail ('Find the genre I've um liked the most'). This constitutes an addition, where the final instruction combines the original intent with the added constraint. The final audible task matches the reference instruction, so the criterion is satisfied.", "heard_transcript": "Songs of which genre have I in my Spotify song library? I left out the detail. Find the genre I've um liked the most." }, "candidate_id": "wav_5485806208e04e4a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a missing detail, the initial phrasing ('Songs of which genre have I...') is slightly awkward and less fluent than audio_01. The addition ('I left out the detail...') is followed by a revision that sounds more like a rephrasing than a natural thought process. The overall delivery is less fluid, with a slightly robotic cadence and less convincing hesitation." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_grouping_attribute", "proposed_instruction": "Songs of which artist have I liked the most in my Spotify song library? Wait, I meant genre, not artist.", "long_pause_offsets": [ 72 ], "source_sha256": "173ede0c3ef6b04fcbd8d3511b5e48cac5700e9d575a3871b05cc29fe6c85441", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial request from 'artist' to 'genre', preserving the rest of the structure and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/revision/revision_01_grouping_attribute/17eb6bfa17e3eb89/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "b39a8190aebe3060c40e43199c081abb90a40839a542a7075454d49aa4054afd", "seconds": 13.28, "spoken_utterances": [ "Songs of which ah... artist have I ... liked the most in my uh... Spotify song library? ", "Wait, I meant genre, not artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'artist' but immediately corrects themselves to 'genre' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the genre, not the artist, which matches the original reference instruction.", "heard_transcript": "Songs of which uh artist have I liked the most in my uh Spotify song library? Wait, I meant genre not artist." }, "candidate_id": "wav_ed9c9bdc9ae736bd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('artist'), self-corrects to 'genre' with a natural-sounding hesitation and intonation shift, and continues the sentence smoothly. The correction is integrated into the flow without awkwardness, and the overall prosody sounds like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the least in my Spotify song library? Make that liked the most, not least.", "long_pause_offsets": [ 72 ], "source_sha256": "7d5fefeed490e667c70d15f07b9ec19c02dabb5850dc8ba621e63056c558ab96", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'liked the least' to 'liked the most', preserving the rest of the query including the app (Spotify), the entity (song library), and the aggregation (genre)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/revision/revision_02_aggregation_order/cf7ec5d60abf5ec8/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "dc3455a8a27fe6ed56deaaee72e91c60f3444f8101590e9858d5686936d59eda", "seconds": 13.8, "spoken_utterances": [ "Songs of which genre have I uh... liked the least in my Spotify song library? ", "Make that liked the ... most, not least." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the genre liked 'the least' but immediately corrects themselves to 'liked the most, not least'. This is a clear revision where the final intent is the corrected version. The original instruction is overridden by the self-correction, and the final goal matches the reference instruction exactly.", "heard_transcript": "Songs of which genre have I uh liked the least in my Spotify song library? Make that liked the most, not least." }, "candidate_id": "wav_f114a3cd1dafab31", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a revision from 'least' to 'most' after completing the initial request. The correction is clear and delivered with a believable change-of-mind intonation. The pacing and fillers ('uh') are natural, though the revision is slightly less fluid than audio_01 due to the pause before the correction." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Songs of which genre have I liked the most in my Apple Music song library? Let me correct the app: use Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "cceee8826c84c75df8a500c921f210c90cc8df4b2ba4bc719d7526219cb817f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the original question about the genre with the most liked songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/revision/revision_03_app/c91f4e92a62c73e8/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "9ae2230167d134ae90cffff79dbdee062c8c6412633ac28fd9db1f26422d25b8", "seconds": 17.64, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my let's see... Apple Music song library? ", "Let me correct the app: use hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction exactly, with the target app corrected from Apple Music to Spotify.", "heard_transcript": "songs of which um genre have I liked the most in my let's see, Apple Music song library. Let me correct that. Use hmm, Spotify, not Apple Music." }, "candidate_id": "wav_a2fe47a916dd9d4b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a filler ('um'), then makes a correction from 'Apple Music' to 'Spotify'. The correction is delivered with a self-aware tone ('let me correct the...'), but the phrasing is slightly less natural and the flow feels a bit more rehearsed. The filler and correction are present but less seamlessly integrated than in the top two." } ] }, "is_default": false }, { "variant_id": "revision_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify playlist? Change the source to song library, not playlist.", "long_pause_offsets": [ 67 ], "source_sha256": "76a79b9e5c91db320a1036350d3919b9fd57b4396321326b6369726e1d52e601", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source from 'playlist' to 'song library', which matches the original instruction. The rest of the query, including the action and the scope of 'genre' and 'liked the most', remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/revision/revision_04_source_collection/8dfbd4e869bad082/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "ae12474c169ffbdee154dc9c17ac2deb9d6d6d1cc5d1bdb7896515c6cadf30aa", "seconds": 12.68, "spoken_utterances": [ "Songs of which uh... genre have I um... liked the most in my Spotify playlist? ", "Change the source to song library, not playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'Spotify playlist' but immediately corrects the source to 'song library' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction, and the control intent is to execute the revised task.", "heard_transcript": "Songs of which uh genre have I um liked the most in my Spotify playlist? Change the source to song library, not playlist." }, "candidate_id": "wav_c14ab14d8515aa34", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker begins with fillers ('uh', 'um'), then corrects 'playlist' to 'song library'. The correction is made after the initial request is fully formed, but the phrasing 'change the source to...' sounds more like a mechanical instruction than a spontaneous thought. The delivery is slightly less fluid and more hesitant overall, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "f691939bf8b9414b16a088270febc9360ca78e9d80dc533d5110f8b570c27b77", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/cancellation/cancel_011/818b5210ff7f1572/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "359474e8d86e5cee8b4b28debb5775926c2bea21a2bd237b788fee2355bbd773", "seconds": 14.12, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?", "Hold on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question about liked genres in their Spotify library. After a brief pause and the phrase 'Hold on, I've changed my mind,' they explicitly state 'Cancel that.' This constitutes a clear cancellation of the initial task. The final intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library? Hold on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "6716a76d5ecd5486635fca476c14b1432a95c49b523b4b1c83b15951019f2926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/ambiguous_halt/halt_010/410943e799ea9864/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "9af26c97313c213ebfd694f8aa36f3a650a5a5f33a477b1fe69f6174eb4268e9", "seconds": 12.24, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Hold on a second,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled, but the speaker indicates a need to pause before proceeding.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library? Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "063f534b7c66a106fca4dcb9a81aa6c6876f9bcaabffdda36c2a7499d1b0654f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_1/audio_natural/filler/original/e4db8d33b59b7174/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "405d9217d6680f16590c2ecbcd1d0aed006518ed3c3e09d9e1b3a487b29768af", "seconds": 8.88, "spoken_utterances": [ "Songs of which um... genre have I liked the most in my ... Spotify song library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Songs of which um genre...' and then continues the original instruction without any correction or change in intent. The 'um' is a filler word and does not alter the task goal. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Songs of which um genre have I liked the most in my Spotify song library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the least in my Spotify playlist—or rather, album library?", "long_pause_offsets": [], "source_sha256": "dec6df954f7d20491961585e15ec33d42d68dd31f4c33a3d27c4d7759c39d7f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects it to 'album library' using 'or rather', preserving the rest of the query including the app (Spotify), the action (finding the genre liked the least), and the scope (library)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/false_start/false_start_04_source_collection/e0a419e938614e46/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "9363de1402175fb4428745a5ca44e9ef246bc810c75b182307ebe91cf3566172", "seconds": 10.44, "spoken_utterances": [ "Songs of which genre have I liked the least in my hmmm... Spotify playlist—uh... or rather, album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects to 'album library' with 'Or rather'. This is a clear false start where the final, corrected target is 'album library'. The rest of the instruction matches the original. The final intent is to query the genre with the least liked songs in the album library, which aligns with the original instruction.", "heard_transcript": "Songs of which genre have I liked the least in my Spotify playlist? Or rather, album library?" }, "candidate_id": "wav_3d32ea07b30d49a9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a clear request, then self-corrects from 'playlist' to 'album library' with a believable 'uh' and 'or rather' transition. The prosody is fluid, with a natural hesitation and a smooth, pragmatic correction that sounds like a genuine moment of recall. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_01_grouping_attribute", "proposed_instruction": "Songs of which artist—I meant genre have I liked the least in my Spotify album library?", "long_pause_offsets": [], "source_sha256": "5bdd5d486278cf56857d9ba789935c052b8bf16f76e55ed1f450a157f265b2cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'artist' to 'genre' within the same utterance, preserving the rest of the structure and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/false_start/false_start_01_grouping_attribute/1857f276dc7d0e7b/speech.wav", "wer": 0.05, "wav_sha256": "f500c2ff5ce3b252be99eac6be418029df3149b3d68b5e80bb340f24b24c39c7", "seconds": 10.08, "spoken_utterances": [ "Songs of which artist—um... I meant genre have I liked the least in my Spotify let's see... album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist' but immediately self-corrects to 'genre' with 'I meant genre'. This is a clear false start where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to find the genre with the least liked songs in the Spotify album library.", "heard_transcript": "Songs of which artist, um, I meant genre, have I liked the least in my Spotify, let's see, album library?" }, "candidate_id": "wav_fcca99b82ea732e0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'artist' corrected to 'genre', followed by a natural 'let's see' pause. The correction is immediate and the flow continues smoothly. The speaker's voice and prosody are consistent, and the hesitation feels authentic. The only minor drawback is that the correction is slightly more abrupt than in audio_04, but still very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the most—no, make that least in my Spotify album library?", "long_pause_offsets": [], "source_sha256": "e1774738e10444a5a23b4a6721a9cf9a1f531a1b413940f6a718b524bf7fad3f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most' but immediately corrects to 'least' within the same utterance. The final intended comparison is 'least', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/false_start/false_start_02_aggregation_order/b8c5256cf2586089/speech.wav", "wer": 0.1, "wav_sha256": "27858adc04d89679497ba74b848e482446d2b4faf9ce4354dd6fcca46a4ac073", "seconds": 10.96, "spoken_utterances": [ "Songs of which genre have I liked the most—uh... no, make that ... least in my Spotify um... album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most' but immediately self-corrects to 'least' with 'No, make that least'. This is a clear false start where the final intent is the corrected version. The rest of the instruction, including the context 'in my Spotify album library', remains unchanged and is preserved. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Songs of which genre have I liked the most of? No, make that least in my Spotify um album library?" }, "candidate_id": "wav_83f65614096fcef2", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a false start where the speaker says 'most of' and then corrects to 'least'. The correction is clear and the speaker continues without a pause, which is plausible. However, the initial error ('most') is less likely to be made in this context than 'artist' or 'playlist', making the correction feel slightly less natural. The delivery is otherwise fluent and realistic." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Songs of which genre have I liked the least in my Apple Music—sorry, I should have said Spotify album library?", "long_pause_offsets": [], "source_sha256": "20bd1c0760942c762a53764e721524e7cf3b98fe2dd0d5d8f574475cacf0e298", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query structure and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/false_start/false_start_03_app/351e0533d5b5d6bb/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "6a5dd94cbdaf00331e79d8ac925018ce725b096bf704823438ec4e0662216aa6", "seconds": 12.72, "spoken_utterances": [ "Songs of ... which genre have I liked the least in my Apple Music—ah... sorry, I should have said Spotify um... album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('Apple Music') is discarded. The rest of the instruction, including 'songs of which genre have I liked the least' and 'album library', remains unchanged and is preserved. The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Songs of which genre have I liked the least in my Apple Music, uh, sorry, I should've said Spotify, um, album library?" }, "candidate_id": "wav_5b008c83b5fb6008", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start where the speaker says 'Apple Music' and corrects to 'Spotify'. However, the correction is followed by a long, unnatural pause ('um') that disrupts the flow. The speaker does not continue the sentence after the pause, making the clip feel incomplete or awkward. The prosody also seems slightly off, with a robotic or rehearsed quality compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Songs have I liked the least in my Spotify album library? Oh, I forgot to mention, I want to know which genre I've liked the least.", "long_pause_offsets": [ 58 ], "source_sha256": "fb090e7a0f0c6fcdf40339157d7d2490fef63f99e59d02d24ec4031c082162b8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'genre' specification to the initial question, clarifying that the user wants to know the genre associated with the least-liked songs. The core intent—identifying the least-liked genre—is preserved and completed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/addition/addition_01/7988c545725234a1/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "f7c0cbb83573643481426487f2c575aa9ce93ffe5b1a2bdc3545ea8b7320fba4", "seconds": 13.28, "spoken_utterances": [ "Songs have I liked the least in my ... Spotify ... album library? ", "Oh, I forgot to mention, I want to know which genre I've liked the least." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Songs of which genre have I liked the least in my Spotify album library?'. This is followed by a longer pause and an addition: 'Oh, I forgot to mention, I want to know which genre I've liked the least.' The addition clarifies and refines the original request, making it more direct and removing the slightly awkward phrasing 'Songs of which genre'. The final intent is clearly to find the genre liked the least, which matches the original instruction's goal. The addition does not introduce any conflicting or contradictory information. Therefore, the final intent is consistent with the original instruction and the addition serves to clarify it, satisfying the 'addition' type criterion.", "heard_transcript": "Songs of which genre have I liked the least in my Spotify album library? Oh, I forgot to mention, I want to know which genre I've liked the least." }, "candidate_id": "wav_df74918a7ad3012d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('Songs have I liked the least in my Spotify album library'), then naturally adds a clarification ('Oh, I forgot to mention...') with a believable filler and a smooth, additive correction. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition feels like a genuine memory or oversight, not a rehearsed or robotic insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Songs of which genre have I liked the in my Spotify album library? I left out a detail: it should be the genre I've liked the least.", "long_pause_offsets": [ 67 ], "source_sha256": "bda8c7676d790a89a7c3430031b9496754471e5926da44e192735b927ae8b91d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'liked the least', which clarifies the original incomplete question. The initial utterance is grammatically flawed but the final goal is unambiguous and matches the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/addition/addition_02/9f07db15d315e5bb/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "35cca96400fb2340573f24bd953ac1e9310f9c4328082921cea0c8d20f991031", "seconds": 13.92, "spoken_utterances": [ "Songs of ... which genre have I liked the in my Spotify ... album library? ", "I left out a detail: it should be the genre I've liked the um... least." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the genre liked the least, then adds a clarification that the request should be for the genre liked the least. The final intent is identical to the original instruction, and the addition serves only to correct a perceived omission. The type is 'addition', and the final goal matches the original instruction exactly.", "heard_transcript": "Songs of which genre have I liked the least in my Spotify album library? I left out a detail, it should be the genre I've liked the um least." }, "candidate_id": "wav_005a68c4dd1e6935", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly awkward phrasing ('Songs of which genre have I liked the...') and then corrects it mid-sentence ('I left out a detail, it should be the genre I've liked the...'). While the self-correction is present, the initial phrasing is less natural and the correction is delivered with a flatter, more robotic intonation. The flow feels less like a spontaneous thought process and more like a rehearsed or slightly forced revision, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Songs of which genre have I liked the least in my Apple Music album library? Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "25f261a0b3c00e607dc0739809ef218886aff9f5fc38a718417f4b0816b107ed", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, preserving the rest of the query including the action, scope, and condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/revision/revision_03_app/3f93ccc20de5a9e1/speech.wav", "wer": 0.04, "wav_sha256": "df1f632b85360a9481fba118e6d0482b7c80ad39c17f1f60eb3e87c3393926c7", "seconds": 13.84, "spoken_utterances": [ "Songs of ... which genre have I liked the least in my hmmm... Apple Music album library? ", "Wait, the app should be Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Songs of which genre have I liked the least in my Apple Music album library? Wait, the app should be Spotify, not Apple Music." }, "candidate_id": "wav_b5be0fc82970dcec", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then hesitates with a realistic 'um' before correcting the app name from 'Apple Music' to 'Spotify'. The correction is delivered with a believable shift in intonation and pacing, sounding like a spontaneous realization rather than a rehearsed edit. The flow is continuous and the self-correction feels integrated into the thought process." } ] }, "is_default": true }, { "variant_id": "revision_01_grouping_attribute", "proposed_instruction": "Songs of which artist have I liked the least in my Spotify album library? I misspoke about the grouping attribute; use genre instead of artist.", "long_pause_offsets": [ 74 ], "source_sha256": "95cd737f8b60f8cb3e051df95ba18d86293ef0330cfa6acb8e3f33decf1a22d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the grouping attribute from 'artist' to 'genre', which matches the original instruction. The rest of the query, including the app (Spotify), the scope (album library), and the action (finding the least-liked genre), remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/revision/revision_01_grouping_attribute/156ff1b412b56472/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "b1a687b8773441636d7ef040522eeb333b81e70c931a3652a2a2e0ef06f5a674", "seconds": 13.68, "spoken_utterances": [ "Songs of which artist have I liked the least in my Spotify uh... album library? ", "I misspoke about the grouping attribute; use ... genre instead of artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'artist' but immediately corrects themselves, stating 'I misspoke about the grouping attribute. Use genre instead of artist.' This is a clear revision where the final intent is to group by 'genre'. The correction is spoken after a pause and directly modifies the earlier instruction, satisfying the revision type criterion. The final goal matches the original reference instruction.", "heard_transcript": "Songs of which artist have I liked the least in my Spotify uh album library? I misspoke about the grouping attribute. Use genre instead of artist." }, "candidate_id": "wav_ef09303ab51caf7e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'uh' before correcting the grouping attribute from 'artist' to 'genre'. The correction is delivered with a natural, slightly self-deprecating tone ('I misspoke'), and the intonation shift is credible. The pacing and filler usage are realistic, though the correction is slightly more formal than in audio_01." } ] }, "is_default": false }, { "variant_id": "revision_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the least in my Spotify playlist collection? Make that album library instead of playlist collection.", "long_pause_offsets": [ 79 ], "source_sha256": "227621fa6efd1d7165d360d9e09a2f6b206cbbb0066b8e812ed6ee91dbe4eea3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the collection type from 'playlist collection' to 'album library', preserving the rest of the query including the app (Spotify), the action (finding the genre liked the least), and the scope (my library)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/revision/revision_04_source_collection/f2912a4acf8121be/speech.wav", "wer": 0.24, "wav_sha256": "a44969c76aa587a7a6f9f355104b83fd7cb2740fc38de654ea56c357f7f23265", "seconds": 13.6, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my hmmm... Spotify um... playlist collection? ", "Make that album library instead of playlist collection." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist collection' but immediately corrects it to 'album library' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the genre liked the least in the 'Spotify album library', which matches the original instruction. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Songs of uh, which genre have I liked the least in my Spotify um playlist collection? Make that album library instead of playlist collection." }, "candidate_id": "wav_644a29061c0e9048", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a hesitant 'uh' and 'huh', then corrects 'playlist collection' to 'album library'. The correction is clear and the intonation shift is believable. However, the initial hesitation is more prolonged and less fluid than in the top two, and the overall delivery feels slightly more fragmented, possibly due to the multiple fillers and the correction being a bit more abrupt." } ] }, "is_default": false }, { "variant_id": "revision_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify album library? Change the aggregation to least, not most.", "long_pause_offsets": [ 72 ], "source_sha256": "47bec7064229aa498ec551c69f76160e8e9b9acc4ca80fea054abfc2e82a2d26", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the aggregation from 'most' to 'least', preserving the rest of the query including the app (Spotify), context (album library), and the question about genres." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/revision/revision_02_aggregation_order/91eadc8befecc99e/speech.wav", "wer": 0.04, "wav_sha256": "ef01bff59bd9ab6333b1de13d73babf350b7ab11dacd91cc44a12d22488829b1", "seconds": 13.12, "spoken_utterances": [ "Songs of ... which genre have I liked the most in my Spotify uh, let me think... album library? ", "Change the aggregation to least, not most." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the genre liked 'the most', but immediately corrects this to 'least' after a pause and self-correction. The revision is clear and final, changing the aggregation from 'most' to 'least'. The rest of the instruction, including the context of 'Spotify album library', remains unchanged. This matches the revision type criterion where the correction spoken after a pause overrides the initial incorrect value.", "heard_transcript": "Songs of which genre have I liked the most in my Spotify uh let me think album library? Change the aggregation to least, not most." }, "candidate_id": "wav_0c03a04dc4047227", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear question, then hesitates with 'uh' and 'let me think', followed by a correction from 'most' to 'least'. While the hesitation and correction are plausible, the phrasing 'let me think' is a bit more explicit and less natural than the other fillers. The correction is delivered with a slightly robotic intonation, and the overall flow feels less spontaneous compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "a87512bd6415c70502f13164a0f1d7fe3b662158793bb70cf87c04e98a23e753", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/cancellation/cancel_024/a2974420db46987e/speech.wav", "wer": 0.0, "wav_sha256": "d5ef4e744836defcdbccfb93452f0beb8fe0bbeefaeea293aef72bd7768181ca", "seconds": 13.08, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?", "Hold on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly states 'Hold on, oh never mind, cancel the whole thing' as the final interjection, which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Songs of uh which genre have I liked the least in my um Spotify album library? Hold on, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "12db936dd09889e079c17b6f0e2dd9399e82a79ebc135cad888e2dc5b2dd92b5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/ambiguous_halt/halt_005/c92a00fb1cc0af83/speech.wav", "wer": 0.0, "wav_sha256": "38682a5f3a95bde19e22ca196602b70d41298d21f67e7b8c476121af710b9ec2", "seconds": 10.92, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'One second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, but the speaker explicitly halts the process.", "heard_transcript": "Songs of uh, which genre have I liked the least in my um, Spotify album library? One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e98bfb3414545339d56d8a884cab320298d43ca823c30fff33f1f278d9f9fa1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_2/audio_natural/filler/original/66baf2865b78e133/speech.wav", "wer": 0.0, "wav_sha256": "50843cd29ddaa805aaa094ec6f3a83b048fc1cba324afdf9c985cd6ba4de8694", "seconds": 7.76, "spoken_utterances": [ "Songs of uh... which genre have I liked the least in my um... Spotify album library?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Songs of uh' and 'in my um', which are fillers. The core instruction is clearly stated as 'which genre have I liked the least in my Spotify album library?'. The final intent matches the original instruction, and the fillers do not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Songs of uh, which genre have I liked the least in my um, Spotify album library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify albums—or rather, playlists?", "long_pause_offsets": [], "source_sha256": "8dde9977084aab9278d963a249597c6b47e3039550c237d15e7f7e3f63b76131", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'albums' to 'playlists' within the same utterance, preserving the rest of the query including the app (Spotify), the action (finding liked songs), and the scope (genre)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/false_start/false_start_04_source_collection/2b35a51415b5b8b1/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "362f127bff3c3d780e4cecd075b176732a68c5a19bf47c705527be338895e999", "seconds": 8.24, "spoken_utterances": [ "Songs of which genre have I uh... liked the most in my Spotify ... albums—ah... or rather, playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with 'uh rather playlists'. This is a false start where the final, corrected target is 'playlists'. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final intent is to query liked genres in Spotify playlists.", "heard_transcript": "Songs of which genre have I uh liked the most in my Spotify albums, uh rather playlists." }, "candidate_id": "wav_f02fb929d916e6a6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding self-correction from 'albums' to 'playlists' with a clear, spontaneous 'uh' and 'rather' as a connector. The prosody is consistent, the pace is conversational, and the repair feels like a genuine moment of recall rather than a rehearsed edit. The intonation and rhythm are the most fluid and human-sounding among all clips." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Songs of which genre have I liked the most in my Apple Music—sorry, make that Spotify playlists?", "long_pause_offsets": [], "source_sha256": "ffe789cf1cea69a5ef11264ba204042a7846f093c0b5f23bf31313b28a01bfc3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the app, the action, and the scope of playlists." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/false_start/false_start_03_app/acee1110820a1a35/speech.wav", "wer": 0.05, "wav_sha256": "2840a484bc2320db0e3b62d34be09b7e04ee4283045b96c1bf99bdba65b09ac7", "seconds": 12.0, "spoken_utterances": [ "Songs of which genre have I ... liked the most in my Apple Music—hmmm... sorry, make that um... Spotify playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Songs of which genre have I liked the most in my Apple Music, sorry, make that um, Spotify playlists?" }, "candidate_id": "wav_10fa69353b1d17df", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable false start with 'Apple Music' followed by a self-correction to 'Spotify'. The 'sorry' and 'make that' are natural hesitation markers, and the repair is integrated smoothly. The prosody is slightly less fluid than audio_03, with a slightly more abrupt transition, but the overall delivery remains realistic and conversational." } ] }, "is_default": false }, { "variant_id": "false_start_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the least—no, make that most in my Spotify playlists?", "long_pause_offsets": [], "source_sha256": "a8bd69061cec57b9d372d9b3172dbebe4755d354135a51f3a94098b19b01d6cc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately corrects it to 'liked the most'. The final intended action is to determine the genre with the most liked songs, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/false_start/false_start_02_aggregation_order/31f80f38d3cac16e/speech.wav", "wer": 0.15, "wav_sha256": "5bf5d6ff86d35ebbbc620a3ce2b769234d9a431322c13be4e2b760d3738dfd3c", "seconds": 8.92, "spoken_utterances": [ "Songs of which genre have I um... liked the least—uh... no, make that um... most in my Spotify playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'liked the least' but immediately self-corrects with 'no, make that most', which is the final intended value. The 'false_start' type requires applying the correction, and the final intent matches the original instruction exactly.", "heard_transcript": "Songs of which genre have I um liked the least uh no, make that um most in my Spotify playlists?" }, "candidate_id": "wav_7c36a4e84443dc39", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start with 'least' and corrects to 'most'. The 'uh' and 'no make that' are plausible, but the correction feels slightly rushed and less integrated than in audio_03 or audio_04. The prosody is somewhat flat, and the overall delivery is less fluid, making it less realistic than the top two." } ] }, "is_default": false }, { "variant_id": "false_start_01_grouping_attribute", "proposed_instruction": "Songs of which artist—I meant genre have I liked the most in my Spotify playlists?", "long_pause_offsets": [], "source_sha256": "80b50737afd5b5ee9ea43552eb25b5b4a4d745e660d5997fa6f7ac63041ea489", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'artist' to 'genre' within the same utterance, preserving the rest of the structure and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/false_start/false_start_01_grouping_attribute/da6ef0bf0ad79155/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "7c043f644f47df06a72cef4ce8f06a8cce5b036ccce90a4ff4a72cd99c0e03cd", "seconds": 8.12, "spoken_utterances": [ "Songs of which artist—um... I meant genre have I uh... liked the most in my Spotify playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist' but immediately self-corrects to 'genre' after a pause and filler. According to the 'false_start' type criterion, the final corrected value ('genre') should be applied, and the original mistaken value ('artist') should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Songs of which artist some, um, genre have I uh liked the most in my Spotify playlists?" }, "candidate_id": "wav_76534fa74883246d", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a very flat, robotic delivery with unnatural pauses and a lack of natural intonation. The speaker seems to be reading a script rather than formulating a spontaneous request. The 'um' and 'uh' are awkwardly placed and do not sound like genuine hesitation. The overall prosody is monotonous, making it the least realistic of the four." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Songs of which genre have I in my Spotify playlists? I left out a detail: find the genre I've liked the most.", "long_pause_offsets": [ 53 ], "source_sha256": "ddc310e3965182fc9b1531007b64bdd90b0f0cf1bf1a1f0c255c97d8d28ca9f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'liked the most' to the initial incomplete question, completing the intended goal. The initial utterance is grammatically incorrect but the final goal is unambiguously reconstructed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/addition/addition_02/521147776fa6a704/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "8d6d53513851f1ebc4e756a48b88daa0ad8cf64b4d81c9c7e30b3a1339e6f530", "seconds": 13.6, "spoken_utterances": [ "Songs of which genre have I in my um, let's see... Spotify playlists? ", "I left out a detail: find the genre I've uh... liked the most." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Songs of which genre have I and my, um, let's see, Spotify playlists?' which is incomplete and unclear. The speaker then adds 'I left out a detail. Find the genre I've uh liked the most.' This addition clarifies and completes the original request. The final intent is to identify the genre of songs the user has liked the most in their Spotify playlists, which matches the original instruction. The addition corrects and completes the initial incomplete thought, satisfying the addition type criterion.", "heard_transcript": "Songs of which genre have I and my, um, let's see, Spotify playlists? I left out a detail. Find the genre I've uh liked the most." }, "candidate_id": "wav_b1b9c7d151ed2439", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('Songs of which genre have I and my...') that transitions into a plausible self-correction ('let's see... I left out a detail'). The addition of 'find the genre I've uh like the most' flows naturally from the prior hesitation, maintaining a consistent rhythm and intonation. The pauses and fillers are well-placed and sound spontaneous, contributing to a realistic, conversational delivery." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Songs have I liked the most in my Spotify playlists? Oh, I forgot to mention, I want this by genre.", "long_pause_offsets": [ 53 ], "source_sha256": "e528458246e4469c7865c592fee28039b4352a9a77e13e06d1befec5f96f18f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'by genre' to the initial question, clarifying that the result should be grouped or reported per genre. The core action of identifying the most-liked genre remains consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/addition/addition_01/f8e7d53455257086/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "b26f2aaaf4e96023ca59b4654df631774200d20f1e3daf4c79d967238a4f322a", "seconds": 11.72, "spoken_utterances": [ "Songs have I let's see... liked the most in my uh... Spotify playlists? ", "Oh, I forgot to mention, I want this by genre." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the songs they like most in their Spotify playlists, then adds the constraint 'by genre' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and complete, matching the original instruction with the added constraint.", "heard_transcript": "Songs have I let's see, like the most in my uh Spotify playlists. Uh, I forgot to mention, I want this by genre." }, "candidate_id": "wav_2c539058da9ac279", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly more abrupt phrasing ('Songs have I let's see...') and includes a filler ('uh') that feels less integrated than in audio_01. The correction ('I forgot to mention I want this by genre') is clear but delivered with a slightly flatter intonation and a more mechanical-sounding pause. While still realistic, the flow feels less organic compared to audio_01, with the self-correction sounding more like a post-hoc addition than a natural thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Songs of which genre have I liked the most in my Apple Music playlists? Let me correct the app: look at Spotify playlists, not Apple Music.", "long_pause_offsets": [ 72 ], "source_sha256": "a1463cce8df18ed1396bba23080b9694141fcedd3b32d982e8514df576ded8d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, preserving the original intent of identifying the genre with the most liked songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/revision/revision_03_app/60970cc44801c9ab/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "de14826d0365f7f6694018df455a1c27dce154e07e9f2ff7c05cda1f0750bba7", "seconds": 18.72, "spoken_utterances": [ "Songs of which uh, let me think... genre have I ... liked the most in my Apple Music playlists? ", "Let me correct the app: look at hmmm... Spotify playlists, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music playlists' but immediately corrects to 'Spotify playlists, not Apple Music' after a pause. This is a revision where the final intent matches the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Songs of which uh let me think genre have I liked the most in my Apple Music playlists? Let me correct the app, look at um Spotify playlists, not Apple Music." }, "candidate_id": "wav_e3a522ca5a7b69c7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker begins with a hesitant, self-correcting thought process ('uh, let me think'), then makes a clear, spontaneous correction from 'Apple Music' to 'Spotify' with a natural pause and intonation shift. The prosody is continuous and human-like, with believable fillers and a smooth transition into the corrected request. The revision is integrated into the same utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_02_aggregation_order", "proposed_instruction": "Songs of which genre have I liked the least in my Spotify playlists? Make that I've liked the most, not the least.", "long_pause_offsets": [ 69 ], "source_sha256": "58bc62f8c631e8314104096897125caff479dbf00ae673fd440abdd2545b165f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'liked the least' to 'liked the most', preserving the rest of the query including the app (Spotify), the entity (genre), and the action (querying liked songs)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/revision/revision_02_aggregation_order/537f7d61c5c6891d/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "0881c3ef1c1bdf08c26c01ec1c6deb7212f42e9c50e173800020c49135f25819", "seconds": 13.12, "spoken_utterances": [ "Songs of which genre have I liked the least in my um... Spotify playlists? ", "Make that I've liked the um, let's see... most, not the least." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the genre liked 'the least' but then corrects themselves after a pause, stating 'most, not the least'. This constitutes a revision where the final intent is the opposite of the initial value. The correction is clear and unambiguous, and the rest of the instruction (genre, liked, Spotify playlists) remains consistent. Therefore, the final intent matches the original instruction, satisfying the revision type criterion.", "heard_transcript": "Songs of which genre have I liked the least in my um Spotify playlists? Make that, I've liked the um, let's see, most, not the least." }, "candidate_id": "wav_af7fdc537da85d74", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision from 'least' to 'most', but the correction is delivered as a separate, slightly abrupt sentence ('Make that I've liked the... most not the least'). While the speaker uses natural fillers ('um') and hesitation, the revision lacks the seamless integration seen in audio_01. The second part feels more like a post-hoc correction than a spontaneous self-repair, making it slightly less realistic in flow." } ] }, "is_default": false }, { "variant_id": "revision_01_grouping_attribute", "proposed_instruction": "Songs of which artist have I liked the most in my Spotify playlists? Wait, the grouping should be by genre, not artist.", "long_pause_offsets": [ 69 ], "source_sha256": "7bc138cf613d7e157cb7bb27db8e9626f91816926280c42c7fb0d1ffd3b18c6f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the grouping attribute from 'artist' to 'genre', aligning with the original instruction. The action and context remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/revision/revision_01_grouping_attribute/d46a26d772179ffa/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "2b2b37655b337b107e818e714441db909f8209a7d5c24647c2c13ff10e4cb5b7", "seconds": 12.72, "spoken_utterances": [ "Songs of which artist have I uh... liked the most in my um... Spotify playlists? ", "Wait, the grouping should be by ... genre, not artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for a grouping by 'artist' but immediately corrects themselves with 'Wait, the grouping should be by genre, not artist.' This constitutes a revision where the final intent is to group by 'genre'. The correction is clear and unambiguous, and the rest of the instruction (songs liked most in Spotify playlists) remains unchanged. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Songs of which artist have I uh liked the most in my um Spotify playlists? Wait, the grouping should be by genre, not artist." }, "candidate_id": "wav_5c06bad37900837b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision from 'artists' to 'genre', but the correction is delivered as a separate, declarative sentence ('Wait, the grouping should be by genre not artists'). The speaker's tone shifts noticeably, and the correction feels more like a sudden realization than a natural part of the initial thought process. The prosody is less continuous, and the revision is less integrated into the original utterance, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_04_source_collection", "proposed_instruction": "Songs of which genre have I liked the most in my Spotify albums? Change the source to playlists, not albums.", "long_pause_offsets": [ 65 ], "source_sha256": "4a141c0b8abf45beb19e2288f4e76c017269544b76bfc108363c6bfbe783efe1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'albums' to 'playlists', preserving the genre inquiry and the 'liked the most' comparison. The final goal matches the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Songs of which um... genre have I hmmm... liked the most in my Spotify albums? ", "Change the source to uh... playlists, not albums." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "a4af815b0b514c68e8b947dc2c2e5536cfe7850755a45462db5f4a2aa35f3ef6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/cancellation/cancel_010/3215b20af115f662/speech.wav", "wer": 0.08, "wav_sha256": "2b5210ec2bdbdcf3fe0150ded4a92b9427f19c85363959788abad6ad10b07915", "seconds": 12.6, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?", "Hold on, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question about liked genres in Spotify playlists, then immediately follows with 'Hold on, cancel the request I just made.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and not followed by any resumption or new instruction.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists? Hold on, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "d18d76ccbb83be32c8bcf18e71ac4506f8f89794f6f95293d2f58daee5728643", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/ambiguous_halt/halt_040/6b2aa2c1b29e7320/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "a98c24c601e2b344c3d18d3c9976d67bf4dcf6d5936c9583ba1821b794670e84", "seconds": 11.8, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Hang on, just a minute,' which is a clear request to pause. The final interjection is a pause directive, not a cancellation or a new instruction. The original task remains unchanged and unexecuted, which is consistent with the 'ambiguous_halt' type. No issues are present.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists? Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "425a494_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f24606e86b966cb7a88f8a62a3bfb507a2cd844b91b51f3543d7e9c14594718c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/425a494_3/audio_natural/filler/original/5c5f8b621d4f512b/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "bc8437deb7fa4c526b17e7fdd98660ba81034b56d9738f1a792956bf0b5f2d2e", "seconds": 8.0, "spoken_utterances": [ "Songs of which genre have I uh, let's see... liked the most in my ah... Spotify playlists?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') and pauses but does not change the core request. The final intent remains the same as the original instruction, and the fillers do not constitute a directive to pause or cancel. Therefore, the task should be executed.", "heard_transcript": "Songs of which genre have I uh let's see liked the most in my uh Spotify playlists?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Tell me the card name I used for my last PayPal—I meant amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "ac77fa7d408cd8333dcbd6ae876788b24a3aa48831bcebf6dd17ecb7d6ab5769", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects it to 'Amazon Prime membership payment' within the same utterance. The final intended service is Amazon Prime, not PayPal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/false_start/false_start_02_app/c157a9f524b3070b/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "4d7b7d619d77b432a41f2f5fa83e484deaddcd27549cb2d9584dc0a24b637821", "seconds": 11.08, "spoken_utterances": [ "Tell me the let me think... card name I used for my um... last PayPal—ah... I meant amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Amazon Prime membership payment'. According to the 'false_start' type criterion, the final corrected intent should be executed, and the initial mistaken value should be discarded. The final audible task matches the original instruction exactly.", "heard_transcript": "Tell me the, let me think, card name I used for my um, last PayPal, I meant Amazon Prime membership payment." }, "candidate_id": "wav_0e270edcd4ada749", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start. The speaker begins with 'tell me the' and then inserts 'let me think' as a believable hesitation filler. The initial request is 'card name I used for my um last PayPal'—a plausible but incorrect memory. The self-correction 'I meant Amazon Prime membership payment' is smoothly integrated with a natural-sounding 'I meant' phrase, and the correction flows directly from the prior error. The prosody is consistent, and the repair feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last—I should say last amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "9b0b24849e7ac51dfa1d7e136e0f6e0f62017601059ed0535fb1bfd512557e61", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'second-to-last' but immediately corrects to 'last' within the same utterance. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/false_start/false_start_01_payment_recency/e0d583b077ac27d8/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "b4fc6d68eb94c2c29e8b6b68b8e03e403e082e31da11a8398e65f5a63b1c5bfd", "seconds": 10.04, "spoken_utterances": [ "Tell me the ... card name I used for my second-to-last—uh... I should say um... last amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'second to last' but immediately self-corrects to 'last' after a pause and repair. According to the 'false_start' type criterion, the final corrected value ('last') overrides the initial mistaken value ('second to last'), and the rest of the instruction remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Tell me the card name I used for my second to last uh, I should say um, last Amazon Prime membership payment?" }, "candidate_id": "wav_37f1dabec48197bb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a false start with 'second to last' followed by 'I should say um last', which is a plausible self-correction. However, the correction is less smooth than in audio_03. The speaker hesitates with 'uh' and 'um' in a way that feels slightly more forced or rehearsed. The overall flow is natural, but the repair lacks the organic transition seen in audio_03. The prosody is consistent, but the correction feels more like a verbal stumble than a genuine recollection." } ] }, "is_default": false }, { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last amazon basic membership—scratch that, make it prime membership payment?", "long_pause_offsets": [], "source_sha256": "3fb17a79d0ab09d2c6f859e5aa2cfbb24cdb9c92d5d71f60e860f717ed67b2fa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately corrects it to 'Prime membership' with 'scratch that, make it prime membership payment'. The final intent matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/false_start/false_start_03_subscription_type/75bb55c2296717da/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "845312d5901daa731e472e29a71fb0360c526c9e7bf42279fa049be95627e7b5", "seconds": 8.68, "spoken_utterances": [ "Tell me the um... card name I used for my last ... amazon basic membership—uh... scratch that, make it prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon basic membership' but immediately self-corrects to 'Prime membership payment'. According to the 'false_start' type, the final corrected instruction should be executed. The correction is clear and immediate, and the final intent matches the original reference instruction with the only difference being the correction of 'basic' to 'Prime'.", "heard_transcript": "Tell me the um card name I used for my last Amazon basic membership uh. Scratch that, make it Prime membership payment." }, "candidate_id": "wav_50b47599cf1c5f35", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision rather than a false start. The speaker says 'Amazon basic membership' and then corrects to 'Prime membership payment'. While the correction is grammatically correct, the phrase 'scratch that' is a very abrupt and unnatural way to signal a correction in casual speech. It sounds more like a script than a spontaneous thought. The prosody is flat, and the correction feels tacked on rather than integrated into the flow of thought. The overall delivery is less realistic compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Tell me the card name I used for my amazon prime membership payment? Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment.", "long_pause_offsets": [ 69 ], "source_sha256": "8dbf5c958484556af645b6be64552894bcd92739f0764c99846d8d39184b375e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the constraint 'most recent', which refines the original request without contradicting it. The core action (retrieve card name) and context (Amazon Prime membership payment) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/addition/addition_01/6fd1c19f5c6df50e/speech.wav", "wer": 0.08108108108108109, "wav_sha256": "f04d451a594eb9605049f6a86219402757233ecc1aff37007f9f84688dbdb879", "seconds": 16.16, "spoken_utterances": [ "Tell me the um... card name I used for my uh, let me think... amazon prime membership payment? ", "Oh, I forgot to mention, I want the card name for my hmmm... most recent Amazon Prime membership payment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete and hesitant, but the speaker adds the clarifying phrase 'most recent' after a pause and self-correction. This addition refines the original instruction to specify recency, which is a narrowing of the scope. The final audible intent matches the addition type criterion: the original request is combined with the added information to form the complete goal.", "heard_transcript": "Tell me the um card name I used for my uh let me think. Amazon Prime membership payment. Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment." }, "candidate_id": "wav_9c95f7dc76e5da41", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with a believable 'let me think' filler, followed by a self-repair ('Oh, I forgot to mention') and a smooth continuation that adds the missing detail ('most recent Amazon Prime membership payment'). The prosody includes hesitations and intonation shifts consistent with spontaneous recall, and the addition feels integrated rather than tacked on." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Tell me the card name I used for my last payment? I should also mention, this is specifically for my Amazon Prime membership.", "long_pause_offsets": [ 50 ], "source_sha256": "787e4ed721a9520aa1ccc61bf38cbd0ad6ac36ad1439a049dc47a1ec8c975e95", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the specific context that the payment was for an Amazon Prime membership, which refines the initial broad request about the 'last payment'. The final goal combines both the action and the specific scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/addition/addition_02/bac027dbfd881fa3/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "b96a3e0a97ac62714ce0205d67d9d673aedd1625023726bc5905475f07b1c7b7", "seconds": 14.24, "spoken_utterances": [ "Tell me the hmmm... card name I used for my uh... last payment? ", "I should also mention, this is specifically for my um... Amazon Prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'Tell me the card name I used for my last amazon prime membership payment?'. The speaker first says 'last payment' and then adds 'this is specifically for my Amazon Prime membership' as a clarifying detail. This is a valid addition that narrows the scope of the initial request. The final intent is a combination of the original and the added information, matching the reference instruction. No errors or contradictions are present.", "heard_transcript": "Tell me the card name I used for my last payment. I should also mention, this is specifically for my Amazon Prime membership." }, "candidate_id": "wav_5243ed94a27a5367", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The delivery is fluent and natural, with a 'um' and 'uh' filler and a pause before 'last payment'. The speaker adds a clarifying detail ('this is specifically for my Amazon Prime membership') after a brief pause, which sounds like a remembered restriction. However, the addition is less integrated than in audio_01, as it's delivered as a separate sentence rather than being woven into the original request, making it slightly less seamless." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last amazon basic membership payment? I misspoke about the subscription type: it should be prime membership, not basic membership.", "long_pause_offsets": [ 74 ], "source_sha256": "473a7e3cd2ab0da55d2f75f2bff3e4e583997a10b2a46e0ad4d26bbdc5e1f652", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the subscription type from 'basic' to 'prime', which is the only change from the original instruction. The action, entity (Amazon), and the request for the card name remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/revision/revision_03_subscription_type/1d6392454551cb12/speech.wav", "wer": 0.03125, "wav_sha256": "1e7c99b89f2a164a501f99a92c35951c4d17f07c656c4e57fef0a892534f5858", "seconds": 14.64, "spoken_utterances": [ "Tell me the card name I used for my last um... amazon basic membership payment? ", "I misspoke about the subscription type: it should be um, let's see... prime membership, not basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon basic membership payment' but then corrects themselves to 'Prime membership, not basic membership' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to retrieve the card name for the 'Prime membership' payment, which matches the original reference instruction.", "heard_transcript": "Tell me the card name I used for my last um Amazon basic membership payment. I misspoke about the subscription type, it should be um let's see, Prime membership, not basic membership." }, "candidate_id": "wav_3754d48da7dd241c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, then hesitates with 'I misspoke' before self-correcting 'subscription type' from 'basic' to 'Prime membership'. The correction is integrated smoothly into the same utterance with a natural 'let's see' filler, and the prosody reflects genuine recollection and correction without robotic pacing. The flow is continuous and the intonation matches a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last amazon prime membership payment? Make that the card name for my last amazon prime membership payment, not second-to-last.", "long_pause_offsets": [ 84 ], "source_sha256": "de325511ac7b3cf345e65e2b056ac30a2a178292dbdd489fe1d85dd0698ceb9e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'second-to-last' to 'last', preserving the app (Amazon Prime), the requested information (card name), and the action (tell me)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/revision/revision_01_payment_recency/bfc052c68dd7fc69/speech.wav", "wer": 0.08823529411764706, "wav_sha256": "5b2d3b0f1d7a01304ec5fb720254406048363553f7c31059805b11aa8086827f", "seconds": 16.52, "spoken_utterances": [ "Tell me the ah... card name I used for my ... second-to-last um... amazon prime membership payment? ", "Make that the card name for my last amazon prime membership payment, not second-to-last." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'second to last' payment but immediately corrects themselves to 'last' after a pause, which is the final intent. The revision is clear and unambiguous, and the final request matches the original instruction. The correction is applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "Tell me the uh card name I used for my second to last um Amazon Prime membership payment. Make that the card name for my last Amazon Prime membership payment, not second to last." }, "candidate_id": "wav_2900339b6f7c0cfd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a revision from 'second to last' to 'last' Amazon Prime membership payment. The initial request is clear, and the correction is delivered as a follow-up sentence with a natural 'Make that' transition. However, the correction is slightly more abrupt and less integrated than audio_01, with a less fluid connection between the two clauses, making it sound slightly more like a post-hoc edit than a spontaneous thought correction." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Tell me the card name I used for my last paypal prime membership payment? Let me correct the app: use amazon, not paypal, for the card name.", "long_pause_offsets": [ 74 ], "source_sha256": "e8f805cf0434be0134fc85e5c2b61645bad571c3d0b06d9d1a79cb264455bcd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from paypal to amazon, preserving the rest of the request including the entity (prime membership), the action (tell me the card name), and the temporal scope (last)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/revision/revision_02_app/59f8dcf6c49323ca/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "69513da13e1ae17040ce8b23085c4dbcdf58668177484abd0dc81d3dc83ce935", "seconds": 13.4, "spoken_utterances": [ "Tell me the card name I used for my uh... last paypal prime membership payment? ", "Let me correct the app: use um... amazon, not paypal, for the card name." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal' but immediately corrects it to 'Amazon' after a pause, which is a revision. The final instruction clearly specifies Amazon as the app, aligning with the original reference. The correction is clear and the final intent matches the original instruction.", "heard_transcript": "Tell me the card name I use for my uh last PayPal Prime membership payment. Let me correct the app, use um Amazon, not PayPal, for the card name." }, "candidate_id": "wav_59206ad9757de031", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a request involving 'PayPal Prime membership', then corrects it to 'Amazon'. The correction is delivered as a separate sentence with 'Let me correct the app use', which sounds more like a procedural correction than a natural thought process. The phrasing 'app use' is awkward and less idiomatic, and the overall delivery feels slightly more rehearsed or less spontaneous compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "03f0f4d2b59a3998f05ec5c9e8a73cc42a2da08aadbac1c3c53da55d2e8c2766", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/cancellation/cancel_021/f36da7be43c4297e/speech.wav", "wer": 0.04, "wav_sha256": "71832664bb0685b88469d225ed7cc0aca353e343660cef9ccd6897603ab20b60", "seconds": 11.76, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly revokes it with 'Hold on. Never mind. I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is a complete cancellation, not a pause or a request to wait.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment. Hold on. Never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "39be7443c28020e4713ceb3e0d435d8f6002d88271d6b4cae1d232f19448ad25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/ambiguous_halt/halt_008/8ec8950fa84ccce0/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "bc5cca8dd8ab055df30102bdbc5f898369176bf8111c53b8f992b208706b10bf", "seconds": 10.08, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but then clearly interjects 'Wait a second' as a final interjection, indicating a temporary halt. There is no cancellation or execution of the task. The final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment? Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c3636fc0150f66b5ad086cc2631a517afe375faaec4fefc911a7c27bfc062f0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_1/audio_natural/filler/original/e4d2fd528b467ec4/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "8317943fd3fae5bf26bba8620e99bc7ba17d616e9a04c5baaed17905a73c2c8f", "seconds": 6.96, "spoken_utterances": [ "Tell me the ... card name I used for my last um... amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the filler type allows for such hesitations without altering the task goal.", "heard_transcript": "Tell me the uh card name I used for my last um Amazon Prime membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last spotify free trial—sorry, make that premium membership payment?", "long_pause_offsets": [], "source_sha256": "22891cbdcf3638077d86a39b36ad88b5151179e8328e4e136032c79383da7a57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'free trial' but immediately corrects it to 'premium membership payment'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/false_start/false_start_03_subscription_type/d9efbd39452d3c22/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b746251565e14c17f0089f8f497b3a4ce45bb777aa32025e63d48788f3be4d2e", "seconds": 12.68, "spoken_utterances": [ "Tell me the um... card name I used for uh... my last spotify free trial—hmmm... sorry, make that premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify free trial' but immediately self-corrects to 'premium membership payment'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Tell me the um card name I used for uh my last Spotify free trial. Hmm. Sorry, make that premium membership payment." }, "candidate_id": "wav_4c6ea933f0a1483e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'Spotify free trial', hesitates with 'hmm', then clearly self-corrects to 'premium membership payment'. The correction is integrated smoothly into the sentence with a natural 'sorry' and a clear rephrasing. The prosody, including the pause and filler, sounds spontaneous and not rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my first—I meant last spotify premium membership payment?", "long_pause_offsets": [], "source_sha256": "82eb6c9ad22533707e59baf4a52ec4563c1620d032f3868758eba89cc0c7f25f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'first' to 'last' within the same utterance, preserving the rest of the request including the app (Spotify Premium), the context (membership payment), and the requested information (card name)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/false_start/false_start_01_payment_recency/3b7571f3bb0096ff/speech.wav", "wer": 0.0, "wav_sha256": "0ee273f01e9f740949819326916bd0a4a8be4a22bf621823064c3f00a30b505d", "seconds": 12.32, "spoken_utterances": [ "Tell me the ... card name I used for um... my first—uh... I meant last spotify premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'first' but immediately self-corrects to 'last' with 'I meant last'. This is a false start where the final value is the corrected one. The rest of the instruction matches the original, so the final intent is to retrieve the card name for the last payment.", "heard_transcript": "Tell me the card name I used for um my first uh I meant last Spotify premium membership payment?" }, "candidate_id": "wav_78be370cfc3ef8e1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a false start with 'first' being corrected to 'last'. The correction is clear and the speaker's intonation suggests a genuine moment of recollection. However, the initial error ('first') is less plausible than 'free trial' in the context of a premium membership, making the correction feel slightly less natural than in audio_03. The delivery is otherwise fluent and realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Tell me the card name I used for my last apple music—or rather, spotify premium membership payment?", "long_pause_offsets": [], "source_sha256": "d3dfeddcb38b033945a19a6e075302707d01f1bc42218c383bc45b63923fabd0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'apple music' but immediately corrects to 'spotify premium' within the same utterance. The final intended app is spotify premium, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/false_start/false_start_02_app/b25b2c187141b35b/speech.wav", "wer": 0.05, "wav_sha256": "a9ba102072fb347ed4d36126e5952e6586baebedff8d30acd1b6f2284c576ae2", "seconds": 8.8, "spoken_utterances": [ "Tell me the ... card name I used for uh... my last apple music—um... or rather, spotify premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify Premium' after a pause and the word 'or rather'. This is a clear false start where the final, corrected target is 'Spotify Premium'. The rest of the instruction, including the request for the card name and the context of the last membership payment, remains consistent. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Tell me the card name I used for uh my last Apple Music um or rather Spotify Premium membership payment?" }, "candidate_id": "wav_4824c83d30b52a7a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start where 'Apple Music' is corrected to 'Spotify Premium'. While the correction is present, the initial error is a different service altogether, which is less likely than a simple 'first/last' mix-up. The correction is delivered with a slight stumble, but the overall flow feels less smooth and the error seems more like a simple slip than a genuine moment of recollection, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Tell me the card name I used for my spotify premium membership payment? Oh, I forgot to mention, I need the card name for my most recent Spotify Premium membership payment.", "long_pause_offsets": [ 72 ], "source_sha256": "d062c59b82a4dfe0719c6fc640149e6fdd8c9adb44d459811ea106d4ba17ab74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the constraint 'most recent', which narrows the scope of the original request. The core action (retrieve card name) and context (Spotify Premium membership payment) remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Tell me the ... card name I used for ... my spotify premium membership payment? ", "Oh, I forgot to mention, I need the card name for my uh... most recent Spotify Premium membership payment." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Tell me the card name I used for my last premium membership payment? One more thing: make sure it's for my Spotify Premium membership.", "long_pause_offsets": [ 69 ], "source_sha256": "d848d83ba43561667d5d04cea1e1ce20c9b1397923f82152693cc889d67a6669", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the membership must be Spotify Premium, which refines the initial ambiguous 'premium membership payment' to match the original instruction's specificity." }, "audio": { "status": "pending", "spoken_utterances": [ "Tell me the uh... card name I used for let me think... my last premium membership payment? ", "One more thing: make sure it's for my ... Spotify Premium membership." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "476b213_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Tell me the card name I used for my last apple music premium membership payment? Make that 'spotify', not 'apple music'.", "long_pause_offsets": [ 81 ], "source_sha256": "b9ba8cd5e76a8eb0b2d3fbedd68115ce73ac1a3c81d10ac9d0978aa60fa91cfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from 'apple music' to 'spotify', preserving the action, the context of the last premium membership payment, and the requested information about the card name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/revision/revision_02_app/dfd38adc164657df/speech.wav", "wer": 0.0, "wav_sha256": "9ad7e00f02a4e2e54fef30a73182e751a07b3292ab50ce8e4478529b39cecada", "seconds": 13.68, "spoken_utterances": [ "Tell me the um... card name I used for ... my last apple music premium membership payment? ", "Make that '... spotify', not 'apple music'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' with 'Make that Spotify, not Apple Music.' This is a clear revision where the final intent is to query the card name for the Spotify payment. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Tell me the um card name I used for my last Apple Music premium membership payment. Make that Spotify, not Apple Music." }, "candidate_id": "wav_dbd033f958830451", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my first spotify premium membership payment? Wait, that should be 'last', not 'first'.", "long_pause_offsets": [ 78 ], "source_sha256": "1b00ab98f8b67ebd85576317f5a5f2e9059da29f1e0a909a5c4ea9d5a54b5b8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects 'first' to 'last', preserving the app, action, and context of the membership payment." }, "audio": { "status": "pending", "spoken_utterances": [ "Tell me the card name I used for um... my first spotify premium membership payment? ", "Wait, that should be '... last', not 'first'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last spotify free trial payment? Use 'premium membership' instead of 'free trial'.", "long_pause_offsets": [ 69 ], "source_sha256": "7ef0d2470c1f6fd499bdeda62ddebd346ea96786f38ab27be13922d45ef74922", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces 'free trial' with 'premium membership' in the context of the payment type, preserving the rest of the request including the app (Spotify), the information sought (card name), and the temporal scope (last payment)." }, "audio": { "status": "pending", "spoken_utterances": [ "Tell me the ... card name I used for ... my last spotify free trial payment? ", "Use 'uh... premium membership' instead of 'free trial'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "c174ff964e267efef1e42d6b6d1365a6c12b0b216371f4da1ca5cdccb080a0bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/cancellation/cancel_034/7c9d8c3b03c8396d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "b8dd34017de0ebe9143f4c7bc817efd97f8706a8785d9a97390afd312f994a77", "seconds": 10.32, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?", "Wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking for information but then clearly revokes the entire request with 'Wait, let's drop the whole task.' This constitutes a direct cancellation of the original task. The final intent is to cancel, not to execute or pause, satisfying the cancellation type criterion.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify Premium membership payment. Wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "2822eb6b12c92427a8244b8f778e96738a01cb463b8d75d280391c73cac136af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/ambiguous_halt/halt_038/016818575c8c9e82/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "0d9585f68ae420943822e3b233def649ce5a90877290358f1595447d0eb12d79", "seconds": 10.16, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original request and then makes a clear, final interjection to pause the process. The final audible intent is to hold the task, not to execute it or cancel it. This matches the 'ambiguous_halt' type criterion, which requires a final request to wait or pause without cancellation.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify premium membership payment. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f5f6ed3406ee6fbaec9b8e11f1a4cf0e173db9cf5578fd5dec5fb953db852160", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_2/audio_natural/filler/original/a49983dbe84f3aac/speech.wav", "wer": 0.0625, "wav_sha256": "25d26aa26010a40290bdb6b5142ab8070e8e56db915ffed4d230a684908686ea", "seconds": 5.68, "spoken_utterances": [ "Tell me the uh... card name I used for um... my last Spotify premium membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is to execute the original request.", "heard_transcript": "Tell me the uh card name I used for um my last Spotify premium membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last—no, I mean last amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "bdbac8ea9465e87ba7e8066eeb84aba0873f5d4aeb1381258b3581e85df9a366", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'second-to-last' but immediately corrects to 'last' within the same utterance. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/false_start/false_start_01_payment_recency/6231e3b24b188604/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "fa357ac95fe4751730590ae06278344af0dcf53365a1a84394e7ca3089a9abaa", "seconds": 12.72, "spoken_utterances": [ "Tell me the ah... card name I used for my second-to-last—um... no, I mean ... last amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'second to last' but immediately self-corrects to 'last' after a pause and a 'no, I mean' repair. According to the 'false_start' type criterion, the final corrected value ('last') should be applied, and the initial incorrect value ('second to last') should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Tell me the uh card name I used for my second to last um, no, I mean last Amazon Prime membership payment?" }, "candidate_id": "wav_e4fa0af1b08a264c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible false start. The initial 'uh' and 'second to last' are followed by a clear self-correction with 'no, I mean last' that sounds like a genuine moment of recollection. The prosody is consistent, the pace is conversational, and the correction flows smoothly within the same utterance without any robotic or disjointed feel." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Tell me the card name I used for my last PayPal—sorry, that should be amazon prime membership payment?", "long_pause_offsets": [], "source_sha256": "9241ebd5664629a6ed613a93929325c9272f325abe24de76ef45b10e57ca4f46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Amazon Prime membership payment' within the same utterance, replacing the incorrect payment method with the correct one. The action and the target of the inquiry remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/false_start/false_start_02_app/24cb5549509bccaf/speech.wav", "wer": 0.0, "wav_sha256": "b152cc1c5fdb245d7b85451073f85594a6c09108186d6cc6775faf1619b15c7c", "seconds": 12.88, "spoken_utterances": [ "Tell me the card name I used for my uh, let's see... last PayPal—um... sorry, that should be amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Amazon Prime membership payment'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Tell me the card name I used for my uh let's see last PayPal um sorry, that should be Amazon Prime membership payment?" }, "candidate_id": "wav_5e889e688ce8cff0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable false start with 'uh' and 'let's see' as fillers, followed by a self-correction from 'PayPal' to 'Amazon Prime membership payment'. The correction is natural, but the initial hesitation feels slightly more rehearsed or less spontaneous than audio_02, and the overall delivery is a bit flatter in prosody." } ] }, "is_default": false }, { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last amazon basic membership—no, that should be prime membership payment?", "long_pause_offsets": [], "source_sha256": "eaf59e286e9f623e432cbfc2d1f94b0f323020aeb9202084532c34d5fa7eb579", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately corrects it to 'Prime membership' within the same utterance. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/false_start/false_start_03_subscription_type/8791abc38d7d0f4f/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "c0bfc42798c6fb978a022dc47f7f065eea0a4e70f83d4918f933cad9a2621e26", "seconds": 12.08, "spoken_utterances": [ "Tell me the ... card name I used for my last amazon uh... basic membership—hmmm... no, that should be prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately self-corrects to 'Prime membership payment'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final audible task matches the original instruction after the correction.", "heard_transcript": "Tell me the card name I used for my last Amazon uh basic membership. Hmm. No, that should be Prime membership payment." }, "candidate_id": "wav_545bde7f8e239003", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a false start ('Amazon uh basic membership'), the correction ('no, that should be Prime membership payment') sounds less natural. The 'uh' is awkwardly placed, and the correction feels more like a mechanical replacement than a spontaneous recollection. The prosody is less fluid, and the overall delivery seems slightly more hesitant and less confident than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Tell me the card name I used for my last payment? I should also mention, this is specifically for my Amazon Prime membership.", "long_pause_offsets": [ 50 ], "source_sha256": "787e4ed721a9520aa1ccc61bf38cbd0ad6ac36ad1439a049dc47a1ec8c975e95", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the specific context that the payment was for an Amazon Prime membership, which refines the initial broad request about the 'last payment'. The final goal combines both the action and the specific scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/addition/addition_02/28fbd888439765a9/speech.wav", "wer": 0.08, "wav_sha256": "2dd10da7c7c697caa35945b039cf21a94e8ca98ab72036a8de3be8f1ceafc10d", "seconds": 13.48, "spoken_utterances": [ "Tell me the card name I used for my um... last payment? ", "I should also mention, this is specifically for my ah... Amazon Prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'Tell me the card name I used for my last payment.' This is followed by an addition specifying 'this is specifically for my Amazon Prime membership.' The addition clarifies the scope of the payment, narrowing the initial broad request. The final intent combines the original request with the added constraint, matching the reference instruction. The type is 'addition', and the final goal is fully aligned with the reference.", "heard_transcript": "Tell me the card name I used for my um last payment. I should also mention this is specifically for my uh Amazon Prime membership." }, "candidate_id": "wav_c2402210116ad60d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a slightly broken flow ('for my um...'), which is followed by a clear, additive clarification ('I should also mention...') that is smoothly integrated into the sentence. The prosody is conversational, with a believable pause and filler before the second clause, and the intonation suggests genuine recollection rather than a rehearsed or robotic delivery. The addition feels like a spontaneous memory." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Tell me the card name I used for my amazon prime membership payment? Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment.", "long_pause_offsets": [ 69 ], "source_sha256": "8dbf5c958484556af645b6be64552894bcd92739f0764c99846d8d39184b375e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the constraint 'most recent', which refines the original request without contradicting it. The core action (retrieve card name) and context (Amazon Prime membership payment) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/addition/addition_01/ec5cecfeef5088f9/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "8ba3dc78253ba44db4e69dba8c5f2eab276acd26ea815568f287aff265db1722", "seconds": 16.04, "spoken_utterances": [ "Tell me the ... card name I used for my uh... amazon prime membership payment? ", "Oh, I forgot to mention, I want the card name for my um... most recent Amazon Prime membership payment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'Tell me the card name I used for my Amazon Prime membership payment.' This is followed by an addition: 'Oh, I forgot to mention, I want the card name for my most recent Amazon Prime membership payment.' The addition clarifies that the user wants the card name for the 'most recent' payment, which is a narrowing of the original request. The final intent is to provide the card name for the most recent payment, which combines the original task with the added constraint. The addition is clear and does not introduce any conflicting or contradictory information. Therefore, the final intent satisfies the 'addition' type criterion.", "heard_transcript": "Tell me the card name I used for my uh Amazon Prime membership payment. Oh, I forgot to mention, I want the card name for my um most recent Amazon Prime membership payment." }, "candidate_id": "wav_da824fc2de795e60", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly more fluent but still hesitant delivery ('uh'), then makes a self-correction ('Oh, I forgot to mention...') that is followed by a revision of the initial request. However, the revision ('I want the card name for my um...') is less smoothly integrated and sounds more like a post-hoc addition than a natural thought process. The prosody is flatter, and the correction feels slightly abrupt, making it less realistic than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_payment_recency", "proposed_instruction": "Tell me the card name I used for my second-to-last amazon prime membership payment? Make that the card name for my last amazon prime membership payment, not second-to-last.", "long_pause_offsets": [ 84 ], "source_sha256": "de325511ac7b3cf345e65e2b056ac30a2a178292dbdd489fe1d85dd0698ceb9e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'second-to-last' to 'last', preserving the app (Amazon Prime), the requested information (card name), and the action (tell me)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/revision/revision_01_payment_recency/98bb03af87382838/speech.wav", "wer": 0.0, "wav_sha256": "fa05868dc2da530e0fa0f021c67ea66d1424e193593f48d57d234b872bd2c735", "seconds": 17.2, "spoken_utterances": [ "Tell me the card name I used for my uh... second-to-last amazon prime membership payment? ", "Make that the card name for my um... last amazon prime membership payment, not second-to-last." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'second to last' but immediately corrects themselves to 'last' after a pause, which is the final and intended instruction. The revision is clear and unambiguous, and the final goal matches the original reference instruction. The correction is applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "Tell me the card name I used for my uh second to last Amazon Prime membership payment. Make that the card name for my um last Amazon Prime membership payment, not second to last." }, "candidate_id": "wav_775b917fbc112cde", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding 'uh' and 'um', then makes a clear revision from 'second to last' to 'last' with a smooth self-correction. The prosody is consistent, the pace is conversational, and the revision feels like a genuine moment of recollection rather than a rehearsed correction. The intonation and rhythm are very natural." } ] }, "is_default": true }, { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Tell me the card name I used for my last amazon basic membership payment? I misspoke about the subscription type: it should be prime membership, not basic membership.", "long_pause_offsets": [ 74 ], "source_sha256": "473a7e3cd2ab0da55d2f75f2bff3e4e583997a10b2a46e0ad4d26bbdc5e1f652", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the subscription type from 'basic' to 'prime', which is the only change from the original instruction. The action, entity (Amazon), and the request for the card name remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/revision/revision_03_subscription_type/66e4fdd878527b12/speech.wav", "wer": 0.0, "wav_sha256": "ba4bca3396cbe2a23b9023341ac4619d18b157915ae15fb507295017c8ea5d0d", "seconds": 15.16, "spoken_utterances": [ "Tell me the ... card name I used for my last ... amazon basic membership payment? ", "I misspoke about the subscription type: it should be prime membership, not basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested information for 'Amazon basic membership' but immediately corrected it to 'Prime membership' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to retrieve the card name for the Prime membership, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "Tell me the card name I used for my last Amazon basic membership payment. I misspoke about the subscription type. It should be Prime membership, not basic membership." }, "candidate_id": "wav_d33b02b717329be7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then makes a revision about the subscription type. The correction is delivered with a natural-sounding 'I misspoke' followed by a smooth self-repair. The prosody is consistent, and the correction feels spontaneous. However, the initial request is slightly more formal and less hesitant than audio_03, making it feel slightly less like a real-time thought process." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Tell me the card name I used for my last paypal prime membership payment? Let me correct the app: use amazon, not paypal, for the card name.", "long_pause_offsets": [ 74 ], "source_sha256": "e8f805cf0434be0134fc85e5c2b61645bad571c3d0b06d9d1a79cb264455bcd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from paypal to amazon, preserving the rest of the request including the entity (prime membership), the action (tell me the card name), and the temporal scope (last)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/revision/revision_02_app/6e788b61fda731e7/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "8bfd60380c8a212f7f5c96ef1556d16453527b80eabde6c38c0908abaf5e6146", "seconds": 15.96, "spoken_utterances": [ "Tell me the ... card name I used for my last ... paypal prime membership payment? ", "Let me correct the app: use uh... amazon, not paypal, for the card name." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal Prime membership payment' but immediately corrects it to 'Amazon, not PayPal' in a revision. The final intent is to retrieve the card name for the Amazon Prime payment, which matches the original instruction. The revision is clear and applies to the app context, preserving all other conditions.", "heard_transcript": "Tell me the card name I used for my last PayPal Prime membership payment. Let me correct the app. Use Amazon, not PayPal, for the card name." }, "candidate_id": "wav_9d141ee9b3625058", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then says 'Let me correct the app' which sounds like a meta-comment rather than a natural self-correction. The correction itself is delivered with a slightly robotic and rehearsed tone, and the phrase 'use uh Amazon now PayPal' sounds like a forced insertion rather than a spontaneous thought. The prosody is less fluid, and the overall delivery feels less authentic compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "d880d12166a9bec4d41321b248ac542fa13ee6d49d218ae4bab04e00f4f114b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/cancellation/cancel_019/339b11d9e0d86095/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "5cd804af27641cfc00d740e8cb6d2cc3c7ce6dbfb2a44fc60168c0eecc906008", "seconds": 11.56, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?", "Wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the original request but immediately self-cancels with 'Wait, never mind. Don't do that.' This constitutes a clear cancellation of the entire task as per the 'cancellation' type criterion. The final interjection explicitly revokes the initial request, and no other task is initiated.", "heard_transcript": "Tell me the card name I used for my last Amazon Prime membership payment. Wait, never mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "e8beef3d69b37c716782d9144b6acc9027aad08850a2bf39b17714befa6f1aa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/ambiguous_halt/halt_018/345326ac443beb7d/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "a38db98f4d731f33b4d453f2f83b7d2109ee427766811344f9c2e00c9559016c", "seconds": 10.76, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Could you pause for a second?', which is a clear instruction to hold the current task. The initial request to 'Tell me the card name...' is not canceled, but the final interjection is a valid pause command. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Tell me the card name I used for my uh last Amazon Prime membership payment. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "476b213_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7c70ae27e55d3190cff1dabf3b5d36ab9f303e829cc4e3e3da1e1ccbc54e3f4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/476b213_3/audio_natural/filler/original/779d658680d67737/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "1d661f6e63eefce5ac18e85b9bf21503906c48a3c101a591ac835e2bf44d2713", "seconds": 6.68, "spoken_utterances": [ "Tell me the card name I used for my uh... last ... amazon prime membership payment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent is to retrieve the card name for the last Amazon Prime payment, which matches the original instruction. The filler does not alter the task goal.", "heard_transcript": "Tell me the card name I used for my uh last Amazon Prime membership payment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_email_state", "proposed_instruction": "Delete all my sent—sorry, I meant archived gmail threads that are from before this calendar month.", "long_pause_offsets": [], "source_sha256": "e31b227ff52caea589865d108b091b6b5261b4acae352b6242bd59395560d4fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'archived' within the same utterance, preserving the app (Gmail), action (delete), scope (all), and time condition (before this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/false_start/false_start_01_email_state/15ca6031f0d944ff/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "e021aef9b822134ac2917c7211091d2e9de3d21d9e54ff4dc57cc7da2d2be58a", "seconds": 6.8, "spoken_utterances": [ "Delete all my sent—um... sorry, I meant archived gmail threads that are ... from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'archive'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final task matches the original instruction's intent, with only the action changed from 'delete' to 'archive'. The target ('Gmail threads') and condition ('from before this calendar month') are preserved. Therefore, the final intent is correctly executed.", "heard_transcript": "Delete all my sent um, sorry, I meant archive Gmail threads that are from before this calendar month." }, "candidate_id": "wav_cbdd49596e8897e0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('delete all my sent them'), immediately self-corrects with a natural-sounding 'sorry, I meant', and smoothly continues with the full, coherent request. The prosody is consistent, the repair is integrated into the flow without awkwardness, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Delete all my archived Outlook—I should say gmail threads that are from before this calendar month.", "long_pause_offsets": [], "source_sha256": "ab7870fc380154be03df4d3b725ad9475c40eb9e8377c5bd13d34afc4f593160", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, action, archive status, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/false_start/false_start_02_app/149ca6d0598ae8ee/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "9537d430564631c118af3ccffc1ecefba0d490d730432077508412ef9d68fc8c", "seconds": 8.76, "spoken_utterances": [ "Delete all my archived Outlook—hmmm... I should say gmail threads that are ... from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook whom whom' as a false start, then self-corrects to 'Gmail'. The final audible intent is to delete archived Gmail threads from before the current calendar month, which matches the original instruction. The correction is immediate and the final goal is clear.", "heard_transcript": "Delete all my archived Outlook whom whom I should say Gmail threads that are from before this calendar month." }, "candidate_id": "wav_0bad67e46170ab3e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start ('delete all my archived outlook'), followed by a self-correction ('I should say gmail threads'). However, the correction is less fluid; the speaker hesitates with 'whom? whom?' before the correction, which sounds like a genuine confusion or stutter rather than a smooth recovery. The overall delivery feels slightly more disjointed and less natural than audio_01, with the filler 'whom?' disrupting the pragmatic flow of the self-correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from last calendar month—no, that should be before this calendar month.", "long_pause_offsets": [], "source_sha256": "2f1fd21b8a0918ed20c7cc110ae8aa35de61896d6351a529679950dd435665e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'before this calendar month'. The final intended time filter is 'before this calendar month', which matches the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my archived gmail threads that are from um, let's see... last calendar month—uh... no, that should be ... before this calendar month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Delete all my archived gmail threads that are. Oh, and one more thing: only delete threads from before this calendar month.", "long_pause_offsets": [ 47 ], "source_sha256": "d03cf5e52808628049c46a7da406fd7720b4b8e2e85db35354dae559896c489a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time condition 'from before this calendar month', which refines the incomplete initial request. The action, target app, and archival status remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/addition/addition_03/d4f24eaecffdc547/speech.wav", "wer": 0.2692307692307692, "wav_sha256": "e59dcb7aadd53efd76de617f93fb90a29c7a2c42f2972bda98ed15cc113c0f0b", "seconds": 13.48, "spoken_utterances": [ "Delete all my um... archived gmail threads that are. ", "Oh, and one more thing: only delete threads from before um, let me think... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Delete all my archived Gmail threads that are from' and then adds 'only delete threads from before this calendar month' after a pause and self-interruption. This matches the 'addition' type, where the final goal combines the initial request with the added constraint. The final intent is clear and matches the original instruction.", "heard_transcript": "Delete all my um archived Gmail threads that are from oh and one more thing only delete threads from before um let me think this calendar month." }, "candidate_id": "wav_55fde216cc129930", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Delete all my gmail threads that are from before this calendar month. Oh, I forgot to mention, only delete archived threads.", "long_pause_offsets": [ 70 ], "source_sha256": "c7e8bdccce69d20e2d3998fb95381e96be187595a822882d84f4552f4ba8d973", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the threads must be archived, which refines the initial broad request. The action, app, time condition, and scope are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my gmail threads that are ... from before um... this calendar month. ", "Oh, I forgot to mention, only delete ... archived threads." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Delete all my archived threads that are from before this calendar month. I should also mention, this is for my Gmail threads.", "long_pause_offsets": [ 73 ], "source_sha256": "ce2bcb47fecc485029aad75d1bec7dfbd61624bcbb7861956b2b61fa5c3cfc65", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the threads are from Gmail, which is consistent with the original instruction's context and does not contradict any other requirement. The action, scope (archived), time condition (before this calendar month), and target (threads) remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my um, let me think... archived threads that are from before ... this calendar month. ", "I should also mention, this is for my Gmail threads." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_email_state", "proposed_instruction": "Delete all my sent gmail threads that are from before this calendar month. Wait, delete all my archived gmail threads that are from before this calendar month.", "long_pause_offsets": [ 75 ], "source_sha256": "a2d8a3288499d6dc874efe9479a1a6f7d4b1eecdb6cbaa9b9b26d478d2600bb3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the thread category from 'sent' to 'archived', while preserving the app (Gmail), action (delete), and time condition ('before this calendar month')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/revision/revision_01_email_state/4461cba76fbad9eb/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "3ac4b8cd43ca5f6c918852f697abfa964e7b4a338bcc62cce12ad2b87bb03ba6", "seconds": 13.84, "spoken_utterances": [ "Delete all my hmmm... sent gmail threads that are let's see... from before this calendar month. ", "Wait, delete all my archived gmail threads that are from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent Gmail threads' but immediately corrects to 'archived Gmail threads' after a pause and self-correction ('Wait, delete all my archived...'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Delete all my hmm, sent Gmail threads that are let's see, from before this calendar month. Wait, delete all my archived Gmail threads that are from before this calendar month." }, "candidate_id": "wav_05c5760c0213d6ea", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. It begins with a false start ('Sent'), self-corrects to 'Archive', then pauses and formulates a new request ('Wait, delete all my archived Gmail threads...'). The correction is integrated into a single, continuous utterance with a believable change-of-mind intonation and a clear reference back to the initial idea. The 'Hmm' and 'let me see' are placed naturally, and the overall prosody is fluid and human." } ] }, "is_default": true }, { "variant_id": "revision_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from last calendar month. That should be 'before this calendar month', not 'last calendar month'.", "long_pause_offsets": [ 72 ], "source_sha256": "55a2d76584f8a87265423b38ac259ebb2b48ee1411e9aaac5e8440a483f6774d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last calendar month' to 'before this calendar month', which is the original instruction's intended condition. The action, target app, and data scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/revision/revision_03_time_filter/2497ab267efd940b/speech.wav", "wer": 0.12, "wav_sha256": "9c964f067d724eb1dc22656bd7b1ce1ecec95c610d624634296a50cb19a00546", "seconds": 14.68, "spoken_utterances": [ "Delete all my um... archived gmail threads that are from ... last calendar month. ", "That should be 'uh... before this calendar month', not 'last calendar month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects themselves to 'before this calendar month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Delete all my um archived Gmail threads that are from last calendar month. That should be uh before this calendar month, not last calendar month." }, "candidate_id": "wav_0961607e64087a0f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear revision where the speaker completes an initial request ('last calendar month') and then corrects it ('before this calendar month'). The correction is delivered as a separate, slightly hesitant sentence with a 'that should be' phrase, which is a plausible way to self-correct. The prosody is natural, and the pause before the correction feels like a genuine moment of reconsideration. However, the flow is less integrated than audio_03, as it's a two-part utterance rather than a single, fluid thought." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Delete all my archived Outlook threads that are from before this calendar month. Change the app to Gmail, not Outlook.", "long_pause_offsets": [ 81 ], "source_sha256": "5a0a408c271f40d0523d65f7b4d772fee8198cfca727d4619a40cb0191aabacf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, preserving the action, archive status, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/revision/revision_02_app/11bc7a97410fc97c/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "69ee6130b340e82407e095d9eb2def1fcec91c519d0de48432062ace8a9b95b2", "seconds": 11.96, "spoken_utterances": [ "Delete all my archived uh... Outlook threads that are from before uh, let's see... this calendar month. ", "Change the app to Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but then corrects to 'Gmail' with the explicit phrase 'Change the app to Gmail, not Outlook.' This is a revision where the final target app is clearly stated as Gmail. The rest of the instruction (delete archived threads from before this calendar month) remains consistent. The revision applies to the target app, and the final intent matches the original instruction.", "heard_transcript": "Delete all my archived uh Outlook threads that are from before uh let's see, this calendar month. Change the app to Gmail, not Outlook." }, "candidate_id": "wav_6e052898b0510599", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a less natural revision. It starts with a request for 'Outlook' threads, then pauses and corrects to 'Gmail'. However, the correction is delivered as a separate, abrupt sentence ('Change the app to Gmail, not Outlook') rather than being integrated into the original request. The prosody of the correction sounds slightly more robotic and less like a spontaneous thought, and the overall flow feels more like a command being issued than a person formulating a request in real-time." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "de5c2a9aa4fe56f1fe2ff72a9eff7c554e8b1c95b331c3bcac50ee32e0bd2584", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/cancellation/cancel_024/e53028960ccecee7/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "f421dfb5f819fc8e977a914a00ea9accaaac3c19379cff7036f2b6cf3c59ff26", "seconds": 12.48, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month.", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'hang on, oh never mind, cancel the whole thing.' This constitutes a final, unambiguous cancellation of the task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "delete all my um archived gmail threads that are um let's see from before this calendar month. hang on, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "178dc546f052a3cf68b23379301cbdcb176a64496410418be76d6365e91f4463", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/ambiguous_halt/halt_002/4a9163fefdd63e86/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "132a5a8c523788dd2906c346768a9a5301a88f5318b78e9334869d46ab03585e", "seconds": 10.52, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month.", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'hold on,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there a new instruction to resume. The final interjection is a pause command, consistent with the 'ambiguous_halt' type. The initial request is not executed, but the final intent is to halt, not to cancel or execute.", "heard_transcript": "delete all my um archived gmail threads that are um let's see from before this calendar month. hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7657680ec4f5dc6c028a853ba02a352fd074a2b161dafe4010dca5bab3467c2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_1/audio_natural/filler/original/1c0f6167e45c1fbc/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "82b60245ddc8d0638ac51052edb788de8c8a825748dff690047b454e43d5dddf", "seconds": 7.48, "spoken_utterances": [ "Delete all my um... archived Gmail threads that are um, let's see... from before this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final intent matches the original instruction, with no correction or cancellation. The filler words do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Delete all my um archived Gmail threads that are um let's see from before this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from last calendar month—no, that should be this calendar month.", "long_pause_offsets": [], "source_sha256": "bc5ed21658f05efbaa646e8d8482627c35fb00eb8f3d4b3b46b2765ffb42df73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'this calendar month'. The final intended time filter is 'this calendar month', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/false_start/false_start_03_time_filter/e3c62b915b8a4118/speech.wav", "wer": 0.1, "wav_sha256": "7674ab4805c17e30e192c6cafa80e6c4fbbda1c482243054430df4fd53403381", "seconds": 11.96, "spoken_utterances": [ "Delete all my archived ... gmail threads that are from last calendar month—hmmm... no, that should be ... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately self-corrects to 'this calendar month'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final audible instruction matches the original reference instruction, so the task is executed as intended.", "heard_transcript": "Delete all my archived Gmail threads that are from last calendar month. No, that should be this calendar month." }, "candidate_id": "wav_068e6f4b497d5e32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'last calendar month', then immediately self-corrects to 'this calendar month' with a clear, spontaneous-sounding 'No, that should be...' followed by a smooth rephrasing. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels like genuine real-time thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_01_email_state", "proposed_instruction": "Delete all my sent—sorry, I meant archived gmail threads that are from this calendar month.", "long_pause_offsets": [], "source_sha256": "3e46aa014f934ee297512fa402974a4c20a7de5614bdd1d8ed0d0f7338265e50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'archived' within the same utterance, preserving the app (Gmail), action (delete), scope (all), and date condition (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/false_start/false_start_01_email_state/73812cb5318c5990/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "c342b6a8993edd833903895234410b894659657efe1d473dbea5287bd48cd243", "seconds": 10.8, "spoken_utterances": [ "Delete all my sent—uh... sorry, I meant um... archived gmail threads that are from ... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'archived'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction with the correction applied.", "heard_transcript": "Delete all my sent uh, sorry, I meant um, archived Gmail threads that are from this calendar month." }, "candidate_id": "wav_d3c6f744cb95b640", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start with 'sent' followed by a self-correction to 'archived'. The correction is natural, but the pause and 'sorry' feel slightly more rehearsed or hesitant than in audio_03. The overall delivery is still very realistic, but the correction is less fluid and integrated than in the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Delete all my archived Outlook—I should say gmail threads that are from this calendar month.", "long_pause_offsets": [], "source_sha256": "d6ec08cfdabe93cad8f2c698bbc602157436b4b8a73d80174db62ede5b0d02e1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, archive status, app context, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/false_start/false_start_02_app/d210dafda733f82c/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "700cba70575531ff6a5fc30326fd9db7023f55985c9a2316041ac85b1ae23374", "seconds": 8.88, "spoken_utterances": [ "Delete all my um... archived Outlook—uh... I should say gmail threads that are from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archived Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, with only the app name corrected from Outlook to Gmail.", "heard_transcript": "Delete all my um archived Outlook uh I should say Gmail threads that are from this calendar month." }, "candidate_id": "wav_a2a4f33b127f8fc0", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'archived Outlook' and a correction to 'Gmail's threads'. However, the correction is less smooth and the initial error is more jarring. The speaker also uses a filler 'uh' and a longer pause, which, while realistic, combined with the less fluid correction, makes the overall delivery feel slightly less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Delete all my archived threads that are from this calendar month. Oh, right, this is specifically for my Gmail threads.", "long_pause_offsets": [ 66 ], "source_sha256": "9d9b5b5041604f01acbdc91aedacc3b10b6f85371a43199bcd2866f38d7aef70", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the initial 'threads' refer specifically to Gmail threads, which matches the original instruction's app context. The action, scope ('all'), archive status, and date condition remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/addition/addition_02/6b842c656d800592/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "85f601a76e2966300accd36a28bb8053d69e5340cbdcf5261850ed713b6b1931", "seconds": 14.44, "spoken_utterances": [ "Delete all my uh... archived threads that are from this calendar month. ", "Oh, right, this is specifically for my let me think... Gmail threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'delete all my archived threads that are from this calendar month'. After a pause and self-correction, the speaker adds 'this is specifically for my gmail threads'. This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "delete all my uh archived threads that are from this calendar month. oh right, this is specifically for my, let me think, gmail threads." }, "candidate_id": "wav_7c44980bfb40326f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural, slightly hesitant 'uh' and a believable self-correction ('Oh right, this is specifically for my... let me think... Gmail threads'). The 'let me think' filler is placed at a plausible moment of cognitive processing, and the correction flows smoothly without awkwardness. The prosody is consistent, and the overall delivery sounds like a spontaneous, real-time formulation of a task." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "Delete all my archived gmail threads that are from. To narrow that down, only delete threads from this calendar month.", "long_pause_offsets": [ 52 ], "source_sha256": "75d3627160314bc905efb4860bb693d7c284579ee5e1e2b924973435a45d5556", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing temporal restriction 'this calendar month' to the incomplete initial request, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/addition/addition_03/8513e0700622ac98/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "8f3c0a2de88b17a1bbff7aa0321c7e81984f2079aede9f8d8071a480286c1d30", "seconds": 13.56, "spoken_utterances": [ "Delete all my uh... archived uh, let's see... gmail threads that are from. ", "To narrow that down, only delete threads from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all my archived gmail threads that are from', which is incomplete. After a pause, the speaker adds 'to narrow that down, only delete threads from this calendar month', which completes and specifies the original request. This matches the 'addition' type, where the final goal is the combination of the initial instruction and the added constraint. The final audible intent is to delete archived Gmail threads from this calendar month, which matches the original instruction.", "heard_transcript": "Delete all my uh archived uh let's see gmail threads that are from to narrow that down, only delete threads from this calendar month." }, "candidate_id": "wav_b8aef89b2b849987", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but includes a noticeable 'uh' and a filler phrase ('let's see') that sounds like genuine hesitation. The addition ('to narrow that down...') is a natural-sounding revision that fits the context. However, the delivery is slightly less fluid than audio_03, with a more abrupt transition into the revision and a slightly less natural intonation on the filler." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Delete all my gmail threads that are from this calendar month. I didn't mention this, but only delete archived threads.", "long_pause_offsets": [ 63 ], "source_sha256": "fed04c294a2b3b16e2b7b0a3674dd8b2e91292dcb225affe04809f93b9793dcd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the threads must be archived, which narrows the scope of the initial request without contradicting it. The action, app, time condition, and deletion intent remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/addition/addition_01/9b7bee8789dd9903/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "dec627fef8f445695c01fdaacdbb5926d8efaf11c049fc776d7caed89a9779fb", "seconds": 14.08, "spoken_utterances": [ "Delete all my um... gmail threads that are from ... this calendar month. ", "I didn't mention this, but only delete ... archived threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all my Gmail threads that are from this calendar month.' This is followed by an addition: 'I didn't mention this, but only delete archived threads.' The addition clarifies that the deletion should be limited to archived threads. The final intent combines the original request with the added constraint, resulting in 'Delete all my archived Gmail threads that are from this calendar month.' This matches the expected behavior for an 'addition' type, where the final goal is the combination of the initial instruction and the added information.", "heard_transcript": "Delete all my um Gmail threads that are from this calendar month. I didn't mention this, but only delete archived threads." }, "candidate_id": "wav_7a1ab72ddb9bd4f0", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then adds a correction ('I didn't mention this, but only delete archived threads') that feels more like a post-hoc clarification than a spontaneous self-correction. The phrase 'I didn't mention this' is an unusual way to introduce a restriction and sounds slightly unnatural. The prosody is also flatter compared to the others, and the correction lacks the organic flow of a true self-repair." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from last calendar month. Make that delete all my archived gmail threads from last calendar month, not this calendar month.", "long_pause_offsets": [ 72 ], "source_sha256": "0144c3d0422b07e3bf66157ebd4fbe33fe59192e2647619f62a7f62e8c3e7e4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the time filter from 'this calendar month' to 'last calendar month', reversing the original instruction's intended time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/revision/revision_03_time_filter/f673d1e5e40cea31/speech.wav", "wer": 0.03225806451612903, "wav_sha256": "06c77acedd2a046535bc96a278b83dac5d0507e717a6744b7f6ebd5072616b83", "seconds": 17.52, "spoken_utterances": [ "Delete all my archived let's see... gmail threads that are from uh... last calendar month. ", "Make that delete all my archived gmail threads from last calendar month, not this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this calendar month' but immediately corrects to 'last calendar month' and further clarifies with 'not this calendar month'. This constitutes a revision where the final, corrected instruction is clear and unambiguous. The final intent matches the revised instruction, so the task should be executed as revised.", "heard_transcript": "Delete all my archived let's see, Gmail threads that are from uh last calendar month. Make that delete all my archived Gmail threads from last calendar month, not this calendar month." }, "candidate_id": "wav_109d525243342fd1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with natural hesitation ('let's see') and a self-repair ('Make that delete all my archived Gmail threads from last calendar month, not this calendar month'). The correction is smoothly integrated into the same utterance, with a believable shift in intonation and pacing. The prosody is consistent, and the speaker sounds like they are formulating and correcting a thought in real time." } ] }, "is_default": false }, { "variant_id": "revision_01_email_state", "proposed_instruction": "Delete all my sent gmail threads that are from this calendar month. Wait, delete all my sent gmail threads that are from this calendar month, not archived.", "long_pause_offsets": [ 68 ], "source_sha256": "01971c8a209f835d2a9bee508fcc29887a9c1ca43d89c6b73e5fca8df63a3b1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly restricts the action to 'sent' threads and excludes 'archived' ones, which contradicts the original instruction that targeted 'archived' threads regardless of sent status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/revision/revision_01_email_state/7f73ee6786bb131a/speech.wav", "wer": 0.0, "wav_sha256": "a0150726fa5c07a076b94374dedb6f6f6110992293b8360b068b69cad7da934a", "seconds": 18.0, "spoken_utterances": [ "Delete all my um... sent gmail threads that are from uh... this calendar month. ", "Wait, delete all my um, let's see... sent gmail threads that are from this calendar month, not archived." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'archived' but then corrects to 'sent' and adds the constraint 'not archived'. The final instruction is a revision of the original, changing the target from archived to sent threads and adding a negation. The final intent matches the revised instruction and satisfies the revision type criterion.", "heard_transcript": "Delete all my um, sent Gmail threads that are from uh, this calendar month. Wait, delete all my um, let's see, sent Gmail threads that are from this calendar month, not archived." }, "candidate_id": "wav_a1f22a18b57bf4c3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear pause and self-correction ('Wait, delete all my... let's see, sent Gmail threads... not archived'). The speaker hesitates, restarts the phrase, and then adds a clarifying restriction. The flow is natural, and the 'wait' signals a genuine change of mind. However, the initial phrase 'sent Gmail threads' is less coherent than the final version, and the correction feels slightly more disjointed than audio_01's seamless repair." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Delete all my archived outlook threads that are from this calendar month. Let me correct the app: delete all my archived threads from Outlook, not Gmail.", "long_pause_offsets": [ 74 ], "source_sha256": "652c65c9224939bdd8a15023fcceed97c282063d8e8b65c4fd5103d50f2cbb2d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Gmail to Outlook, while retaining the action, archive status, and calendar month condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/revision/revision_02_app/36adec2b85f8c725/speech.wav", "wer": 0.09375, "wav_sha256": "9e4dad255edd69147cb870fa878a4964a82e5a52b4a24c003f36bd6a745f74c4", "seconds": 18.32, "spoken_utterances": [ "Delete all my archived uh... outlook threads that are from um, let me think... this calendar month. ", "Let me correct the app: delete all my archived threads from Outlook, not uh... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'gmail' but immediately corrects to 'Outlook' after a pause, which is a revision. The final instruction matches the original intent of deleting archived threads from the current calendar month, but now correctly specifies the app as Outlook instead of Gmail. The revision is clear and final, so the task should be executed with the corrected target.", "heard_transcript": "Delete all my archived uh Outlook threads that are from um, let me think, this calendar month. Let me correct the app. Delete all my archived threads from Outlook, not uh Gmail." }, "candidate_id": "wav_351031ffe7ee36b9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a clear request but then introduces a significant error ('Outlook threads') and a long, unnatural pause ('let me think'). The correction ('Let me correct the app') is awkward and sounds like a meta-comment rather than a spontaneous self-repair. The final correction ('delete all my archived threads from Outlook, not Gmail') is delivered with a flat, robotic tone and lacks the natural prosody of the others. The overall delivery feels less like a spontaneous thought process and more like a rehearsed or synthesized correction." } ] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_020", "hide_variant_original": true, "source_sha256": "792dc5b968d8d0c63d605efdea97ed26068f4b047eb104e343433c51c45ede46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/cancellation/cancel_020/e82f0887963cd3d6/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "82e6c1421fbcb02624c4e55c60a293c171ce5eae3207e601088a2fef2add6517", "seconds": 12.0, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month.", "Hang on, never mind, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete request, then immediately follows with 'Hang on. Never mind. Cancel that request.', which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the prior instruction, satisfying the cancellation type criterion.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month. Hang on. Never mind. Cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "bd955588bcbc1a1dd834c3b6b98c4b3609f5adea2920485d12473856b6f9f9e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/ambiguous_halt/halt_028/c8af004c540dd40e/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "bb4732baf97ded1525201c6bb1406188357ad83e38efcadfa73d2a9b8c5194a4", "seconds": 11.32, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption. The initial instruction is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8fa56f255e99e4ce765474f01043e52628620a6ee1afc25331416525f58b2f0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_2/audio_natural/filler/original/2ab36f6c76502965/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "e3c696d968644974c528ae53fb9c2635405637b57b1a3891e7f6d300a3a5ac6e", "seconds": 6.68, "spoken_utterances": [ "Delete all my uh... archived Gmail threads that are from um... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, with no correction or cancellation. The task is executed as intended.", "heard_transcript": "Delete all my uh archived Gmail threads that are from um this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from this or the last two calendar months—no, that should be this or the last calendar month.", "long_pause_offsets": [], "source_sha256": "11ce37bfc6c97f85a0f7b7cbff34b20dea774ea849e3f5e4f3959524c1465941", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'two calendar months' but immediately corrects it to 'one calendar month'. The final value is the corrected one, and all other conditions (archived status, app, action, and time scope) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/false_start/false_start_03_time_filter/07fe40159b04fe6e/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "1c387209045de8543e72370edf7aee41db1feb74c5d7574133472f3a7a79ff81", "seconds": 10.8, "spoken_utterances": [ "Delete all my archived um... gmail threads that are from ... this or the last two calendar months—uh... no, that should be this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last two calendar months' but immediately self-corrects to 'last calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Delete all my archived um Gmail threads that are from this or the last two calendar months uh no, that should be this or the last calendar month." }, "candidate_id": "wav_bd6c30ab72138b5f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then self-corrects 'two calendar months' to 'calendar month' with a natural hesitation ('uh') and a smooth, immediate repair. The prosody is consistent, and the correction feels spontaneous rather than forced. The pacing and intonation suggest genuine recall and correction." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Delete all my archived Outlook—I should say gmail threads that are from this or the last calendar month.", "long_pause_offsets": [], "source_sha256": "95dc6508f7711daa7e423c9346ae51da31915b110fd4a53114dfff8262f4b248", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, archive status, thread scope, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/false_start/false_start_02_app/3e0ab4767af9193e/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "41579725a77c66b423713d613b82b8fc8cd863edb9ad8772dc5da3535cee4e18", "seconds": 9.88, "spoken_utterances": [ "Delete all my let's see... archived Outlook—ah... I should say gmail threads that are from this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value ('Gmail') should be applied, and the original mistaken value ('Outlook') should be discarded. The rest of the instruction, including the action ('Delete'), target ('archived Gmail threads'), and condition ('from this or the last calendar month'), remains consistent and correctly understood. Therefore, the final intent matches the original instruction after the correction.", "heard_transcript": "Delete all my let's see, archived Outlook uh, I should say Gmail threads that are from this or the last calendar month." }, "candidate_id": "wav_035c7e176d9e13ad", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a false start ('Let's see'), then corrects 'Outlook' to 'Gmail' mid-sentence. The correction is clear and the flow is mostly natural, though the initial hesitation and the 'uh' sound slightly more hesitant than in audio_01. The overall delivery is realistic but slightly less fluid due to the initial stumble and the correction being more of a full rephrase than a simple word swap." } ] }, "is_default": false }, { "variant_id": "false_start_01_email_state", "proposed_instruction": "Delete all my sent—sorry, I meant archived gmail threads that are from this or the last calendar month.", "long_pause_offsets": [], "source_sha256": "9581d173714b65b1fc79548466c0b5ac0e8497abdea46868ec5f5a85678f819b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'archived' within the same utterance, preserving the app (Gmail), action (delete), and time condition. The final intended target is the archived threads." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/false_start/false_start_01_email_state/fc5e6d2079b0d721/speech.wav", "wer": 0.16, "wav_sha256": "f3f389917e4aeec2a553d9d79eefb098d6c434387850cdbcbd80a1f89f95c377", "seconds": 12.32, "spoken_utterances": [ "Delete all my sent—uh... sorry, I meant archived um... gmail threads that are from uh, let me think... this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent up' but immediately self-corrects to 'archived'. The rest of the instruction, including the target ('archived Gmail threads') and the time constraint ('this or the last calendar month'), remains consistent. The false start is corrected, and the final intent matches the original instruction.", "heard_transcript": "Delete all my sent up, sorry, I mean archived um Gmail threads that are from uh, let me think, this or the last calendar month." }, "candidate_id": "wav_7240039ca84b4df4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with 'sent', then corrects to 'archived' with a clear self-repair. However, the pause and 'let me think' after 'from uh' feels more like a deliberate pause to formulate the sentence rather than a spontaneous hesitation. The correction is less integrated into the flow, and the overall delivery sounds slightly more rehearsed or less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all my gmail threads that are from this or the last calendar month. Oh, I forgot to mention, only delete archived threads.", "long_pause_offsets": [ 75 ], "source_sha256": "640163df137e50cc2cc63ebd65ce588e727dd1f4d11131b522e7d5c0488c0e61", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the threads must be archived, which refines the initial broad request. The time range and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/addition/addition_01/8446c6ee654fa1b0/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "a5604eec03d462cf14a230803e3530447685fea7094018e7a51cca47b33bdc7e", "seconds": 14.04, "spoken_utterances": [ "Delete all my uh, let me think... gmail threads that are from ... this or the last calendar month. ", "Oh, I forgot to mention, only delete um... archived threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all my Gmail threads that are from this or the last calendar month.' This is followed by an addition: 'only delete um, archived threads.' The addition correctly narrows the scope of the original request to only archived threads, which matches the original instruction. The final intent is a combined version of the initial and added constraints, and the control intent is to execute the task as revised.", "heard_transcript": "Delete all my uh, let me think, uh, Gmail threads that are from this or the last calendar month. Oh, I forgot to mention, only delete um, archived threads." }, "candidate_id": "wav_b8c32cf8d53286c6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-initiated pause ('let me think'), which is followed by a clear, coherent initial request. The addition of a new condition ('only delete um... archive threads') is introduced with a believable 'oh, I forgot to mention' phrase, and the correction is smoothly integrated into the ongoing utterance with a natural pause and filler ('um'). The prosody is consistent, and the flow reflects a realistic thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Delete all my archived threads that are from this or the last calendar month. I left out a detail: do this for my Gmail threads.", "long_pause_offsets": [ 78 ], "source_sha256": "c1fff5a8d3ad2cd09e490ca449d875fc856468a5a3b55418002f4723774b5ebd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Gmail' to the initial request, which was otherwise complete. The action, scope ('all archived threads'), and time condition ('this or the last calendar month') are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/addition/addition_02/79b6be34580807a1/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "c7fec66d8d71bff1de966e6440379969f9aad3ead8caf43ac8e74e9e926d5444", "seconds": 13.96, "spoken_utterances": [ "Delete all my ... archived threads that are from uh... this or the last calendar month. ", "I left out a detail: do this for my ah... Gmail threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all my archived threads that are from this or the last calendar month.' This is followed by a self-correction: 'I left out a detail. Do this from my Gmail threads.' The addition specifies the email platform as Gmail, which narrows the scope of the original request. The final intent combines the original action with the added detail, matching the reference instruction. The addition type requires combining the earlier instruction with the added information, which is done correctly.", "heard_transcript": "Delete all my uh archived threads that are from uh this or the last calendar month. I left out a detail. Do this from my uh Gmail threads." }, "candidate_id": "wav_e6078562dc72face", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar hesitation ('uh') and a clear initial request. However, the follow-up correction ('I left out a detail. Do this from my uh... Gmail threads') is less integrated; it's a separate sentence with a more abrupt transition and a longer, less natural pause before the final word. The prosody is slightly flatter, and the correction feels more like a post-hoc addition than a spontaneous thought correction, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_filter", "proposed_instruction": "Delete all my archived gmail threads that are from this month. Make that 'this or the last calendar month' instead of 'this month'.", "long_pause_offsets": [ 63 ], "source_sha256": "70ca894169abc67ba12c74e2583588e6767d7f49d94dea8900aea88dbb404df8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the time filter to include both the current and previous calendar month, which matches the original instruction. The action, target app, and data type remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/revision/revision_03_time_filter/3030fce9c9283409/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "95b67cd043ba2177ad3cb2292eebcc98aa0d60aca4e55b5d332a65cd487b0ba8", "seconds": 14.24, "spoken_utterances": [ "Delete all my archived uh, let's see... gmail threads that are from um... this month. ", "Make that 'uh... this or the last calendar month' instead of 'this month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this month' but then corrects it to 'this or the last calendar month' after a pause. According to the revision type criterion, the correction spoken after a longer pause should be applied to the earlier instruction. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "delete all my archived uh let's see gmail threads that are from um this month. Make that uh this or the last calendar month instead of this month." }, "candidate_id": "wav_c23427695f40ed89", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and spontaneous speech pattern. It begins with a clear request, includes a realistic self-correction ('this month' to 'this or the last calendar month'), and uses natural fillers ('uh', 'let's see') that are well-placed and not overused. The intonation and pacing are consistent with someone formulating a task in real time, and the revision flows smoothly without sounding rehearsed or robotic." } ] }, "is_default": true }, { "variant_id": "revision_01_email_state", "proposed_instruction": "Delete all my sent gmail threads that are from this or the last calendar month. Wait, the email state should be archived, not sent.", "long_pause_offsets": [ 80 ], "source_sha256": "86380dd9a7bd7021ebd0bcf4e592d683262c8ac0e1b9660aced69083fba5729e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email state from 'sent' to 'archived', preserving the app, action, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/revision/revision_01_email_state/1ad334f216b0aa8e/speech.wav", "wer": 0.04, "wav_sha256": "31c2ef1d2d66edd75e89df8988f452c220c94558f933d2330503bb1dcb89bbce", "seconds": 11.52, "spoken_utterances": [ "Delete all my ... sent uh... gmail threads that are from this or the last calendar month. ", "Wait, the email state should be archived, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately corrects to 'archived' after a pause and 'Wait'. This is a revision where the final intent is the corrected version. The original instruction was to delete archived threads, and the final audible intent matches this exactly. The revision is clear and unambiguous.", "heard_transcript": "Delete all my sent uh Gmail threads that are from this or the last calendar month. Wait, the email state should be archived not sent." }, "candidate_id": "wav_94fd15464840a453", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a believable revision where the speaker initially states 'sent' threads, then corrects to 'archived'. The pause and 'wait' are natural, and the correction is integrated into the same utterance. However, the initial request is slightly less fluent than audio_03, with a slightly faster pace and less organic use of fillers, making it feel a bit more rehearsed." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Delete all my archived Outlook threads that are from this or the last calendar month. Change the app to Gmail, not Outlook.", "long_pause_offsets": [ 86 ], "source_sha256": "4172138e565a18a6fc7d3764653277bbcee52200c7bba5977c1ccb80eae1f0b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the archive status, thread scope, and date condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/revision/revision_02_app/4f5d1d18b6172210/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "2c166d05fb429a4a34271b784c75cc394d2968ded6cae8c7ee97e7a399da5626", "seconds": 13.8, "spoken_utterances": [ "Delete all my ... archived hmmm... Outlook threads that are from this or the last calendar month. ", "Change the app to ... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests deleting archived Outlook threads, then revises the app from Outlook to Gmail after a pause. The revision is clear and final, and the rest of the instruction (archived, this or last calendar month) remains unchanged. The final intent matches the original reference instruction.", "heard_transcript": "Delete all my archived Outlook threads that are from this or the last calendar month. Change the app to Gmail, not Outlook." }, "candidate_id": "wav_9a9456022990bd20", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a flat, robotic delivery with unnatural pauses and a lack of natural fillers. The speaker says 'delete all my archived... Outlook threads...' and then abruptly changes the app to 'Gmail', but the transition is jarring and lacks the prosodic cues of a real-time correction. The overall prosody is monotonous, and the delivery sounds more like a read-aloud than spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "3d32972090fc9cb893c27e93285a3d5ba47d6bf8a48003e23ad1fdf8997fa5a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/cancellation/cancel_034/c656f69e72dd6ae1/speech.wav", "wer": 0.08, "wav_sha256": "c67cf4b2e3f6313dd0904dd56c68bdce3aac4745f02b2f646fee5a726cd28406", "seconds": 13.32, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month.", "Hold on, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly revokes it with 'Hold on, let's drop the whole task,' which constitutes a cancellation. The final audible intent is to cancel, not to execute the original request.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month. Hold on, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "03413733556ed31b648dcbc09f8cbee2b63c29e04f574d35082ed41147485c32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/ambiguous_halt/halt_003/b30e1fb5dcaeca4b/speech.wav", "wer": 0.1, "wav_sha256": "284e06399268cfb9995b60e932bd41a45768022d237607cb03916ac08b379876", "seconds": 11.68, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month.", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final interjection indicates a temporary hold on the task. The initial instruction is not canceled, and the speaker does not provide a new command or correction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "4d12842_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1113391fb259d84b3cd7c0fbaa0d1b63e141c4acff0877d606f48b28fd3ae5d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/4d12842_3/audio_natural/filler/original/dd56c7dc0f82cdbf/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "04abbd14db47756a5b6dd940e82a5d5457573e842da3469ca1dd2b9e2d8dcfee", "seconds": 8.92, "spoken_utterances": [ "Delete all my let's see... archived Gmail threads that are from uh... this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('let's see', 'uh') but does not change the core task. The final intent matches the original instruction. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "Delete all my let's see, archived Gmail threads that are from uh this or the last calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from last month—sorry, that should be this month.", "long_pause_offsets": [], "source_sha256": "c751fd60760cdb5c57d207e0a0febc76dd6d2b9a2fbcb83ed163f6fc134fe88b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/false_start/false_start_03_time_reference/a928af1c5bf86c4a/speech.wav", "wer": 0.15, "wav_sha256": "000063cc598f180667551f4ba160d36093d60c474af043c54af524d47b0e4ad1", "seconds": 9.76, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent or received money from last month—um... sorry, that should be this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'sorry, that should be this month.' This is a clear false start where the final intent is the corrected version. The correction is immediate and unambiguous, so the task should be executed with the corrected time frame.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from last month. Um, sorry, that should be this month." }, "candidate_id": "wav_44af174854d4b526", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complete false-start pattern. The speaker begins with 'last month', then self-corrects with 'sorry, that should be this month'—a clear, spontaneous revision with a believable tone of correction. The initial phrase is fully formed, and the correction is integrated smoothly into the same utterance, maintaining a natural flow and prosody. The filler 'um' is used appropriately to signal hesitation, and the overall delivery sounds like a genuine moment of recollection and correction." } ] }, "is_default": true }, { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to—no, I mean sent or received money from this month.", "long_pause_offsets": [], "source_sha256": "864766dcd75632077a74bfcbf864c9699edbd317e4ae3cb21b27a7fe97c005d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately corrects to 'sent or received money from', which matches the original instruction. The app (Venmo), action (befriend), and time frame (this month) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/false_start/false_start_02_transaction_relation/63a474a3a05d34bb/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "62af87f3f12259891415c7c80eb22e7a3d90ffb6f80d6e423fb710aacd305105", "seconds": 10.32, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to—uh... no, I mean um... sent or received money from ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately self-corrects to 'sent or received money from'. This is a false start where the final corrected version matches the original instruction. The correction is immediate and unambiguous, so the final intent is to execute the original task.", "heard_transcript": "Befriend on Venmo anyone I have sent money to uh no, I mean um sent or received money from this month." }, "candidate_id": "wav_81267399f18ae097", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'be friend on Venmo anyone I have sent money to' followed by 'no, I mean sent or received money from this month'. The correction is clear and the speaker rephrases the intent, but the initial phrase is less coherent and the transition feels slightly abrupt. The 'no, I mean' structure is natural, but the correction is less fluid than in audio_03, and the prosody lacks the same level of organic hesitation and recovery." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I should say Venmo anyone I have sent or received money from this month.", "long_pause_offsets": [], "source_sha256": "744bda314cf99809ffed871315de62002fd24726fdf2f47dc899bafddd95e038", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, target scope, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/false_start/false_start_01_app/751472353e1bf99e/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "848b3efd89bd67cc8a37f32f53385983d86631781e4e511b6d09cae1bf704591", "seconds": 7.84, "spoken_utterances": [ "Befriend on CashApp—um... I should say Venmo anyone I have sent or received money from uh... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The rest of the instruction remains unchanged. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Befriend on Cash App, I should say Venmo, anyone I have sent or received money from uh this month." }, "candidate_id": "wav_14b74b067082cd69", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio starts with 'be friend on Cash App' and then corrects to 'Venmo'. The correction is less natural because it changes the platform mid-sentence without a clear reason or hesitation. The speaker says 'I should say Venmo' which is a plausible self-correction, but the initial error ('Cash App') is not contextually related to the rest of the sentence, making the correction feel less integrated. The prosody is also flatter and less expressive, with less variation in intonation and pacing." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month. I left out a detail: only befriend people I've sent or received money from.", "long_pause_offsets": [ 44 ], "source_sha256": "745c6a0fb1e1770dd8fa2c58fac2c1ba7653aca1827a83e3e1b037fabdd57c76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only befriend people I've sent or received money from', which refines the initial broad request. The time frame 'this month' is preserved from the first utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/addition/addition_02/b894284e9c1bbea0/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "3e9f47b275a667e962019e1fca9214e3270e1526d845efdb950f054321c475d2", "seconds": 12.48, "spoken_utterances": [ "Befriend on Venmo anyone I have uh... this month. ", "I left out a detail: only befriend people I've ... sent or received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Befriend on Venmo anyone I have uh this month' and then corrects with 'I left out a detail. Only befriend people I've sent or received money from.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. The addition type criterion is satisfied.", "heard_transcript": "Befriend on Venmo anyone I have uh this month. I left out a detail. Only befriend people I've sent or received money from." }, "candidate_id": "wav_1bfd6b93c6cb5dca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('bfriend on Venmo anyone I have uh... this month'), then naturally adds a clarifying detail ('I left out a detail. Only bfriend people I've sent or received money from'). The self-correction is embedded within the same utterance, with a believable pause and filler ('uh') before the initial request. The addition flows smoothly, maintaining consistent prosody and a natural conversational rhythm. The speaker sounds like they are formulating the request in real time, making the delivery highly realistic." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from. Oh, I forgot to mention, only include people from this month.", "long_pause_offsets": [ 61 ], "source_sha256": "d3aed9440774da7d3d067941e281fe0feb52572efaf3c93ac2b3dc1c131717ae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction that narrows the scope of the initial request; the core action and target remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/addition/addition_01/48af1a324198fc94/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "65d46acecb196a985d54069e3eeb1f3e93d172efefe4d6f9ff1b67ead2e3c098", "seconds": 13.0, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have sent or received money from. ", "Oh, I forgot to mention, only include people from hmmm... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'Befriend on Venmo anyone I have sent or received money from this month.' The speaker first says the same instruction, then adds a constraint 'only include people from this month' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added constraint. The final audible intent is complete and matches the reference instruction exactly.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from. Oh, I forgot to mention, only include people from this month." }, "candidate_id": "wav_62cb74e144c88478", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a more fluent initial request ('bfriend on uh Venmo anyone I have sent or received money from'), but the follow-up ('Oh, I forgot to mention only include people from this month') feels slightly more abrupt and less integrated. The correction is a separate sentence with a noticeable pause and a filler ('uh') that seems to mark a shift in thought rather than a spontaneous self-repair. While still natural, the flow is less seamless than audio_01, making it slightly less realistic in terms of spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money to this month. Make that 'sent or received money from', not 'sent or received money to'.", "long_pause_offsets": [ 70 ], "source_sha256": "322303f7af6db3617aa396f8e778092fed5a046d8bf7f9650c49e46116349274", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the preposition from 'to' to 'from', which matches the original instruction. The action, app, and time scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/revision/revision_02_transaction_relation/bee7f605c892d72c/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "dc5ba95400d76940f4c44885c7c2a95c13107992a1774fb7f1ccc5485306fe7e", "seconds": 14.8, "spoken_utterances": [ "Befriend on Venmo anyone I have sent or received money to ... this month. ", "Make that 'um... sent or received money from', not 'sent or received money to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent or received money to', then corrects it to 'sent or received money from' after a pause and self-correction. This matches the revision type criterion: the final intent is the corrected version, and the original incorrect clause is replaced. The rest of the instruction, including 'this month', remains unchanged and is preserved.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money to this month. Make that um, sent or received money from, not sent or received money to." }, "candidate_id": "wav_ee72a4121125e367", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with a natural-sounding self-correction. The pause before 'Make that' is brief and plausible, and the correction ('sent or received money from') is smoothly integrated into the same utterance with a slight pitch shift and emphasis, mimicking a spontaneous mental correction. The prosody is consistent and the pacing realistic." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have sent or received money from this month. Wait, the app should be Venmo, not PayPal.", "long_pause_offsets": [ 73 ], "source_sha256": "fbcf2e19e194f8eddebc55dec60f8ff993979c66d40d1735a0716c2120f811b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the action, the condition about money transactions this month, and the scope of 'anyone' involved in those transactions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/revision/revision_01_app/38533c4b5f964f32/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "4d18efd1ecbe8f51515eeff6feaa7b7b739aedceb6a8f41c3e2be438256913c7", "seconds": 13.84, "spoken_utterances": [ "Befriend on uh... PayPal anyone I have sent or received money from um... this month. ", "Wait, the app should be ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects themselves to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to execute the task on Venmo, which matches the original reference instruction. All other constraints (action: befriend, target: people money was sent/received from, time: this month) are preserved.", "heard_transcript": "Befriend on uh PayPal anyone I have sent or received money from um this month. Wait, the app should be Venmo, not PayPal." }, "candidate_id": "wav_00e4f6c6a8167bd8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'um' and 'wait,' followed by a correction of the app name. The correction is delivered with a natural shift in intonation and a brief pause, indicating a genuine change of mind. The flow is coherent, though the correction is slightly more abrupt than in audio_03, and the initial 'uh' filler is less integrated." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from last month. I misspoke about the time: it should be 'this month', not 'last month'.", "long_pause_offsets": [ 72 ], "source_sha256": "a21fa81a86a36f05bd4f2727f9dc52d51eca08e4f2f950e33d47f886b9147a4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this month', while retaining the app, action, and condition about money transactions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/revision/revision_03_time_reference/ff73d811cff8e7dd/speech.wav", "wer": 0.1, "wav_sha256": "e60f1acdfa90b4a4b8904dadd6c72b218313c494ff88c3da46569f055cbf49ec", "seconds": 17.96, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent or received money from um, let's see... last month. ", "I misspoke about the time: it should be 'um... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects themselves with 'I misspoke about the time, it should be um, this month, not last month.' This constitutes a revision where the final intent clearly overrides the initial incorrect value. The correction is spoken after a pause and explicitly states the intended time frame. The rest of the instruction remains unchanged and matches the original. Therefore, the final audible intent satisfies the revision criterion.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from, um, let's see, last month. I misspoke about the time, it should be um, this month, not last month." }, "candidate_id": "wav_bae81bdb57a2acd0", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then hesitates with 'um, let's see,' followed by a correction of the time period. The correction is delivered as a separate sentence rather than being integrated into the original utterance, making it sound more like a post-hoc edit than a spontaneous self-correction. The prosody is slightly flat, and the pause before the correction feels longer and less organic than in the other clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_013", "hide_variant_original": true, "source_sha256": "ccb04f4a84f315dec4eb408e4aa3e36763e9b64b9947c137cbb012e660f961c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/cancellation/cancel_013/def44e422cc5862e/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "090b78c6f02f7994744b293f8e7bc95169f61031c2a344b770dc785419b74201", "seconds": 13.72, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month.", "Hold on, I changed my mind. Please cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then clearly says 'Hold on, I changed my mind. Please cancel the task.' This constitutes a direct and unambiguous cancellation of the entire task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month. Hold on, I changed my mind. Please cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "339270912248d5a19c7dd6c188e09aeaee4bc7d71c372bf6f2f6afda7817fae3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/ambiguous_halt/halt_001/9007ac3a22d13ae2/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "aa10e1aeec9960490e0c8f373a0be61b7a8145640494ff71c284cee4a805524f", "seconds": 11.44, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' which is a clear request to pause or hold. The initial instruction is fully stated before the pause command. Since the task is not canceled and the speaker does not resume or correct the instruction, the final control intent is to pause. This matches the 'ambiguous_halt' type criterion, which requires an audible request to hold/wait/pause with no permanent cancellation or permission to resume.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "628192f84aba661f56bec3004604790c82ac7d1141542dca23c76c8c2035b676", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_1/audio_natural/filler/original/0d1a4f88748ff1ff/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "89fb173a475e619f7eab9be293f6e85578513068f3585c75e6e0cf28b2608507", "seconds": 8.72, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have sent or received money from ah... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um, let me think' which is a filler, but then clearly states the original instruction without any correction or change. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Befriend on Venmo anyone I've sent or received money from this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last week—sorry, make that this month.", "long_pause_offsets": [], "source_sha256": "4d56a468659b03d3cae060ca9fcbc9b3b227cacd13e0ad74c29f22031f8564b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'last week' with 'this month'; the action, app, and recipient scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/false_start/false_start_03_time_reference/312e789eef17b1ba/speech.wav", "wer": 0.0, "wav_sha256": "54a65f2f4618310acbbd3eee692ad29aff694f331f6de04b947d7b26ac5efe0b", "seconds": 12.96, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have sent money to last week—uh... sorry, make that ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last week' but immediately self-corrects to 'this month'. According to the 'false_start' type criterion, the final corrected value ('this month') overrides the initial mistaken value ('last week'), and the rest of the instruction remains unchanged. The final audible intent matches the original instruction.", "heard_transcript": "Befriend on uh let's see Venmo anyone I have sent money to last week uh sorry, make that this month." }, "candidate_id": "wav_d7e1fe75af8be90d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a filler ('uh') and a moment of thought ('let's see'), followed by a clear false start ('last week') and a self-correction ('this month'). The correction is delivered with a believable shift in intonation and pacing, and the overall prosody is consistent and human-like. The pauses and fillers are well-placed and contribute to the impression of real-time formulation." } ] }, "is_default": true }, { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have received money from—no, I mean sent money to this month.", "long_pause_offsets": [], "source_sha256": "71a82241b18de832346feac5fea0a4949dfd51f7ef651f84c9d73557d7509b6f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'received money from' but immediately corrects to 'sent money to'. The final intended action matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/false_start/false_start_02_transaction_relation/13bf99e997517726/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "caac060bdc7f986ee48b1f6b6efec58b4cb85fca4c8ef520cf2e3097ca43f2bb", "seconds": 11.64, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from—ah... no, I mean sent money to ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received money from' but immediately self-corrects to 'sent money to'. According to the 'false_start' type criterion, the final corrected intent should be executed. The correction is immediate and unambiguous, and the rest of the instruction matches the original. Therefore, the final intent is correctly captured and should be executed.", "heard_transcript": "Befriend on um Venmo anyone I have received money from uh no, I mean sent money to this month." }, "candidate_id": "wav_8ca9e35a08862304", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear self-correction from 'received' to 'sent', which is a plausible repair. The speaker uses a filler ('um') and a pause before the correction, and the intonation shift when saying 'no, I mean' sounds natural. However, the correction is slightly more abrupt and less smoothly integrated than in audio_03, making it feel a bit more rehearsed or less fluid in its delivery." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I should say Venmo anyone I have sent money to this month.", "long_pause_offsets": [], "source_sha256": "2bdec94552f6bb2702b8f616351a5079933c18951a6025edeee7cf539c454c84", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, target scope, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/false_start/false_start_01_app/6c1387b584b1fed0/speech.wav", "wer": 0.29411764705882354, "wav_sha256": "fb1940103e1240234ff94c48edba6057154a4da176f7e87b2c7f965f15c76581", "seconds": 9.08, "spoken_utterances": [ "Befriend on CashApp—ah... I should say Venmo anyone I have sent money to hmmm... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'I should say Venmo'. This is a false start where the final intent is the corrected version. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final goal is to befriend on Venmo.", "heard_transcript": "Befriend on Cash App. I should say Venmo, anyone I have sent money to this month." }, "candidate_id": "wav_1849bbfeb9d27045", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start ('Cash App') and a self-correction ('Venmo'), but the correction is delivered with a flat, almost robotic intonation. The pause after 'to' is very long and unnatural, and the overall prosody lacks the dynamic variation of the other two. The correction feels tacked on rather than spontaneously integrated, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month. I left out a detail: only befriend people I have sent money to.", "long_pause_offsets": [ 44 ], "source_sha256": "f956d8dac6f84f093f3f1a4b0c05744e0174b02233038bf2d84cd8b8a9fb382d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'only befriend people I have sent money to', which narrows the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/addition/addition_02/bbddb8d3e64b5b0d/speech.wav", "wer": 0.3448275862068966, "wav_sha256": "4ee8028ee3844b530470eff157312860a3211483ab3042d5297ea406168bdb5b", "seconds": 15.92, "spoken_utterances": [ "Befriend on um, let's see... Venmo anyone I have uh, let me think... this month. ", "I left out a detail: only befriend people I have um... sent money to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Befriended on Venmo anyone I have sent money to this month' but then corrects with 'I left out a detail. Only befriend people I have send money to.' This is a clear addition that narrows the scope of the original instruction. The final intent is to befriend only those people to whom money has been sent, which matches the reference instruction. The addition refines the original request without contradicting it, satisfying the 'addition' type criterion.", "heard_transcript": "Befriended on um, let's see. Venmo anyone I have uh, let me think. This month. I left out a detail. Only befriend people I have um, send money to." }, "candidate_id": "wav_191d9cd813c64eed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with plausible fillers ('um', 'let me see', 'let me think') and a self-repair ('I left out a detail') that smoothly transitions into a more complete and specific instruction. The prosody reflects genuine hesitation and recollection, and the correction is integrated naturally into the flow without sounding rehearsed or robotic. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have sent money to. Oh, I forgot to mention, only for people I've sent money to this month.", "long_pause_offsets": [ 47 ], "source_sha256": "366e499851b0d8d703897d91c3ee2ea09b896c942644c98d6551e192fe6cdcee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction that narrows the initial broad request; the final goal is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/addition/addition_01/fb18840c8729d63d/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "06d8406b5649ac1522a362e89f0b3d1ff1d73a55fa359d0da5a8763ba3057cbc", "seconds": 11.56, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to. ", "Oh, I forgot to mention, only for people I've sent money to ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'Befriend on Venmo anyone I have sent money to this month.' The speaker first says the same instruction, then adds a clarifying constraint 'only for people I have sent money to this month.' This addition narrows the scope of the original request, which is consistent with the 'addition' type. The final intent is a more specific version of the original, and the addition is clearly spoken after a pause, satisfying the revision/addition criterion. The final control intent is to execute the combined task.", "heard_transcript": "Befriend on Venmo anyone I have sent money to. Oh, I forgot to mention, only for people I have sent money to this month." }, "candidate_id": "wav_c939c183ab1d6473", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker delivers a clear and grammatically correct request, the initial statement is too fluent and lacks the natural hesitation or fillers seen in spontaneous speech. The follow-up ('Oh, I forgot to mention...') is a plausible afterthought, but the transition feels slightly abrupt and the overall delivery is more rehearsed than the first clip. The prosody is relatively flat, and the correction is delivered as a separate, almost parenthetical remark rather than a continuation of the original thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last month. That should be 'this month', not 'last month'.", "long_pause_offsets": [ 58 ], "source_sha256": "77ff8ce937338cdbb65e04d2d4f3aafc2d33fd03d802eaff55894a6073b619c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'last month' to 'this month', preserving the action, app, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/revision/revision_03_time_reference/bfc15d8476b8df3a/speech.wav", "wer": 0.15, "wav_sha256": "47ad3a7551785f901a25de789d23f225278a1dd08140eab426505468f728ca2b", "seconds": 12.44, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to um... last month. ", "That should be '... this month', not 'last month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' with 'That should be this month, not last month.' This is a clear revision where the final intent is the corrected version. The original instruction matches the final intent after the revision, so the task is executed as intended.", "heard_transcript": "Befriend on Venmo anyone I have sent money to last month. That should be this month, not last month." }, "candidate_id": "wav_63667e1a7d05ca46", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody. The self-correction ('That should be this month, not last month') is a believable, smoothly integrated revision with appropriate intonation and pacing. The pause before the correction is realistic, and the correction flows naturally within the same utterance, showing a plausible moment of realization." } ] }, "is_default": true }, { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month. Change 'received money from' to 'sent money to'.", "long_pause_offsets": [ 64 ], "source_sha256": "dcbcfe2786af99aa4904e1849741ad5761cbcfb0a2eab64d282634749da7f6ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the phrase 'received money from' to 'sent money to', restoring the original instruction's intended meaning. The app, action, and time frame remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/revision/revision_02_transaction_relation/4879175284c24e1f/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "33725d24056a9311faf1c682e3ea83084b28f45ab9c6f878d082ce588e9d883e", "seconds": 13.84, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have received money from ... this month. ", "Change 'received money from' to 'ah... sent money to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the incorrect condition 'received money from' but immediately corrects it to 'sent money to' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Befriend on Venmo anyone I have received money from this month. Change 'received money from' to uh 'sent money to'." }, "candidate_id": "wav_49ff1c1ba2bf2f64", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then makes a revision ('Change received money from to sent money to'). The correction is stated as a separate sentence, which is less fluid than a self-repair within one utterance. The 'uh' filler and slight hesitation are natural, but the revision feels more like a post-hoc instruction than a spontaneous correction, making it less realistic than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have sent money to this month. Use Venmo instead of PayPal.", "long_pause_offsets": [ 59 ], "source_sha256": "41ae42d16635f388d65272b981d1ab3c8ebc1af0e1f38ed12bca171391728b01", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the action, recipient scope, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/revision/revision_01_app/f6516f7c86a9b965/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "b21a9da31f119b2cb5e37d687662e73bdf0f25051ff311ba207f112f7f80e38b", "seconds": 12.4, "spoken_utterances": [ "Befriend on hmmm... PayPal anyone I have sent money to this month. ", "Use uh, let me think... Venmo instead of PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction ('let me think'). This is a revision where the final audible intent matches the original instruction. The correction is clear and applied to the entire task, so the final goal is to befriend on Venmo.", "heard_transcript": "Befriend on hmm, PayPal anyone I have sent money to this month. Use uh, let me think, Venmo instead of PayPal." }, "candidate_id": "wav_f4dfe6d97deaa0a9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then introduces a revision ('Use uh let me think Venmo instead of PayPal'). The phrase 'let me think' is a filler, but the correction is awkwardly phrased as a separate command rather than a self-repair. The prosody is less natural, and the revision feels more like a task instruction than a spontaneous thought process, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_034", "hide_variant_original": true, "source_sha256": "44c222156ab9665c3c29f49b830ff6461e37c436a7721399ead9ac66f85cd12d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/cancellation/cancel_034/1e7de43e9e43129a/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "9e1b217cbde3ea3f57801929db803c0563a8df7c10701b94d26a9a45c31db1af", "seconds": 10.56, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month.", "Wait, let's drop the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then clearly says 'Wait, let's drop the whole task,' which is an unambiguous cancellation of the entire request. The final interjection is a direct command to cancel, not a pause or hesitation. Therefore, the control intent is correctly identified as 'cancel' and the task is fully revoked.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month. Wait, let's drop the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "845000bdbd0b67c21390918fb6d16f8813bb10955efc9604d275bcf23d2f77de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/ambiguous_halt/halt_028/aeb05b21f8128983/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "71d14ea37115f7f2e004e24dd0cab4c0cb7e18ad50a8c28805b8e360e1d4d82a", "seconds": 10.96, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command, not a cancellation or a new instruction. The initial request is not executed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2929fd157401811b0e7601e3fff247fb29f3d2de0ccbc43345cc2859da077c93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_2/audio_natural/filler/original/e4af33a2c0fb9b78/speech.wav", "wer": 0.36363636363636365, "wav_sha256": "0e39632eba05f623c97ab90c4e9d1f09a5d35ba0c13e20d714501eda8d75c10c", "seconds": 6.56, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for such minor hesitations or pauses that do not alter the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to—wait, I meant received money from this month.", "long_pause_offsets": [], "source_sha256": "68a2423d7fd60fdc28658d4c7c14ef874e7ff7906e6e372f9da3ab9fae92b035", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately corrects to 'received money from', which matches the original instruction. The app (Venmo), action (befriend), and time condition (this month) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/false_start/false_start_02_transaction_relation/0a4a98cdc1e587db/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "e6612af679d89868ec6e440042b006ac4a69fbdfeda00333472ade4d4aa4e81f", "seconds": 9.48, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to—uh... wait, I meant um... received money from ... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately self-corrects to 'received money from' with 'wait, I meant'. This is a false start where the final corrected value is the intended goal. The correction is immediate and unambiguous, and the rest of the instruction ('this month') remains unchanged. The final intent matches the original instruction.", "heard_transcript": "Befriend on Venmo anyone I have sent money to uh wait I meant um received money from this month." }, "candidate_id": "wav_9df411672b6a2f16", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then self-corrects with a natural-sounding 'wait, I meant' followed by a smooth revision of the verb. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels spontaneous rather than rehearsed. The pause before 'wait' is brief and plausible, fitting a moment of reconsideration." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—sorry, make that Venmo anyone I have received money from this month.", "long_pause_offsets": [], "source_sha256": "5fcb6392f631813394d29b03fa25a33705547f08b88c735ea1d497299b978b76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, recipient scope, and time condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/false_start/false_start_01_app/4f110b200288f9db/speech.wav", "wer": 0.35, "wav_sha256": "d3e29ffaa7afce12e60f893fad5e78ccc95a22fd82d15f2d08ab48e7904737e7", "seconds": 12.16, "spoken_utterances": [ "Befriend on CashApp—um... sorry, make that hmmm... Venmo anyone I have received money from um, let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' with 'Sorry, make that whom. Venmo...'. This is a false start where the initial incorrect value is self-corrected. The final instruction matches the original reference, so the task should be executed as corrected.", "heard_transcript": "Befriend on Cash App. Sorry, make that whom. Venmo anyone I have received money from. Um, let's see. This month." }, "candidate_id": "wav_b3662e7480bbfe0d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a false start ('befriend on cash app'), corrects to 'Venmo', and then continues. The correction is immediate and the flow is mostly natural. However, the initial error ('cash app') is more jarring than in audio_02, and the filler 'um' after 'from' feels slightly less integrated. The overall delivery is still realistic but slightly less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have received money from last month—no, that should be this month.", "long_pause_offsets": [], "source_sha256": "7df542a57e7fee640822cb37a2edadbb31d20be9dafd40ec29f0b2045db5f438", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/false_start/false_start_03_time_reference/a96efd027481ef72/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "95464528e1d814a7482a493ff77bd1b7a8dd57cd36218496d603bf311831013f", "seconds": 9.08, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from last month—uh... no, that should be this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'no, that should be this month.' This is a clear false start where the final value is the corrected one. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "befriended on um Venmo anyone I have received money from last month uh no, that should be this month." }, "candidate_id": "wav_9467dede61559e53", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a hesitant 'um' and a slightly awkward phrasing ('befriended on um Venmo'). The correction ('no, that should be this month') is clear but feels more like a post-hoc realization than a spontaneous self-correction. The pause before the correction is longer and less naturally integrated, and the overall prosody is less continuous, giving it a slightly more rehearsed or disjointed quality compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have received money from. Oh, I forgot to mention, only include people from this month.", "long_pause_offsets": [ 53 ], "source_sha256": "c25077207fe7783e53d4c9d8e268aa537a2baff162708e5912b8f79833832b21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint, narrowing the initial broad request to only those people from 'this month'. The action, app, and recipient scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/addition/addition_01/b6f161caf9ca4435/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "cd1ee08381a8ca4a83d0507fa8bee613ad5ef55cc50fd080de4dcd31b81ddbe5", "seconds": 13.56, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from. ", "Oh, I forgot to mention, only include people from uh, let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to befriend everyone from whom money was received, then adds a constraint specifying 'this month' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final audible intent is clear and complete, matching the reference instruction exactly.", "heard_transcript": "Befriend on Venmo anyone I have received money from. Oh, I forgot to mention, only include people from uh, let's see, this month." }, "candidate_id": "wav_e2ea8e660a189018", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a believable self-correction. The initial phrase 'befriended on um Venmo anyone I have received money from' is followed by a plausible filler ('oh I forgot to mention') and a revision ('only include people from uh let's see this month'). The intonation, pacing, and use of 'uh' and 'let's see' sound spontaneous and conversational, with a clear change-of-mind flow. The addition of the time restriction is integrated naturally into the ongoing thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone this month. I left out a detail: only befriend people I have received money from.", "long_pause_offsets": [ 37 ], "source_sha256": "ca42c37946f912b927573072480495dbd8befcf6a7f18ba6329cb5455b7b9259", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only those people from whom money was received, preserving the app and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/addition/addition_02/aa735169b9c4e02e/speech.wav", "wer": 0.1, "wav_sha256": "972e68c42c09ec8716eac415a54b1806ca53620c2f1f91733fe00c3dc2271ec5", "seconds": 12.32, "spoken_utterances": [ "Befriend on Venmo anyone uh... this month. ", "I left out a detail: only befriend people I have ... received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Befriend on Venmo anyone this month' and then adds a clarifying constraint: 'Only befriend people I have received money from.' This addition narrows the scope of the original instruction, which is consistent with the 'addition' type. The final intent combines the original request with the added constraint, matching the reference instruction. No errors or contradictions remain.", "heard_transcript": "Befriend on Venmo anyone uh this month. I left out a detail. Only befriend people I have received money from." }, "candidate_id": "wav_db4682390a79f4b6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip begins with a clear request but includes a noticeable pause and a self-repair ('I left out a detail') that feels slightly more rehearsed or segmented than truly spontaneous. The correction ('only befriend people I have received money from') is delivered with a flat, declarative tone, lacking the natural intonation and hesitation of a real-time thought process. While the content is coherent, the delivery is less fluid and more like a recited instruction than a natural spoken request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "522e5e5_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have received money from this month. Let me correct the app: befriend on Venmo, not PayPal.", "long_pause_offsets": [ 65 ], "source_sha256": "2d1f8e8a372d8a6a0c2e59a3acddf1c4fc6a3e0df25d7dbfe7c382a934e863b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from PayPal to Venmo, while retaining the action, condition, and scope. The final goal matches the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on PayPal anyone I have received money from uh... this month. ", "Let me correct the app: befriend on um... Venmo, not PayPal." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month. Change the relation to 'received money from', not 'sent money to'.", "long_pause_offsets": [ 58 ], "source_sha256": "0ddb32c48637beb6bca9d7216c47b06eed5f2edf02593e3a58222ea99f800c2f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the relation from 'sent money to' to 'received money from', aligning with the original instruction. The app and time scope remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent money to uh... this month. ", "Change the relation to '... received money from', not 'sent money to'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have received money from last month. That should be 'this month', not 'last month'.", "long_pause_offsets": [ 64 ], "source_sha256": "ee35d678a194662968f6f6e67c46428427f3d61853b4c39d1677bf8eb224e702", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'last month' to 'this month', preserving the action, app, and recipient condition." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on Venmo anyone I have received money from um... last month. ", "That should be 'uh... this month', not 'last month'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "522e5e5_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "cddfbba2ee7fc5c5118f74c07d3a065a823938734a2b7f77957933ab6aa4d270", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/cancellation/cancel_002/a5ae24c5bdc5ea31/speech.wav", "wer": 0.2, "wav_sha256": "22e5f67ae21027fff337e352b0a515b69e8c6b65691103d842eb583d9926167d", "seconds": 11.6, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month.", "Wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly and explicitly says 'cancel the whole task' as the final interjection, which constitutes a direct cancellation of the entire request. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Befriend on um Venmo anyone I have received money from uh let's see this month. Wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "1712f2f10d421239610e95af70c6287b7ca1707dbac42d8dae66c30a37e24f31", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/ambiguous_halt/halt_001/d0dc739b22244f54/speech.wav", "wer": 0.25, "wav_sha256": "8f87458be6e4142cf9c3ab29d3030d449a6bd32dfed6e03fed9d22657633da05", "seconds": 10.28, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'wait' after completing the full instruction. This constitutes a clear request to pause or hold the current task, with no indication of cancellation or permission to resume. The instruction is fully audible and understood, and the final interjection is a pause command. Therefore, the final intent matches the expected behavior for an 'ambiguous_halt' type.", "heard_transcript": "befriend on um Venmo anyone I have received money from uh let's see this month. wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "522e5e5_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3be549474f2bf811512be14ddf9911e597cd5f5a11bce3d7cdee2ccf3866f4c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/522e5e5_3/audio_natural/filler/original/e3e02160e1ae4292/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "6691de3727d70c759c1fb7f1b77433e3eebc9dc02a5d2d7aaef2cdd1a294ecbd", "seconds": 7.72, "spoken_utterances": [ "Befriend on um... Venmo anyone I have received money from uh, let's see... this month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not change the core request. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "Befriend on um Venmo anyone I have received money from uh let's see this month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all treadmills—no, I meant weightlifting benches in my amazon cart.", "long_pause_offsets": [], "source_sha256": "e96dc248e418a8cd40deb1f110d12fd4d2b08b2183ce0514b8f9c8d6213cac50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'treadmills' to 'weightlifting benches' within the same utterance, preserving the action, scope ('all'), and destination ('amazon cart')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/false_start/false_start_01_product_category/f5c2ffed56c476e0/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "eaa6dcb0fbffc369e4e97544441c401238b20d4fa9ee16e63e6a63fa0a3eeb77", "seconds": 10.84, "spoken_utterances": [ "Place an order for all ... treadmills—uh... no, I meant hmmm... weightlifting benches in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'treadmills' but immediately self-corrects to 'weightlifting benches' after a hesitation. According to the 'false_start' type criterion, the final corrected value should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly, with no additional or missing constraints.", "heard_transcript": "Place an order for all treadmills, uh, no, I meant, uh, weightlifting benches in my Amazon cart." }, "candidate_id": "wav_5cc310ec0a210c81", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. The speaker begins with 'treadmills', then immediately self-corrects to 'weightlifting benches' with a believable 'no, I meant' structure. The correction is integrated smoothly into the ongoing sentence, and the prosody includes a natural hesitation and intonation shift. The flow feels like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all weightlifting benches in my amazon wishlist—sorry, that should be cart.", "long_pause_offsets": [], "source_sha256": "9aa766af20ac72ac9f3fb41bb05f8ed291d0441c584c17774da9208cb4237c13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects it to 'cart'. The final target is the cart, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/false_start/false_start_03_source_collection/df165d3b31cf4030/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "64fd810068a11723a0b2f6489c8aa4ef9080e088b4baaf99cc9cf7c9a024e694", "seconds": 9.32, "spoken_utterances": [ "Place an order for all weightlifting benches in uh... my amazon wishlist—um... sorry, that should be cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately self-corrects to 'cart'. According to the 'false_start' type criterion, the final corrected value ('cart') should be applied, and the original incorrect value ('wishlist') should be discarded. The final intent matches the original instruction with the corrected target.", "heard_transcript": "Place an order for all weightlifting benches in uh my Amazon wishlist um sorry, that should be cart." }, "candidate_id": "wav_50ab2ada02945613", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'wish list' followed by a self-correction to 'cart'. The correction is immediate and grammatically correct, but the prosody is slightly less fluid than audio_03. The 'um' filler is placed naturally, but the overall delivery is more straightforward and less dynamically varied in intonation, making it slightly less spontaneous in feel." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all weightlifting benches in my eBay—I should say amazon cart.", "long_pause_offsets": [], "source_sha256": "bdc96d660402ebcde8270016c94525888ee336bfea25906e949746c4350ae24a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately corrects to 'Amazon cart'. The final target app is Amazon, and the action and item scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/false_start/false_start_02_app/5890e02df3d52a5c/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "7a474f85542240c72b1a202ae334a3867345179bd8d07f4047ea431fe6b0f892", "seconds": 9.0, "spoken_utterances": [ "Place an order for all weightlifting benches in uh... my eBay—hmmm... I should say amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon cart'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for all weightlifting benches in uh my eBay hmm I should say Amazon cart." }, "candidate_id": "wav_035f6ba55d0f8008", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'eBay' and a correction to 'Amazon cart'. However, the correction is less integrated and more abrupt. The speaker says 'I should say Amazon cart' as a separate clause rather than a smooth self-repair within the sentence. The prosody is flatter, and the transition feels more like a correction than a natural thought process, reducing its realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for all weightlifting benches. Oh, I forgot to mention, place the order from my amazon cart.", "long_pause_offsets": [ 46 ], "source_sha256": "151b3a5b3a1373ffb1b0bab489087f0e400f30c03b75e431a4e368f9c5b58d7a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source context 'from my Amazon cart' without contradicting or removing any prior requirement. The action, item, and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/addition/addition_01/263a90a100777c70/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "309c215b9d7fbc423fc7a6c120f634f261c4a0a3764e2639766cd4c8366bfc31", "seconds": 13.08, "spoken_utterances": [ "Place an order for all uh, let me think... weightlifting benches. ", "Oh, I forgot to mention, place the order from ... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Place an order for all weightlifting benches' is followed by a longer pause and an addition: 'place the order from my Amazon cart'. This addition clarifies the source of the order, which is a relevant constraint. The final intent combines the original request with the added detail, matching the reference instruction. The hesitation and self-correction do not change the core task, and the addition is clearly intended to refine the original request.", "heard_transcript": "Place an order for all, uh, let me think, weightlifting benches. Oh, I forgot to mention, place the order from my Amazon cart." }, "candidate_id": "wav_2676213390faba67", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon cart. I should also mention, it's specifically for weightlifting benches.", "long_pause_offsets": [ 42 ], "source_sha256": "b938e22895387d5e9a9ef1d237a6c9338a997e470e93eb24c4aefaa5bbbee8b1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing item type 'weightlifting benches' as a clarification, which combines with the initial instruction to form the complete goal." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for all in uh... my amazon cart. ", "I should also mention, it's specifically for hmmm... weightlifting benches." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all treadmills in my amazon cart. Make that all weightlifting benches, not treadmills.", "long_pause_offsets": [ 53 ], "source_sha256": "0ca3202e7fdfd8671c6a2316a8a493ac287cf712ce52ad878a9b445aae086987", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial product category from treadmills to weightlifting benches, preserving the action, app, and scope of 'all' items." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/revision/revision_01_product_category/a1b3856776c5fa89/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "70a052f4fd3651b9b5c8446271b8e613741491f3f5732a5609ecb5598550fd11", "seconds": 11.88, "spoken_utterances": [ "Place an order for all treadmills in uh... my amazon cart. ", "Make that all ... weightlifting benches, not treadmills." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'treadmills' but immediately corrects to 'weightlifting benches' after a pause and the phrase 'Make that all'. This constitutes a revision where the final audible intent is to order weightlifting benches, matching the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Place an order for all treadmills and uh, my Amazon cart. Make that all, weightlifting benches, not treadmills." }, "candidate_id": "wav_7a2fc608329fd162", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('treadmills and uh...'), then smoothly corrects the item to 'weightlifting benches' with a natural-sounding self-repair ('Make that all...'). The prosody is consistent, and the correction flows logically within the utterance without awkward pauses or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all weightlifting benches in my Walmart cart. Let me correct the app: place the order in Amazon, not Walmart.", "long_pause_offsets": [ 65 ], "source_sha256": "0a3a412c283aa39595bbb8043cf9a4450d97d22e4c0782fa36aea7a8240eb221", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, while retaining the action, item type, and scope of 'all weightlifting benches'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/revision/revision_02_app/270ee5aee212ce3a/speech.wav", "wer": 0.0, "wav_sha256": "18e3ff2bcaa920269125c6c53392794efc438c773334acc91a609f7ba2121e74", "seconds": 11.76, "spoken_utterances": [ "Place an order for all let's see... weightlifting benches in ... my Walmart cart. ", "Let me correct the app: place the order in Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to place an order in Walmart, but immediately corrects it by stating 'Place the order in Amazon, not Walmart.' This constitutes a revision where the final intent is to execute the original task but with the correct target app. The correction is clear and unambiguous, and the rest of the instruction (ordering all weightlifting benches) remains unchanged. Therefore, the final intent matches the original instruction with the corrected target.", "heard_transcript": "Place an order for all, let's see, weightlifting benches in my Walmart cart. Let me correct that. Place the order in Amazon, not Walmart." }, "candidate_id": "wav_152758364677fb63", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, includes a filler ('let's see'), and then corrects the platform from 'Walmart' to 'Amazon'. The correction is delivered with a natural-sounding self-repair ('Let me correct that...'), and the prosody remains consistent. However, the initial hesitation and filler are slightly less smooth than in audio_03, making it slightly less realistic." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all weightlifting benches in my amazon wishlist. Wait, the order should be from the cart, not the wishlist.", "long_pause_offsets": [ 68 ], "source_sha256": "e65669b8c6e0eb102f9409bb8144df23a829bb335135e80b1bbf7356690f25ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from wishlist to cart, preserving the action, item type, and scope of 'all weightlifting benches'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/revision/revision_03_source_collection/7bbe41aae10d2af1/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "24676199925fbf80f693375e9a72fe3a906646a73ad15eb29d45fcd3c83e1f27", "seconds": 13.04, "spoken_utterances": [ "Place an order for all ah... weightlifting benches in ... my amazon wishlist. ", "Wait, the order should be from the ... cart, not the wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests placing an order from the wish list but immediately corrects themselves after a pause, specifying that the order should be from the cart. This is a revision where the final intent is the corrected version, which matches the original instruction. The correction is clear and unambiguous.", "heard_transcript": "Place an order for all uh weightlifting benches in my Amazon wish list. Wait, the order should be from the cart, not the wish list." }, "candidate_id": "wav_9336f02784c67389", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then corrects the source from 'wishlist' to 'cart'. The correction is delivered with a natural-sounding hesitation ('Wait, the order should be...'), but the overall delivery is flatter and less prosodically varied than the others. The correction feels slightly more abrupt and less integrated into the flow of speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "fcff850aab9d81701df2d5a227a448a71129645938c174f762f201b47b75663d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/cancellation/cancel_011/35877ae69cbb5797/speech.wav", "wer": 0.15, "wav_sha256": "936aa60cdb1611345c6e058682131d5a1341babeced04884378c85368d25c2d6", "seconds": 11.8, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart.", "Oh, wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then immediately corrects with 'Oh wait, I've changed my mind. Cancel that.' This constitutes a clear cancellation of the entire task. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for all um weightlifting benches in my amazon cart. Oh wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "d1d331144ad80d399bd0df6630ed0327f12fe754cd6b716bca082479e9c7f57c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/ambiguous_halt/halt_024/336d7d99b319fc25/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "2b9962bd52b3c12454ed7db45208b8cbc2ef9278028f01c2b3873bc7f8235612", "seconds": 10.2, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not resume or complete the task. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Place an order for all um weightlifting benches in my amazon cart. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "12c667f4f24b47b06917011d17734da074fd7f71b2cff63be963e9f731d155eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_1/audio_natural/filler/original/bbfc2b16b4546000/speech.wav", "wer": 0.25, "wav_sha256": "0b696832fbb8a5ad65827677f3ff1b20a4cb6bacb3d62028b1d3bb8382a9d832", "seconds": 6.88, "spoken_utterances": [ "Place an order for all hmmm... weightlifting benches in ... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task with a filler word ('um') but completes the original instruction without any correction or change in intent. The final audible goal matches the original instruction, so the task should be executed.", "heard_transcript": "Place an order for all um weightlifting benches in my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all beach towels—I meant beach balls in my amazon cart.", "long_pause_offsets": [], "source_sha256": "7ea67262c32e47106978f5fb1e0f071de7db338df30833b35737f002b38d844f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately corrects to 'beach balls' using 'I meant'. The final intended action is to order beach balls, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/false_start/false_start_01_product_category/757a8adb27cc888a/speech.wav", "wer": 0.0625, "wav_sha256": "3d7a75d2d24318c6a8cbe0ebf5f82d1adeca2cf10442e48dc69d61dc55a291ed", "seconds": 7.8, "spoken_utterances": [ "Place an order for all beach towels—uh... I meant ... beach balls in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately self-corrects to 'beach balls' with 'I meant'. According to the false_start type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Place an order for all beach towels. Uh, I meant beach balls in my Amazon cart." }, "candidate_id": "wav_b75334cc68e949a7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a clear false start ('beach towels') followed by a natural self-correction ('I meant beach balls') and a smooth continuation. The intonation and pacing are very natural, with a believable hesitation and a seamless transition. The correction is integrated into the flow of the sentence, making it sound like a spontaneous thought process rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all beach balls in my amazon wishlist—no, that should be cart.", "long_pause_offsets": [], "source_sha256": "c132357b5494069d39833e6b8b662f15a9ccdfe3dc619a5f557fb6c94cc9cf03", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects it to 'cart'. The final intended target is the cart, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/false_start/false_start_03_source_collection/49cc2f278df768a5/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "2d7e7cd2fa9e9f87326fa7384dfd3fc39a760787c90b418c74302fefb7cb2cbb", "seconds": 10.2, "spoken_utterances": [ "Place an order for um... all beach balls ... in my amazon wishlist—uh... no, that should be cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately self-corrects to 'cart'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for um all beach balls in my amazon wishlist uh no, that should be cart." }, "candidate_id": "wav_76ae15ba9d5799cc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start ('wishlist') corrected to 'cart', but the correction is abrupt and the sentence is cut off mid-word ('no, that should be cart'). The prosody is less fluid than audio_03, and the correction feels more like a mechanical fix than a natural thought process. The trailing 'uh' and the incomplete final word reduce the sense of spontaneity." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all beach balls in my Etsy—sorry, make that amazon cart.", "long_pause_offsets": [], "source_sha256": "93d0ffa7764fa0ce30a567ab1eabc6695317387eb3a1cec268b8c424a8bc8f8c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon' within the same utterance, preserving the action, item, and scope of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/false_start/false_start_02_app/38b8accc08004211/speech.wav", "wer": 0.1, "wav_sha256": "c5f9cb84ff2178022206541a2d40a07b3deb9e3a0fa12fdba10541fa6ea4cfc0", "seconds": 10.12, "spoken_utterances": [ "Place an order for all beach balls uh, let me think... in my Etsy—hmmm... sorry, make that amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Etsy but immediately self-corrects to Amazon cart. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction with the correct target (Amazon cart).", "heard_transcript": "Place an order for all beach balls, uh, let me think. In my Etsy, uh, sorry, make that Amazon cart." }, "candidate_id": "wav_f690694f67934e1f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip includes a filler ('let me think') and a revision ('Etsy' to 'Amazon cart'), but the revision is delivered with a noticeable pause and a hesitant tone that sounds less like a natural thought process and more like a forced correction. The 'uh' and 'sorry' feel slightly over-enunciated, and the overall flow is less continuous than the others, making it sound less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for beach balls in my amazon cart. Oh, right, place an order for all beach balls in my amazon cart.", "long_pause_offsets": [ 50 ], "source_sha256": "e7284f0d97269c610b65be7925214f72cb7bf5eaadc58e179d25efa9f29d3473", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the word 'all', which was missing in the initial utterance, completing the original instruction's scope. The app, action, and item remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/addition/addition_01/42e41fd791120de3/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "77c69c78275851bd385ebc9e0b2fc86cf3f4df971ff78987c2c5c804d7891022", "seconds": 11.36, "spoken_utterances": [ "Place an order for beach balls ... in my amazon cart. ", "Oh, right, place an order for uh... all beach balls in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for beach balls in my Amazon cart.' This is followed by a self-correction where 'all' is added before 'beach balls' after a pause. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent is to place an order for 'all beach balls' in the Amazon cart, which matches the reference instruction. The addition of 'all' is a narrowing of the initial request, and the final combined goal is identical to the reference.", "heard_transcript": "Place an order for beach balls in my Amazon cart. Oh right, place an order for uh all beach balls in my Amazon cart." }, "candidate_id": "wav_7b762531d313723d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident request, then naturally self-corrects with a filler ('uh') and a plausible revision of the item ('all beach balls'), maintaining a smooth, continuous flow. The prosody is natural, with appropriate pauses and intonation that suggest spontaneous thought. The correction feels integrated into the utterance rather than tacked on." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for all in my amazon cart. I left out a detail: only place an order for beach balls.", "long_pause_offsets": [ 42 ], "source_sha256": "a01cc0d3869490e1b8e1960557ca011760cf4fc5b740dca1982eb2a7ad81607d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only beach balls, preserving the action and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/addition/addition_02/92e4b1afd7af6b18/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "165fb33c24ed208187a572d4d4f2c966ea438956dee1d9108b0dc2687d9e868e", "seconds": 12.0, "spoken_utterances": [ "Place an order for um... all in my amazon cart. ", "I left out a detail: only place an order for hmmm... beach balls." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Place an order for all of my Amazon cart,' which is incomplete and grammatically incorrect. After a pause, the speaker corrects this by adding 'Only place an order for beach balls.' This addition narrows the scope of the original request, which matches the 'addition' type criterion. The final intent is to order only the beach balls, which is a valid and complete instruction. The original instruction's 'all' is overridden by the addition, and the final goal is clear and executable.", "heard_transcript": "Place an order for um all of my Amazon cart. I left out a detail. Only place an order for beach balls." }, "candidate_id": "wav_86f4b93d046ad1aa", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a hesitant 'um' and a slightly awkward phrasing ('all of my Amazon cart'), then pauses and revises the request. While the revision is logically coherent, the delivery feels more like a disjointed sequence of thoughts rather than a fluid, natural correction. The second part of the utterance sounds more like a separate thought than a spontaneous follow-up, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all beach towels in my amazon cart. Wait, the product category should be beach balls, not beach towels.", "long_pause_offsets": [ 55 ], "source_sha256": "1bdf9b95709689c4c2bf166988bb3d92ac82b45accfd9491aac8ce5e347b67d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the product category from beach towels to beach balls, while retaining the action, scope ('all'), and destination ('amazon cart')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/revision/revision_01_product_category/dff85e80a2cdab1d/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "1cd5b4b5f32f69a0b113da9f5b21a783017168cb177001792dd3c0f808c77219", "seconds": 11.64, "spoken_utterances": [ "Place an order for all beach towels ... in my amazon cart. ", "Wait, the product category should be uh... beach balls, not beach towels." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'beach towels' but immediately corrects to 'beach balls' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to order beach balls, which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "Place an order for all beach towels in my Amazon cart. Wait, the product category should be uh beach balls, not beach towels." }, "candidate_id": "wav_85951f46e93d950f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and continuous prosody with a believable hesitation ('wait') and a clear, self-corrected revision. The speaker's intonation rises slightly on 'wait' and then drops on the correction, matching a genuine moment of realization. The correction is smoothly integrated into the same utterance without awkward phrasing or unnatural pauses, making it sound like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all beach balls in my Walmart cart. Let me correct the app: place the order on Amazon, not Walmart.", "long_pause_offsets": [ 55 ], "source_sha256": "341e7fec39f407d991d2885af75876463ac28049e51414a4f0807d459a7a867c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, while retaining the action, item, and scope of 'all beach balls'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/revision/revision_02_app/c7df61b3de25a9ec/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "3640a924e79d0e7dfed7921da7267c3e9f6d4106132f2e85578f91f14baf0233", "seconds": 12.96, "spoken_utterances": [ "Place an order for um... all beach balls um, let's see... in my Walmart cart. ", "Let me correct the app: place the order on Amazon, not Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart cart' but immediately corrects it to 'Amazon, not Walmart'. This is a clear revision where the final instruction overrides the initial incorrect value. The final goal matches the original reference instruction, so the revision is correctly applied.", "heard_transcript": "Place an order for all beach balls, um, let's see, in my Walmart cart. Let me correct that. Place the order on Amazon, not Walmart." }, "candidate_id": "wav_7a2392592886522e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear self-correction ('let me correct the app, place the order on Amazon not Walmart'), but the initial request is slightly less fluent with filler words ('um') and a longer pause. The correction is delivered in a separate sentence, which is less fluid than audio_02's integrated revision. The prosody is still natural, but the flow is slightly more segmented." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all beach balls in my amazon wishlist. Change the source collection to cart, not wishlist.", "long_pause_offsets": [ 58 ], "source_sha256": "e60cc9dd2c75a0faa4a82c926f0433917aafe27a7738744ffe09c8734e510872", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source collection from wishlist to cart, preserving the action, item type, and scope of 'all beach balls'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/revision/revision_03_source_collection/2af14c809a0388fd/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "5aa4e636500124a243eecbda2be483e688117d9398e777d996d76b3981d30b9e", "seconds": 12.96, "spoken_utterances": [ "Place an order for hmmm... all beach balls um... in my amazon wishlist. ", "Change the source collection to cart, not wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects it by stating 'Change the source collection to cart, not wishlist.' This constitutes a revision where the final intent is to use the cart, not the wishlist. The correction is clear and unambiguous, and the rest of the instruction matches the original. Therefore, the final intent satisfies the revision criterion.", "heard_transcript": "Place an order for um all beach balls um in my Amazon wishlist. Change the source collection to cart, not wishlist." }, "candidate_id": "wav_4e1f52f8a3004564", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision, but the delivery is less natural due to multiple fillers ('um') and a more abrupt, less conversational tone. The correction ('change the source collection to cart not wishlist') is phrased more like a command than a spontaneous realization, and the pause before it feels slightly too long and disconnected. The overall prosody is flatter and less expressive compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_036", "hide_variant_original": true, "source_sha256": "501ed6244f43c42e3923d9a467c9e020ce6fdb6d240925cc7c62e83fb62ecea8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/cancellation/cancel_036/3fbffc82c98c1a07/speech.wav", "wer": 0.16, "wav_sha256": "2a7dbf3f8115dd9b1dfb91d4e1bb211906d2bb98dcffe865aac55d27c1c87951", "seconds": 12.28, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart.", "Hang on, don't continue with that request. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to state the original task, then after a pause and hesitation, explicitly says 'Hang on, don't continue with that request. Cancel it.' This constitutes a clear cancellation of the task. The final audible intent is to cancel, not to execute the order. The cancellation is complete and unambiguous.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart. Hang on, don't continue with that request. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "7137395e3e059953155e9a3aad967e75aaaea4524b3c359325c094df7d5cb3d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/ambiguous_halt/halt_038/ef173eb10d81198d/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "37062185f7b3a608a300d1b0a8a93bb3c2ccc8f66664ec6f895fe8529a6c6845", "seconds": 11.48, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly requests to pause with 'Could you pause there for a second?'. This constitutes a clear, final request to hold the task, matching the 'ambiguous_halt' type criterion. The initial task is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "95baae1bee359e73121e863a6600a5264e0b75ec2accba24f828a2e0da40b5c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_2/audio_natural/filler/original/b12e26214005cf46/speech.wav", "wer": 0.25, "wav_sha256": "cd50beba53bfef8bd2ae8bbad635fffce1b5488153a9e28c4030d0f4331569b7", "seconds": 7.56, "spoken_utterances": [ "Place an order for ah... all beach balls uh, let me think... in my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let me think') but does not change the core request. The final intent remains to place an order for all beach balls in the Amazon cart, matching the original instruction. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Place an order for uh all beach balls uh let me think in my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Place an order for all staplers in my ebay—sorry, I meant amazon cart.", "long_pause_offsets": [], "source_sha256": "4235f9e4fbe9a1386f0308ae58dde1b7eac913cb47ac811f69e914b3745ef94e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the action, item, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/false_start/false_start_02_app/3ea1075d346afc43/speech.wav", "wer": 0.0, "wav_sha256": "90051e4b8c61b69129701ee393e5d75575191596f6b0bbafd87375093f3409bd", "seconds": 7.52, "spoken_utterances": [ "Place an order for all staplers in my ebay—um... sorry, I meant ... amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon cart'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('eBay') is discarded. The final goal matches the original instruction with the corrected target.", "heard_transcript": "Place an order for all staplers in my eBay um, sorry, I meant Amazon cart." }, "candidate_id": "wav_bca9e8e347364f4b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_01_product_category", "proposed_instruction": "Place an order for all pens—make that staplers in my amazon cart.", "long_pause_offsets": [], "source_sha256": "8322ff37028660644fb72ca5338db5909d92fba36390e0c7ce36480d9fbccdc5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pens' but immediately corrects to 'staplers' within the same utterance. The final intended category is staplers, matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for all pens—um... make that staplers in uh, let's see... my amazon cart." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Place an order for all staplers in my amazon wishlist—no, that should be cart.", "long_pause_offsets": [], "source_sha256": "e8c4180541cb156684db80128a5d81f6150f000143b374c3072462989c1152b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects it to 'cart'. The final target collection is 'cart', matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for uh... all staplers in ... my amazon wishlist—um... no, that should be cart." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for staplers in my amazon cart. Oh, right, this should apply only to all staplers.", "long_pause_offsets": [ 47 ], "source_sha256": "29d3c47149a69918a1924d88a13ac04e0fb71b2df28104176929758c51ab182b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the initial request should be restricted to 'all staplers', which matches the original instruction. The initial utterance was broad ('staplers'), but the addition refines it to the exact scope specified in the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/addition/addition_01/e8b17bc3afc0e6c9/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "0505d9a852ff5d21cce1e7d7c4a484f2502d766422910b4a9b8615b98b844105", "seconds": 11.84, "spoken_utterances": [ "Place an order for staplers in uh... my amazon cart. ", "Oh, right, this should apply only to uh, let me think... all staplers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Place an order for staplers in my Amazon cart.' The speaker then adds the clarification 'this should apply only to all staplers' after a pause and hesitation. This addition narrows the scope of the initial request to 'all staplers', which is consistent with the original instruction. The final intent is to place an order for all staplers, matching the original instruction. The addition refines the request without contradicting it, satisfying the addition type criterion.", "heard_transcript": "Place an order for staplers in uh my Amazon cart. All right, this should apply only to uh let me think, all staplers." }, "candidate_id": "wav_46b52b88ed2c33df", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for all staplers. I should also mention, place the order in my Amazon cart.", "long_pause_offsets": [ 33 ], "source_sha256": "dca1053507c758a904926cf0a023b8569296ff18906e2b713f718252c11f3938", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context of the Amazon cart, which was implied in the original instruction but not explicitly stated. The final goal combines the initial action with the added constraint." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for ... all staplers. ", "I should also mention, place the order in uh, let's see... my Amazon cart." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "Place an order for all staplers in my amazon wishlist. Change the source to cart, not wishlist.", "long_pause_offsets": [ 55 ], "source_sha256": "97bc6c186ef129b52de24152e4e1bb1c5907fece9bd2c515376406cdfe1796f9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source from wishlist to cart, which aligns with the original instruction. The action, item scope, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/revision/revision_03_source_collection/51a6760a835b285e/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b7874aada24939ce788802079556b8528907423de8801bf62606f1dce4af907f", "seconds": 14.68, "spoken_utterances": [ "Place an order for um, let me think... all staplers in um... my amazon wishlist. ", "Change the source to uh... cart, not wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests placing an order for staplers in the wishlist, but then corrects the source to the cart. According to the revision type criterion, the correction spoken after a pause must be applied to the earlier instruction. The final audible intent is to order all staplers from the cart, which matches the original instruction's goal and target, despite the initial error.", "heard_transcript": "Place an order for um, let me think, all staplers in um, my Amazon wishlist. Change the source to uh, cart, not wishlist." }, "candidate_id": "wav_7a187ab5e5892e39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a self-initiated thought process ('let me think'), which is followed by a clear, coherent request. The revision ('change the source to cart not wishlist') is delivered with a smooth, conversational tone and a believable shift in focus. The prosody is consistent, and the correction flows naturally without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Place an order for all pens in my amazon cart. Sorry, I meant all staplers, not pens.", "long_pause_offsets": [ 47 ], "source_sha256": "0540b755022e8eee99686b110d4031d539ca3f0f1bf0e8e22826735780870100", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the item category from 'pens' to 'staplers', while retaining the action, scope ('all'), and context ('in my amazon cart')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/revision/revision_01_product_category/6e9e0d30bcac7bd1/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "eed28d0c72da47af61251b3e234a13da8e3d2348d2f3afc1a8231c7556f151ba", "seconds": 12.72, "spoken_utterances": [ "Place an order for hmmm... all pens in uh... my amazon cart. ", "Sorry, I meant um, let me think... all staplers, not pens." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all pens' but immediately corrects to 'all staplers, not pens' after a self-correction. According to the revision type criterion, the correction spoken after a pause (or self-correction) overrides the initial instruction. The final audible intent is to order all staplers, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints are altered.", "heard_transcript": "Place an order for whom? All pens in uh my Amazon cart. Sorry, I meant um let me think. All staplers, not pens." }, "candidate_id": "wav_0bbe15aa16288c54", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a hesitant 'hmm' and a slightly awkward 'uh' before the request. The initial request is somewhat unclear, and the revision ('sorry, I meant... all staplers, not pens') is delivered with a noticeable shift in intonation and a self-correction that sounds slightly rehearsed. While the content is understandable, the delivery is less fluid and the pauses feel more forced compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Place an order for all staplers in my eBay cart. Wait, place the order on Amazon, not eBay.", "long_pause_offsets": [ 49 ], "source_sha256": "29bcfbeb64c0c4dc199b009093226b4d396dd9b4973f926138c07db8723206b4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, while retaining the action, item type, and scope of 'all staplers'." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for uh... all staplers in let me think... my eBay cart. ", "Wait, place the order on ... Amazon, not eBay." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "44fd4b3867dd6adc6cd57aea46604c1cb0d1325daf6546de1fc48ed23175b286", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/cancellation/cancel_030/af0cee1894c9fec5/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "2920a69fff86da375678666262371f67762042562da3cae792f60e3d67f521b6", "seconds": 11.72, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart.", "Hold on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by requesting to place an order for all staplers in the Amazon cart, but then clearly revokes the entire task by stating, 'Hold on, you can drop that task. I no longer need it.' This constitutes a direct cancellation of the original instruction. The final intent is to cancel, not to execute or pause, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for all staplers in um my Amazon cart. Hold on, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "ffc55789a11a26bdcaf52711f8e47152580a9493c7313753312ee82dbb6ecd56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/ambiguous_halt/halt_037/373c81935c183265/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "1e79b91a15549ebe097c1ee009aa8376a2dd177eed17c2f6daa05d23d9ba98a8", "seconds": 9.72, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to place an order is not executed, but the final control intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Place an order for all staplers in um my Amazon cart. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5238afc_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "037f095e0bebaf5107d7b1c51e946a6c3abebbe9401ec3e0404d78223433c1e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5238afc_3/audio_natural/filler/original/265c1ce484ce601b/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6851e5fbdcdcf9e13736374542e2e368831f2e5d453413e73b2a7eb6bc9b6aca", "seconds": 6.52, "spoken_utterances": [ "Place an order for ... all staplers in um... my amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. The task should be executed as requested.", "heard_transcript": "Place an order for all staplers in um my Amazon cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill—sorry, I mean phone bill on venmo this year so far?", "long_pause_offsets": [], "source_sha256": "d48c0fa03874aaf2ade2fd102f3dcddddef56167e842015cd3369d9943b10ab5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'electricity bill' to 'phone bill' within the same utterance, preserving the app (Venmo), time frame ('this year so far'), and the action of querying the total amount paid." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/false_start/false_start_01_bill_category/535062db11addd11/speech.wav", "wer": 0.05, "wav_sha256": "01faa72b71acb50e8ec75b82f6d1b15543bc33953129674e81c045406f5921ff", "seconds": 9.2, "spoken_utterances": [ "How much have I paid in electricity bill—um... sorry, I mean ... phone bill on venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately self-corrects to 'phone bill'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the app (Venmo) and time frame ('this year so far'), remains unchanged and matches the reference. Therefore, the final audible intent matches the original instruction after the correction.", "heard_transcript": "How much have I paid in electricity bill um sorry I mean phone bill on Venmo this year so far?" }, "candidate_id": "wav_0687e50f20115cca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with a clear self-correction from 'electricity bill' to 'phone bill'. The filler 'um' is placed at a plausible point of hesitation, and the correction is integrated smoothly into the sentence without awkwardness. The prosody is consistent, and the overall flow mimics a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in phone bill on PayPal—no, that should be venmo this year so far?", "long_pause_offsets": [], "source_sha256": "731e2b52ce4575c5df878b20c703dd0d29e82029ef58bf2894cd2c4a20a1a7df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the rest of the query including the category (phone bill), time frame (this year so far), and the intent to retrieve a monetary amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/false_start/false_start_02_app/5b6c25fb3374363c/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "5b74da87334d26be0218339bdb091ff95c675bedab5e2c8c4bf6b2e55a38e52d", "seconds": 11.04, "spoken_utterances": [ "How much have I paid in uh, let me think... phone bill on ... PayPal—ah... no, that should be venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'no, that should be Venmo'. This is a false start where the final value is corrected. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in uh let me think phone bill on PayPal uh no, that should be Venmo this year so far?" }, "candidate_id": "wav_09114ec2e181b015", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a more complex self-correction with multiple fillers ('uh', 'let me think') and a switch from 'PayPal' to 'Venmo'. While the repair is realistic, the multiple hesitations and the slightly drawn-out correction make the delivery feel a bit more rehearsed or less fluid than audio_02. The speaker's voice remains consistent, but the pacing is slightly less natural." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much have I paid in phone bill on venmo last month—wait, I meant this year so far?", "long_pause_offsets": [], "source_sha256": "2c6674b532065bf39d5906380c31e9328cea9da08fb1596ec6e19cbbb260607d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects to 'this year so far'. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/false_start/false_start_03_time_reference/5a502fb06ffe48dd/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "e06bb77c0fbc8117aee12870384beca15c1f31bf15229c32b9c5cf1486b225e0", "seconds": 8.8, "spoken_utterances": [ "How much have I paid in uh... phone bill on venmo last month—hmmm... wait, I meant this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects with 'Wait, I meant this year so far?' which clearly indicates the intended final goal. This matches the 'false_start' type criterion where the immediate self-correction overrides the initial mistaken value. The final intent is clear and executable.", "heard_transcript": "How much have I paid in uh phone bill on Venmo last month? Hmm. Wait, I meant this year so far?" }, "candidate_id": "wav_15beb89ad9982ca8", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start from 'last month' to 'this year so far', but the correction is abrupt and lacks the natural flow of the other two. The pause before 'Wait' feels slightly too long and disconnected, and the intonation on 'Wait' sounds more like a script than spontaneous speech. The overall delivery is less convincing in its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in phone bill on venmo so far? I didn't mention this, but it's so far this year.", "long_pause_offsets": [ 52 ], "source_sha256": "6712e6a07828b92362f95e9530c5685eff78f918c78cebbfc36fe8ea6c48e950", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time frame 'this year' to the initial question, refining the scope without changing the core request about the phone bill amount on Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/addition/addition_03/46f0b5c95443059f/speech.wav", "wer": 0.125, "wav_sha256": "fcd1b42b55951fa59951c7651fcc72f82ad814feedb88f7feb2dddcedee4c823", "seconds": 12.08, "spoken_utterances": [ "How much have I paid in uh... phone bill on hmmm... venmo so far? ", "I didn't mention this, but it's so far this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the phone bill paid on Venmo 'so far'. After a pause, they add the constraint 'this year', which is a clarifying addition to the original instruction. The final intent combines the original request with the added time frame. The addition is clear and does not contradict the original request, so the task is executed with the combined scope.", "heard_transcript": "How much have I paid in uh phone bill on Venmo so far? I didn't mention this, but it's so far this year." }, "candidate_id": "wav_4e53ab433c9b2430", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant question, then seamlessly adds a revision ('I didn't mention this but it's so far this year') that corrects the time frame. The intonation and pacing are very human, with a believable 'hmm' and a natural-sounding correction that integrates smoothly into the overall request." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in on venmo this year so far? Oh, I forgot to mention, it's specifically for the phone bill.", "long_pause_offsets": [ 51 ], "source_sha256": "0a4501d46499cbb616202e44b5cb73b07bd57c2cd6ab8cd9dae8c67e1b081180", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'specifically for the phone bill' to the initial question, which already included the app (Venmo), time frame ('this year so far'), and action ('how much have I paid'). The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/addition/addition_01/dd4ef9b47ba4bf1c/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "8d584c4e86e3a3d13d3f8383abd16619e13d3f3e50de12375d31cd6e6db69c84", "seconds": 12.76, "spoken_utterances": [ "How much have I paid in on venmo hmmm... this year so far? ", "Oh, I forgot to mention, it's specifically for the ... phone bill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How much have I paid in on Venmo this year so far?' which is a partial version of the original instruction. After a pause, the speaker adds the missing constraint 'specifically for the phone bill'. This addition correctly narrows the scope of the original request to match the reference instruction. The final intent is a complete and correct execution of the original task.", "heard_transcript": "How much have I paid in on Venmo this year so far? Oh, I forgot to mention it's specifically for the phone bill." }, "candidate_id": "wav_00bb66176cf6ffa0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is a strong second, featuring a clear initial request followed by a believable addition ('Oh I forgot to mention it specifically for the phone bill'). The speaker's tone is natural, and the addition is phrased as a remembered detail. The only slight drawback is that the initial request is a bit more clipped and less fluid than audio_02, but the overall delivery remains very realistic." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in phone bill on this year so far? One more thing: make sure it's on Venmo.", "long_pause_offsets": [ 56 ], "source_sha256": "e299e3c163fbbbf518caa7f2ef28323f3b5f7c4657585f151f38e6e8947587a6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app constraint 'on Venmo', which was omitted in the initial utterance. The final goal combines the original intent with the added condition, resulting in the same meaning as the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/addition/addition_02/74b0da6b772416ad/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "08cc84c88e628c17b4456e57cfb1dc5c3041fdd7197c1d9bab8d53e6f952b9a6", "seconds": 13.24, "spoken_utterances": [ "How much have I paid in um... phone bill on um, let's see... this year so far? ", "One more thing: make sure it's on um... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How much have I paid in phone bill on um let's see this year so far', which is a partial and hesitant version of the original instruction. The speaker then adds 'One more thing, make sure it's on um Venmo' after a pause, which adds the specific condition that the payment must be on Venmo. This matches the 'addition' type, where the final goal combines the original request with the added constraint. The final audible intent is fully aligned with the original instruction, with the addition of the Venmo condition, which is already part of the original instruction. Therefore, the final intent is correctly executed.", "heard_transcript": "How much have I paid in um phone bill on um let's see this year so far. One more thing, make sure it's on um Venmo." }, "candidate_id": "wav_f0a4ff1363b178f8", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic of the three. The speaker starts with a clear question but then makes a disjointed addition ('One more thing, make sure it's on um Venmo') that sounds tacked on rather than naturally integrated. The prosody is less fluid, and the transition between the initial request and the addition feels abrupt and less spontaneous compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in phone bill on PayPal this year so far? Wait, use Venmo, not PayPal.", "long_pause_offsets": [ 63 ], "source_sha256": "7823b68ac7896c30801f081ddf1c57171b618c45dfb73510021e053cbddd1992", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, preserving the rest of the query including the expense category, time frame, and intent to retrieve a total amount." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/revision/revision_02_app/23c00eb14c17eb5f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3c5bcee66f0a9f0dfe37dd5a9140b94c082d4c46ba36f35bc3f105d5678514ff", "seconds": 11.64, "spoken_utterances": [ "How much have I paid in uh... phone bill on hmmm... PayPal this year so far? ", "Wait, use Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query the phone bill amount on Venmo, which matches the original instruction. All other constraints (amount, time frame 'this year so far') remain unchanged.", "heard_transcript": "How much have I paid in uh phone bill on PayPal this year so far? Wait, use Venmo, not PayPal." }, "candidate_id": "wav_3d276454a79f722d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant phrasing ('uh') and a brief throat clear, which sounds like genuine speech preparation. The correction from 'PayPal' to 'Venmo' is delivered with a clear change in intonation and a self-correcting 'wait' that sounds spontaneous. The flow is continuous and the prosody remains consistent throughout, with no artificial pauses or robotic delivery." } ] }, "is_default": true }, { "variant_id": "revision_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill on venmo this year so far? Sorry, I meant phone bill, not electricity bill.", "long_pause_offsets": [ 68 ], "source_sha256": "4f910cf71cf2fd9adcbdf318266c56e3d99a23408bac5af9c11566af85f5af5d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'electricity bill' to 'phone bill', preserving the app (Venmo), time frame ('this year so far'), and action (querying total amount paid)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/revision/revision_01_bill_category/12c86a67c28673f2/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "8aeac0b68a6f52b9438a99a0d9313058e243a89012f97fe1dad151f8c9621b82", "seconds": 15.12, "spoken_utterances": [ "How much have I paid in hmmm... electricity bill on ... venmo this year so far? ", "Sorry, I meant ... phone bill, not electricity bill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately corrects it to 'phone bill' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to inquire about the phone bill, which matches the original reference instruction. All other conditions (Venmo, this year so far) are preserved.", "heard_transcript": "How much have I paid in electricity bill on Venmo this year so far? Sorry, I meant phone bill, not electricity bill." }, "candidate_id": "wav_0edc035eae2c13df", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also shows a natural hesitation and a self-correction, the initial error ('electricity bill') is corrected to 'phone bill' after a longer pause and a more formal 'sorry, I meant'. The correction feels slightly more rehearsed or deliberate compared to audio_01, and the overall delivery has a slightly flatter prosody. The correction is less fluid and more like a formal edit than a spontaneous thought correction." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much have I paid in phone bill on venmo last month so far? Make that this year, not last month.", "long_pause_offsets": [ 63 ], "source_sha256": "17c815be4a7317aa69f16fcc6e35e599c547cc509046df2b5343aca46a9391df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this year', preserving the action, app, and category. The phrase 'so far' is retained and interpreted as part of the time scope." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in um, let's see... phone bill on um... venmo last month so far? ", "Make that ... this year, not last month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "48f79c86063c50eb103cfb6cc1002449afe7b2391e5f19aac00f0d480c7238e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/cancellation/cancel_039/9d39e59c0361ec71/speech.wav", "wer": 0.0, "wav_sha256": "aa83b83996572bdf90fbd4a02b233e2539555cf329bebc62aafe3f2c34ba4c39", "seconds": 10.96, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?", "Oh, wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately corrects themselves by saying 'Oh wait, don't go ahead with that. Cancel it.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute the original query.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far? Oh wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "b8e251aa925415912e844bb08911836f9bc45f3d87c4b4eb3fb7ac512cb85220", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/ambiguous_halt/halt_016/cb0fb4c79ae737ff/speech.wav", "wer": 0.0, "wav_sha256": "b61ba3e43bd7536f5028a4d555ac4000eb463a902947bbaed197df6370dfef2c", "seconds": 9.36, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not revoked or corrected, and the final interjection is a temporary hold command, not a cancellation. Therefore, the final control intent is correctly interpreted as 'pause'.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far? Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bed073b402dc6f826da094b5cb1dc3a48a0d42b51b17e458d877a28696de1e6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_1/audio_natural/filler/original/89c4096991a3ad8c/speech.wav", "wer": 0.0, "wav_sha256": "98a26c49b95a3553f08aad09547a953af8e87928d14ed53e42214691527ef27b", "seconds": 5.48, "spoken_utterances": [ "How much have I paid in ... phone bill on Venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker delivers the full original instruction without any hesitation, self-correction, or change in intent. The audio contains no fillers that alter the task goal, and the final utterance matches the original instruction exactly. Therefore, the task is executed as intended.", "heard_transcript": "How much have I paid in phone bill on Venmo this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in electricity bill on cashapp—no, I mean venmo this year so far?", "long_pause_offsets": [], "source_sha256": "259ef0f5c39526fef5fb640a761b8fea1a56acd03af07f2aa8e771eb9d6f5ca7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the action, category, time frame, and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/false_start/false_start_02_app/c94948cc7f70aa6b/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "3f5b3701eb6e94d1ac1cd1e249c540bc0419db04a8100c57b2a9b6ff11308c08", "seconds": 9.96, "spoken_utterances": [ "How much have I paid in electricity bill on cashapp—hmmm... no, I mean venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, with only the corrected app name.", "heard_transcript": "How much have I paid in electricity bill on Cash App? Hmm. No, I mean Venmo. This year so far?" }, "candidate_id": "wav_a29437664a34cf9c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with a clear, slightly hesitant question, then self-corrects from 'cash app' to 'venmo' with a smooth, conversational tone and a natural pause. The intonation and rhythm are consistent with spontaneous speech, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "false_start_01_bill_category", "proposed_instruction": "How much have I paid in water bill—sorry, I meant electricity bill on venmo this year so far?", "long_pause_offsets": [], "source_sha256": "456f5a07be3d0e2e2eba703633063c473c6a33e2b471e206bc5efcd85cbce7b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'water bill' to 'electricity bill' within the same utterance, preserving the app (Venmo), time frame ('this year so far'), and the action of querying the total amount paid." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/false_start/false_start_01_bill_category/b77d4bcd3b070cc3/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d7b8cab6bd5eea658bf09c094b31bcf67c6292e3e21fbee28fd84112d92e24c5", "seconds": 9.08, "spoken_utterances": [ "How much have I paid in water bill—hmmm... sorry, I meant uh... electricity bill on venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'water bill' but immediately self-corrects to 'electricity bill'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final audible task matches the original instruction exactly, so the control intent is to execute.", "heard_transcript": "How much have I paid in water bill, um, sorry, I meant uh, electricity bill on Venmo this year so far?" }, "candidate_id": "wav_408a6bef5b521e57", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a plausible false start with a correction from 'water bill' to 'electricity bill'. However, the correction is slightly less fluid than audio_01, with a more abrupt shift and a filler word ('uh') that feels a bit more forced. The overall delivery is still realistic, but the repair lacks the seamless flow of the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much have I paid in electricity bill on venmo last month—wait, make that this year so far?", "long_pause_offsets": [], "source_sha256": "6e6fda9964de12c3ba265a1c35209002a7cb4b5e7488f40d61b03887fc711ae2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this year so far'. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/false_start/false_start_03_time_reference/4d158d780497484c/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ddaa6a910739abd745dcd30b90b3a1198b5027d21ce71e284dafaab938e88edb", "seconds": 9.2, "spoken_utterances": [ "How much have I paid in electricity bill on venmo um... last month—ah... wait, make that um... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this year so far'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in electricity bill on Venmo um last month uh wait make that um this year so far?" }, "candidate_id": "wav_b381f28e794f5df5", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a more complex and less natural repair. The speaker starts with 'last month', then hesitates with 'uh', and corrects to 'this year so far'. The multiple pauses and the awkward phrasing ('make that') sound less spontaneous and more like a rehearsed or artificially constructed edit. The prosody is less consistent, and the correction feels less like a genuine recollection." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in electricity bill on venmo so far? And just to be specific, it's so far this year.", "long_pause_offsets": [ 58 ], "source_sha256": "f59bf12dbd3b5843df7d36d72452375ca86bc50c436f1ac875022febbf915be7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the time constraint 'this year' to clarify the scope of 'so far', which is consistent with the original instruction's intent. The final goal combines the initial query with the added specificity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/addition/addition_03/15224bb20bd8c83b/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "d3a4eb31733aa82a240bfa4cdb18fc93a2edd54c932d22c768a907102dce8344", "seconds": 14.44, "spoken_utterances": [ "How much have I paid in ... electricity bill uh, let's see... on venmo so far? ", "And just to be specific, it's so far uh... this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the core question and then adds the clarifying constraint 'this year' after a pause, which is the defining characteristic of an 'addition' type. The final intent combines the original request with the added temporal scope, matching the reference instruction exactly. The hesitation and filler words do not alter the final goal.", "heard_transcript": "How much have I paid in electricity bill uh let's see on Venmo so far and just to be specific it's so far uh this year" }, "candidate_id": "wav_92a745941753f403", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker's delivery is highly natural, with a realistic flow of thought. The initial question is followed by a plausible 'uh let's see' hesitation, and the speaker then adds clarifying details ('on Venmo so far') and a time frame ('this year') in a continuous, conversational manner. The prosody is varied and the pacing reflects genuine formulation of the request, making it sound like a spontaneous, real-time inquiry." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in electricity bill this year so far? One more thing: make sure it's on Venmo.", "long_pause_offsets": [ 59 ], "source_sha256": "e66ac93482f2c56f725560a7e58a0dce8a93be2f2113459700d83a6243530af5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing payment method constraint, completing the original request with the correct app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/addition/addition_02/14f396de959f57ff/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "05529ba3972b7bbc33f65e835ed6ee46a7a21482a20f81f3dee46430c714a91e", "seconds": 13.6, "spoken_utterances": [ "How much have I paid in electricity bill let's see... this year so far? ", "One more thing: make sure it's uh... on Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How much have I paid in electricity bill on venmo this year so far?'. The speaker first asks the same question without specifying Venmo, then adds the constraint 'on Venmo' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added detail. The final audible intent is to find the total amount paid for electricity bills on Venmo this year so far, which matches the reference instruction.", "heard_transcript": "How much have I paid in electricity bill? Let's see. This year so far. One more thing, make sure it's uh on Venmo." }, "candidate_id": "wav_7fee635dc0d6c204", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker's delivery is also natural, the addition of 'one more thing' after the initial request feels slightly less integrated than in audio_01. The phrase 'one more thing' introduces a new clause that is somewhat abruptly tacked on, and the intonation shift is less smooth. The overall prosody is good, but the addition lacks the seamless flow of the first clip, making it sound slightly more like a post-hoc correction than a naturally evolving thought." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in on venmo this year so far? Oh, I forgot to mention, it's specifically for the electricity bill.", "long_pause_offsets": [ 51 ], "source_sha256": "e4a3084a54609e0e45ece6f17470d1b88f911bd8e04a5f2f42e99c2e1364522f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing category 'electricity bill' to the initial question about Venmo payments, refining the scope without changing the app, time frame, or action." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in hmmm... on venmo ... this year so far? ", "Oh, I forgot to mention, it's specifically for the ... electricity bill." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_bill_category", "proposed_instruction": "How much have I paid in water bill on venmo this year so far? Wait, I misspoke about the bill category; it should be electricity bill, not water bill.", "long_pause_offsets": [ 62 ], "source_sha256": "6c588e6ed08ec021a265a237286dfd66cb0c82eafa8d6a10f17e3cd02ecd393b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the bill category from 'water' to 'electricity', while retaining the app (Venmo), time frame ('this year so far'), and the action of querying the total amount paid." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in let me think... water bill on venmo uh... this year so far? ", "Wait, I misspoke about the bill category; it should be um... electricity bill, not water bill." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in electricity bill on PayPal this year so far? Make that on Venmo instead of PayPal.", "long_pause_offsets": [ 69 ], "source_sha256": "b0cc331abc0b55b3e19dd00a82cbfcce5c2ee289730dd949292a721eb58bd96d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, preserving the other conditions: electricity bill, time frame (this year so far), and the action of querying the total amount paid." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in electricity bill um, let's see... on PayPal this year so far? ", "Make that on ... Venmo instead of PayPal." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much have I paid in electricity bill on venmo last month so far? That should be this year, not last month.", "long_pause_offsets": [ 69 ], "source_sha256": "4dc9101db4485516e7526ba2969ae806432dc838154d9a68e203a7f182b20d50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this year', preserving the action, app, and category. The original instruction's time scope is restored." }, "audio": { "status": "pending", "spoken_utterances": [ "How much have I paid in electricity bill ... on venmo last month so far? ", "That should be hmmm... this year, not last month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "552869a_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "0333d10af265031e27f9746c6b99815891881f85f8c3abdd49da714c9d26fd07", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/cancellation/cancel_008/2d2ec9c0ed5daf22/speech.wav", "wer": 0.04, "wav_sha256": "99954a113e1bd0532d9b3d43885510e24af8e72e40ba4957bdb588884878f1cb", "seconds": 14.2, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question, then after a pause says 'Hang on. I want to cancel that task.' This is a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or new instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo? Let's see. This year, so far. Hang on. I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "646903618916c1f06bfdd433e71be9ed235b6660d9286cc0998616e51d2e9fa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/ambiguous_halt/halt_031/9a8587aeceba6b4f/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "6c500fa91f47d4850e51ca6f6795215b276aaf60383c190b6751d06d2ee00351", "seconds": 13.56, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on. I just', which is a clear request to pause the current task. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo? Let's see. This year, so far. Hang on. I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "21791b91e8079af3e169fcf207e31f644071a68bb95591005a1903889fa435bb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_2/audio_natural/filler/original/cee59151694622cd/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "a4f61b1707091ad6fb597a41d29c9787e2f5b4f9ea99a30f2caee41469dbe2ed", "seconds": 9.72, "spoken_utterances": [ "How much have I paid in uh... electricity bill on Venmo let's see... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core request. The final intent matches the original instruction exactly. The 'filler' type allows for hesitation and backchannels as long as the remaining task meaning is preserved.", "heard_transcript": "How much have I paid in uh electricity bill on Venmo, let's see, this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill—I meant internet bill on venmo this year so far?", "long_pause_offsets": [], "source_sha256": "050cc27c63cd1e923bb5be8fdb67cd481bc7afb88a90f6bb98789515fa588d2d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'electricity bill' to 'internet bill' within the same utterance, preserving the app (Venmo), time frame (this year so far), and action (query total amount paid)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/false_start/false_start_01_bill_category/38b68689928fa8e3/speech.wav", "wer": 0.0, "wav_sha256": "eda67366a06a301b6ec1c43539ca982259343227b28cbe405a8084e6f07d43a0", "seconds": 8.04, "spoken_utterances": [ "How much have I paid in electricity bill—uh... I meant um... internet bill on venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately self-corrects to 'internet bill'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final request matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in electricity bill, uh, I meant um, internet bill on Venmo this year so far?" }, "candidate_id": "wav_c2fc5ac1dd729bcd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('electricity bill'), immediately self-corrects with 'I meant' to the correct term ('internet bill'), and smoothly continues with the rest of the request. The correction is natural and integrated into the flow, with a believable hesitation and filler ('uh') before the initial error. The prosody is consistent and the pacing realistic." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How much have I paid in internet bill on PayPal—sorry, make that venmo this year so far?", "long_pause_offsets": [], "source_sha256": "1bd882d19d149f1290066b5768b090fe7cde447606bd36a6efd2a4eef9fd0fc3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says PayPal but immediately corrects to Venmo. The rest of the query, including the bill type and time frame, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/false_start/false_start_02_app/436a554d2c9d4fe6/speech.wav", "wer": 0.1, "wav_sha256": "9e8d4f2f2ac686e73a88eedac6add2454e05319328f3a210677726a3a2bf29ff", "seconds": 8.64, "spoken_utterances": [ "How much have I paid in uh... internet bill on ... PayPal—um... sorry, make that venmo this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'sorry, make that Venmo'. This is a clear false start where the final value replaces the initial incorrect one. The rest of the instruction, including the action (how much paid), category (internet bill), and time frame (this year so far), remains unchanged. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "How much have I paid in uh internet bill on PayPal um sorry, make that Venmo this year so far?" }, "candidate_id": "wav_a656c3f919cc2c67", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a false start from 'PayPal' to 'Venmo' mid-sentence, which is a credible error. The correction is immediate and the sentence continues smoothly. The 'um' filler and slight hesitation sound natural, and the overall delivery is fluent and continuous without awkward pauses or robotic intonation." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "How much have I paid in internet bill on venmo last month—no, that should be this year so far?", "long_pause_offsets": [], "source_sha256": "a42bc08fe9ea57215b5ddf631ea024b38fed0337c006ab531af6872849650665", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this year so far'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/false_start/false_start_03_time_reference/a35952dab6a24842/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "6fcae52096afc09cb9540b3364467f8f3f0721121fcd603c84e9efa63d0f930d", "seconds": 8.72, "spoken_utterances": [ "How much have I paid in internet bill on venmo um, let's see... last month—uh... no, that should be this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this year so far'. According to the 'false_start' type, the final corrected value should be applied, and the original instruction is preserved. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much have I paid in internet bill on Venmo um let's see last month uh no, that should be this year so far." }, "candidate_id": "wav_c132792514d3e986", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While the speaker starts with a plausible request and then self-corrects 'last month' to 'this year so far', the correction is less integrated. The 'uh' and 'no' feel slightly abrupt, and the overall delivery has a flatter, more hesitant prosody. The pause before 'no' is longer than in the other clips, making the correction feel more like a disjointed revision than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "How much have I paid in internet bill on venmo so far? And just to be specific, only include payments made this year.", "long_pause_offsets": [ 55 ], "source_sha256": "bce2ba051039ac243904840c95036f5f07a8c6dc89eb93af3b14937b81543926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this year') to the initial query, which is consistent with the original instruction's intent to ask about payments so far this year. The final goal combines the original scope with the added specificity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/addition/addition_03/43b98a5688153ffb/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "080f10d640a619d50aafbb37f5ca3f29ff609077e87601fb1595d9c5715e83e5", "seconds": 13.4, "spoken_utterances": [ "How much have I paid in ... internet bill on let me think... venmo so far? ", "And just to be specific, only include payments made um... this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How much have I paid in internet bill on venmo this year so far?'. The speaker begins with a similar request, then adds the constraint 'only include payments made this year' after a pause. This matches the 'addition' type, where the final goal combines the original request with the added constraint. The final audible intent is complete and matches the reference instruction with the added specificity.", "heard_transcript": "How much have I paid in internet bill on Venmo so far? And just to be specific, only include payments made um this year." }, "candidate_id": "wav_3e92d4304dfc2bdf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous speech patterns. The speaker uses a filler ('let me think') at a plausible point in the sentence, indicating genuine formulation. The addition of 'and just to be specific' followed by a clear, self-contained restriction ('only include payments made this year') flows naturally as a remembered detail. The prosody is varied and conversational, with appropriate pauses and intonation that suggest real-time thought processing." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How much have I paid in internet bill on this year so far? One more thing: make sure it's on Venmo.", "long_pause_offsets": [ 59 ], "source_sha256": "3148b4df4bfe813a71657105de720b9b0e37b9948f262bc4b0e0475a241fe05e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app constraint 'on Venmo', which was omitted in the initial utterance. The final goal combines the original intent with the added condition, resulting in the same meaning as the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/addition/addition_02/b9e54f3488ba2dca/speech.wav", "wer": 0.125, "wav_sha256": "12ac880509aa35444043845502ae4f729b2c0f2d1efdd9d1668c9ce7c93e4e6e", "seconds": 12.28, "spoken_utterances": [ "How much have I paid in uh... internet bill on um... this year so far? ", "One more thing: make sure it's on uh... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully preserved and the addition 'make sure it's on uh Venmo' clarifies the payment method, which is consistent with the original. The final intent combines the initial request with the added constraint, matching the reference instruction. No contradictions or omissions are present.", "heard_transcript": "How much have I paid in uh internet bill on um this year so far? One more thing, make sure it's on uh Venmo." }, "candidate_id": "wav_4fa7b931c038ff4d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear addition after the initial request. The speaker first asks about the internet bill on Venmo, then adds a clarification. The transition is marked by 'One more thing,' a common conversational phrase. However, the second part of the sentence ('Make sure it's on uh Venmo') is slightly awkward and less fluent than the first, with a filler ('uh') and a slightly unnatural phrasing that suggests a minor repair or hesitation rather than a fully formed thought." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "How much have I paid in on venmo this year so far? Oh, I forgot to mention, it's specifically for the internet bill.", "long_pause_offsets": [ 51 ], "source_sha256": "321926d3768bb75b612912de175eb765ee137e74bb43c306ecdfe7f75ae5567b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing category 'internet bill' to the initial question, which was otherwise complete. The app (Venmo), time frame ('this year so far'), and action ('how much have I paid') are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/addition/addition_01/a97e53a1ffa71800/speech.wav", "wer": 0.0, "wav_sha256": "8ef5fdc21062df47e9456adc9246662e0b1e6afb08ba4b0ee842e5d91db6d85c", "seconds": 10.8, "spoken_utterances": [ "How much have I paid in on ... venmo this year so far? ", "Oh, I forgot to mention, it's specifically for the um... internet bill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How much have I paid in on Venmo this year so far?' which is a partial version of the original instruction. The speaker then adds the missing constraint 'it's specifically for the internet bill' after a pause, completing the original instruction. The addition clarifies the payment category, and the final goal matches the original instruction exactly. Therefore, the addition type is satisfied with a pass.", "heard_transcript": "How much have I paid in on Venmo this year so far? Oh, I forgot to mention, it's specifically for the um, internet bill." }, "candidate_id": "wav_09b435008393826f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio presents a revision rather than an addition. The speaker first asks about Venmo payments, then corrects themselves to specify the internet bill. While the correction is plausible, the phrasing 'Oh, I forgot to mention' is a bit too formal and self-aware for a spontaneous utterance. The second part of the sentence is delivered with a slightly flat intonation and lacks the natural flow of the other two, making it sound more like a rehearsed or read statement than a genuine thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "How much have I paid in internet bill on venmo last month so far? Make that this year, not last month.", "long_pause_offsets": [ 66 ], "source_sha256": "305713e261eeab8155288b090c45f109825c07d3279a1f3f2bc9322dfabee051", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this year', preserving the action, app, and category. The phrase 'so far' is retained, aligning with the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/revision/revision_03_time_reference/0ebb4e961ea7161d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "1e996291a1a05fa1ce806f54f00cab812915a19a135991a66b21175f44d7ba52", "seconds": 12.28, "spoken_utterances": [ "How much have I paid in ... internet bill on ... venmo last month so far? ", "Make that ah... this year, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this year' with 'Make that uh this year, not last month.' This is a clear revision where the final intent is to query the amount paid this year. The correction is spoken after a short pause and directly modifies the time frame of the original request. The rest of the instruction remains unchanged, so the final goal matches the reference.", "heard_transcript": "How much have I paid in internet bill on Venmo last month so far? Make that uh this year, not last month." }, "candidate_id": "wav_2a3d475d7274c7e7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('uh') before correcting the time frame from 'last month' to 'this year'. The correction is smoothly integrated with a self-repair ('Make that uh...') and a believable change of mind. The prosody is consistent, and the pacing reflects genuine thought processing without robotic pauses or unnatural word boundaries." } ] }, "is_default": true }, { "variant_id": "revision_01_bill_category", "proposed_instruction": "How much have I paid in electricity bill on venmo this year so far? Sorry, I meant internet bill, not electricity bill.", "long_pause_offsets": [ 68 ], "source_sha256": "d224443c94d2e1d691525b03948bc112e738c7e91f6820f0f3d8e795af9e7e77", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'electricity bill' to 'internet bill', preserving the app (Venmo), time frame ('this year so far'), and the action of querying the total amount paid." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/revision/revision_01_bill_category/a2e1b328d37e7613/speech.wav", "wer": 0.125, "wav_sha256": "a5e28436bdd4aa8ce231100cdd77fcff36b647173d0f741698e2394828f442d2", "seconds": 11.28, "spoken_utterances": [ "How much have I paid in um... electricity bill on ah... venmo this year so far? ", "Sorry, I meant internet bill, not electricity bill." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'electricity bill' but immediately corrects to 'internet bill' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly, with only a minor grammatical paraphrase ('on Venmo' vs 'in internet bill on venmo') that does not change the meaning.", "heard_transcript": "How much have I paid in um electricity bill on uh Venmo this year so far? Sorry, I meant internet bill, not electricity bill." }, "candidate_id": "wav_bec31293ef81ac61", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a plausible request, then corrects the bill type from 'electricity' to 'internet'. The self-repair ('Sorry, I meant...') is natural and the intonation reflects a genuine correction. However, the initial mention of 'electricity bill' is slightly more abrupt and less smoothly integrated than in audio_03, and the overall flow feels a bit more rehearsed or less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How much have I paid in internet bill on PayPal this year so far? Wait, use Venmo, not PayPal.", "long_pause_offsets": [ 66 ], "source_sha256": "21e1d60622eaf07db7290cf914a651b17ee80fa4e8684a62721eb19c77e86105", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, preserving the bill type, time frame, and query intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/revision/revision_02_app/3caf76e82267381a/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "38102e220e15bc609b23b761e901661632e453fc1e1ef45f4fabdcbded25dd1e", "seconds": 12.72, "spoken_utterances": [ "How much have I paid in um... internet bill on hmmm... PayPal this year so far? ", "Wait, use um... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final intent matches the original instruction exactly, with only the target app corrected from PayPal to Venmo.", "heard_transcript": "How much have I paid in um internet bill on whom? PayPal this year so far? Wait, use um, Venmo, not PayPal." }, "candidate_id": "wav_a4b3628f1c094128", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but makes a correction from 'PayPal' to 'Venmo'. The correction is delivered with a 'Wait' and a self-repair, but the initial mention of 'PayPal' is followed by a longer, more awkward pause and the correction sounds slightly forced. The prosody is less fluid, and the overall delivery feels less natural compared to the other two, possibly due to the longer hesitation and less seamless integration of the correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "940dec84dd4e7fd290e4650054e9ab7885f46b7c567aab13ed882a3e62f34c99", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/cancellation/cancel_030/c636876e8e0cd3f2/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "143d61b52423bfde927aa82bf11722ab8c3c3334efd45fdbe940537fcc42cd9e", "seconds": 13.32, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?", "Hold on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Hold on, you can drop that task. I no longer need it.' This is a clear and unambiguous cancellation of the initial request. The final interjection explicitly revokes the task, satisfying the cancellation type criterion.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far? Hold on, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "97bc1e1ff3256b5315eb046d62d55978b04f42e448a60be2a2e0cc4384af8714", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/ambiguous_halt/halt_001/80f348f3440e00d4/speech.wav", "wer": 0.0625, "wav_sha256": "2d531f4a92347da0ca309fdfb48e0ce5169644f39bb3095758079484e3314b33", "seconds": 11.16, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' which is an audible request to hold or pause. There is no cancellation of the task, nor is there a new instruction to resume. The final interjection clearly indicates a temporary halt, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far? Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "552869a_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e907e897cd459e11ddf4fda5012944c17e3a2611ea4463184e8ad5d2297b26a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/552869a_3/audio_natural/filler/original/25b550a003a43f32/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "48072c714f0caedc196d4e1a089e4d9b2872d6709d56549a15da598c1359d54e", "seconds": 8.4, "spoken_utterances": [ "How much have I paid in um... internet bill on ... Venmo ... this year so far?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How much have I paid in um internet bill on Venmo this year so far?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_artist_reference", "proposed_instruction": "Like all the songs from the bands I follow—I should say artists I follow on Spotify.", "long_pause_offsets": [], "source_sha256": "0160f12b7d6b4473fd7061ac3b103a4b8d5ed2efc95556c95b3a065abf89c40d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands' but immediately corrects to 'artists' within the same utterance. The final intended target is 'artists', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/false_start/false_start_01_artist_reference/cd6bb5401db57f1e/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "f07ba3982e7e8965921615dc6dd70db3f4030c599045c8b4261045102f9ec75b", "seconds": 8.04, "spoken_utterances": [ "Like ... all the songs from the um... bands I follow—uh... I should say artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands I follow' but immediately self-corrects to 'artists I follow'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final goal matches the reference instruction exactly.", "heard_transcript": "Like all the songs from the bands I follow uh, I should say artists I follow on Spotify." }, "candidate_id": "wav_525d56714b6025f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural-sounding false start with 'bands' followed by a self-correction to 'artists' and a complete, coherent request. The intonation and pacing are very conversational, with a believable hesitation and a smooth, pragmatic flow. The correction feels spontaneous and integrated into the sentence, not mechanically inserted." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music—sorry, make that Spotify.", "long_pause_offsets": [], "source_sha256": "9d5fa4f35182cc3456bbadec42899a226b8332709696e0318e1fd6fcea8761ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The final app is Spotify, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/false_start/false_start_02_app/e83a9b5fee83489b/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "bceb7f7e8fb9e78affcd0436bd3509804408c3fe103b99ea503774c3635a2deb", "seconds": 8.24, "spoken_utterances": [ "Like all the songs from the artists I follow on Apple Music—ah... sorry, make that uh... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Like all the songs from the artists I follow on Apple Music. Sorry, make that uh Spotify." }, "candidate_id": "wav_9a735536a2b9274f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start with 'Apple Music' corrected to 'Spotify', but the correction is less fluid. The speaker says 'uh, sorry, make that uh, Spotify' with a slightly awkward, hesitant phrasing and a longer, less natural pause. The overall delivery feels a bit more robotic and less like a spontaneous correction compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like all the songs from the artists I follow on Spotify.", "long_pause_offsets": [ 43 ], "source_sha256": "7f5b781ae1e46c5f6ff0ea7ac924bc7018311b394e51c548bd08d2b11091a9c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The second utterance adds the missing phrase 'all the songs from' to complete the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/addition/addition_01/193e97ea06659b0b/speech.wav", "wer": 0.0, "wav_sha256": "c2d4f82ae947c67f108b29b0e41eb5c82c46769f337120c9beaf5ee1bb263483", "seconds": 13.72, "spoken_utterances": [ "Like from the artists I follow on um... Spotify. ", "Oh, I forgot to mention, like let's see... all the songs from the artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance 'Like from the artists I follow on Spotify' is incomplete and grammatically awkward. The speaker then adds 'all the songs from the artists I follow on Spotify' after a pause and self-correction. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final intent is a complete and coherent request to like all songs from the artists the user follows on Spotify, which matches the original instruction. The addition clarifies and completes the initial incomplete thought, resulting in a valid and executable task.", "heard_transcript": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like, let's see, all the songs from the artists I follow on Spotify." }, "candidate_id": "wav_9c1615fb34d3750a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a natural, spontaneous flow with a believable 'oh, I forgot to mention' revision. The speaker starts with a partial request, hesitates with 'um', then corrects and completes the full task. The prosody is conversational, with appropriate pauses and intonation shifts that suggest genuine thought processing. The correction is integrated smoothly into the utterance, making it sound like a real-time formulation of a request." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Like all the songs from the on Spotify. To narrow that down, only like songs from the artists I follow.", "long_pause_offsets": [ 40 ], "source_sha256": "df4b5bb09cb64e4b962daa1a94643890647acdc05f76200c651872896b41319a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'from the artists I follow' to the initial incomplete request, completing the intended scope. The app 'Spotify' is preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/addition/addition_02/83a619bca459b17c/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "0390e40265a994b0acf2463b4d9ab7afa52150cc4d75cd9c92e2dd53a9633fff", "seconds": 11.6, "spoken_utterances": [ "Like ... all the songs from the on hmmm... Spotify. ", "To narrow that down, only like songs from the artists I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all the songs from the on, um, Spotify', which is a filler and false start. They then correct and add the constraint 'only like songs from the artists I follow'. This addition narrows the original request, and the final intent matches the reference instruction. The addition type requires combining the earlier instruction with the added information, which results in the correct final goal.", "heard_transcript": "Like all the songs from the on, um, Spotify. To narrow that down, only like songs from the artists I follow." }, "candidate_id": "wav_3393e6f3cb3cc8d4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is less realistic due to its fragmented and hesitant delivery. The speaker begins with 'like all the songs from the on', which is grammatically incomplete and awkward. The long pause and 'um' sound forced rather than natural, and the subsequent 'Spotify' is spoken as if it's a separate thought. The final phrase 'to narrow that down only like songs from the artists I follow' is delivered in a flat, robotic tone with no natural intonation or flow, suggesting a rehearsed or synthesized utterance rather than spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like on Spotify. Wait, the songs should be from artists I follow, not artists I like.", "long_pause_offsets": [ 55 ], "source_sha256": "a57080480da39f72c6ef2d43476c243b63cc252f884c7dd8e0044a01dbcec86a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source of the artists from 'I like' to 'I follow', preserving the action (like), the object (all songs), and the platform (Spotify)." }, "audio": { "status": "pending", "spoken_utterances": [ "Like ... all the songs from the artists I like on um... Spotify. ", "Wait, the songs should be from artists I follow, not artists I like." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music. Let me correct the app: like the songs on Spotify, not Apple Music.", "long_pause_offsets": [ 61 ], "source_sha256": "4b5a6abd4d0d46bde8168c3a4b82d7a89a388b26f169bf2db4856ae9f2970644", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining the action and the source of artists (those the user follows)." }, "audio": { "status": "pending", "spoken_utterances": [ "Like uh, let's see... all the songs from the artists I follow on Apple Music. ", "Let me correct the app: like the songs on um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "57c3486_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "2a6a738f162b6548dc0a4767a7719186b5215edfb1ef61b8ca477c2d7fd635a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/cancellation/cancel_010/507846610848f644/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "750ea30746b733a3c72dee951cb895f03e4096eae360f382c9e57c710ef66400", "seconds": 11.76, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify.", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the user stating the original request, then immediately follows with a clear cancellation command: 'Oh wait, cancel the request I just made.' This constitutes a final, unambiguous cancellation of the task. The cancellation is not merely a pause or hesitation, but a direct revocation of the prior instruction. Therefore, the final control intent matches the expected 'cancel' behavior for this type.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "0f1307b9f7ad7ec917f6661f459c544fa116b719aa7169c326ce00c395ef943f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/ambiguous_halt/halt_010/f908cfe7bcc8e6cc/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "19722c61e8e2525b2d88c002b072f66cd379906bfc3d73c29e795028cee35b0d", "seconds": 10.56, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly says 'Hold on a second,' which is a standard verbal cue to pause or wait. There is no cancellation or instruction to resume, and the final interjection is a request to hold. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "da8240354f9ee4f9db7524dd6b80136d9cdcac1ef10076c4b1ccac82eed2bb71", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_1/audio_natural/filler/original/194de8afae53fb83/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "d780f708b5b62a7b09679c08ca7d4e59e9e872a26393046e1bb791762871bc22", "seconds": 7.2, "spoken_utterances": [ "Like ... all the songs from the artists I follow on um, let's see... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let's see') and a brief pause but does not change the core request. The final audible intent is to like all songs from the artists followed on Spotify, matching the original instruction. The hesitation is a filler and does not alter the task goal.", "heard_transcript": "Like all the songs from the artists I follow on um, let's see, Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music—sorry, I meant Spotify.", "long_pause_offsets": [], "source_sha256": "56db02317ddfe8f0ed4fdd5a717ca6b95252724b0c3789835f135ef63852d4e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, scope, and target app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/false_start/false_start_02_app/b447fac66f02b24f/speech.wav", "wer": 0.05, "wav_sha256": "b8d03b09ad5a4b8c737a229e133aa35437849d9da6e17edc066cfc872eb0396e", "seconds": 8.52, "spoken_utterances": [ "Like all the songs from the artists I follow on uh, let's see... Apple Music—uh... sorry, I meant ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the reference instruction exactly.", "heard_transcript": "Like all the songs from the artists I follow on uh let's see, Apple Music uh sorry, I meant Spotify." }, "candidate_id": "wav_34dbd589933492dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start and self-correction. The speaker begins with 'Apple Music' and then immediately corrects to 'Spotify' with a believable 'sorry, I meant' phrase. The prosody includes a slight hesitation and a natural pitch drop on 'sorry,' followed by a smooth correction. The pacing and intonation are consistent with spontaneous speech, and the correction feels integrated into the flow rather than tacked on." } ] }, "is_default": true }, { "variant_id": "false_start_01_artist_reference", "proposed_instruction": "Like all the songs from the bands I follow—or rather, artists I follow on Spotify.", "long_pause_offsets": [], "source_sha256": "bb05495e03a162f1a63f2baeed4df1081955e6f7e4e59c535e079005624d3f9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands' but immediately corrects to 'artists' within the same utterance. The final intended target is 'artists', which matches the original instruction. The app (Spotify) and action (like all songs) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/false_start/false_start_01_artist_reference/523ee657c82f0a47/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "058a507179fca1312d51218863cf021113a83269f6970129a70a9d5ae1eab034", "seconds": 8.72, "spoken_utterances": [ "Like uh... all the songs from the bands I follow—um... or rather, artists I follow on hmmm... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bands I follow' but immediately self-corrects to 'artists I follow'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only a minor paraphrase ('artists' instead of 'bands') that does not change the meaning.", "heard_transcript": "Like uh all the songs from the bands I follow um or rather artists I follow on Spotify." }, "candidate_id": "wav_85c5fd234ccc95a4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a false start, but the correction is less natural. The speaker says 'bands I follow' and then corrects to 'artists I follow,' which is semantically plausible but feels more like a word substitution than a genuine slip. The correction is abrupt, with no filler or hesitation before the change, and the overall delivery is flatter and less prosodically varied than audio_02. The correction lacks the spontaneous, self-correcting quality seen in the higher-ranked clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Like all the songs from the on Spotify. I didn't mention this, but only like songs from the artists I follow.", "long_pause_offsets": [ 40 ], "source_sha256": "ceed268105e5489194f3dabe7415bd81d5d50df7f93bc8da9fa6f543a63a676d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'from the artists I follow' to the initial incomplete instruction, which already specified 'on Spotify'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/addition/addition_02/b33ff9d96f58820e/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "1052029569f710775606787f3f4e7e36e5717e91ca623a47e21f15673af40e11", "seconds": 13.4, "spoken_utterances": [ "Like ... all the songs from the on uh, let me think... Spotify. ", "I didn't mention this, but only like songs from the hmmm... artists I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all the songs from the um uh let me think. Spotify.' and then adds 'I didn't mention this, but only like songs from the hmm artists I follow.' The addition clarifies that the action should be limited to songs from the artists the user follows. This matches the 'addition' type criterion, where the final goal combines the original instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Like uh all the songs from the um uh let me think. Spotify. I didn't mention this, but only like songs from the hmm artists I follow." }, "candidate_id": "wav_57d5878881575982", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-correction ('let me think'), followed by a plausible addition of a restriction ('only like songs from the artists I follow'). The prosody is conversational, with a believable pause and filler ('hmm') before the final clarification. The flow from initial request to addition feels spontaneous and coherent." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like all the songs.", "long_pause_offsets": [ 43 ], "source_sha256": "0dbc16c62e80ea65ea73c4ed3348afebb26b0a46949dd22f7b479cfa49413ae5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing phrase 'all the songs' to the initial partial request, completing the intended goal without contradicting or removing any prior elements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/addition/addition_01/4a8e27228e82debf/speech.wav", "wer": 0.0, "wav_sha256": "247f4bf195e478fe99def02870cf757ba74dc63b1f1f1c90ac7b2e79386dad40", "seconds": 10.24, "spoken_utterances": [ "Like from the artists I follow on ... Spotify. ", "Oh, I forgot to mention, like uh... all the songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Like from the artists I follow on Spotify.' This is followed by a self-correction where the speaker says, 'Oh, I forgot to mention, like uh all the songs.' This is a clear addition that completes the original request by specifying 'all the songs.' The final intent combines the original instruction with the added detail, matching the reference instruction. The addition is not a revision or cancellation, but a clarification that expands the scope of the original task. Therefore, the final intent is to execute the combined task.", "heard_transcript": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like uh all the songs." }, "candidate_id": "wav_49b14c9993812f55", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear, fluent request, then adds a correction ('Oh, I forgot to mention...') with a natural-sounding hesitation and a follow-up phrase. However, the addition is less integrated than in audio_01—the initial request is more complete, and the correction feels slightly more abrupt. The prosody is still natural, but the overall flow is less organic compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like on Spotify. I misspoke about the artist reference: it should be 'artists I follow', not 'artists I like'.", "long_pause_offsets": [ 55 ], "source_sha256": "0e808e173dcc56a3d589869a448485a6a890aaec0f914aad32f9ffdea68f12ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'artists I like' to 'artists I follow', which matches the original instruction. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/revision/revision_01_artist_reference/9facf2ed6d714bf3/speech.wav", "wer": 0.09375, "wav_sha256": "aa4a8c52988e777c33d50b6c50a1a4694e2b0fb6c7017a701bd2bd73437ac60c", "seconds": 15.96, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I like on ... Spotify. ", "I misspoke about the artist reference: it should be 'uh... artists I follow', not 'artists I like'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artists I like' but immediately corrects themselves to 'artists I follow' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to like all songs from the artists the user follows, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "Like um, let me think. All the songs from the artists I like on Spotify. I misspoke about the artist reference. It should be uh, artists I follow, not artists I like." }, "candidate_id": "wav_8157273a0de827da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic and natural sequence of spontaneous speech: it begins with a filler ('like um') and a hesitation ('let me think'), followed by a clear initial request. The speaker then self-corrects a specific term ('artists I like' to 'artists I follow') with a natural false start and repair, including a brief pause and a soft 'uh' before the correction. The prosody is conversational, with appropriate stress and pacing, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The flow is continuous and psychologically plausible." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music. Change the app to Spotify, not Apple Music.", "long_pause_offsets": [ 61 ], "source_sha256": "57bd6323ec84d61fa808bfd71e33c6f17e9e1948af650f10bf474109d7f251c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the action, scope (all songs from followed artists), and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/revision/revision_02_app/4c5d093d9dfc9640/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "971940a62d38f570e6c0d9bb6fd8bedb5a016c1939ef2c85ba499c56730fcb10", "seconds": 12.72, "spoken_utterances": [ "Like ... all the songs from the artists I follow on ... Apple Music. ", "Change the app to uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' with 'Change the app to uh Spotify, not Apple Music.' This is a clear revision where the final audible intent matches the original instruction. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Like all the songs from the artists I follow on Apple Music. Change the app to uh Spotify, not Apple Music." }, "candidate_id": "wav_6b474d5130f54555", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a revision, but the delivery is less natural. It starts with 'like' and proceeds with a clear initial request, but the correction ('change the app to uh Spotify not Apple Music') is delivered in a more abrupt, almost robotic manner. The 'uh' before 'Spotify' is less integrated and sounds more like a placeholder than a genuine hesitation. The overall prosody is flatter and the pacing is more deliberate, lacking the organic rhythm and subtle intonation shifts of a truly spontaneous speaker. While the content is semantically clear, the delivery feels less authentic and more like a rehearsed or edited version." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "06e3022a0f2b18d001dca911668a1ce285851f95cded48bb4c5ddbe066cbe4d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/cancellation/cancel_023/81181076d327a014/speech.wav", "wer": 0.0, "wav_sha256": "76d9ee81f1afe7b2a82809c1d3baa195f9a215ce99491168095d0d523fc74e07", "seconds": 12.76, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify.", "Hold on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to retrieve songs from followed artists on Spotify, then clearly revokes it with 'Hold on, never mind about that. Don't carry out the request.' This constitutes a direct cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify. Hold on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "22d5f76dc4d87f9428ff9c661722cdc197666dd649c8f4b32e743def7c009b1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/ambiguous_halt/halt_011/23cde13909b5ac92/speech.wav", "wer": 0.0, "wav_sha256": "05b4aaf268cc0b8756d251f21d31fd13f3bb5890d7cddfc2687e2c09e3a8d365", "seconds": 10.48, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request, then says 'Hang on a second' as a final interjection. This is a clear, audible request to pause or wait, with no indication of cancellation or resumption. The initial request is not executed, but the final intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "fc9b5e8a10ab4fba49a72d7ab42d97f180abd3c429e6815423b6f3cce9452432", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_2/audio_natural/filler/original/e8d2304c5618070c/speech.wav", "wer": 0.0, "wav_sha256": "5063d5e1503322cb469a3fda1bb416a40a1ec0284152a86201803d78ff82dc8b", "seconds": 7.36, "spoken_utterances": [ "Like um, let me think... all the songs from the artists I follow on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('um', 'let me think') but proceeds to state the original instruction without any correction or change in intent. The final utterance matches the original instruction, and the hesitation is a typical filler that does not alter the task goal. Therefore, the final intent is to execute the original request.", "heard_transcript": "Like um, let me think. All the songs from the artists I follow on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like—I should say artists I follow on Spotify.", "long_pause_offsets": [], "source_sha256": "2599adeed34efc126fa83283479e114489bd7570bc8e7ee7c75f4c668eab7d09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'artists I like' to 'artists I follow' within the same utterance, preserving the app (Spotify) and the action (like all songs). The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/false_start/false_start_01_artist_reference/34e76204e3cb7802/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "0d17efe0f42c0db84933c33125e74dff45b279f40f298f4eb7d0be71f3814ccf", "seconds": 8.88, "spoken_utterances": [ "Like hmmm... all the songs from the ... artists I like—uh... I should say artists I follow on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artists I like' but immediately self-corrects to 'artists I follow on Spotify'. This is a false start, where the final corrected version matches the original instruction exactly. The correction is immediate and unambiguous, so the final intent is to execute the original task.", "heard_transcript": "Like, hmm, all the songs from the artists I like, uh, I should say artists I follow on Spotify." }, "candidate_id": "wav_e4db9da8c5e94833", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('hmm'), then formulates a request that is slightly off ('artists I like'), followed by a believable self-correction ('I should say artists I follow on Spotify'). The correction is smoothly integrated into the same utterance with a natural intonation shift and a plausible rephrasing. The pacing and prosody are consistent with spontaneous speech, and the false start feels authentic rather than forced." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music—sorry, that should be Spotify.", "long_pause_offsets": [], "source_sha256": "5ef9cd683976795388c05553a03a7e5fd9862f1823a47963698abf4298652224", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, scope, and target app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/false_start/false_start_02_app/96c4610351fa1cd7/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "dbd3a6333a73d6faac6939d50d73b7f6266562334fd6a98958639f6db0ca82b2", "seconds": 11.0, "spoken_utterances": [ "Like all the songs from the artists I follow on Apple Music—hmmm... sorry, that should be ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify' after a brief pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Like all the songs from the artists I follow on Apple Music. Hmm. Sorry, that should be Spotify." }, "candidate_id": "wav_8ea36dacba3d8ac9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start, but the correction ('Spotify') is delivered as a separate, isolated phrase after a long pause. The initial request is cut off mid-sentence, and the correction lacks the natural flow of being embedded within the same thought. The prosody feels more like a read-aloud with a correction tacked on, rather than a single, continuous thought process. The pause between the initial error and the correction is too long and abrupt, reducing the sense of spontaneity." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Like all the songs from the on Spotify. I left out a detail: only like songs from the artists I follow.", "long_pause_offsets": [ 40 ], "source_sha256": "4dedf076dfe6b349bc666d1c93476af677125ac81443f951adb325b107788249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'only like songs from the artists I follow', which clarifies the scope of the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/addition/addition_02/a4212fb9ce49b765/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "0e3882cb641d864f7d4aaad338a277f5b589d09e5b4492fad6f208c549b1863b", "seconds": 12.92, "spoken_utterances": [ "Like um... all the songs from the on Spotify. ", "I left out a detail: only like songs from the hmmm... artists I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all the songs from the on Spotify' which is a malformed version of the original instruction. They then correct this by adding the constraint 'only like songs from the hmm artists I follow'. This addition narrows the scope of the original request, which is the expected behavior for the 'addition' type. The final audible intent is to like only the songs from the artists the user follows, which matches the reference instruction after the addition.", "heard_transcript": "Like um all the songs from the on Spotify. I left out a detail. Only like songs from the hmm artists I follow." }, "candidate_id": "wav_7112fc3e9168200f", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Like from the artists I follow on Spotify. Oh, I forgot to mention, like all the songs.", "long_pause_offsets": [ 43 ], "source_sha256": "0dbc16c62e80ea65ea73c4ed3348afebb26b0a46949dd22f7b479cfa49413ae5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing phrase 'all the songs' to the initial partial request, completing the intended goal without contradicting or removing any prior elements." }, "audio": { "status": "pending", "spoken_utterances": [ "Like from the artists I follow on um... Spotify. ", "Oh, I forgot to mention, like uh... all the songs." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Like all the songs from the artists I follow on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 61 ], "source_sha256": "89dce863ff2ee65a3580dc2ccc186c56fd838efccd97942901bc3ae44b410d84", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the action and the scope of artists (those the user follows)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/revision/revision_02_app/0aeaaf6b22cbe4ba/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "983d353bf167f2c6897891b8b22fb6d049918a28287fa1d7891d139738e3c9de", "seconds": 13.72, "spoken_utterances": [ "Like um... all the songs from the artists I follow on um, let's see... Apple Music. ", "Make that ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify, not Apple Music' after a pause. This is a clear revision where the final intent is to use Spotify. The correction is unambiguous and the original instruction is fully preserved except for the app name, which is corrected. The final control intent is to execute the revised task.", "heard_transcript": "Like um all the songs from the artists I follow on um let's see Apple Music. Make that Spotify, not Apple Music." }, "candidate_id": "wav_970e7629cbe0e024", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('Like um...') and a plausible self-correction from 'Apple Music' to 'Spotify'. The correction is immediate and smoothly integrated into the sentence, with a clear shift in intonation and a brief pause before the new value. The flow feels spontaneous and unscripted, with no signs of artificiality or forced pacing." } ] }, "is_default": true }, { "variant_id": "revision_01_artist_reference", "proposed_instruction": "Like all the songs from the artists I like on Spotify. Wait, it should be 'artists I follow', not 'artists I like'.", "long_pause_offsets": [ 55 ], "source_sha256": "d95489bc72a11431503bcef84b3c0420755b006144783d2636e0c68623102b93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the phrase from 'artists I like' to 'artists I follow', which matches the original instruction. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/revision/revision_01_artist_reference/d80539b8baf8e5d5/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "5097746eaa43ef06f794c88fa1f907ceb9cd03663d1810551a8cee90860da228", "seconds": 14.92, "spoken_utterances": [ "Like um... all the songs from the artists I like on um, let me think... Spotify. ", "Wait, it should be 'um... artists I follow', not 'artists I like'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artists I like' but immediately corrects themselves to 'artists I follow' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original instruction exactly, so the task should be executed as intended.", "heard_transcript": "Like um all the songs from the artists I like on um let me think, Spotify. Wait, it should be um artists I follow, not artists I like." }, "candidate_id": "wav_f905dfafde9407b6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also shows hesitation ('let me think') and a self-correction ('artists I like' to 'artists I follow'), the correction is less fluid. The phrase 'wait, it should be...' introduces a slightly more abrupt and self-aware tone, and the correction is delivered with a noticeable shift in pitch and pacing that sounds slightly rehearsed. The overall delivery is still realistic but less seamless than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_035", "hide_variant_original": true, "source_sha256": "919d67324f548d19976bdf4c3f049f1cc9f7e4104888a10cbfc5ff26f01816d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/cancellation/cancel_035/a3680eaf8c015c67/speech.wav", "wer": 0.0, "wav_sha256": "277967b56d6dd0eff01d9580caddbf7f49736a6b4f3b1acd753d9f1497c15100", "seconds": 11.88, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify.", "Wait, stop working on that and cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction, followed by a clear self-correction and cancellation command: 'Wait, stop working on that and cancel the task.' This constitutes a final, unambiguous cancellation of the initial request. The cancellation is not merely a pause or hesitation, but a direct instruction to abort the task. Therefore, the final control intent matches the expected 'cancel' behavior.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify. Wait, stop working on that and cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "8713c535da949ef058097bf64fd3ffcdb89d6876f47e3ff9dc0c8df115730f3c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/ambiguous_halt/halt_022/159d2805d7c65a45/speech.wav", "wer": 0.0, "wav_sha256": "8f5e46a38ae6ea6664a20cc41f9fdf3e2480e50576c191b5b95efca3cff26ef1", "seconds": 10.8, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on. Let me think.', which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify. Hang on. Let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "57c3486_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4ea28eee411cf06f9440b7ea2d2929b3b6d6d72100c492d9083634d9533141d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/57c3486_3/audio_natural/filler/original/1b19b81c0ab59330/speech.wav", "wer": 0.0, "wav_sha256": "bcb6affdb93163ee14adbf161de3e9378a8e90bd313b4cfbbe0c828d6e010d2e", "seconds": 6.76, "spoken_utterances": [ "Like um... all the songs from the artists I follow on let me think... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'um' and 'let me think' but does not change the core request. The final audible intent is to like all songs from the artists the user follows on Spotify, which matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Like um all the songs from the artists I follow on let me think Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_folder_reference", "proposed_instruction": "Delete all PDF files from the Documents folder—sorry, I meant downloads folder in my home directory.", "long_pause_offsets": [], "source_sha256": "7c90d4ddacfc11331b3d247a6569ad99f503d2d4c4ddae2539fb9b63e4db7820", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Documents folder' to 'downloads folder in my home directory', preserving the file type (PDF), action (delete), and scope (all files)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/false_start/false_start_02_folder_reference/93b51eca9526e270/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "7e6b11927e7af89a2017aad282e705ae8a20d60a821cf754a505e91d5f484570", "seconds": 9.4, "spoken_utterances": [ "Delete all let me think... PDF files from the Documents folder—um... sorry, I meant downloads folder in my home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'documents folder' but immediately self-corrects to 'downloads folder' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Delete all, let me think, PDF files from the documents folder. Um, sorry, I meant downloads folder in my home directory." }, "candidate_id": "wav_2def786313aa2cc5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly realistic false start and self-correction. The speaker begins with 'delete all' and then hesitates with 'let me think', which is a natural filler. They then state an incorrect folder ('documents'), immediately self-correct with 'sorry, I meant downloads folder', and complete the request. The intonation, pacing, and natural flow of the correction, including the back-reference to the earlier mistake, make this sound like a spontaneous, real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_01_file_filter", "proposed_instruction": "Delete all DOC—make that PDF files from the downloads folder in my home directory.", "long_pause_offsets": [], "source_sha256": "8c6806621cfba77d252159b0fe9a15d807f7a196affab0711a6811866b12c16a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'DOC' but immediately corrects to 'PDF' within the same utterance. The final intended file type is PDF, and all other conditions (action, location) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/false_start/false_start_01_file_filter/82db0e345a84459f/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "2ae193fd563966b4cfc9e27c41751af42617f32553c66a33e51f21f16816b553", "seconds": 8.64, "spoken_utterances": [ "Delete all DOC—um... make that PDF files from the uh... downloads folder in my ... home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'DOC' but immediately self-corrects to 'PDF' with 'make that PDF files'. This is a false start where the final value is the corrected one. The rest of the instruction matches the original, so the final intent is to delete all PDF files from the downloads folder in the home directory, which aligns with the original instruction.", "heard_transcript": "Delete all DOC um, make that PDF files from the uh downloads folder in my home directory." }, "candidate_id": "wav_d089c02a89deefcd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is less realistic because it lacks a natural self-correction. The speaker says 'delete all DOC um' and then immediately corrects to 'PDF files', but the correction is abrupt and lacks the typical prosodic cues of a spontaneous error. The 'um' filler is present, but the transition from 'DOC' to 'PDF' feels more like a read-aloud correction than a genuine moment of realization. The overall delivery is flatter and less conversational than audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all files from the downloads folder in my home directory. Oh, I forgot to mention, only delete PDF files.", "long_pause_offsets": [ 65 ], "source_sha256": "5d710589306d3a55e2507438a102d8d3ed241fd3cb62be82c98cf38676680c72", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad request to only PDF files, preserving the folder and location." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/addition/addition_01/be8af8b6de7d7290/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "9384b76801adccd123a3aea8a33fee70f34b2c2e96110ad8ac987570d4654a0a", "seconds": 11.96, "spoken_utterances": [ "Delete all files from the uh... downloads folder in my um... home directory. ", "Oh, I forgot to mention, only delete ... PDF files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all files from the downloads folder. The speaker then added a constraint specifying that only PDF files should be deleted. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Delete all files from the uh downloads folder in my um home directory. Oh, I forgot to mention, only delete PDF files." }, "candidate_id": "wav_125ae414af9e2dc9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('delete all files from the uh downloads folder...') and then naturally adds a crucial restriction ('only delete PDF files') after a brief pause. The addition is grammatically and semantically coherent, with a plausible shift in focus. The prosody is natural, with appropriate fillers ('uh', 'um') and a believable conversational rhythm. The correction is integrated smoothly into the same utterance, making it sound like a spontaneous clarification rather than a rehearsed or disjointed edit." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Delete all PDF files from the. To narrow that down, delete them only from the downloads folder in my home directory.", "long_pause_offsets": [ 31 ], "source_sha256": "660b4067561f2bf36c58b9e02f848ea3b5bc26b53b856e722fb2f8c4e5308a89", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a location restriction that narrows the initial broad request, preserving the file type and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/addition/addition_02/f0e31d1e3f182ba2/speech.wav", "wer": 0.08, "wav_sha256": "ff00edbf07b64e231a299c77989c7d63bf26efe6580a39179727b14ffc865b62", "seconds": 12.0, "spoken_utterances": [ "Delete all um... PDF files from the. ", "To narrow that down, delete them only from the um, let's see... downloads folder in my home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Delete all PDF files from the' is followed by a clarification 'to narrow that down, delete them only from the... downloads folder in my home directory.' This is a clear addition that narrows the scope of the original task. The final intent matches the original instruction exactly, with the addition specifying the location. The hesitation and filler words do not change the final goal.", "heard_transcript": "Delete all um PDF files from the to narrow that down, delete them only from the um let's see downloads folder in my home directory." }, "candidate_id": "wav_5d228de81f8f4836", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request ('delete all um PDF files...') but then attempts to narrow the scope by adding a location. However, the phrasing 'to narrow that down delete them only from the um let's see downloads folder...' is awkward and less natural. The phrase 'to narrow that down' is an unusual way to introduce a restriction, and the 'let's see' filler feels like a moment of hesitation that doesn't lead to a smooth continuation. The overall flow is less coherent than audio_01, and the addition sounds more like a forced or rehearsed edit than a spontaneous thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_folder_reference", "proposed_instruction": "Delete all PDF files from the Documents folder. Let me correct the folder: delete from the downloads folder in my home directory, not the Documents folder.", "long_pause_offsets": [ 48 ], "source_sha256": "0e1101ff7db18c1917d02d4f17c7ce33625ccc4c9ef59478e2787ee78dd65ddd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder from 'Documents' to 'downloads folder in my home directory' and negates the initial incorrect folder, preserving the file type and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/revision/revision_02_folder_reference/4cc4cd55a1a1c358/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "352c8ebcd9a9c9a4353911975327d1a056ab56a89ecd5d268de57bebb6bb4932", "seconds": 14.0, "spoken_utterances": [ "Delete all uh... PDF files from the uh, let's see... Documents folder. ", "Let me correct the folder: delete from the uh... downloads folder in my home directory, not the Documents folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the folder as 'documents folder' but immediately self-corrects with 'Let me correct that' and specifies the correct folder as 'downloads folder in my home directory'. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Delete all uh PDF files from the uh documents folder. Let me correct that. Delete from the uh downloads folder in my home directory, not the documents folder." }, "candidate_id": "wav_3e0171b37f8163da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('Delete all uh PDF files from the uh... documents folder'), then self-corrects with a natural-sounding revision ('Let me correct the folder: delete from the uh... downloads folder in my home directory not the documents folder'). The correction is integrated smoothly into the same utterance, with plausible filler words ('uh') and a realistic prosodic shift indicating a change of mind. The flow is continuous and the repair feels spontaneous." } ] }, "is_default": true }, { "variant_id": "revision_01_file_filter", "proposed_instruction": "Delete all DOC files from the downloads folder in my home directory. Change the file filter to PDF, not DOC.", "long_pause_offsets": [ 69 ], "source_sha256": "1636d5ee24bf4d08c8471ca0c7cc5068c67ea3efe230b0c26505a2b0da667239", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the file type from DOC to PDF, while retaining the folder and action. The original instruction's file type is replaced by the corrected value." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/revision/revision_01_file_filter/76fe28dcbe898020/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "e0c70fe88237a148df5009532a9f1f72acabfe5030d941980eb1528259299dde", "seconds": 12.96, "spoken_utterances": [ "Delete all DOC files from the ... downloads folder in my uh... home directory. ", "Change the file filter to ... PDF, not DOC." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'DOC files' but immediately corrects to 'PDF, not DOC' after a pause, which is a revision. The final instruction is to delete PDF files, matching the original reference. The correction is clear and applied to the earlier instruction.", "heard_transcript": "Delete all DOC files from the downloads folder in my home directory. Change the file filter to PDF, not DOC." }, "candidate_id": "wav_1eb9551058af9618", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a complete initial request ('Delete all DOC files from the downloads folder in my uh home directory'), followed by a pause and a revision ('Change the file filter to PDF not DOC'). While the revision is grammatically correct and semantically coherent, the transition between the two parts feels slightly abrupt. The second sentence is delivered with a more declarative, less conversational tone, making the revision seem more like a separate command than a natural self-correction within a single thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "cddb521ec352a8410f61eb3603c5005def931fca4ae04211ac6f66282b0c145c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/cancellation/cancel_010/a094884b938bfb30/speech.wav", "wer": 0.1, "wav_sha256": "5b018516cf3a9e53c01aa4ed306da28953728f538fb5585f6457dd652c463f07", "seconds": 9.68, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory.", "Wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a deletion request and then immediately follows with a clear cancellation command ('cancel the request I just made'). This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final audible intent is to cancel, not to execute the deletion.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory. Wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "e24dfe50657f205074ecc7f1146e2f4809d4ac3c41d6fec85ed709cb4e687c6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/ambiguous_halt/halt_014/97e876a8bd1fb429/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5646db6f2e9aebf5d4bff35a5c093c9491a8f8504cab739fa5289c816f176693", "seconds": 9.0, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the initial instruction and then clearly requests a pause with 'Can you hold on a second?'. This constitutes a final, audible request to pause, which matches the 'ambiguous_halt' type's expected behavior. No cancellation or execution is requested.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3f730ec818fe6dd56610ed5af779d032d577e4b30acf74a4d4219b5dd40e4924", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_1/audio_natural/filler/original/05727aafef9bae49/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "7e7c9719042aa0d3ff005da61c8414d0b6b13ce67e6e0946f1132a92ce009753", "seconds": 5.28, "spoken_utterances": [ "Delete all PDF files from the uh... downloads folder in my ... home directory." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler word 'uh' but the final intent is clear and matches the original instruction. The user does not change the goal or request a pause/cancellation. The task is executed as intended.", "heard_transcript": "Delete all PDF files from the uh downloads folder in my home directory." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_file_filter", "proposed_instruction": "Delete all .png—no, that should be .jpg files from my file system ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "7b62303e728a28259ea966c5a44790082cd6d91358a479e581206d480e783309", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says .png but immediately corrects to .jpg; the action, scope, and destination folder remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/false_start/false_start_01_file_filter/6abae793e3ccc311/speech.wav", "wer": 0.375, "wav_sha256": "0961c42719e84bfbdba08ca534e03b05ca2e8cadbe6875b0bca94f4d63f356b3", "seconds": 9.08, "spoken_utterances": [ "Delete all .png—um... no, that should be ... .jpg files from my file system ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.png files' but immediately self-corrects to '.jpeg files'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The rest of the instruction, including the target folder, remains unchanged. The final intent matches the original instruction with the corrected file extension.", "heard_transcript": "Delete all .png files. No, that should be .jpeg files from my file system ~/downloads folder." }, "candidate_id": "wav_ecf1abd9914537da", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, confident command ('Delete all .png files'), then immediately self-corrects with a natural-sounding hesitation ('No, that should be...') before smoothly continuing with the corrected instruction. The prosody includes a believable shift in intonation and pacing during the correction, and the final phrase is delivered at a steady, conversational pace. The repair feels spontaneous and integrated into the flow of speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_folder_reference", "proposed_instruction": "Delete all .jpg files from my file system ~/Desktop—scratch that, make it ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "2ec3d7a9c5ee67a42c18f544d331fd0f6e3e112b4c3c4c8b06b01da9c96b243e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions ~/Desktop but immediately corrects it to ~/downloads. The action, file type, and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/false_start/false_start_02_folder_reference/1588eb60554abb24/speech.wav", "wer": 0.1875, "wav_sha256": "cd77efc19abee8ee1997db62b50bde5f1c8bd5a3ad91871d4db0d600f6cdd8d6", "seconds": 9.0, "spoken_utterances": [ "Delete all .jpg files from ... my file system ~/Desktop—um... scratch that, make it ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Delete all dot jpeg files from my file system' but immediately corrects with 'Scratch that, make it ~/downloads folder.' This is a false start where the final intent is the corrected version. The original instruction is replaced by the new one, which matches the reference instruction in action and target, with the only difference being the file extension format (.jpeg vs .jpg), which is phonetically equivalent and not a semantic error.", "heard_transcript": "Delete all dot jpeg files from my file system. Scratch that, make it ~/downloads folder." }, "candidate_id": "wav_4ef7cd20427ff5f0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request ('Delete all .jpeg files...'), but the self-correction ('Scratch that, make it...') is less integrated. The initial phrase trails off with a long, slightly unnatural 'um' sound, and the correction is delivered with a more abrupt shift in tone and pace. While the false start is present, the delivery feels slightly more hesitant and less fluid than audio_01, making it less realistic in terms of natural speech flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Delete all .jpg files from my file system folder. I left out a detail: do this only in the ~/downloads folder.", "long_pause_offsets": [ 50 ], "source_sha256": "8cec0f7d6753fbc063a8849360a821ed24502132fd9fbf0840fdab57a2f855ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint, narrowing the scope from the general file system folder to the specific ~/downloads folder. The action and file type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/addition/addition_02/50e3ae950cd4562b/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "98a201f7cf82519991dd2b6c3c730d9a42fc0b5e887b91ae04dd0206a44857f1", "seconds": 12.08, "spoken_utterances": [ "Delete all .jpg files from ... my file system folder. ", "I left out a detail: do this only in the ... ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Delete all .jpg files from my file system ~/downloads folder.' The speaker then adds a clarifying constraint: 'Do this only in the ~/downloads folder.' This addition refines the scope of the original task, narrowing it to a specific folder. The final intent is to delete .jpg files only from the ~/downloads folder, which matches the original instruction's intent but with a more precise scope. Since the addition is a refinement and not a contradiction, the task is executed as specified.", "heard_transcript": "Delete all .jpg files from my file system ~/downloads folder. I left out a detail. Do this only in the ~/downloads folder." }, "candidate_id": "wav_9776e974237c7c65", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody and a believable pause before adding a restriction. The addition ('I left out a detail') is phrased conversationally, and the correction ('only in the Downloads folder') is smoothly integrated with a slight rise in pitch and emphasis, matching how a person would recall a missing detail. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Delete all files from my file system ~/downloads folder. Oh, I forgot to mention, only delete .jpg files.", "long_pause_offsets": [ 57 ], "source_sha256": "df4bfca7cc7db39a37ec122d0d67f82906e165fe4712edb564b382750f29a12e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad deletion to only .jpg files, preserving the location and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/addition/addition_01/85f0c356b9de8f91/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "08a47e0c4d869846bfd2e7f0b411c7559f29a326b2f37e5bd885b93490389324", "seconds": 13.08, "spoken_utterances": [ "Delete all files from um... my file system hmmm... ~/downloads folder. ", "Oh, I forgot to mention, only delete uh, let's see... .jpg files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all files from the downloads folder. After a pause and self-correction, the speaker added the constraint to only delete .jpeg files. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Delete all files from um my file system downloads folder. Oh, I forgot to mention, only delete uh let's see .jpeg files." }, "candidate_id": "wav_c2b9fd8480f9dfaa", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a plausible addition ('Oh, I forgot to mention'), the delivery is less natural. The initial request is slightly rushed and unclear ('delete all files from um my file system'), with an awkward laugh ('hehe') that disrupts the flow. The addition is also less smoothly integrated, with a hesitant 'uh let's see' that sounds more like a filler than a genuine recollection. The overall prosody is less consistent, and the correction feels tacked on rather than naturally flowing from the prior statement." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_file_filter", "proposed_instruction": "Delete all .png files from my file system ~/downloads folder. Wait, delete all .png files, not .jpg files.", "long_pause_offsets": [ 62 ], "source_sha256": "1431762c17cd7370633f5c0cfa712918c8851ddd33ae0f11919cc6ae0b2e057b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the file extension from .jpg to .png, while retaining the action, location, and scope. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/revision/revision_01_file_filter/46545c4b0cac4be6/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "9ab9daa61946e2b3c850b9f9fa1bce0231d8ffb2fa316a995031a3927f06dbd7", "seconds": 15.56, "spoken_utterances": [ "Delete all hmmm... .png files from um, let's see... my file system ~/downloads folder. ", "Wait, delete all um... .png files, not .jpg files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.jpg files' but immediately corrects to '.png files' after a pause and the word 'wait'. The final instruction is a clear revision of the file type to delete. The folder and action remain consistent. The revision criterion is satisfied as the correction is applied to the earlier instruction.", "heard_transcript": "delete all um .png files from um let's see my file system uh downloads folder. wait, delete all um .png files, not .jpeg files." }, "candidate_id": "wav_e0a3e9f918490ae9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with natural fillers ('um'), then self-corrects the file type from '.jpeg' to '.png' in a spontaneous, believable manner. The correction is integrated into the same utterance with a natural pause and intonation shift, showing a plausible moment of realization. The prosody is consistent, and the delivery sounds like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_folder_reference", "proposed_instruction": "Delete all .jpg files from my file system ~/Documents folder. Let me correct the folder: delete from ~/downloads, not ~/Documents.", "long_pause_offsets": [ 62 ], "source_sha256": "5c3e42e231a3e249fe2be8df41bba025bdd57fc76c60457bede5413293f98893", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder path from ~/Documents to ~/downloads, while retaining the file type and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/revision/revision_02_folder_reference/121a6c379a9bfca8/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "7880d0fbef8e03133f8a19451332963a3b6376229acb6e0a4a72a6d27608c94c", "seconds": 15.56, "spoken_utterances": [ "Delete all uh... .jpg files from my file system um... ~/Documents folder. ", "Let me correct the folder: delete from ... ~/downloads, not ~/Documents." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'documents folder' but immediately corrects themselves to 'downloads, not downloads, not', indicating a revision to the target folder. The final audible intent is to delete from the ~/downloads folder, matching the original instruction. The revision is clear and the final target is correctly specified.", "heard_transcript": "delete all uh jpeg files from my file system um documents folder. let me correct the folder, delete from downloads, not downloads, not" }, "candidate_id": "wav_c7855c37a656328d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then attempts a revision that is less coherent. The correction 'delete from downloads not downloads not' is awkward and repetitive, suggesting a stumble or confusion rather than a smooth self-correction. The intonation and pacing are less natural, and the repetition of 'downloads' sounds like a verbal error rather than a believable thought process. The overall delivery feels less fluent and more disjointed compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "7f73ad04c2308a732b7c745edff2b0ad8729091d6339e56047be577ed4c25b8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/cancellation/cancel_028/75667405a738556f/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "882603de6e73754f26379a257e67405639d1f6a7ad3a1ec5487cb5890195d28a", "seconds": 14.36, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder.", "Hold on, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to delete .jpeg files, then immediately follows with 'hold on, disregard that instruction, don't carry it out,' which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder. hold on, disregard that instruction, don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_020", "hide_variant_original": true, "source_sha256": "48073c8dd9970f28af6dafd618a94365815b93e1522d4457c2076836b5bd4b4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/ambiguous_halt/halt_020/0fc4062910551b6b/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "2627c95f79843df179eca9f687590ece26f494ee3102cb4e48532fb8a68b40e5", "seconds": 13.16, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder.", "Hold on just a moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly says 'hold on just a moment please' as a final interjection, which is a standard request to pause or wait. There is no cancellation of the task, and the initial request is not completed. The final audible intent is to halt execution temporarily, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder. hold on just a moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ccf4e5ce1cfce68d47d9e7c517adbfbb827dc618ae7fd44385d37192eab91485", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_2/audio_natural/filler/original/32cf3d9f7d971ae4/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "df2ab298611e39b72906cc5e9970f52294c413ba6ff4b4c9485f81f75b58f913", "seconds": 9.04, "spoken_utterances": [ "Delete all um... dot j p e g files from uh, let me think... my file system ... tilde slash downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let me think') but does not change the core task. The final intent is to delete .jpeg files from the specified folder, matching the original instruction. The file extension is pronounced as '.jpeg' but is semantically equivalent to '.jpg'. No correction or change of intent is made.", "heard_transcript": "delete all um .jpeg files from uh let me think my file system tilde slash downloads folder" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_file_filter", "proposed_instruction": "Delete all .txt—sorry, that should be .pdf files from my file system ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "46ddebb860bce88a0e79f0a63e93704e09f05b696210ce0855c793b7878db805", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects .txt to .pdf within the same utterance, preserving the action, scope, and target directory." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/false_start/false_start_01_file_filter/9c0b72deb95b99c3/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "2a8f9a1e7fb87144c0f8b4541ea4a32ed2b7b3396b3210fb448ea045d7619573", "seconds": 12.08, "spoken_utterances": [ "Delete all .txt—um... sorry, that should be uh... .pdf files from my file system ... ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says '.txt' but immediately self-corrects to '.pdf'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction matches the reference, so the final task is to delete all .pdf files from the ~/downloads folder.", "heard_transcript": "Delete all .txt um, sorry, that should be uh .pdf files from my file system downloads folder." }, "candidate_id": "wav_3cfe500dfdee515e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('delete all .txt'), immediately self-corrects with a natural 'sorry' and 'that should be', then smoothly continues with the correct command ('pdf files') and specifies the location ('downloads folder'). The prosody includes a brief pause and filler ('um') at the start, and the correction flows without awkwardness or repetition. The delivery is continuous and sounds like a spontaneous, real-time correction." } ] }, "is_default": true }, { "variant_id": "false_start_02_folder_reference", "proposed_instruction": "Delete all .pdf files from my file system ~/Documents—no, that should be ~/downloads folder.", "long_pause_offsets": [], "source_sha256": "a1aa9db51166547bdf4d718ea9e75c9ac67aea64973602651893d60d5256c7a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions ~/Documents but immediately corrects it to ~/downloads. The action, file type, and scope remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/false_start/false_start_02_folder_reference/c43c29ec3ed45a09/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "c996e262eab360a1908ce29635bf7e624f9a8207d12880e3ecc41321f30c42bf", "seconds": 13.36, "spoken_utterances": [ "Delete all uh... .pdf files from my file system ~/Documents—hmmm... no, that should be ... ~/downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Documents' but immediately self-corrects to 'Downloads folder'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction with the correct folder path.", "heard_transcript": "Delete all uh PDF files from my file system. Documents. No, that should be Downloads folder." }, "candidate_id": "wav_e7dd4135c4cb60d4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with the correct command ('delete all pdf files') but then hesitates with multiple 'um' fillers and an incorrect location ('documents'). After a pause, they self-correct with 'no, that should be' and state the correct folder. However, the correction is less fluid than audio_01, with a longer pause and a more abrupt shift. The repeated 'um' and the backtracking feel slightly more hesitant and less naturally integrated than the first clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all files from my file system ~/downloads folder. Oh, I forgot to mention, only delete .pdf files.", "long_pause_offsets": [ 57 ], "source_sha256": "565c2d99567bc008795bd328a53f52c18bc94bd77f80db182796fc440acfc773", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the initial broad deletion to only .pdf files, preserving the location and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/addition/addition_01/986d69cb586eb7fa/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "24cb742aa17566eb185bc8902a4746c324d77b01f7f237e9790683d1cd4a7953", "seconds": 13.72, "spoken_utterances": [ "Delete all files um... from my file system hmmm... ~/downloads folder. ", "Oh, I forgot to mention, only delete um... .pdf files." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all files from the downloads folder. After a pause, the speaker added a constraint to delete only .pdf files. This is a clear addition that narrows the scope of the original request. The final intent is to delete only .pdf files, which matches the reference instruction. The addition is spoken after a longer pause, satisfying the revision/addition type criterion.", "heard_transcript": "delete all files um from my file system uh downloads folder. oh, I forgot to mention, only delete um pdf files." }, "candidate_id": "wav_14e826b6b658ef5a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Delete all .pdf files from my file system folder. One more thing: do this only in the ~/downloads folder.", "long_pause_offsets": [ 50 ], "source_sha256": "225a2136a2836a4a065607eb14065e26d215d99250d0ceddc796df0fb6c0ac08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint, narrowing the scope from the general 'file system folder' to the specific '~/downloads folder'. The action and file type remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all .pdf files ... from my file system folder. ", "One more thing: do this only in the ... ~/downloads folder." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_folder_reference", "proposed_instruction": "Delete all .pdf files from my file system ~/Documents folder. Let me correct the folder: delete from ~/downloads, not ~/Documents.", "long_pause_offsets": [ 62 ], "source_sha256": "43e60675b9ad1d7184e45acd00f0509373d8fb585cfbee51402807050dc9e78e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder path from ~/Documents to ~/downloads, while retaining the action and file type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/revision/revision_02_folder_reference/bb447d3d100c2f9e/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "cc6ceb9c88da29ae83f01acefae05ef1c754453e5dd62a1cff1d3aaaf931d654", "seconds": 18.32, "spoken_utterances": [ "Delete all .pdf files ... from my file system ... ~/Documents folder. ", "Let me correct the folder: delete from let's see... ~/downloads, not ~/Documents." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives an incorrect folder path ('~/documents') but immediately self-corrects to the intended folder ('~/downloads') after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction.", "heard_transcript": "Delete all .pdf files from my file system ~/documents folder. Let me correct the folder. Delete from, let's see, ~/downloads, not ~/documents." }, "candidate_id": "wav_8e21e2f7898782ea", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_file_filter", "proposed_instruction": "Delete all .txt files from my file system ~/downloads folder. Wait, delete all .txt files, not .pdf files.", "long_pause_offsets": [ 62 ], "source_sha256": "65331c44f318194853cb54f70aef204e8dfbd46c53cc073674eafa464e3b2adc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the file type from .pdf to .txt, reversing the original instruction's intent." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all .txt files from my file system ... ~/downloads folder. ", "Wait, delete all um... .txt files, not .pdf files." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "131e160a27bab30e428ba0af2b0b5642e1a6ade39d907ae0d80486bb62b88c41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/cancellation/cancel_009/bc5a66e773ae550d/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "bb7612651c31eb699fb8a853a22159a2bf3c5488115516293995f9afad672d3e", "seconds": 13.48, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder.", "Wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a task to delete .pdf files from a specific folder, then immediately follows with 'wait, cancel that altogether,' which is a clear and final cancellation of the entire task. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "delete all .pdf files uh from my file system uh let me think tilde slash downloads folder. wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "9751a888897945a301e39506dccab4b8751fa243c2ac7c93976c147fb63cbacb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/ambiguous_halt/halt_025/618210feab81b63e/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "2a18c6fd3b0febd47826e1c84162e8a2393547cc873f843bd728cbcfb2d3cac3", "seconds": 14.0, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder.", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'wait, I need to think for a second,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request to delete PDF files is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "delete all let's see dot pdf files uh from my file system uh let me think tilde slash downloads folder. wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5a83b05_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ee2906fe09910f13c5d1b47b07ad4d07d867b26f8c130e7d87fbf2d153be7bd9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5a83b05_3/audio_natural/filler/original/949ac18e17d2844c/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "b1e777b38b98942ce074db6dc50dd36105c93f4ecac39d41b40a049e34d627f9", "seconds": 9.36, "spoken_utterances": [ "Delete all let's see... dot pdf files ... from my file system uh, let me think... tilde slash downloads folder." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler phrase 'let's see' and a hesitation 'uh', then corrects the initial 'all' to 'all' again, and includes a self-pause 'let me think'. Despite these fillers and hesitations, the final, complete instruction matches the original task. The speaker does not change the action, target, or value. The 'filler' type allows for such hesitations and self-corrections as long as the final goal is clear and matches the original instruction.", "heard_transcript": "Delete all let's see dot pdf files uh from my file system uh let me think tilde slash downloads folder." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject—no, that should be empty subject and body.", "long_pause_offsets": [], "source_sha256": "bd0f788a30873acb9af8f09f82a268fad3372429f5fe8aedbbf3fc1a5b5ab7e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions only 'empty subject' but immediately corrects to 'empty subject and body', which matches the original instruction exactly. The final goal includes both conditions as required." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/false_start/false_start_03_empty_field_condition/51ab03b3747f4b1d/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a5d8d85b4acb0574366e589602809d2a21eae5da283a54182470a0651e11f002", "seconds": 12.28, "spoken_utterances": [ "Delete all my Gmail drafts that have um... empty subject—uh... no, that should be ... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but immediately self-corrects to 'empty subject and body' after a pause and 'no'. This is a false start where the final corrected instruction matches the original reference. The correction is clear and the final intent is fully specified.", "heard_transcript": "deletes all my Gmail drafts that have um empty subject uh no, that should be empty subject and body." }, "candidate_id": "wav_3aad08d04ce65cc5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'empty subject' to 'empty subject and body' with a plausible hesitation ('uh') and a smooth, continuous flow. The correction is integrated into the same utterance without awkwardness, and the prosody reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "Delete all my Gmail sent—scratch that, make it drafts that have empty subject and body.", "long_pause_offsets": [], "source_sha256": "33915d6239cdbbdf33dad7b200db7e5410131b06496366234f31d5a4437255a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately corrects it to 'drafts' within the same utterance. The final target is correctly specified as drafts, and all other conditions (empty subject and body) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/false_start/false_start_02_email_state/60dd25ceab50066c/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "1c4e931739b99f56b263c99396464c4b6cfd75b6d713200cc013e82068296709", "seconds": 11.28, "spoken_utterances": [ "Delete all my Gmail sent—ah... scratch that, make it uh... drafts that have ah... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent ta' but immediately self-corrects with 'scratch that' to 'drafts'. This is a clear false start where the final intent is the corrected version. The final instruction matches the original reference exactly, so the control intent is to execute the corrected task.", "heard_transcript": "Delete all my Gmail sent ta... scratch that, make it uh drafts that have uh empty subject and body." }, "candidate_id": "wav_6c29df20447d82e7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but makes a more abrupt and less integrated correction from 'sent' to 'drafts'. The phrase 'scratch that' is a common filler for self-correction, but the subsequent 'make it uh drafts' sounds slightly disjointed and less fluent than audio_01. The prosody is less continuous, and the correction feels more like a forced edit than a spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Delete all my Outlook—I should say Gmail drafts that have empty subject and body.", "long_pause_offsets": [], "source_sha256": "3bf3ec0e780a16abaf99e8fd7399a30677f05a4a117722faa1ed33b45561c823", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, action, and all other conditions." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my um... Outlook—uh... I should say Gmail drafts that have empty subject ... and body." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have empty subject and body.", "long_pause_offsets": [ 28 ], "source_sha256": "6e2e7f0920e89feb1f903dfc366472f1678f7ac71f1a7eb37b8010d612998c9c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'empty subject and body' to the initial request to delete drafts, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/addition/addition_01/8aacecc8c1837025/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "710e938f2a10f842a37ed3713d7850121c26fd4d8686614e2de70fe4e6ffbc78", "seconds": 11.88, "spoken_utterances": [ "Delete all my ... Gmail drafts. ", "Oh, I forgot to mention, only delete drafts that have ... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to delete all Gmail drafts. After a pause and an interjection, the speaker added a constraint: only delete drafts that have empty subject and body. This matches the addition type criterion, where the final goal combines the initial instruction with the added condition. The final intent is fully specified and matches the original reference instruction.", "heard_transcript": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have empty subject and body." }, "candidate_id": "wav_f8eac92e12f96fd7", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Delete Gmail drafts that have empty subject and body. I left out a detail: delete all my Gmail drafts.", "long_pause_offsets": [ 54 ], "source_sha256": "e395ab2a766883f4dbe599152581a2583d97a728518bb616594cccf0c1e43926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the scope 'all my Gmail drafts', which is consistent with the original instruction and does not contradict any other requirement. The final goal remains the same as the original." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete hmmm... Gmail drafts that have um... empty subject and body. ", "I left out a detail: delete um, let me think... all my Gmail drafts." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Delete all my Outlook drafts that have empty subject and body. Wait, the app should be Gmail, not Outlook.", "long_pause_offsets": [ 63 ], "source_sha256": "4d9eeff6695299eef65003466f1f1107575e9b18a27817ee5c02fa31a623e52b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the rest of the instruction including the target (drafts), condition (empty subject and body), and scope (all my)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/revision/revision_01_app/96593b007d5a456d/speech.wav", "wer": 0.15, "wav_sha256": "99f99d8942c7663962dcdbe9a986810c41bc4c029dce8b855c33a041917990f2", "seconds": 12.44, "spoken_utterances": [ "Delete all my Outlook drafts that have empty subject ... and body. ", "Wait, the app should be hmmm... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail' after a pause, which is a revision. The final intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "to lead all my Outlook drafts that have empty subject and body. Wait, the app should be, Gmail, not Outlook." }, "candidate_id": "wav_be5af2c113082b50", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, coherent request, then pauses naturally with 'wait' to self-correct the application from 'Outlook' to 'Gmail'. The correction is delivered with a believable hesitation and a slight upward inflection on 'huh', followed by a smooth, self-contained revision. The prosody and pacing are consistent with spontaneous speech, and the revision feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_02_email_state", "proposed_instruction": "Delete all my Gmail sent that have empty subject and body. Change the email state to drafts, not sent.", "long_pause_offsets": [ 59 ], "source_sha256": "befba26b9898734cdee46386f66eda36c0429c13441b0f3fd54203e9fb994dd1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email state from 'sent' to 'drafts', preserving the rest of the instruction including the condition of empty subject and body." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/revision/revision_02_email_state/263ea89c5cb38c41/speech.wav", "wer": 0.0, "wav_sha256": "ff41157489d2f72be9ce6852f2051e8b28511096afe6dddc84b7f1a8fa40c1e6", "seconds": 11.44, "spoken_utterances": [ "Delete all my uh... Gmail sent that have um... empty subject and body. ", "Change the email state to drafts, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction incorrectly refers to 'Gmail sent' emails, but the speaker immediately corrects this by revising the target to 'drafts, not sent'. The revision is clear and applies to the same condition (empty subject and body). The final intent is to delete Gmail drafts with empty subject and body, which matches the original instruction. The revision is applied correctly, and no other constraints are changed.", "heard_transcript": "Delete all my uh Gmail sent that have um empty subject and body. Change the email state to drafts, not sent." }, "candidate_id": "wav_be487da11dd1182b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then abruptly cuts off mid-sentence after 'body'. The delivery is fluent and natural up to that point, but the sudden, unexplained termination makes the utterance feel incomplete and less realistic as a standalone spontaneous request. There is no audible hesitation, self-correction, or follow-up, which reduces its credibility as a natural speech sample." } ] }, "is_default": false }, { "variant_id": "revision_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject. Make that empty subject and body, not just empty subject.", "long_pause_offsets": [ 52 ], "source_sha256": "885255b8220acb5645b26cc1f0e5f225df5cbe217dd17a398a3917437f582706", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision adds the body condition to the initial subject-only condition, resulting in the conjunction of both empty fields as required by the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my ... Gmail drafts that have um, let me think... empty subject. ", "Make that ... empty subject and body, not just empty subject." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "060fcee094fa43169629432f09785992f9364b13195cc02bc1e744dc1c5aa6b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/cancellation/cancel_030/9a1b3c6f645ef241/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "811c86799818f9db780bef7ecf7850a2cec6db3bb43ae5abe905a7ea699bdc68", "seconds": 12.76, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body.", "Wait, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the deletion task but then clearly revokes it by stating, 'Wait, you can drop that task. I no longer need it.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is a complete cancellation, not a pause or a revision.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body. Wait, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "72fe94f82523b48a7c99885ddca48064f1f22fcffa67b862a37f32acbe3e07a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/ambiguous_halt/halt_012/1bf6a6db6e83cbe3/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "07a4ddaf938dbbc2179323e0ff4724813f90626e72ae938cf1482c6cceec9ca1", "seconds": 10.6, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The recording ends with the explicit interjection 'Give me a second,' which is a clear request to pause or wait. This is consistent with the 'ambiguous_halt' type, where the final audible intent is to pause, not to execute or cancel. The initial request is not completed, but the final control intent is to pause, which matches the expected behavior for this type.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "057e681fd43c97b6f5f3eeb175902e7134e1bedc54568b51478cadd7b8b7d90d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_1/audio_natural/filler/original/7c458b6a7d1984bb/speech.wav", "wer": 0.2, "wav_sha256": "d68d8652c0d1f24705c98c32cf6dc054f11f9be9c717fc8b81d50ded53bb5096", "seconds": 7.2, "spoken_utterances": [ "Delete all my uh... Gmail drafts that have um, let's see... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') and a self-repair ('what's the') but does not change the core request. The final audible intent matches the original instruction. The filler words and hesitation do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Delete all my uh Gmail drafts that have um what's the empty subject and body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject—sorry, I meant empty subject or body.", "long_pause_offsets": [], "source_sha256": "fe4fa502b711da9535d615d1508ebab0a516df39398d69fa2306579f002050ca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but immediately corrects to 'empty subject or body'. The final intended condition includes both criteria, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/false_start/false_start_03_empty_field_condition/613a902557980793/speech.wav", "wer": 0.3888888888888889, "wav_sha256": "5b3a4502921c8a01ce274f2a247d0f65352c4df4082374b5ba9d70ae52188dfa", "seconds": 7.16, "spoken_utterances": [ "Delete all my um... Gmail drafts that have empty subject—uh... sorry, I meant ... empty subject or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'empty subject' but then self-corrects to 'empty subject or body' after a pause and an apology. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is correctly executed.", "heard_transcript": "Delete all my um Gmail drafts that have empty subject uh uh sorry, I meant empty subject or body." }, "candidate_id": "wav_cc508beaca8adbbb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Delete all my um Gmail drafts that have empty subject' and then self-corrects with 'sorry, I meant empty subject or body' in a smooth, conversational manner. The repair is immediate, the intonation reflects a genuine moment of correction, and the pacing feels spontaneous rather than rehearsed. The filler 'um' and the self-correction are well-integrated into the flow." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Delete all my Outlook—I should say Gmail drafts that have empty subject or body.", "long_pause_offsets": [], "source_sha256": "c75bdf3598bbdd27db8f9ded31fea1e316373153cd76dee8d6298608fd417c8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, action, and conditions on subject/body. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/false_start/false_start_01_app/a8b29ba1acf13896/speech.wav", "wer": 0.15, "wav_sha256": "b6a346a885cf3babdff706dfcf59876759a668a2d4b9928824ca1f14fe236c50", "seconds": 8.08, "spoken_utterances": [ "Delete all my Outlook—hmmm... I should say ... Gmail drafts that have empty subject uh, let me think... or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. The final audible intent matches the original instruction exactly. The 'false_start' type requires applying the correction, which is done here. No other constraints are changed or lost.", "heard_transcript": "Delete all my Outlook [inaudible] I should say Gmail drafts that have empty subject uh let me think or body." }, "candidate_id": "wav_372e5f22ab69efd1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start with a self-correction ('Outlook' to 'Gmail'), but the repair is less fluid. The speaker says 'I should say Gmail drafts' before continuing, which creates a slight break in the flow. The 'uh let me think' pause feels more like a hesitation to formulate the rest of the request rather than a natural correction of a prior word. The overall delivery is still realistic but slightly less seamless than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "Delete all my Gmail sent—no, make that drafts that have empty subject or body.", "long_pause_offsets": [], "source_sha256": "b24829975aa083229d3d09f41ef62b6f1f716794fac394e8acf259736eabb3b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'drafts' within the same utterance. The final target is correctly specified as drafts, and the condition about empty subject or body remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/false_start/false_start_02_email_state/ee57f981d5a460c5/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "3d4212e3829d34d9a3a609efff9ebdf4ff3c200ef908eec943c8e42605f35929", "seconds": 7.24, "spoken_utterances": [ "Delete all my Gmail um, let's see... sent—uh... no, make that drafts that have empty subject or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'drafts' after a pause and the word 'no'. This is a false start where the final corrected value is 'drafts'. The rest of the instruction matches the original. The final intent is clear and executable.", "heard_transcript": "Delete all my Gmail, um, let's see, sent, uh, no, make that drafts that have empty subject or body." }, "candidate_id": "wav_3f5c7435491a5f7e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a more complex revision with multiple interruptions and a change of mind ('sent' to 'drafts'). The speaker says 'Delete all my Gmail um let's see sent uh no make that drafts', which sounds less like a spontaneous correction and more like a rehearsed or awkwardly constructed sentence. The 'um' and 'uh' fillers are present, but the overall structure feels less natural and more disjointed compared to the other two. The revision is less integrated into the speech flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have an empty subject or body.", "long_pause_offsets": [ 28 ], "source_sha256": "926c14f7af540ad460adace35949887db91bc0c1c798616b013cb43d98410864", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition about empty subject or body, which refines the initial broad request. The final goal combines both the scope (all Gmail drafts) and the constraint (empty subject or body)." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my ... Gmail drafts. ", "Oh, I forgot to mention, only delete drafts that have an uh... empty subject or body." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Delete all my drafts that have empty subject or body. Oh, right, this is specifically for my Gmail drafts.", "long_pause_offsets": [ 54 ], "source_sha256": "ec6c4a8c745709a2d122d3b85ec675bf3f2a464b7b3e452bd3ec635ed524811a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Gmail' to the initial draft deletion request, preserving the condition about empty subject or body." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my drafts that have ... empty subject or body. ", "Oh, right, this is specifically for my ... Gmail drafts." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Delete all my Outlook drafts that have empty subject or body. Wait, delete drafts in Gmail, not Outlook.", "long_pause_offsets": [ 62 ], "source_sha256": "9c3af6fc0e4aed8847a1d436922a5ada5be650e652226467e11ce4a3e0a5e048", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the rest of the instruction: deleting drafts with empty subject or body." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/revision/revision_01_app/ee3e962854fd1023/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "0859d69473d928d3888b2504f6cc70dd4750d57b2e92f49713b83dbcc538e60c", "seconds": 12.04, "spoken_utterances": [ "Delete all my ... Outlook drafts that have empty subject uh... or body. ", "Wait, delete drafts in uh, let's see... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the email service as 'Outlook' but immediately corrects it to 'Gmail' after a pause and self-correction. The revision applies the correction to the earlier instruction, preserving the action (delete), the condition (drafts with empty subject or body), and the target (Gmail). The final intent matches the original instruction exactly.", "heard_transcript": "Delete all my Outlook drafts that have empty subject uh or body. Wait, delete drafts in uh let's see, Gmail, not Outlook." }, "candidate_id": "wav_74949f50177104db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then self-corrects with a natural hesitation ('uh') and a plausible revision of the email service from 'Outlook' to 'Gmail'. The correction is integrated smoothly into the same utterance, with a believable shift in intonation and pacing. The pause and filler ('let's see') before the correction sound like genuine recollection, not a rehearsed or robotic edit. The prosody remains consistent and the delivery is continuous, making it sound like a spontaneous, real-time correction." } ] }, "is_default": true }, { "variant_id": "revision_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have has attachments. I misspoke about the condition: delete drafts that have empty subject or body, not those that have attachments.", "long_pause_offsets": [ 54 ], "source_sha256": "ebd9d383160fd3bc2565d38c372298da0773051b6a70e018df7a1ff0e17aa72c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'has attachments' to 'empty subject or body', which matches the original instruction. The app and scope (all drafts) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/revision/revision_03_empty_field_condition/69533eedeb7ec114/speech.wav", "wer": 0.1, "wav_sha256": "aea516ba8cfb7b7cdd8208c1fbd26aec2cf270b20da9778d627ee783e72a2a1d", "seconds": 14.88, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have um... has attachments. ", "I misspoke about the condition: delete drafts that have uh... empty subject or body, not those that have attachments." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the condition as 'drafts that have attachments' but immediately corrects themselves with 'I misspoke about the condition' and clearly states the intended goal as 'delete drafts that have empty subject or body'. This is a revision where the final audible intent overrides the initial incorrect statement. The final instruction matches the original reference.", "heard_transcript": "Delete all my um Gmail drafts that have um has attachments. I misspoke about the condition. Delete drafts that have uh empty subject or body, not those that have attachments." }, "candidate_id": "wav_545b56f3cac19387", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'um' before stating the condition. After a pause, they correct themselves with a full revision of the condition from 'has attachments' to 'empty subject or body'. While the correction is grammatically and semantically coherent, the delivery feels slightly more segmented and less fluid than audio_01. The second part of the utterance sounds like a separate thought or a delayed realization, with a longer pause and a more abrupt shift in focus. The prosody is less continuous, and the correction feels more like a post-hoc addition than a spontaneous self-repair." } ] }, "is_default": false }, { "variant_id": "revision_02_email_state", "proposed_instruction": "Delete all my Gmail sent that have empty subject or body. Make that delete all my Gmail drafts, not sent emails.", "long_pause_offsets": [ 58 ], "source_sha256": "b35a6fbfcd36713d6c99230e74fac3fcf5351c4be5ea4ea0381475255a853aa0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email folder from 'sent' to 'drafts' and explicitly excludes 'sent emails', preserving the condition about empty subject or body." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my Gmail sent that have um, let me think... empty subject or body. ", "Make that delete all my Gmail um, let's see... drafts, not sent emails." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "151237c26ff7771b3a02291062552f604476e03b476ae377f0e9d95aa27efa2b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/cancellation/cancel_019/bff7412e722971ee/speech.wav", "wer": 0.0, "wav_sha256": "4ea0800af5d5f2757202973c66ec1c717d2d58fa8fc289ee1bf7616314098c8e", "seconds": 10.6, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body.", "Oh, wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Oh wait, never mind, don't do that.' This constitutes a cancellation of the entire task as per the type-specific criterion. The final interjection is a direct cancellation, not a pause or ambiguous halt.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body. Oh wait, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "84b895f1da36b8bad25576646ee75ef1749d5cbf85b6f39d913fda8148787897", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/ambiguous_halt/halt_014/ad65773cea4711a2/speech.wav", "wer": 0.0, "wav_sha256": "dcc6ed6c6f655b39318af516ea39fdd0680575a947ba00c66b8387f9ab286fc5", "seconds": 9.76, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Can you hold on a second?' as the final interjection, which is a clear request to pause. The initial part of the sentence is a filler and hesitation, not a final directive to execute. The final audible intent is to pause, not to execute or cancel. Therefore, the audio matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c6faa857ec79608c7f0a751ac47db3b72b35eb33ccf4ebfe8ed5648c73624f12", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_2/audio_natural/filler/original/0dbaa977a5f8748b/speech.wav", "wer": 0.0, "wav_sha256": "4657ec0cd54eb87b63980501e339174d18ab7b77b9044054060de5ca793f4f03", "seconds": 6.32, "spoken_utterances": [ "Delete all my uh, let me think... Gmail drafts that have empty subject uh... or body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let me think') and a pause but does not change the core task. The final intent remains to delete Gmail drafts with empty subject or body, matching the original instruction. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Delete all my uh let me think Gmail drafts that have empty subject uh or body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Delete all my Outlook—I should say Gmail drafts that have empty subject and body.", "long_pause_offsets": [], "source_sha256": "3bf3ec0e780a16abaf99e8fd7399a30677f05a4a117722faa1ed33b45561c823", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the app, action, and all other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/false_start/false_start_01_app/9bcf97a754f3ead6/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "70860c1e0a457bda3fe84b5a9ca1930694430264743ab31f22f21f079ba6e128", "seconds": 10.2, "spoken_utterances": [ "Delete all my uh... Outlook—um... I should say Gmail drafts that have empty subject hmmm... and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Delete all my uh Outlook um, I should say Gmail drafts that have empty subject and body." }, "candidate_id": "wav_ddd6157dbdd2ac71", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('Outlook'), self-corrects to 'Gmail drafts' with a natural hesitation ('I should say'), and continues with a smooth, coherent request. The pacing is conversational, with a believable pause and filler ('hmm') before the final clause. The correction feels spontaneous and integrated into the flow." } ] }, "is_default": true }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "Delete all my Gmail sent—no, make that drafts that have empty subject and body.", "long_pause_offsets": [], "source_sha256": "522bef86cdebb810b76b30583bea96212c2ec7cacdbdc6f35709b290b6314072", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'sent' to 'drafts' within the same utterance. The final target is correctly specified as drafts, and all other conditions (empty subject and body) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/false_start/false_start_02_email_state/dfeca1231b4f379e/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "e999801d62264ba3ccdb340ef4822dbbf75dbc774da2fd58fcc9a15a429ffdc3", "seconds": 14.04, "spoken_utterances": [ "Delete all my Gmail sent—um... no, make that uh... drafts that have empty subject uh, let me think... and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent items' but immediately self-corrects to 'drafts'. The correction is clear and immediate, with no additional conditions or changes to the rest of the instruction. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Delete all my Gmail sent items. No, make that uh uh drafts that have empty subject uh let me think and body." }, "candidate_id": "wav_4db14e27aef4e014", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a false start ('Gmail sent'), followed by a self-correction to 'drafts'. However, the correction is less fluid, with multiple stutters ('uh uh') and a longer, more awkward pause ('let me think') that disrupts the natural rhythm. The final phrase 'and body' is delivered abruptly, making the overall flow feel more hesitant and less integrated than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have empty subject or body—sorry, I meant empty subject and body.", "long_pause_offsets": [], "source_sha256": "58105fda9d3854117a5a95e01b33dd50553184596a3935227534b8cffe172eec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially uses 'or' but immediately corrects to 'and' within the same utterance. The final conjunction is 'and', which matches the original instruction's requirement for both conditions to be satisfied." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my um... Gmail drafts that have empty subject or body—uh... sorry, I meant empty subject and body." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that have an empty subject and body.", "long_pause_offsets": [ 28 ], "source_sha256": "3dd8cd3c7b6d7e48c38a72dc945100fcdaaa5e4cce849f3ede3717395f938249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition about empty subject and body, which refines the initial broad request. The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/addition/addition_01/64f75c2e4a881d69/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "0c5eae76d8d27d0b4c519a392f8c3a80679e044c834c64618a84a64783e50f98", "seconds": 13.24, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts. ", "Oh, I forgot to mention, only delete drafts that have an let me think... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Delete all my Gmail drafts' and then adds a constraint 'only delete drafts that haven't... empty subject and body'. The phrase 'haven't' is likely a self-correction or hesitation for 'have', but the final audible intent is to delete drafts with empty subject and body. The addition refines the initial request, and the final goal matches the original instruction. The hesitation and self-correction do not change the final intent.", "heard_transcript": "Delete all my Gmail drafts. Oh, I forgot to mention, only delete drafts that haven't, let me think, empty subject and body." }, "candidate_id": "wav_32d2a56f21c01d3e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, direct command. After a natural pause, they add a clarifying detail ('only delete drafts that have...') with a believable hesitation ('let me think') and a smooth, self-corrected continuation. The prosody is consistent, and the addition feels like a spontaneous afterthought rather than a rehearsed or robotic insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Delete all my Gmail drafts that have. I left out a detail: the drafts must have an empty subject and body.", "long_pause_offsets": [ 38 ], "source_sha256": "c3fc1df40d13acc7f820a94b52a264b9646fe68d33939444cc9fa10d5461df2a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition about empty subject and body, completing the initial incomplete instruction. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/addition/addition_02/06c90d92bd37df15/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "5e9a5fa89c9548489f152752c8415c13d5d1b352e31047b071326a9f7780f1a1", "seconds": 11.68, "spoken_utterances": [ "Delete all my ... Gmail drafts that have. ", "I left out a detail: the drafts must have an ah... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to delete drafts with an unspecified condition, then adds the missing constraint 'empty subject and body' after a pause. This matches the 'addition' type, where the final goal is the combination of the initial instruction and the added detail. The final audible intent is clear and complete, matching the original reference instruction.", "heard_transcript": "Delete all my Gmail drafts that have... I left out a detail, the drafts must have an empty subject and body." }, "candidate_id": "wav_15d68e6d90e84a90", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial command is clear, but the addition is less natural. The speaker says 'I left out a detail' followed by a pause and then 'the drafts must have an uh empty subject and body.' The 'uh' is awkwardly placed and sounds like a filler that doesn't fit the flow of the sentence. The prosody is slightly off, and the phrase 'an uh empty subject' sounds like a stumble rather than a genuine thought process, reducing the overall realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_empty_field_condition", "proposed_instruction": "Delete all my Gmail drafts that have has a subject and body. That should be empty subject and body, not has a subject and body.", "long_pause_offsets": [ 61 ], "source_sha256": "6f7f028c5dcee171deab8ecda05a883133598e5203a1f8107ab2ced930eaa076", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the condition from 'has a subject and body' to 'empty subject and body', which reverses the intended meaning. The original instruction requires deleting drafts with both subject and body empty, but the initial utterance incorrectly requested deletion of drafts that have both fields filled." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/revision/revision_03_empty_field_condition/c81dce4e6dfed042/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "48ab1ce9c5852bdaea64370b0b6b0ad4a4dff61427d169d41ca9edcd83992c41", "seconds": 12.96, "spoken_utterances": [ "Delete all my Gmail drafts that have ... has a subject and body. ", "That should be um... empty subject and body, not has a subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'has a subject and body' but immediately corrects themselves to 'empty subject and body' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed.", "heard_transcript": "Delete all my Gmail drafts that have has a subject and body. That should be um empty subject and body, not has a subject and body." }, "candidate_id": "wav_4cea3726076ed96f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural prosody and a single, plausible self-correction ('has a subject and body' to 'empty subject and body'). The correction is smoothly integrated into the sentence, and the overall flow is coherent and spontaneous, suggesting genuine task formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": false }, { "variant_id": "revision_02_email_state", "proposed_instruction": "Delete all my Gmail sent that have empty subject and body. Change the email state to drafts, not sent.", "long_pause_offsets": [ 59 ], "source_sha256": "befba26b9898734cdee46386f66eda36c0429c13441b0f3fd54203e9fb994dd1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email state from 'sent' to 'drafts', preserving the rest of the instruction including the condition of empty subject and body." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/revision/revision_02_email_state/b33fc32a079b1a29/speech.wav", "wer": 0.05, "wav_sha256": "1ee7d017b1bea45372377ef807caf784bfabed4b0004efd9169703eef346a80f", "seconds": 11.52, "spoken_utterances": [ "Delete all my um... Gmail sent that have ... empty subject ... and body. ", "Change the email state to drafts, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail sent' but immediately corrects to 'Gmail drafts' in the revision. The final instruction is a revision of the original request, changing the target folder from 'sent' to 'drafts'. The condition of empty subject and body is preserved. The additional instruction to 'change the email state to drafts, not sent' is consistent with the correction and clarifies the desired outcome. The final intent matches the revised goal.", "heard_transcript": "Delete all my um Gmail sent that have empty subject and body. Change the email state to drafts, not sent." }, "candidate_id": "wav_7bd1724d4f7f18e5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a filled pause ('um') and a slightly hesitant delivery. The correction ('change the email state to drafts, not sent') is abrupt and delivered with a flat, declarative tone, lacking the natural intonation and flow of a spontaneous thought process. The transition between the initial request and the correction feels more like a post-hoc edit than a genuine self-repair, reducing its realism." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Delete all my Outlook drafts that have empty subject and body. Wait, delete drafts in Gmail, not Outlook.", "long_pause_offsets": [ 63 ], "source_sha256": "b398099c226b7a592cf26aea1afb43c01aee60e5a1e2b26c0edf1b7be8f4fd6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Outlook to Gmail, while retaining the other conditions: deleting drafts with empty subject and body. The final goal matches the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Delete all my ... Outlook drafts that have empty subject ... and body. ", "Wait, delete drafts in um... Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "d3325902550232cfae05906c3249cc8715bb1ec67c7cab1c206675b82259761a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/cancellation/cancel_019/f772be97ed91b88a/speech.wav", "wer": 0.1, "wav_sha256": "cdf7bd94c784c8bf86bb256bc581db937b694104da975f9b09680b5068734abe", "seconds": 12.4, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body.", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to delete drafts with empty subject and body, then immediately follows with 'Hang on, never mind, don't do that,' which is a clear and unambiguous cancellation of the entire task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body. Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "531eb1e908612a2af25da05bc8e16410b900c91f38941176377a359cb6c54703", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/ambiguous_halt/halt_011/86fe3617dd84d2ce/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "b8600c04e21ec091b0c1ecded41a7018279441ba847cc2153f1ddae7d11f88ab", "seconds": 10.76, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then makes a final interjection 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final audible intent is to pause, not to execute or cancel. No correction or new instruction is given after the pause request.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "5e27cd7_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "133f3b49de7b50eeb508ebe85b15cb02816ad258fe15685c677c7a3a18c77a49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/5e27cd7_3/audio_natural/filler/original/28ffe785b9e1e44d/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "1fbdc6e009f47bef80d4099f44ca71df68fa6cf1bdbd109d3d4f98d4a9bfa0b4", "seconds": 7.36, "spoken_utterances": [ "Delete all my hmmm... Gmail drafts that have uh... empty subject and body." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Delete all my Gmail drafts that have uh empty subject and body.' The 'uh' is a filler sound that does not alter the task goal. The final intent matches the original instruction, and the filler does not constitute a change in control intent. Therefore, the task is executed as intended.", "heard_transcript": "Delete all my Gmail drafts that have uh empty subject and body." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_email_state", "proposed_instruction": "Send all my scheduled—I meant future-scheduled emails on Gmail right away.", "long_pause_offsets": [], "source_sha256": "4a67e7799688d31814425e9abd555e1da658a919251730b8a316dbc3c54daf3c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'scheduled' to 'future-scheduled' within the same utterance, clarifying the intended scope of emails to be sent. The app (Gmail) and action (send right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/false_start/false_start_01_email_state/34361990f2f331d9/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "6ed6a665684eb7c553e101508dbe89dff5318164e6c0f3dce0262026d775d734", "seconds": 9.52, "spoken_utterances": [ "Send all my scheduled—um... I meant hmmm... future-scheduled emails on Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'scheduled' but immediately self-corrects to 'future scheduled' with 'I meant'. This is a false start where the final corrected value replaces the initial mistaken one. The final intent matches the original instruction, and the correction is immediate and clear. Therefore, the task should be executed as per the final corrected intent.", "heard_transcript": "Send all my scheduled um, I meant future scheduled emails on Gmail right away." }, "candidate_id": "wav_d876dac4e7d4f6a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('send all my scheduled um...'), followed by a natural self-correction ('I meant...') and a smooth continuation. The prosody is consistent, the repair is integrated into the flow, and the intonation reflects genuine hesitation and correction. The pacing and rhythm are very natural, with a believable pause and filler ('um') before the correction." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook—sorry, I should have said Gmail right away.", "long_pause_offsets": [], "source_sha256": "40b3935b3f45195f71e4b5b0cd8097de2e83e61afa8cd3ded61f0b7bb0d1ccc2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, scope, and urgency. The final intended app is Gmail." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/false_start/false_start_02_app/b0a52bfed37b3b29/speech.wav", "wer": 0.0, "wav_sha256": "ba3d9964d09f210d851c9bfba7325f599969a19217c12684c3ff928dec36e985", "seconds": 8.68, "spoken_utterances": [ "Send all my future-scheduled emails on Outlook—uh... sorry, I should have said ... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with no other changes or omissions.", "heard_transcript": "Send all my future scheduled emails on Outlook uh uh sorry, I should have said Gmail right away." }, "candidate_id": "wav_837a5d9485a03d85", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false start and correction, but the correction ('sorry, I should have said...') is delivered as a separate sentence rather than being integrated into the original utterance. The prosody is slightly less fluid, and the pause before the correction feels a bit longer and more abrupt. While still realistic, the repair is less seamlessly woven into the initial request compared to audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later—no, that should be right away.", "long_pause_offsets": [], "source_sha256": "5d8bd1557ac4ff246907c5a293da22ad2c0100c558cce479acaefc3596b316b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'later' but immediately corrects it to 'right away' within the same utterance. The final intended action is to send the emails immediately, matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Send all my let's see... future-scheduled emails on Gmail later—uh... no, that should be um... right away." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Send all my future-scheduled emails on Gmail. Oh, and one more thing: send them right away.", "long_pause_offsets": [ 46 ], "source_sha256": "7761cab5df22c696df33843ffe5be85a41746b923803ca44b67d3cc62f5569a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the urgency condition 'right away' to the initial instruction, which is consistent with the original goal and does not contradict or remove any existing requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/addition/addition_03/e34eb3d023366eed/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "4b43b684c6e31949f1499ee9dfef5b67ef74b3aca8ef478be6867be888ddaacd", "seconds": 13.12, "spoken_utterances": [ "Send all my uh... future-scheduled emails on let's see... Gmail. ", "Oh, and one more thing: send them uh... right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'future scheduled emails on let's see, Gmail' and then adds 'send them uh right away' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final intent is to send the emails immediately, which is consistent with the reference instruction.", "heard_transcript": "Send all my uh future scheduled emails on let's see, Gmail. Oh and one more thing, send them uh right away." }, "candidate_id": "wav_18de1b7cae2172a9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear initial request, includes a natural hesitation ('uh') and a plausible self-correction ('let's see'), then adds a new detail ('send them uh right away') after a short pause. The prosody is conversational, with appropriate stress and pacing. The addition feels spontaneous and integrated into the flow, not tacked on." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future-scheduled.", "long_pause_offsets": [ 40 ], "source_sha256": "8c8860353e97f886a01a020700735667e43fdb92c5c3a401083edea1ad740c17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'future-scheduled', which narrows the initial broad request to only those emails. The app (Gmail), action (send), and timing (right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/addition/addition_01/b18b9727afa48a4a/speech.wav", "wer": 0.0, "wav_sha256": "f51744a0f6b784be9a7c9f422c429cec8707c7ecede5162d0048d22caf7de995", "seconds": 13.36, "spoken_utterances": [ "Send all my emails on ... Gmail right away. ", "Oh, I forgot to mention, only send emails that are ... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to send all emails on Gmail right away. After a pause and an interjection, the speaker added a constraint: 'only send emails that are future scheduled.' This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the new constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future scheduled." }, "candidate_id": "wav_7eda8cb41d05d15a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a clear initial request, then adds a new restriction after a pause. However, the addition ('only send emails that are future scheduled') is delivered with a slightly more formal, less conversational tone. The prosody is flatter, and the transition feels more like a correction than a natural afterthought, making it less realistic than audio_01." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Send all my future-scheduled emails on right away. I left out a detail: send them on Gmail.", "long_pause_offsets": [ 51 ], "source_sha256": "c93d75795e15d3a59df296adb44db8c7dc958c5bec8aabc3440bad382eebf9be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app 'Gmail' to the initial instruction, which otherwise matches the original goal. The action, scope, and timing remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on right away. ", "I left out a detail: send them on ... Gmail." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_email_state", "proposed_instruction": "Send all my draft emails on Gmail right away. Wait, the emails should be future-scheduled, not draft.", "long_pause_offsets": [ 46 ], "source_sha256": "16d105f7307ddb6ac69e3ea3be8bff718eba8ca8fe937e2a4656077d9c3a973a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision corrects the email type from 'draft' to 'future-scheduled', which reverses the intended category of emails to be sent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/revision/revision_01_email_state/07ee9e5d09fcba38/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "7e8a1ebdd4c7628a2840a66f3e0ffb91f9c9a3ccc7f2f0b6d17a7ecdaa83244f", "seconds": 14.32, "spoken_utterances": [ "Send all my draft emails on uh... Gmail right away. ", "Wait, the emails should be uh, let's see... future-scheduled, not draft." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'draft emails' but immediately corrects to 'future scheduled' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Send all my draft emails on the Gmail right away. Wait, the emails should be uh, let's see, future scheduled, not draft." }, "candidate_id": "wav_e685c72c7afc91e6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then naturally hesitates with 'wait' and 'let's see' before self-correcting from 'draft' to 'future scheduled'. The correction is integrated smoothly into the same utterance with a plausible shift in intonation and pacing, mimicking a real-time thought process. The prosody is consistent and the delivery sounds spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later. That should be right away, not later.", "long_pause_offsets": [ 52 ], "source_sha256": "f46e58a5e49d45350f2f3b995b197edabe925ddf44ad635771268e3fde33a311", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the timing from 'later' to 'right away', reversing the original instruction's intended execution time." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/revision/revision_03_execution_timing/6ae6852999742710/speech.wav", "wer": 0.0, "wav_sha256": "a4a013c4b891554b8261602bdf406d40ceb7a781e9aa6b6b4ff92804030b3483", "seconds": 12.08, "spoken_utterances": [ "Send all my uh, let me think... future-scheduled emails on ... Gmail later. ", "That should be right away, not later." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'later' but immediately corrects themselves with 'That should be right away, not later.' This constitutes a revision where the final intent is to send the emails right away, which matches the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Send all my uh, let me think, future scheduled emails on Gmail later. That should be right away, not later." }, "candidate_id": "wav_d4e41eecf71ed163", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a filler 'uh' and 'let me think', which is realistic, but the correction from 'later' to 'right away' is delivered as a separate, abrupt statement rather than a fluid self-repair. The shift in intonation feels more like a post-hoc correction than a natural thought process, and the overall delivery is slightly less cohesive than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook right away. Change the app to Gmail, not Outlook.", "long_pause_offsets": [ 59 ], "source_sha256": "88a1d3244c92b58bc146f928331f153c58dd514747c462a9b967bbc71fcc5774", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Outlook to Gmail, preserving the action, scope, and urgency." }, "audio": { "status": "pending", "spoken_utterances": [ "Send all my ah... future-scheduled emails on let me think... Outlook right away. ", "Change the app to uh... Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "6a5e690_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_016", "hide_variant_original": true, "source_sha256": "0fdd17f481aafaa033dbb5761502bf9ed1a2bc82a5752527a7a0a2fc45eeb0cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/cancellation/cancel_016/fcd5145c3acc0d60/speech.wav", "wer": 0.0, "wav_sha256": "afd22bded9b1685aa351ea96403dcacbae398a19c22c11d6008c7c77bdbc9d44", "seconds": 10.68, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away.", "Hold on, you know what? Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues the task to send all future-scheduled emails on Gmail, then immediately corrects themselves by saying 'Hold on. You know what? Cancel that.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. Therefore, the final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Send all my future scheduled emails on Gmail right away. Hold on. You know what? Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "871ff112f63674c12fa516cd7adf01fbcef86589f628b8f74109d126893198d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/ambiguous_halt/halt_031/c477fcc29f56a996/speech.wav", "wer": 0.0, "wav_sha256": "37ba48f6160ce242fc33ffa7ac9c4d79cbe9a3723f8c967bb9f16af3c1aec571", "seconds": 9.44, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or hold. The final interjection indicates a temporary halt, not a cancellation or a new instruction. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause.", "heard_transcript": "Send all my future scheduled emails on Gmail right away. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3ecef42de2604667d6406be2709ad5fa20a9072b7b6c5cefc7536d4be7f1276c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_1/audio_natural/filler/original/d524167b16f5877e/speech.wav", "wer": 0.0, "wav_sha256": "aaf98f057adb4cf0c4edd14311deaeff3a8873749318226c3e59d3b225828e31", "seconds": 5.76, "spoken_utterances": [ "Send all my ... future-scheduled emails on ... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Send all my future scheduled emails on Gmail right away.' There are no hesitations, fillers, or self-corrections that alter the task. The final intent matches the original instruction, and the 'filler' type allows for minor hesitations or backchannels that do not change the goal. Since no such elements are present, the task is executed as stated.", "heard_transcript": "Send all my future scheduled emails on Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook—sorry, I should have said Gmail right away.", "long_pause_offsets": [], "source_sha256": "40b3935b3f45195f71e4b5b0cd8097de2e83e61afa8cd3ded61f0b7bb0d1ccc2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, scope, and urgency. The final intended app is Gmail." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/false_start/false_start_02_app/4a88ca0fa2b70803/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "9093304fb2120ca2a1054dfb181612754ff7faf49f95eecc7ec72e55d4202e69", "seconds": 8.56, "spoken_utterances": [ "Send all my ... future-scheduled emails on Outlook—hmmm... sorry, I should have said ... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Send all my future scheduled emails on Outlook. Sorry, I should've said Gmail right away." }, "candidate_id": "wav_e592307641e8ef8e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request, then pauses with a thoughtful 'hmm' before a natural self-correction. The correction is smoothly integrated with 'sorry I should have said' and a clear rephrasing of the intended task. The prosody and pacing feel spontaneous and conversational, with a believable moment of reconsideration." } ] }, "is_default": true }, { "variant_id": "false_start_01_email_state", "proposed_instruction": "Send all my scheduled—I meant future-scheduled emails on Gmail right away.", "long_pause_offsets": [], "source_sha256": "4a67e7799688d31814425e9abd555e1da658a919251730b8a316dbc3c54daf3c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'scheduled' to 'future-scheduled' within the same utterance, clarifying the intended scope of emails to be sent. The app (Gmail) and action (send right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/false_start/false_start_01_email_state/85f5bfa97c011669/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "21c57a513d9609eb86835b55462d36699fc0b2e6b077125393bbbf7acd5b6616", "seconds": 8.68, "spoken_utterances": [ "Send all my scheduled—uh... I meant future-scheduled emails on uh, let's see... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'scheduled emails' but immediately self-corrects to 'future scheduled emails'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send all my scheduled, uh, I meant future scheduled emails on, uh, let's see, Gmail right away." }, "candidate_id": "wav_82b685b92f074dcd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a false start ('scheduled'), self-corrects to 'future scheduled', and then hesitates with 'uh' and 'let's see'. While the repair is plausible, the 'let's see' filler feels slightly more like a planning pause than a genuine moment of recall, and the overall delivery is a bit more fragmented than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later—no, I mean right away.", "long_pause_offsets": [], "source_sha256": "b1f17cfaf6df0d5e7484a200bcc640198cc68f070f229edf484bd8bb89499ebe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'later' but immediately corrects with 'I mean right away', replacing the time instruction. The app (Gmail), subject (future-scheduled emails), and action (send) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/false_start/false_start_03_execution_timing/55b6a393c837b9a5/speech.wav", "wer": 0.0625, "wav_sha256": "c4c576398c22ef3527dd593028bb01d8aef25c014b4a8d6ca1446b37827b2add", "seconds": 8.72, "spoken_utterances": [ "Send all my future-scheduled emails on Gmail ... later—ah... no, I mean uh... right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Later up' as a false start, then immediately corrects with 'No, I mean uh, right away.' The final audible intent is 'right away', which corrects the initial mistaken value. The type is 'false_start', so the immediate self-correction is applied. The final goal matches the original instruction with the corrected timing.", "heard_transcript": "Send all my future scheduled emails on Gmail. Later up. No, I mean uh, right away." }, "candidate_id": "wav_c4a33a4104c39dfc", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then immediately self-corrects from 'later' to 'right away'. However, the correction is abrupt and lacks the natural hesitation or filler that would make it sound like a spontaneous thought process. The delivery is very direct and almost robotic, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future-scheduled.", "long_pause_offsets": [ 40 ], "source_sha256": "8c8860353e97f886a01a020700735667e43fdb92c5c3a401083edea1ad740c17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'future-scheduled', which narrows the initial broad request to only those emails. The app (Gmail), action (send), and timing (right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/addition/addition_01/717d53e3c09c09e5/speech.wav", "wer": 0.0, "wav_sha256": "c93974f48fbe3746da596ea5963c7aef4c3edadd15b9b97d27bfe6fb922cd78d", "seconds": 10.96, "spoken_utterances": [ "Send all my emails on ... Gmail right away. ", "Oh, I forgot to mention, only send emails that are ... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to send all emails on Gmail right away. After a pause and an interjection, the speaker added a constraint to only send 'future scheduled' emails. This is a clear addition that narrows the scope of the original request. The final intent is to execute the original task with the added constraint, matching the expected behavior for an 'addition' type.", "heard_transcript": "Send all my emails on Gmail right away. Oh, I forgot to mention, only send emails that are future scheduled." }, "candidate_id": "wav_5f2a56cf9d080fe2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial command with natural prosody. After a brief pause, they add a restriction ('only send emails that are future scheduled') in a way that sounds like a spontaneous afterthought, with a slight shift in intonation and a natural, conversational tone. The flow feels like a real person remembering a detail mid-sentence, making it highly realistic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send all my future-scheduled emails on right away. I should also mention, do this on Gmail.", "long_pause_offsets": [ 51 ], "source_sha256": "564b31e2d4d05295fa3066e63b43e61391c373caa39a4a203a90c49f4fa44050", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the app as Gmail, which was missing in the initial utterance. The action, scope (future-scheduled emails), and timing (right away) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/addition/addition_02/6a0f0969aabcaa2c/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "03b1e0c87998f67e6dea0666461104ff125297b481f7341728a88928c413c8a1", "seconds": 11.0, "spoken_utterances": [ "Send all my um... future-scheduled emails on right away. ", "I should also mention, do this on ... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('on right away'), but the speaker immediately adds the missing context ('do this on Gmail') after a longer pause, which is the defining characteristic of an 'addition' type. The final combined goal matches the original instruction exactly, including the action, target, and condition. The initial incompletion is resolved by the addition, so the final intent is clear and executable.", "heard_transcript": "Send all my um future scheduled emails on right away. I should also mention do this on Gmail." }, "candidate_id": "wav_e32f9e1f9b0ce8ae", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant 'um' and delivers the request in a more deliberate, almost rehearsed manner. The addition ('I should also mention do this on Gmail') sounds like a later correction or clarification rather than a natural afterthought. The intonation is flatter and less conversational compared to audio_01, making it feel less spontaneous and slightly less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later. That should be 'right away', not 'later'.", "long_pause_offsets": [ 52 ], "source_sha256": "5935e45afbd045ba484e1be094f8ee5c5c254f25e263a24ef3ecc46cbf3e09bb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time instruction from 'later' to 'right away', while retaining the app (Gmail), action (send), and target (all future-scheduled emails)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/revision/revision_03_execution_timing/31d9bdb72f64c983/speech.wav", "wer": 0.0, "wav_sha256": "fd26ff4ab490dc317ec0779d6df27da11620050b21cdc7959d75050de036b9d4", "seconds": 12.04, "spoken_utterances": [ "Send all my future-scheduled emails on um, let's see... Gmail later. ", "That should be '... right away', not 'later'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail later' but immediately corrects themselves with 'That should be right away, not later.' This constitutes a revision where the final intent is to send the emails right away, which matches the original instruction. The correction is clear and unambiguous, and the final control intent is to execute the revised task.", "heard_transcript": "Send all my future scheduled emails on um, let's see, Gmail later. That should be right away, not later." }, "candidate_id": "wav_5727b6f4646ccee0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow. The speaker begins with a clear intent, uses a plausible filler ('um') and a self-correction ('let's see'), then smoothly revises the original request from 'Gmail later' to 'right away'. The correction is integrated into the same utterance with natural intonation and pacing, showing a believable moment of reconsideration without any robotic or disjointed transitions." } ] }, "is_default": true }, { "variant_id": "revision_01_email_state", "proposed_instruction": "Send all my draft emails on Gmail right away. Change the email state from 'draft' to 'future-scheduled'.", "long_pause_offsets": [ 46 ], "source_sha256": "c5b4e133e498a69f91c2f36a8dca431d57b7330920ff0306b1df7112dc1cc944", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the email state condition from 'draft' to 'future-scheduled', which reverses the original intent of sending future-scheduled emails. The final goal is to modify the state of emails, not to send them." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/revision/revision_01_email_state/8f6052e3f6f35346/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "4df4f4bee8a8c199bcd8796d7d0b705a5e019c1706abef4667b1811d501040ea", "seconds": 12.12, "spoken_utterances": [ "Send all my uh... draft emails on uh, let me think... Gmail right away. ", "Change the email state from 'draft' to 'future-scheduled'." ] }, "audio_semantic": { "status": "different", "reason": "The original instruction was to send all future-scheduled emails. The speaker initially misstated the source state as 'draft' but then corrected it to 'future scheduled' in the revision. The final instruction is to change the state from draft to future scheduled, which is a different action than sending. This revision changes the core action from 'send' to 'change state', violating the revision criterion that the correction must apply to the earlier instruction. The final intent is not aligned with the original task.", "heard_transcript": "Send all my uh draft emails on uh let me think Gmail right away. Change the email state from draft to future scheduled." }, "candidate_id": "wav_04946c06c9c7326a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a realistic speaker formulating a complex request with multiple components. The speaker uses natural fillers ('uh', 'let me think') and pauses, and the request evolves from sending draft emails to changing their state. The second part of the utterance is a clear addition rather than a correction, which is plausible. However, the delivery is slightly more hesitant and less fluid than audio_01, with a longer pause and a less seamless transition between the two parts of the request." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook right away. Use Gmail instead of Outlook for sending the emails.", "long_pause_offsets": [ 59 ], "source_sha256": "ccb4c5e993fcbdfe27cb7700a7c1f26c2992329fa1bf35ef979a28e74a953108", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email platform from Outlook to Gmail, preserving the action, scope, and urgency." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/revision/revision_02_app/0ae1425c94c220e2/speech.wav", "wer": 0.04, "wav_sha256": "d80ef1362c8bd641e4e6ea7c6eab60d228e9dc5f6e026bfc0e841042339d1803", "seconds": 14.32, "spoken_utterances": [ "Send all my ... future-scheduled emails on let's see... Outlook right away. ", "Use uh, let me think... Gmail instead of Outlook for sending the emails." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but then corrects to 'Gmail' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with no other conflicting or missing constraints.", "heard_transcript": "Send all my future scheduled emails on let's see, Outlook right away. Use uh let me think, Gmail instead of Outlook for sending the emails." }, "candidate_id": "wav_f5c70e0e06cd1b2d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with a clear request but then introduces a revision ('use uh let me think Gmail instead of Outlook') that feels artificially inserted. The phrase 'let me think' is used as a filler but is immediately followed by a correction that doesn't naturally follow the prior sentence. The prosody is less continuous, and the revision sounds like a mechanical replacement rather than a spontaneous change of mind, making it less believable as a real-time spoken request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "86ddf384a238caa83338c9775e030a23841962c235f68f74d01e30f59d9e637a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/cancellation/cancel_024/4589b7deb9ee15ab/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "0bf42247d69871c481c5cbe2401b954af67ccb0238b8ad2fd97b0bb3f33b0efe", "seconds": 13.16, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away.", "Oh, wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to send scheduled emails, but immediately corrects themselves with 'Oh wait, oh never mind, cancel the whole thing.' This constitutes a clear cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away. Oh wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "12627eafac1b42f7308b386d51b0771619bcffe1ae6db281c36a237f8bfe5f41", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/ambiguous_halt/halt_036/c760b059785a389c/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d7e37477c9624e825a51a29daaac7468c0b7cd312bf692a5590c34298d59b70d", "seconds": 12.4, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second, before you go ahead,' which is a clear request to pause the task. This matches the 'ambiguous_halt' type, where the final intent is to halt execution temporarily. The initial instruction to send emails is not canceled, but the speaker explicitly requests a pause before proceeding.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "519e494ae9064ebcdb5613f44d491dddd5ed1e60eb817442dbc3c75f25b32e35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_2/audio_natural/filler/original/529301c9ecc21dfe/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "f3283c3efd59aad61ef793589be10adc5fff87908342635dd364328c89a66618", "seconds": 8.08, "spoken_utterances": [ "Send all my uh... future-scheduled emails on um, let's see... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core task. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored, and the remaining task meaning must match the original instruction. The final intent is clear and matches the reference.", "heard_transcript": "Send all my uh future scheduled emails on um let's see Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook—sorry, I meant Gmail right away.", "long_pause_offsets": [], "source_sha256": "3a7cec03f8df2405b2334e2fad0c88b06f9429f9c31a2637f9ce4b9e57c32f81", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the action, scope, and urgency." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/false_start/false_start_02_app/e3f23fc2c2011317/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a9dc60ee3b0e20ebcf01365d18deb28ed5a7f8a36357aedd3b1af5abb2e33120", "seconds": 9.88, "spoken_utterances": [ "Send all my um, let me think... future-scheduled emails on Outlook—um... sorry, I meant Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail' after a pause and apology. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Send all my um, let me think, future scheduled emails on Outlook um, sorry, I meant Gmail right away." }, "candidate_id": "wav_4385a5b86a16c020", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and complex self-correction pattern. It begins with a filler ('um') and hesitation ('let me think'), then makes a clear false start ('Outlook'), immediately self-corrects to 'Gmail', and finally adds 'right away' as a follow-up clarification. The flow is continuous, with realistic pauses and intonation shifts that mimic genuine thought processing. The correction is embedded within the utterance, not a separate sentence, making it sound spontaneous and credible." } ] }, "is_default": true }, { "variant_id": "false_start_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later—no, I mean right away.", "long_pause_offsets": [], "source_sha256": "b1f17cfaf6df0d5e7484a200bcc640198cc68f070f229edf484bd8bb89499ebe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'later' but immediately corrects with 'I mean right away', replacing the time instruction. The app (Gmail), subject (future-scheduled emails), and action (send) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/false_start/false_start_03_execution_timing/658e3b906b3a5a82/speech.wav", "wer": 0.0, "wav_sha256": "4f6658cb0e6a3d3912a026f5d1bdfdd5ccaf02a99538d8481712e244979d62d9", "seconds": 7.92, "spoken_utterances": [ "Send all my future-scheduled emails on Gmail ah... later—um... no, I mean right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'later' but immediately self-corrects with 'no, I mean right away.' This is a clear false start where the final intent is the corrected version. The original instruction is 'Send all my future-scheduled emails on Gmail right away,' and the final audible intent matches this exactly. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Send all my future scheduled emails on Gmail. Ah, later. Um, no, I mean right away." }, "candidate_id": "wav_67401c7ba612165c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start with 'Gmail' followed by a self-correction to 'right away'. The speaker uses a natural 'uh' and 'no' to signal the correction, and the intonation reflects a genuine change of mind. However, the correction is slightly more abrupt and less integrated than in audio_03, and the initial phrase 'send all my future scheduled emails on Gmail' is delivered without hesitation, making the correction feel a bit more like a post-hoc edit than a fully spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "false_start_01_email_state", "proposed_instruction": "Send all my scheduled—or rather, future-scheduled emails on Gmail right away.", "long_pause_offsets": [], "source_sha256": "dff2289d2b723131b1a13abf1d404789c338c1bb7d058579905b3a7cd926278f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'scheduled' to 'future-scheduled' within the same utterance, clarifying that only emails scheduled for a future time should be sent. The app (Gmail) and action (send right away) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/false_start/false_start_01_email_state/6aa0a7903acdc16e/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "949e206602609e3cc7ed2d1e5d6b2f1e5c4aef8d5a722b6d5b73e744f8168ce6", "seconds": 12.0, "spoken_utterances": [ "Send all my scheduled—uh... or rather, future-scheduled emails on uh, let me think... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'scheduled' but immediately self-corrects to 'future-scheduled' with 'uh rather'. The final control intent is clear and matches the original instruction. The hesitation and filler words ('uh', 'let me think') do not alter the core task. The final intent is to execute the task as corrected.", "heard_transcript": "Send all my scheduled uh, uh rather future-scheduled emails on uh, let me think, Gmail right away." }, "candidate_id": "wav_467d97a56ae927a7", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains multiple fillers ('uh', 'rather', 'let me think') and hesitations, but the self-correction is less natural. The phrase 'send all my scheduled uh uh rather future scheduled emails on uh let me think Gmail right away' is fragmented and lacks smooth transitions. The repetition of 'uh' and the long pause before 'let me think' sound more like a stalling tactic than a genuine thought process. The correction is also less integrated, with 'Gmail right away' tacked on at the end rather than being part of a continuous thought flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Send all my emails on Gmail right away. Oh, right, these are only the emails that are future-scheduled.", "long_pause_offsets": [ 40 ], "source_sha256": "5515efc6be5c9fd57dbf96a57e7f33447e00edffb46769fb9e80f0e28b88655f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'future-scheduled' to the initial broad request 'all my emails', narrowing the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/addition/addition_01/f7ea983b8c0a865b/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "6e4787f913ebbf4514d71dbdcae7e0757986628a70170ab37fb4823f4f08d64e", "seconds": 14.52, "spoken_utterances": [ "Send all my emails on uh, let me think... Gmail right away. ", "Oh, right, these are only the emails that are uh, let's see... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Send all my emails on Gmail right away' but then corrects themselves with 'Oh right, these are only the emails that are future scheduled.' This is a clear addition that narrows the scope of the original instruction. The final intent is to send only the future-scheduled emails, which matches the original instruction with the added constraint. Therefore, the task is executed with the corrected scope.", "heard_transcript": "Send all my emails on uh let me think Gmail right away. Oh right, these are only the emails that are uh let's see future scheduled." }, "candidate_id": "wav_ec5a7a14fe596663", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous thought process with a believable 'let me think' pause, followed by a self-correction ('Oh right') and a clear, additive clarification ('These are only the emails that are... future scheduled'). The prosody is conversational, with appropriate intonation and pacing. The addition of the detail about 'future scheduled' emails feels like a remembered restriction, not a forced insertion, and the flow is continuous and pragmatic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Send all my future-scheduled emails on right away. I should also mention, do this on Gmail.", "long_pause_offsets": [ 51 ], "source_sha256": "564b31e2d4d05295fa3066e63b43e61391c373caa39a4a203a90c49f4fa44050", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the app as Gmail, which was missing in the initial utterance. The action, scope (future-scheduled emails), and timing (right away) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/addition/addition_02/4c830d2a65d54191/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d58f2eb589362b47cec8ef9946613d6822e80e4ef66c34e340382b24f26847f9", "seconds": 12.56, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on right away. ", "I should also mention, do this on let me think... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Send all my future scheduled emails on right away', which is incomplete and lacks the platform. After a pause and self-correction, the speaker adds 'do this on, let me think, Gmail', which clarifies the platform as Gmail. The final intent combines the original action with the added platform, matching the reference instruction. The addition type requires combining the earlier instruction with the added information, which is done correctly here.", "heard_transcript": "Send all my um future scheduled emails on right away. I should also mention do this on, let me think, Gmail." }, "candidate_id": "wav_7778c4f42b65d6c8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a filler ('um') and a 'let me think' pause, the overall delivery is less fluid. The initial request is somewhat fragmented ('send all my... future scheduled emails on right away'), and the addition ('I should also mention do this on... Gmail') sounds more like a pasted-on afterthought than a naturally flowing thought. The prosody is flatter, and the transition into the addition feels less integrated, making it sound less spontaneous than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_email_state", "proposed_instruction": "Send all my draft emails on Gmail right away. Wait, send all my draft emails on Gmail right away, not future-scheduled.", "long_pause_offsets": [ 46 ], "source_sha256": "1ac2be659e498a73bb555ff37ea91bfd094a62ef0b50f4bbd0f48922837870c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the action should apply to draft emails, not future-scheduled ones, which removes the original ambiguity and aligns with the final instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/revision/revision_01_email_state/2467679f8e593208/speech.wav", "wer": 0.0, "wav_sha256": "f8a12d00def460b3e39d66799d6a082a5ea20f1468fb1c9d04eb47d9d95b4158", "seconds": 15.96, "spoken_utterances": [ "Send all my uh... draft emails on um... Gmail right away. ", "Wait, send all my draft emails on Gmail right away, not ... future-scheduled." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'draft emails' but then corrects to 'not future scheduled', which clarifies the intended scope. The revision narrows the original instruction from 'future-scheduled emails' to 'draft emails' while excluding future-scheduled ones. The final intent is a clear revision of the original instruction, and the correction is applied as per the revision type criterion.", "heard_transcript": "Send all my uh draft emails on um Gmail right away. Wait, send all my draft emails on Gmail right away, not future scheduled." }, "candidate_id": "wav_03a119f8df112a33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a slightly mumbled phrase ('send all my uh'), then corrects to 'send all my draft emails on Gmail right away'. The correction is immediate and smooth, with a clear shift in intonation and pacing. The speaker then adds a clarifying revision ('not future scheduled') with a natural pause and a slightly lower pitch, indicating a self-correction. The prosody is consistent, and the flow feels spontaneous and conversational." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Send all my future-scheduled emails on Outlook right away. Let me correct the app: send all my future-scheduled emails on Outlook right away, not Gmail.", "long_pause_offsets": [ 59 ], "source_sha256": "2c1167438203555f8dedf5699e558382fc75695a6700e8120bbcaaaa1a654e42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Gmail to Outlook, changing the target application from the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/revision/revision_02_app/5b4f5036f3c27858/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "82b446b756a8ddde5ede7cc752960be0ac6c2b459caf5ecd0c7875ead428f6d0", "seconds": 19.0, "spoken_utterances": [ "Send all my ... future-scheduled emails on uh... Outlook right away. ", "Let me correct the app: send all my future-scheduled emails on Outlook right away, not um... Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail' but immediately corrects it to 'Outlook' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to send the emails on Outlook, which matches the expected control intent of executing the revised task.", "heard_transcript": "Send all my future scheduled emails on uh Outlook right away. Let me correct the app. Send all my future scheduled emails on Outlook right away, not um Gmail." }, "candidate_id": "wav_8f966a7ccf278cca", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear but slightly hesitant delivery ('send all my future scheduled emails on uh Outlook right away'), then explicitly pauses and self-corrects ('let me correct the app'). The correction is delivered with a natural shift in tone and pacing, and the revised sentence is spoken with confidence. The addition of 'not um Gmail' at the end is a plausible follow-up clarification. The overall delivery is realistic, though the explicit 'let me correct' makes it slightly less spontaneous than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_03_execution_timing", "proposed_instruction": "Send all my future-scheduled emails on Gmail later. Change the timing to 'later' instead of 'right away'.", "long_pause_offsets": [ 52 ], "source_sha256": "20f7f22b9de95ab5ddd375e8210fc9b6f3276430a617f729511d302fc3cdcd00", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly confirms the timing as 'later', replacing the original 'right away'. The action, app, and scope of emails remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/revision/revision_03_execution_timing/977c69dc1cc7aa1b/speech.wav", "wer": 0.1, "wav_sha256": "9ab36dc61795705ecaf0cad1e966bf6532159831ad85862340491d913bb5f879", "seconds": 14.2, "spoken_utterances": [ "Send all my um... future-scheduled emails on ... Gmail later. ", "Change the timing to 'later' instead of 'hmmm... right away'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'right away' but immediately corrects it to 'later' after a pause, which is a clear revision. The final intent is to send the emails with the timing set to 'later', matching the revised instruction. The original instruction is overridden by the correction.", "heard_transcript": "Send all my um future scheduled emails on Gmail later. Change the timing to later instead of um right away." }, "candidate_id": "wav_8665cada8543272d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a hesitant 'send all my um future scheduled emails on Gmail later', then explicitly states 'change the timing to later instead of um right away'. The phrasing 'change the timing to later instead of um right away' is more of a meta-instruction than a natural self-correction. The delivery is less fluid, with a noticeable shift in tone and a slightly robotic phrasing. The hesitation and filler words are present but feel less integrated into a spontaneous thought process compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6a5e690_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "736a85264a13200cf9561ac3fbac4f3409e7a975b1e8d75e0c5194170ca3a36d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/cancellation/cancel_024/ea627da849fe6319/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "bb0c6e0d530400ef18bb7f355ea48503ececb28882116dfdd85d70b2e99c374b", "seconds": 16.2, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on um, let's see... Gmail right away.", "Wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a request to send scheduled emails, but immediately corrects and cancels the entire task with 'Wait, oh never mind. Cancel the whole thing.' This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Send all my future scheduled emails on emails on um, let's see, Gmail right away. Wait, oh never mind. Cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6a5e690_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "363d952c28b2bb8b1605cfc8c2df8e61ecf1993209a2ca5dde07afb6c2b8bc12", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/ambiguous_halt/halt_021/97e8896eea2702fb/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "b1519efc66ef1b529706d46c1db62877761e28fce05e8c0d7c66da4f9f3477c9", "seconds": 15.28, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on um, let's see... Gmail right away.", "Wait, let me think..." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "6a5e690_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "88503b9932cdbc06f7ca16996b7b4e138423bddc4764fd6379b845489a039da8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6a5e690_3/audio_natural/filler/original/201f373f1592f364/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "068cb34f9bbffd7b79ed896ad00930eaa1e9bfbcf762c6a0656bd8b42543dfc7", "seconds": 11.16, "spoken_utterances": [ "Send all my hmmm... future-scheduled emails on um, let's see... Gmail right away." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the platform as 'emails' but immediately self-corrects to 'Gmail'. The phrase 'let's see' is a filler and does not alter the final goal. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Send all my future scheduled emails on emails on um, let's see, Gmail right away." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_reference", "proposed_instruction": "How many people follow the artist of the last played song—sorry, I meant currently playing song on Spotify?", "long_pause_offsets": [], "source_sha256": "f35346b96a952f048f28da97aa42341519decb15038514885f3b81efeb52a734", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately corrects to 'currently playing song'. The final intended query matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/false_start/false_start_01_song_reference/35dab78ab6a62a4d/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "da6e4a713efd73a605cdd8e529d04705d93728ab26e51a23ce38a59cb07b9976", "seconds": 10.2, "spoken_utterances": [ "How many people follow uh... the artist of the um... last played song—ah... sorry, I meant currently playing song on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately self-corrects to 'currently playing song'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved otherwise. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow uh the artist of the um last played song that, sorry, I meant currently playing song on Spotify?" }, "candidate_id": "wav_937d012999e6e825", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous false start with 'last played song' followed by a self-correction to 'currently playing song'. The filler words 'uh' and 'um' are placed at plausible points of hesitation, and the correction flows smoothly without awkwardness. The prosody is consistent, and the overall delivery sounds like a genuine, unscripted attempt to formulate the request." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music—no, that should be Spotify?", "long_pause_offsets": [], "source_sha256": "2b21fadb87c72b9cda40c0e8807133b2e608a9bdbc6213df5969f1732a817b2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/false_start/false_start_02_app/0c42284b5648c159/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "46c84e33e3e5b21a7619ccbca2dd80e9ede3aebffd6f7413652f437398762431", "seconds": 11.08, "spoken_utterances": [ "How many people follow the artist of the currently playing song let me think... on Apple Music—um... no, that should be ah... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'On Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow the artist of the currently playing song? Let me think. On Apple Music. Um, no, that should be uh Spotify." }, "candidate_id": "wav_b3911979d7944ca9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While this clip also contains a false start and correction ('Apple Music' to 'Spotify'), the correction is less fluid. The speaker says 'No, that should be uh Spotify' with a slightly unnatural, hesitant tone and a longer pause before the correction. The prosody is less continuous, and the correction feels more like a mechanical switch than a natural thought process, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to mention, it's about the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "f1734efda5df0b51f6412b1336bd06d73b6924db7eb0487a9c5ef1fe95b6b87f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context 'of the currently playing song' to the initial question, completing the reference to the correct artist. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/addition/addition_01/56385a454ad0d202/speech.wav", "wer": 0.04, "wav_sha256": "497896e09aef2c147737f81a588e1bedd45d23fbfb5e7cc6d262dcdea54e4c64", "seconds": 13.8, "spoken_utterances": [ "How many people follow let's see... the artist on Spotify? ", "Oh, I forgot to mention, it's about the artist of the ah... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the artist on Spotify, then adds a clarifying detail about the 'currently playing song' after a pause. This matches the 'addition' type, where the final goal combines the initial request with the added information. The final intent is fully aligned with the original instruction.", "heard_transcript": "How many people follow, let's see, the artist on Spotify? Oh, I forgot to mention, it's about the artist of the uh currently playing song." }, "candidate_id": "wav_dfb0e4a71fd83dda", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many people follow the artist of the currently playing song? I left out a detail: do this on Spotify.", "long_pause_offsets": [ 65 ], "source_sha256": "4da2e0dd1bee948ab40c78e62ab709bf576589460f0fc7503733489ba64db5b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify', which was omitted in the initial utterance but is present in the original instruction. The core question and all other conditions remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "How many people follow let me think... the artist of the currently playing song? ", "I left out a detail: do this ah... on Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_song_reference", "proposed_instruction": "How many people follow the artist of the latest released song on Spotify? Wait, I meant the artist of the currently playing song, not the latest released song.", "long_pause_offsets": [ 74 ], "source_sha256": "b57e284e4313fc6138f1733a7cf24376935da64cd5ff7ed8f4f65369f2851b50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'latest released song' to 'currently playing song', preserving the rest of the query including the action and app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/revision/revision_01_song_reference/0f348da374122483/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "9860ec223f6c8be0115972c194ec20d8e94c7afe7d7bee021660583802cc27dd", "seconds": 14.68, "spoken_utterances": [ "How many people follow uh, let me think... the artist of the latest released song um... on Spotify? ", "Wait, I meant the artist of the currently playing song, not the latest released song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'latest release song' but immediately corrects themselves with 'Wait, I meant the artist of the currently playing song, not the latest release song.' This constitutes a revision where the final intent is the corrected version. The original instruction matches this final intent, so the task is executed as revised.", "heard_transcript": "How many people follow uh let me think the artist of the latest release song um on Spotify? Wait, I meant the artist of the currently playing song, not the latest release song." }, "candidate_id": "wav_1b8fc9b7ffc51537", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh, let me think'), formulates a slightly incorrect initial request ('latest release song'), then self-corrects with a plausible revision ('Wait, I meant the artist of the currently playing song, not the latest release song'). The correction is integrated smoothly into the same utterance with a believable shift in intonation and emphasis, demonstrating a realistic thought process and repair." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music? Let me correct the app: the question should be about Spotify, not Apple Music.", "long_pause_offsets": [ 80 ], "source_sha256": "4b030d7c037d55342f4c74d12286c1e218a573c41c743dbc58ca877a98e95b37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, aligning with the original instruction. All other components of the query remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/revision/revision_02_app/903d3e06c5e6f648/speech.wav", "wer": 0.0, "wav_sha256": "c50858ab49d776ad896929dd131180573de243c41d823bc1e6164e672d282159", "seconds": 14.04, "spoken_utterances": [ "How many people follow the artist of the currently playing song ... on Apple Music? ", "Let me correct the app: the question should be about ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects the app to Spotify. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent is to query Spotify, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints are altered.", "heard_transcript": "How many people follow the artist of the currently playing song on Apple Music? Let me correct the app. The question should be about Spotify, not Apple Music." }, "candidate_id": "wav_608ca4d06540a999", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a clear initial request, then pauses and corrects the platform from 'Apple Music' to 'Spotify'. While the correction is natural, the phrasing 'Let me correct the app' is slightly awkward and less idiomatic than a spontaneous self-correction. The overall delivery is fluent but lacks the organic hesitation and self-repair flow of audio_01, making it slightly less realistic in terms of spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "003c3a608518a0a331d084f7d1a3de7256d6338dcd80ecc2458ee48d987d1434", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/cancellation/cancel_026/e88a2f69855ef378/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "5ba3d3b2af13d879b138483e537b794768c8eafffdcaf7628649b3bbdbda5dc1", "seconds": 11.56, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?", "Hold on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial request. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify? Hold on, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "f703e1c35d678752a396dc125a5a5a3536f3d614bd5075b630c084abf205e5a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/ambiguous_halt/halt_034/66f011b93e69cd83/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d0095f6503bde23b677d58245da046c38e45c4f1944497050ca7c40c021afb92", "seconds": 10.44, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or resumption. The initial request is not executed, but the intent to pause is audible and final.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify? Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e44e996850acbcf17a821dafc35733a1aba600bc27a648a90ab85b4d347582fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_1/audio_natural/filler/original/07b30797f8b164f1/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "5118bfef84536ac5b5561a2a5c11b58ee30dcd8fb0564c97c53f2483a87cb527", "seconds": 6.4, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song ah... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker hesitated with 'uh' but did not change the core request. The final intent matches the original instruction, and the hesitation is a filler, not a change in control. The task is to execute the original request.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music—no, that should be Spotify?", "long_pause_offsets": [], "source_sha256": "2b21fadb87c72b9cda40c0e8807133b2e608a9bdbc6213df5969f1732a817b2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/false_start/false_start_02_app/9eae7d234ec4c6bc/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "54437956d65fa25b074c6208c7b6c15df950e8be5b085c1c49469aa3b7997acd", "seconds": 9.16, "spoken_utterances": [ "How many people follow uh... the artist of the currently playing song on Apple Music—um... no, that should be uh... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow uh the artist of the currently playing song on Apple Music? Um, no, that should be uh Spotify." }, "candidate_id": "wav_01d0d93cc226e46d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then self-corrects 'Apple Music' to 'Spotify' with a plausible false start and immediate repair. The intonation and pacing are consistent with spontaneous speech, and the correction flows smoothly within the same utterance without awkwardness or hesitation." } ] }, "is_default": true }, { "variant_id": "false_start_01_song_reference", "proposed_instruction": "How many people follow the artist of the last played song—sorry, I mean currently playing song on Spotify?", "long_pause_offsets": [], "source_sha256": "bb6284b885ddabb1d04d1f589cb5957fb6fca9246a3e3fef28dfd467179d6c34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately corrects it to 'currently playing song'. The final intended query matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/false_start/false_start_01_song_reference/800c2009bda8cb7b/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "09cff1253ae81db6bf85bfd4f9742cb40ed9f01520c528cdd1e36df12b2fed9e", "seconds": 11.52, "spoken_utterances": [ "How many people follow the artist of the hmmm... last played song—uh... sorry, I mean currently playing song hmmm... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'whom last played song' but immediately self-corrects to 'currently playing song'. The final intent matches the original instruction, and the correction is immediate and clear. The 'whom' is likely a grammatical error but does not change the meaning of the query. The task is executed as intended.", "heard_transcript": "How many people follow the artist of the whom last played song, sorry, I mean currently playing song whom on Spotify?" }, "candidate_id": "wav_779f7785133b0703", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear question but then makes a revision ('last played song' to 'currently playing song') after a pause. While the correction is natural, the initial phrase 'the whom' is grammatically incorrect and sounds like a slip of the tongue, which is less realistic than a full sentence with a minor factual correction. The overall delivery is slightly less fluent than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to say, it's the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "9727f2c5b16ae95768910e3d2ea734c95ef89a6d2a54743ba0fcc95ab1b906d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specification 'of the currently playing song' to the initial broad request about the artist, completing the intended target." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/addition/addition_01/ac10c0f368885605/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "1f68bfad70c352f630f5ab1c23ed36761ed001334c81319007173ca23906f66f", "seconds": 11.52, "spoken_utterances": [ "How many people follow the artist uh... on Spotify? ", "Oh, I forgot to say, it's the artist of the ... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, missing the crucial qualifier 'of the currently playing song'. The speaker then adds this missing constraint after a pause, completing the original instruction. The addition is clear and directly resolves the ambiguity, resulting in a complete and executable task that matches the reference. The type is 'addition', and the final intent is fully aligned with the original instruction.", "heard_transcript": "How many people follow the artist uh on Spotify? Oh, I forgot to say, it's the artist of the currently playing song." }, "candidate_id": "wav_91a2b83002b33cbc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then hesitates with 'uh' before continuing. The addition of 'Oh I forgot to say it's the artist of the currently playing song' is delivered with a believable, slightly self-correcting intonation and a natural pause before the clarification. The flow feels spontaneous and conversational, with the filler and follow-up fitting seamlessly into the task formulation." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many people follow the artist of the currently playing song? I forgot to add, this should be on Spotify.", "long_pause_offsets": [ 65 ], "source_sha256": "e623876cd1d5ceef1db9e323bed38ee805b11c589d2695317fd0f011aaad7254", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app requirement 'on Spotify', which was omitted in the initial utterance but is present in the original instruction. The core question about follower count for the artist of the currently playing song remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/addition/addition_02/3a9247e0070a1d25/speech.wav", "wer": 0.0, "wav_sha256": "050595c26e9022edc158f7257d8179bcb86bc666c9027ac4f087e4833870826d", "seconds": 10.6, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song? ", "I forgot to add, this should be ... on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is followed by a clear addition specifying the platform as 'on Spotify'. The addition is not a correction of a prior error but a clarification that narrows the scope of the original request. The final intent combines both parts, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "How many people follow the artist of the currently playing song? I forgot to add, this should be on Spotify." }, "candidate_id": "wav_5c42bc071531ef32", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is spoken fluently and clearly. However, the follow-up 'I forgot to add this should be on Spotify' sounds slightly more rehearsed or less naturally integrated. The phrase 'this should be' introduces a minor awkwardness in phrasing, and the correction feels a bit more abrupt and less conversational than in audio_01. The prosody is still natural, but the overall delivery lacks the subtle hesitation and organic flow of the top-ranked clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_song_reference", "proposed_instruction": "How many people follow the artist of the latest played song on Spotify? Wait, I meant the artist of the currently playing song, not the latest played song.", "long_pause_offsets": [ 72 ], "source_sha256": "c1b8758e5f734e55359b62ab59e2de1b9721ade8a7b72eaa30ed46501dbfb70f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'latest played song' to 'currently playing song', which is the same as the original instruction. The correction clarifies the intended reference, and the final goal matches the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/revision/revision_01_song_reference/5e6231a0952dc391/speech.wav", "wer": 0.0625, "wav_sha256": "0bb5612e3ed25a2d62d22b9d65054e4e1dc9e9fe379900dca631f53956a0d120", "seconds": 14.4, "spoken_utterances": [ "How many people follow hmmm... the artist of the latest played song on Spotify? ", "Wait, I meant the artist of the let me think... currently playing song, not the latest played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'latest played song' but immediately corrects themselves to 'currently playing song' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to ask about the currently playing song, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "How many people follow, um, the artist of the latest played song on Spotify? Wait, I meant the artist of the, let me think, currently playing song, not the latest played song." }, "candidate_id": "wav_07013d997d225ad3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' and 'wait,' followed by a self-correction that is smoothly integrated into the flow. The revision from 'latest played song' to 'currently playing song' is delivered with a believable change-of-mind intonation and a brief pause ('let me think') that sounds spontaneous. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music? Let me correct the app: check the followers on Spotify, not Apple Music.", "long_pause_offsets": [ 80 ], "source_sha256": "289432f6da39483e20a9d378b9a46b02be223ad71fe76bb67285b887384c9b83", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/revision/revision_02_app/6fc1a2ab55f1700b/speech.wav", "wer": 0.03125, "wav_sha256": "28ffe197823193dcd7a8933fc07064f5c5098f4154238538b01b59151286a40f", "seconds": 14.36, "spoken_utterances": [ "How many people follow let's see... the artist of the currently playing song um, let's see... on Apple Music? ", "Let me correct the app: check the followers on Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the follower count on Apple Music but immediately corrects themselves to Spotify. This is a clear revision where the final intent is to execute the task on Spotify, matching the original instruction. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "How many people follow, let's see, the artist of the currently playing song, um, let's see, on Apple Music. Let me correct the app, check the followers on Spotify, not Apple Music." }, "candidate_id": "wav_a891dce560fdf8dc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar structure but includes more fillers ('let's see') and a slightly more disjointed flow. The initial request is interrupted by 'let's see,' then the correction is delivered with a second 'let's see' and a more abrupt switch from Apple Music to Spotify. The prosody is less fluid, and the repetition of 'let's see' feels less natural and more like a filler loop. The correction is clear but less seamlessly integrated than in audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "0827f2d32123c9ad4b81f335a8535bd15c1e1465c5980912a602c1662cd3ed58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/cancellation/cancel_009/7679f91f7d33b552/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "4369f2b771b5cd20f7b017fcaad2d8f757db3f130ee7ba293b63035d587de01f", "seconds": 9.84, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?", "Oh, wait, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately follows with 'Oh wait, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify? Oh wait, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "08175044958ad991186731486f312dc6743eb258e67daaac8f8dcdb9f37f24e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/ambiguous_halt/halt_039/a36d947bd5bf6195/speech.wav", "wer": 0.05, "wav_sha256": "7bd905f04512a76725322850c074ca7ab6840036cc279b17d568b1e7c4ea9113", "seconds": 9.4, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, give me a second first,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial request is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify? Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "58cfa44345fee02b95be3551e1b2976cba6c6ec4d9b892305ec7df5068adec07", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_2/audio_natural/filler/original/4b845533ea363ba7/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "30ee3b0aa7819a0bc952d0ef74126d2ff688f68b15ee42a7d3962bcbfb622984", "seconds": 5.36, "spoken_utterances": [ "How many people follow um... the artist of the currently playing song ... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'How many people follow' and hesitates with 'um' before continuing the original instruction. The hesitation and filler word do not alter the task goal. The final intent matches the original instruction, and the control intent is to execute the task.", "heard_transcript": "How many people follow um the artist of the currently playing song on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music—sorry, that should be Spotify?", "long_pause_offsets": [], "source_sha256": "168d93cc7ea2a47663e0e06f07d9172ecf4c56b2ceb5070261e69fc2f7f9e603", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the rest of the query including the action, subject, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/false_start/false_start_02_app/54c40f2477776c74/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "2f53cde40c13587e990d15c294871620ef96ba8203d23599fbdae715dadd5463", "seconds": 9.92, "spoken_utterances": [ "How many people follow the artist of the currently playing song uh... on Apple Music—um... sorry, that should be Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Apple Music um sorry, that should be Spotify." }, "candidate_id": "wav_ea4e83ff0e53134d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a clear and natural false start with 'uh' and 'um' fillers, a plausible self-correction ('sorry that should be Spotify'), and a smooth, continuous flow. The prosody is conversational, with appropriate pauses and intonation that suggest spontaneous speech rather than a rehearsed or robotic delivery. The correction is integrated naturally into the utterance, making it highly realistic." } ] }, "is_default": true }, { "variant_id": "false_start_01_song_reference", "proposed_instruction": "How many people follow the artist of the last played song—I should say currently playing song on Spotify?", "long_pause_offsets": [], "source_sha256": "0f4e6918c5b8e758b0906a0a4e08a7836fc231cf89962d44f95baef0a5e6bf43", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately corrects to 'currently playing song'. The final intended query matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/false_start/false_start_01_song_reference/7782badb43a0f8fc/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "857e7bef0a3f1d1eb42ec533780625ff629ac8c7586f3925bb8504373a40fe2c", "seconds": 10.8, "spoken_utterances": [ "How many people follow um... the artist of the last played song—ah... I should say currently playing song on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last played song' but immediately self-corrects to 'currently playing song' in the same utterance. According to the 'false_start' type criterion, the immediate self-correction must be applied, and the mistaken value discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "How many people follow um the artist of the last played song? I should say currently playing song on Spotify." }, "candidate_id": "wav_e3f7bb3a36d7113f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also shows a false start with 'um' and a self-correction ('I should say currently playing song on Spotify'), but the correction is delivered as a separate sentence rather than being integrated into the original phrase. The prosody is slightly less fluid, and the correction feels more like a post-hoc addition than a seamless repair. While still realistic, it lacks the natural continuity of audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many people follow the artist on Spotify? Oh, I forgot to mention, it's the artist of the currently playing song.", "long_pause_offsets": [ 46 ], "source_sha256": "c4c1a0a30db3517dc0c5282972744737025f8f3647dc68c5ec2b5b2a32691d13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specification 'of the currently playing song' to the initial question, completing the reference to the correct artist. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/addition/addition_01/c2444c45ade2c956/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "f2ef2bd852db2d83601c433eb3565aab20715ded2b399981c2893625c5a18b48", "seconds": 15.36, "spoken_utterances": [ "How many people follow let's see... the artist uh... on Spotify? ", "Oh, I forgot to mention, it's the artist of the ... currently playing song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the number of followers of 'the artist' on Spotify, then adds the clarifying detail 'it's the artist of the currently playing song' after a longer pause. This is a clear addition that narrows the scope of the initial request. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many people follow, let's see, the artist uh, on Spotify? Oh, I forgot to mention, it's the artist of the currently playing song." }, "candidate_id": "wav_45a0855307f83c78", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How many people follow the artist of the currently playing song on Apple Music? Let me correct the app: the question should be about Spotify, not Apple Music.", "long_pause_offsets": [ 80 ], "source_sha256": "4b030d7c037d55342f4c74d12286c1e218a573c41c743dbc58ca877a98e95b37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, aligning with the original instruction. All other components of the query remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/revision/revision_02_app/fd97a8c1588792a6/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "e0d8485200922d740716a2ee8159d9b10614a2475be94d87187e60c3c140d6c8", "seconds": 13.68, "spoken_utterances": [ "How many people follow ... the artist of the currently playing song uh... on Apple Music? ", "Let me correct the app: the question should be about Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects the app to Spotify. This is a revision where the final intent is to query Spotify, matching the original instruction. The correction is clear and the final goal is fully specified.", "heard_transcript": "How many people follow the artist of the currently playing song uh on Apple Music. Let me correct the app. The question should be about Spotify, not Apple Music." }, "candidate_id": "wav_a1844f550f4d652d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects the app name from 'Apple Music' to 'Spotify' in a natural, conversational manner. The correction is smoothly integrated with a filler word ('uh') and a clear, audible rephrasing ('Let me correct the app'). The prosody is consistent, and the pacing reflects genuine thought processing, making the revision sound spontaneous and realistic." } ] }, "is_default": true }, { "variant_id": "revision_01_song_reference", "proposed_instruction": "How many people follow the artist of the latest released song on Spotify? Wait, I meant the artist of the currently playing song, not the latest released song.", "long_pause_offsets": [ 74 ], "source_sha256": "b57e284e4313fc6138f1733a7cf24376935da64cd5ff7ed8f4f65369f2851b50", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'latest released song' to 'currently playing song', preserving the rest of the query including the action and app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/revision/revision_01_song_reference/a1183eeb13a656ca/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "cc65a08e599bd7dfcb9ecc1f47eb3b65e0ff0508e2197d85b4ded2d2b676c478", "seconds": 15.4, "spoken_utterances": [ "How many people follow uh... the artist of the latest released song um... on Spotify? ", "Wait, I meant the artist of the uh... currently playing song, not the latest released song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'latest release song' but immediately corrects themselves to 'currently playing song' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to inquire about the artist of the currently playing song, which matches the original instruction.", "heard_transcript": "How many people follow uh the artist of the latest release song um on Spotify? Wait, I meant the artist of the uh currently playing song, not the latest release song." }, "candidate_id": "wav_c82b6e4e0ab4b95c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant phrasing ('uh', 'um') and then corrects the song type from 'latest release song' to 'currently playing song'. While the correction is plausible, the initial request is less fluent, and the correction is delivered with a slightly abrupt shift in intonation. The filler words are present but feel a bit more forced, and the overall flow is less smooth than in audio_01, making it slightly less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "9320583425994037af95ea31be1de123f6f05b589eac1e44de5a5765aac9418a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/cancellation/cancel_025/a3306c27fc13f000/speech.wav", "wer": 0.0, "wav_sha256": "baf69b5bae807009196705a07bfa281e508b6ad5d302682d784122e3f4daa088", "seconds": 14.6, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?", "Oh, wait, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking a question but then clearly revokes the entire request with 'Oh wait, forget that request, don't do it.' This constitutes a direct cancellation of the task. The final interjection is a complete and unambiguous cancellation, not a pause or a revision. Therefore, the control intent is correctly identified as 'cancel' and the verdict is 'pass'.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify? Oh wait, forget that request, don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "5bade18a62f260e91af90dce0f1af07d466f41f0b4d7ed80425c90d599cfaf46", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/ambiguous_halt/halt_016/5c20ff28cecf3727/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "a2e83e80a513f3a32c3498a2a2272d42538d4929ab9e231c012f42447fcb26b9", "seconds": 12.88, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a second jeez', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify just a second jeez" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6bdbc26_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1836c7063558f7f5129c7e02174520f444a309321684d34ef7c3b263b8143156", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6bdbc26_3/audio_natural/filler/original/fe280ec82867431c/speech.wav", "wer": 0.0, "wav_sha256": "a649383334417c13d645268a564ad2e5cfb6c0535e902ed5f8337675c5f3bca8", "seconds": 9.48, "spoken_utterances": [ "How many people follow uh, let's see... the artist of the currently playing song let me think... on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see', 'let me think') but does not change the core request. The final audible intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored, and the remaining task meaning must match the original instruction.", "heard_transcript": "How many people follow uh let's see the artist of the currently playing song let me think on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family and friends—no, I mean roommates and coworkers.", "long_pause_offsets": [], "source_sha256": "2869d11a62d3cef038a6d8803fc66cd17b84a7d874b8498da589246edfc2b419", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'family and friends' with 'roommates and coworkers'; the action, app, and request type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/false_start/false_start_04_requester_group/cf271c8a77dc40f0/speech.wav", "wer": 0.2631578947368421, "wav_sha256": "9caff59366e50c5fc9558ecbd85f042182cb5e19c2ea6a9c3bac4d1358274f03", "seconds": 10.52, "spoken_utterances": [ "Accept all pending ... Venmo payment requests from um... my family and friends—uh... no, I mean roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family and friends' but immediately self-corrects to 'roommates and coworkers'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Accept all pending Venmo payment requests from um my family and friends uh no, I mean roommates and coworkers." }, "candidate_id": "wav_a823d5558a10e6ae", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural false-start pattern. The speaker begins with 'family and friends', then self-corrects with 'no, I mean roommates and coworkers' using a believable shift in intonation and a clear, spontaneous-sounding revision. The initial phrase is fully formed, the pause before the correction is natural, and the correction flows smoothly without awkwardness. The prosody and rhythm are consistent with a person recalling and adjusting a detail in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Accept all pending CashApp—I should say Venmo payment requests from my roommates and coworkers.", "long_pause_offsets": [], "source_sha256": "99f188b2b3ce8b7c1cb020e371c55e80093c26b17d6371cf12240155cd29fa53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, scope of requests, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/false_start/false_start_03_app/39a434d4c59f5b90/speech.wav", "wer": 0.47058823529411764, "wav_sha256": "e655c2f2a20a111ec516ebb2522971a534b3265b9d7483b6b12175f025c0ebf1", "seconds": 9.16, "spoken_utterances": [ "Accept all uh... pending CashApp—um... I should say Venmo payment requests from my roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash app' but immediately self-corrects to 'Venmo' with 'I should say'. This is a false start where the final, corrected value is 'Venmo'. The rest of the instruction matches the original. The final intent is clear and executable.", "heard_transcript": "Accept all uh pending cash app um, I should say Venmo payment requests from my roommates and coworkers." }, "candidate_id": "wav_45affcb7800d7b38", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'cash app' corrected to 'Venmo'. The correction is immediate and the speaker continues without hesitation. However, the initial error ('cash app') is less plausible than 'family and friends' in the context of roommate/co-worker payments, and the correction feels slightly rushed. The prosody is natural, but the overall flow is less convincing than audio_02 due to the less contextually fitting initial error." } ] }, "is_default": false }, { "variant_id": "false_start_02_request_status", "proposed_instruction": "Accept all unpaid—sorry, that should be pending Venmo payment requests from my roommates and coworkers.", "long_pause_offsets": [], "source_sha256": "79f7ed0f398aebd201ad21dd1c66901d031723919338c863508dd1058ad04324", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'unpaid' to 'pending' within the same utterance, preserving the app, action, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/false_start/false_start_02_request_status/3d24c8ca96bb476a/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "ae5ac5702637ab6df1eee1e2317e27a432a51269438babe794e817c78aed40e0", "seconds": 9.16, "spoken_utterances": [ "Accept all unpaid—um... sorry, that should be pending Venmo payment requests from uh... my roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unpaid' but immediately self-corrects to 'pending', which matches the original instruction. The final intent is clear and aligns with the original task. The correction is immediate and does not introduce any new constraints or changes to the target or action.", "heard_transcript": "Accept all unpaid um, sorry, that should be pending Venmo payment requests from uh my roommates and coworkers." }, "candidate_id": "wav_92a4ec2f8a3729ff", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'unpaid' corrected to 'pending'. The correction is clear, but the initial word 'unpaid' is semantically opposite to the intended action ('accept'), making the error less plausible. The speaker's tone is slightly more hesitant, and the correction feels more like a mechanical fix than a natural recollection. The prosody is generally good, but the semantic mismatch reduces realism." } ] }, "is_default": false }, { "variant_id": "false_start_01_request_action", "proposed_instruction": "Decline—no, I meant Accept all pending Venmo payment requests from my roommates and coworkers.", "long_pause_offsets": [], "source_sha256": "4dbfa6443b6e376e052b373e7e4e863690f66484c6988bdc4cac1f639bcc27d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Decline' to 'Accept' within the same utterance, preserving the app, action scope, and recipient groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/false_start/false_start_01_request_action/c912b832d6b9bb55/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "310fe5e0cee9c8df616371c32ac6edc1ae3736e87cb088579866d7fd1bb89786", "seconds": 10.24, "spoken_utterances": [ "Decline—um... no, I meant Accept all hmmm... pending Venmo payment requests from my roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Decline them' but immediately self-corrects with 'No, I meant accept all...'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistaken value should be discarded. The final intent matches the original instruction exactly, with no additional or missing constraints.", "heard_transcript": "Decline them. No, I meant accept all pending Venmo payment requests from my roommates and coworkers." }, "candidate_id": "wav_48c58286c049d9f2", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip begins with 'decline', then corrects to 'accept'. While the correction is present, the initial command is semantically contradictory to the final one, making the self-correction feel less natural. The speaker's hesitation ('hmm') and the abrupt shift from 'decline' to 'accept' sound more like a task reset than a spontaneous correction. The prosody is somewhat flat, and the overall delivery feels less like a real-time thought process and more like a rehearsed or forced correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Accept all pending Venmo payment requests. I left out a detail: only accept requests from my roommates and coworkers.", "long_pause_offsets": [ 43 ], "source_sha256": "2b67cdbac427b0332b2d92e560ff31ee341de616c85f6aaad2b8184c80fca266", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the initial broad request to only include requests from roommates and coworkers, which matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/addition/addition_02/40547f8e7cbd5988/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "0a81d34b22f12b41808fc62decd851d3a141071521e7f8b3f8769511f3a1b6c1", "seconds": 11.8, "spoken_utterances": [ "Accept all uh... pending um... Venmo payment requests. ", "I left out a detail: only accept requests from my roommates and coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Accept all pending Venmo payment requests'. The speaker then adds a clarifying constraint: 'Only accept requests from my roommates and coworkers.' This addition narrows the scope of the original request, and the final intent is to execute the original task with the added constraint. The final control intent is to execute the combined task.", "heard_transcript": "Accept all uh pending um Venmo payment requests. I left out a detail. Only accept requests from my roommates and coworkers." }, "candidate_id": "wav_0fb8dab50fea7826", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding false start ('accept all uh pending um'), followed by a clear self-correction ('I left out a detail') and a smooth, additive revision that specifies the restriction ('only accept requests for my roommates and coworkers'). The prosody is conversational, with appropriate pauses and filler words that reflect genuine thought processing. The correction is integrated into the same utterance, making the flow feel spontaneous and realistic." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Accept all Venmo payment requests from my roommates and coworkers. Oh, I forgot to mention, only accept pending payment requests.", "long_pause_offsets": [ 67 ], "source_sha256": "7a2f8ea1080234cdd2f176a69620f1b7909c0c742a67410a6f9d43e54401ed77", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'pending' to the initial request, which is a refinement that does not contradict the original. The recipient groups (roommates and coworkers) and the app (Venmo) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/addition/addition_01/c091e91c963a3133/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "389d38c1aee0f759682f67281ad0193f246a23309b988eb27e354bb218a922c1", "seconds": 15.32, "spoken_utterances": [ "Accept all Venmo payment requests from ... my roommates and let me think... coworkers. ", "Oh, I forgot to mention, only accept ... pending payment requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to accept all Venmo payment requests from roommates and coworkers, but the speaker paused and corrected the scope to 'only accept pending payment requests' after a longer pause. This is a clear addition that narrows the original request. The final intent combines the original target group (roommates and coworkers) with the new constraint (pending requests), matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Accept all Venmo payment requests from my roommates and let me think, coworkers. Oh, I forgot to mention, only accept pending payment requests." }, "candidate_id": "wav_818fe296281ec114", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a more fluent initial request but includes a filler ('let me think') that sounds slightly rehearsed rather than spontaneous. The subsequent correction ('Oh, I forgot to mention...') is delivered with a natural-sounding 'oh' and a pause, but the phrasing 'only accept pending payment requests' feels like a pasted-on addition rather than a fluid thought process. The prosody is less varied, and the overall delivery is slightly more robotic compared to audio_01, despite containing a plausible revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family. I misspoke about the requester group: it should be roommates and coworkers, not family.", "long_pause_offsets": [ 58 ], "source_sha256": "3e75983dfcb37c004c74bea4751a7ce972db8694d1ab047ec81afee148238c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the requester group from 'family' to 'roommates and coworkers', while retaining the action, app, and request status. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/revision/revision_04_requester_group/cfb5ee89520f5c7b/speech.wav", "wer": 0.1, "wav_sha256": "5af535e07530be7c80b11ad89f81bda4d030e93fc831af054167c9b523448d4e", "seconds": 15.44, "spoken_utterances": [ "Accept all uh, let's see... pending Venmo payment requests from uh, let me think... my family. ", "I misspoke about the requester group: it should be roommates and coworkers, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately corrects themselves with 'I misspoke about the requester group. It should be roommates and coworkers, not family.' This is a clear revision where the final intent is to accept requests from 'roommates and coworkers,' matching the original instruction. The revision is spoken after a pause and explicitly corrects the earlier value, satisfying the revision type criterion.", "heard_transcript": "Accept all uh let's see pending Venmo payment requests from uh let me think my family. I misspoke about the requester group. It should be roommates and coworkers, not family." }, "candidate_id": "wav_2ecd656115ebe846", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker starts with a clear initial request, then uses a believable 'let me think' filler pause before stating 'my family'. The self-correction is smooth and pragmatic: 'I misspoke about the requester group, it should be roommates and coworkers, not family.' The intonation and pacing of the correction sound spontaneous and not rehearsed, with a clear reference back to the earlier mistake. The prosody is consistent and the delivery is continuous without robotic or stilted qualities." } ] }, "is_default": true }, { "variant_id": "revision_02_request_status", "proposed_instruction": "Accept all completed Venmo payment requests from my roommates and coworkers. Wait, the requests should be pending, not completed.", "long_pause_offsets": [ 77 ], "source_sha256": "09f824d0690c988734ae56a16265d3299da3b64ada645d9e1c134e596651fd39", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the status from 'completed' to 'pending', which reverses the original condition and results in a different final goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/revision/revision_02_request_status/90803fa5c4c59bc1/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "3fd5b77726b6f5961af1efcd43f48cf0abee939ce9ee5d635d6b00b7750c298e", "seconds": 11.56, "spoken_utterances": [ "Accept all uh, let's see... completed Venmo payment requests from my roommates and um... coworkers. ", "Wait, the requests should be pending, not completed." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'completed' but immediately corrects themselves to 'pending' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Accept all uh let's see completed Venmo payment requests from my roommates and um coworkers. Wait, the requests should be pending, not completed." }, "candidate_id": "wav_85f314a6797fd931", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a credible revision. The speaker begins with a clear request, then hesitates with 'wait' before correcting 'completed' to 'pending'. The correction is immediate and the intonation shift is natural, indicating a genuine change of mind. The prosody is consistent, and the flow is continuous. While the correction is less detailed than audio_03's, it is still a believable and spontaneous self-repair within the same utterance." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Accept all pending CashApp payment requests from my roommates and coworkers. Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 77 ], "source_sha256": "0a977614f3b1d8f2028375c76460cd714ed5a3c3e4546ac2ef4843dffb32c28a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from CashApp to Venmo, while retaining the action, request status, and recipient groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/revision/revision_03_app/6249df3d0ff471a2/speech.wav", "wer": 0.3181818181818182, "wav_sha256": "1e547f20fe8f32a378d52cc6059c12f6b912b4958403ad956c5d550bce44d5bd", "seconds": 12.76, "spoken_utterances": [ "Accept all hmmm... pending CashApp payment requests from uh... my roommates and coworkers. ", "Let me correct the app: use Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo, not Cash App' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to use Venmo, which matches the original instruction. All other constraints (pending requests, roommates and coworkers) are preserved.", "heard_transcript": "Accept all pending Cash App payment requests from my roommates and coworkers. Let me correct that. Use Venmo, not Cash App." }, "candidate_id": "wav_ab7188156f8d127d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision, but the self-correction feels slightly less natural. The speaker says 'let me correct the app, use Venmo not Cash App' after initially saying 'Cash App'. The phrasing 'correct the app' is slightly awkward and less common in spontaneous speech than a direct self-correction of the app name. The delivery is otherwise fluent and the prosody is consistent, but the phrasing of the correction makes it sound a bit more like a scripted fix than a natural thought process." } ] }, "is_default": false }, { "variant_id": "revision_01_request_action", "proposed_instruction": "Decline all pending Venmo payment requests from my roommates and coworkers. Change the action to Accept, not Decline.", "long_pause_offsets": [ 76 ], "source_sha256": "728e55b35be1a5a16eff8f6cae76339a424dda0c19e7a56c4ad1edacd11152c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the action from 'Decline' to 'Accept', while retaining the app, request type, and recipient groups." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/revision/revision_01_request_action/92e02c5b95934178/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "a881e0d61191f72274f1af4600ab7b4cf75fd05a05f46e969208b5c5629199ac", "seconds": 17.4, "spoken_utterances": [ "Decline all pending uh... Venmo payment requests from hmmm... my roommates and let me think... coworkers. ", "Change the action to Accept, not Decline." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Decline' but immediately corrects themselves with 'Change the action to accept, not decline.' This constitutes a revision where the final audible intent is to accept, not decline. The target group 'roommates and coworkers' remains unchanged. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Decline all pending uh Venmo payment requests from my roommates and let me think, coworkers. Change the action to accept, not decline." }, "candidate_id": "wav_6833b0715bd8e47b", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. It begins with a clear request, then includes a 'let me think' pause, but the correction is not a self-repair of the initial utterance. Instead, it's a separate command: 'Change the action to accept, not decline.' This sounds more like a post-hoc instruction to a system than a spontaneous revision of one's own speech. The prosody is consistent, but the pragmatic flow is unnatural because the speaker does not correct their own words but rather issues a new command, breaking the illusion of a single, continuous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_035", "hide_variant_original": true, "source_sha256": "cc3398e891a9af7e9f7d5b3a3eafb3c1d541440d59ad7a89f9c2402a05414977", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/cancellation/cancel_035/29e71453de3a9e38/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "fab3c8733b08976364625efe8aa2b079b2fec6d9413adcd23663d1fc6bd1efdc", "seconds": 17.4, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers.", "Hold on, stop working on that and cancel the task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hold on, stop working on that and cancel the task.' This constitutes a direct cancellation of the entire task, satisfying the 'cancellation' type criterion. The initial hesitation and self-correction do not alter the final intent.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers. Hold on, stop working on that and cancel the task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "69d8556f6604c5c364e9886bf9ba3c7c807759afc78aa0e97b3d52ae549d59cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/ambiguous_halt/halt_040/7c94d6cde9974463/speech.wav", "wer": 0.125, "wav_sha256": "2bbb63f4597028592f20532e33b507f7182617dd53b65a925138a99a5863996e", "seconds": 15.36, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request with 'Accept all pending Venmo payment requests from my roommates and coworkers' but then immediately interjects 'Hang on, just a minute,' which is a clear request to pause. The final interjection is not a cancellation or a revision, but a temporary halt. The type is 'ambiguous_halt', and the final audible intent matches the expected control intent of pausing.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bc966380685f328a80c04390eefdfeb0d2d1f60eff132bbf4cdcc9ff16f3d36c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_1/audio_natural/filler/original/e479d91bd84ccc41/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "a91baf429294a7bf3e55d9feb0849844f4da2640d80d5a3c6d17e7228201b350", "seconds": 11.64, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from let me think... my roommates and um... coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitations ('um', 'let me think') and pauses but does not change the core request. The final intent matches the original instruction exactly, with no self-correction or modification. The hesitations are filler and do not alter the task goal.", "heard_transcript": "Accept all pending, um, let me think, Venmo payment requests from, let me think, my roommates and um, coworkers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Reject all pending Venmo payment requests from my family—no, I mean friends and roommates.", "long_pause_offsets": [], "source_sha256": "8f4a137c80203c2dabda147d9c2ad68d4e73cb353969df1c7cabaeca4917518b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'friends and roommates', which matches the original instruction. The app and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/false_start/false_start_04_requester_group/a65eb1d3f46dd032/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1ae17a077160da011fcea7afb1e0f8f310a05aeb8f7bc67aaae41bbcb7170b19", "seconds": 9.04, "spoken_utterances": [ "Reject all pending Venmo payment requests from uh... my family—um... no, I mean friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately self-corrects to 'friends and roommates'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Reject all pending Venmo payment requests from uh my family. Um, no, I mean friends and roommates." }, "candidate_id": "wav_481d11b862ec55c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'my family', hesitates with 'um', then clearly self-corrects to 'no, I mean friends and roommates'. The intonation and pacing of the correction sound spontaneous and conversational, with a believable shift in focus. The pause and filler are used appropriately to signal a change of mind, and the overall flow is continuous and realistic." } ] }, "is_default": true }, { "variant_id": "false_start_02_request_status", "proposed_instruction": "Reject all approved—sorry, that should be pending Venmo payment requests from my friends and roommates.", "long_pause_offsets": [], "source_sha256": "19d2f51dbd6d2528d3d13212d72a9073cc7260e491ba859525887ce04f3088ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'approved' to 'pending' within the same utterance, preserving the app, action, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/false_start/false_start_02_request_status/4bfe5325e4b3c2bc/speech.wav", "wer": 0.0, "wav_sha256": "38264df3c6ef593c79f66645a9114a6c8de8f02d1aa084eab7a0fcea7e62b75b", "seconds": 7.68, "spoken_utterances": [ "Reject all approved—uh... sorry, that should be um... pending Venmo payment requests from my friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'approved' but immediately self-corrects to 'pending' after a pause and repair. According to the 'false_start' type criterion, the final corrected value ('pending') should be applied, and the original incorrect value ('approved') should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Reject all approved uh, sorry, that should be um, pending Venmo payment requests from my friends and roommates." }, "candidate_id": "wav_ce22ed59c9fea14c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'approved' followed by a self-correction to 'pending'. The correction is clear and the speaker's tone suggests a moment of realization. However, the initial error ('approved') is less plausible than 'family' in the context of Venmo requests, and the correction feels slightly more mechanical. The prosody is natural, but the flow is less organic than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_01_request_action", "proposed_instruction": "Accept—no, I meant Reject all pending Venmo payment requests from my friends and roommates.", "long_pause_offsets": [], "source_sha256": "c34ea6cb68f8c0a4b9017cfe0bc3d3708378bae55c03d643446d266e1b317129", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Accept' to 'Reject' within the same utterance, preserving the app, action scope, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/false_start/false_start_01_request_action/89c5ae5b9d28ba0d/speech.wav", "wer": 0.1, "wav_sha256": "6331b792b6c99d0a856611d827325ec4912cc9b7f0798274f8babe1d950d2223", "seconds": 11.16, "spoken_utterances": [ "Accept—uh... no, I meant Reject all um, let me think... pending Venmo payment requests from my friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Accept' but immediately self-corrects with 'no, I meant reject'. This is a clear false start where the final audible intent is to reject, matching the original instruction. The subsequent hesitation ('um let me think') does not alter the final goal, which remains to reject all pending requests from the specified group.", "heard_transcript": "Accept uh no, I meant reject all um let me think pending Venmo payment requests from my friends and roommates." }, "candidate_id": "wav_06a02326297cbc6b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a cancellation and a later correction. The speaker starts with 'accept', then immediately corrects to 'reject'. The 'let me think' pause is realistic, but the correction is abrupt and the speaker seems to be restarting the entire request rather than refining a single part. The prosody is natural, but the structure feels less like a spontaneous thought process and more like a rehearsed correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Reject all pending CashApp—I should say Venmo payment requests from my friends and roommates.", "long_pause_offsets": [], "source_sha256": "27a79f841edea20663f2dc19ed9d27612159ce303b8deb1f8f657c6765e8e692", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, target group, and request status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/false_start/false_start_03_app/f13fc9a3b06259df/speech.wav", "wer": 0.29411764705882354, "wav_sha256": "4a592c8735b22ca9233e64c88e3c3a0cb96a211bb183ba6f377a47d072060f46", "seconds": 10.12, "spoken_utterances": [ "Reject all pending CashApp—um... I should say uh... Venmo payment requests from ... my friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cash up' but immediately self-corrects to 'Venmo payment requests'. The final intent matches the original instruction, and the correction is immediate and clear. The 'false_start' type requires executing the corrected version, which is fully aligned with the reference.", "heard_transcript": "Reject all pending cash up um, I should say uh Venmo payment requests from my friends and roommates." }, "candidate_id": "wav_74ef64e9cfbfca62", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has a false start with 'cash up' followed by a correction to 'Venmo payment requests'. However, the correction is awkward and the phrase 'cash up' is not a plausible misstatement in this context. The speaker's tone is flat and the correction feels forced rather than natural. The prosody lacks the organic flow of the other clips, making it the least realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Reject all Venmo payment requests from my friends and roommates. Oh, I forgot to mention, only reject pending requests.", "long_pause_offsets": [ 65 ], "source_sha256": "e679c1f9f391358c7d9dfc8d0c58cda271851d08d17556782afe64db41eb039d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'pending' to the initial request, refining the scope without changing the app, action, or recipients. The final goal is a more specific version of the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/addition/addition_01/04bf816217bfd324/speech.wav", "wer": 0.05, "wav_sha256": "bfd084aca6b63f44e427916c202634d74bfb1bcf3e76a46690ff29861bb2dd6b", "seconds": 11.04, "spoken_utterances": [ "Reject all Venmo payment requests from ... my friends and roommates. ", "Oh, I forgot to mention, only reject um... pending requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Reject all Venmo payment requests from my friends and roommates.' This is followed by a longer pause and an addition: 'only reject pending requests.' This addition narrows the scope of the original instruction to only 'pending' requests. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type is correctly applied and the final goal is fully satisfied.", "heard_transcript": "Reject all Venmo payment requests from my friends and roommates. Oh, I forgot to mention, only reject um pending requests." }, "candidate_id": "wav_f8bea58fa9deaf60", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural pacing and a believable, additive follow-up. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to introduce a restriction, and the correction 'only reject pending requests' is smoothly integrated with a natural pause and filler ('um'). The prosody is consistent, and the addition feels spontaneous rather than forced." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Reject all pending payment requests from my friends and roommates. Oh, right, this is specifically on Venmo.", "long_pause_offsets": [ 67 ], "source_sha256": "af304dac6d494ef3e1d4b29d2cac6af4df6a28b9ee65323971d3723afe4b42d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Venmo' without changing any other requirement. The initial instruction already specifies the action, target group, and request status, so the final goal is fully specified and consistent with the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/addition/addition_02/7e37342cdff3b617/speech.wav", "wer": 0.0, "wav_sha256": "9a70fea19ecf3d18c7a02641f6b853e9d95fda23c0cbb0a8f3bac81a775da5b3", "seconds": 14.96, "spoken_utterances": [ "Reject all um, let me think... pending payment requests from uh, let me think... my friends and roommates. ", "Oh, right, this is specifically on Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with fillers ('um', 'let me think') but proceeds to state the original instruction. After a pause, they add the clarifying detail 'this is specifically on Venmo', which is an addition to the original instruction. The final intent combines the original request with the added platform specificity. The addition is clear and does not contradict the original, so the task is executed with the added constraint.", "heard_transcript": "Reject all um, let me think, pending payment requests from uh, let me think, my friends and roommates. Oh right, this is specifically on Venmo." }, "candidate_id": "wav_0e639d58f08b8929", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a clear request but includes multiple hesitations ('um', 'let me think') that disrupt fluency. The self-correction 'Oh, right, this is specifically on Venmo' is somewhat abrupt and lacks the natural backtracking intonation seen in more realistic speech. The overall delivery feels slightly rehearsed or hesitant, with less fluid transitions between ideas compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_requester_group", "proposed_instruction": "Reject all pending Venmo payment requests from my my family. Change the requester group to my friends and roommates, not my family.", "long_pause_offsets": [ 61 ], "source_sha256": "6a6a343e01b916bf55a57998a786bea646d398dac7d2e918c8caf286094ca4ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the requester group from 'my family' to 'my friends and roommates', which matches the original instruction. The action and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/revision/revision_04_requester_group/c09536f1cdd850a6/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "337436db9b44b508052cf7b9b3ef67ebbbf3a8d275e8063de9fff2665b16d120", "seconds": 14.68, "spoken_utterances": [ "Reject all um, let me think... pending Venmo payment requests from um... my my family. ", "Change the requester group to my ... friends and roommates, not my family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'from my family' but immediately corrects to 'Change the requester group to my friends and roommates, not my family.' This is a clear revision where the final intent overrides the initial incorrect value. The final instruction matches the original reference exactly, so the revision criterion is satisfied.", "heard_transcript": "Reject all um, let me think, pending Venmo payment requests from um, my family. Change the requester group to my, friends and roommates, not my family." }, "candidate_id": "wav_a43c56ecdbef2c33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic flow. The speaker begins with a clear request, uses a believable filler ('let me think'), hesitates with a pause and 'um', then corrects the group from 'my family' to 'my friends and roommates' with a smooth, self-correcting intonation. The revision is pragmatic and the prosody reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "revision_02_request_status", "proposed_instruction": "Reject all completed Venmo payment requests from my friends and roommates. Make that all pending requests, not completed ones.", "long_pause_offsets": [ 75 ], "source_sha256": "6f8a9e7665a1a83caa566cc2d55899d8b5fb60a1303bdf6a86b40b93d5c465b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the status from 'completed' to 'pending', preserving the app, recipient group, and action. The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/revision/revision_02_request_status/6d1eecb75348fa0d/speech.wav", "wer": 0.1, "wav_sha256": "feeabde51fe6e984b9a42b9e69b640dd947cb5a01d4ccfe6fe6ffb2fa320dae9", "seconds": 11.96, "spoken_utterances": [ "Reject all um... completed Venmo payment requests from my friends and roommates. ", "Make that all ... pending requests, not completed ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'completed' but immediately corrects to 'pending' after a pause, which is the revision type. The final intent is to reject pending requests, matching the original instruction. The correction is clear and unambiguous.", "heard_transcript": "Reject all um completed Venmo payment requests from my friends and roommates. Make that all pending requests, not completed ones." }, "candidate_id": "wav_560dfb90224cf228", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision where the speaker corrects 'completed' to 'pending'. The delivery is fluent with a natural pause and filler ('um'), and the correction is integrated smoothly. However, the initial phrasing is slightly more robotic than audio_02, and the correction lacks the same level of conversational hesitation, making it slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Reject all pending CashApp payment requests from my friends and roommates. Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 75 ], "source_sha256": "e8f12d6064961d17faca33d6ec3c07ac3dc20956355d490e9495af19d0981238", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from CashApp to Venmo, while retaining the action, request status, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/revision/revision_03_app/ea9cb4c0c52888a6/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "94baf0b83367f53b6adaad410ed77f166169c3357be44e633a1ad6787454ffca", "seconds": 11.32, "spoken_utterances": [ "Reject all uh... pending CashApp payment requests from my friends and roommates. ", "Let me correct the app: use ... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' after a short pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Reject all pending Cash App payment requests from my friends and roommates. Let me correct the app, use Venmo not Cash App." }, "candidate_id": "wav_7501dc83d65c6abb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker corrects 'Cash App' to 'Venmo'. The delivery is clear, but the correction is abrupt and lacks the natural hesitation or filler that would make it sound more spontaneous. The phrase 'let me correct the app' sounds slightly rehearsed, reducing its realism compared to the more organic flow of the others." } ] }, "is_default": false }, { "variant_id": "revision_01_request_action", "proposed_instruction": "Accept all pending Venmo payment requests from my friends and roommates. Wait, change the action to Reject, not Accept.", "long_pause_offsets": [ 73 ], "source_sha256": "6d607b07914b35e9063721c9414fb322da783f87d2682f15e0b01ea3577fb3de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the action from 'Accept' to 'Reject', while retaining the app, scope, and target group. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/revision/revision_01_request_action/f75690a5565292b1/speech.wav", "wer": 0.1, "wav_sha256": "0b02697d2cdc6c276c3ad1ae9d1ec6a333ec4ca1597095a1dc92a31674153154", "seconds": 11.84, "spoken_utterances": [ "Accept all um... pending ... Venmo payment requests from ... my friends and roommates. ", "Wait, change the action to Reject, not Accept." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Accept all pending Venmo payment requests' but immediately corrects themselves with 'Wait, change the action to reject, not accept.' This constitutes a revision where the final intent is to reject, not accept. The correction is clear and unambiguous, and the rest of the instruction (targeting 'pending Venmo payment requests from my friends and roommates') remains consistent. Therefore, the final intent matches the original reference instruction.", "heard_transcript": "Accept all um pending Venmo payment requests from my friends and roommates. Wait, change the action to reject, not accept." }, "candidate_id": "wav_f02da833be7147f0", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a false start where the speaker says 'accept' and then corrects to 'reject'. While the correction is present, the initial word 'accept' is spoken with a flat, robotic tone, and the correction is delivered without any hesitation or filler. The overall delivery sounds rehearsed or read aloud rather than spontaneously generated, making it the least realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "1becefa59432356162028c58d6984707677906941c4b2469803cecc7001455e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/cancellation/cancel_021/81fe76edecf70316/speech.wav", "wer": 0.08, "wav_sha256": "6e520aab6e3550f27d223fe265da3a7feb6d031413408e978c6dcd5b71148257", "seconds": 12.44, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates.", "Oh, wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task but then clearly revokes it with 'Oh wait, never mind. I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion. The final interjection is a complete and unambiguous cancellation, not a pause or a revision.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates. Oh wait, never mind. I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "d6b94be10f09c1b0540f41c27cb718716271ef8c5c7d92592be595e0dbe04b47", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/ambiguous_halt/halt_001/ad428fca460bdc05/speech.wav", "wer": 0.125, "wav_sha256": "daa13c2cafade1c374589a6cc8f8c6d9cce5f5b2304d5a5cf81aa01f433f4ace", "seconds": 10.48, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait' after a brief hesitation, which constitutes a clear request to pause or hold the current task. The initial instruction to reject payment requests is not canceled, but the speaker explicitly halts execution to reconsider. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "83f3d47e43555e3a2d42ce2ad10ae6f085b22c1f5add44c591edd1462f357e0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_2/audio_natural/filler/original/3ab1e684dbce76be/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "149a234fdabbe9cdad99ae24ab81ff0c3791cb05ce2e7c893806dcfa24a782a3", "seconds": 7.72, "spoken_utterances": [ "Reject all uh... pending Venmo payment requests from um, let's see... my friends and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let's see') but does not change the core request. The final intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "Reject all uh pending Venmo payment requests from um let's see my friends and roommates." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family—sorry, I mean coworkers and friends.", "long_pause_offsets": [], "source_sha256": "daedf01420f56fb165f1095b554024274361420fd4a9728f17535974ea0d7839", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'coworkers and friends'. The final intended group matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/false_start/false_start_04_requester_group/cd82dd943de36c2f/speech.wav", "wer": 0.1875, "wav_sha256": "eb369dd6f8f3914cd06eae214c2c95a1d47d06de3a8c3fa80faa3e43a04db141", "seconds": 10.04, "spoken_utterances": [ "Accept all pending ... Venmo payment requests from my family—hmmm... sorry, I mean coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'coworkers and friends'. This is a false start, and the final intent matches the original instruction. The correction is clear and immediate, so the task should be executed as stated in the correction.", "heard_transcript": "Accept all pending Venmo payment requests from my family, whom, sorry, I mean coworkers and friends." }, "candidate_id": "wav_30c2d1eb12fc737a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'family' followed by a self-correction to 'co-workers and friends'. The pause and filler 'hmm' before the correction sound spontaneous, and the correction is delivered with a slight shift in intonation that mimics genuine recollection. The overall prosody is smooth and conversational, with no signs of artificiality or forced pacing." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Accept all pending CashApp—I should say Venmo payment requests from my coworkers and friends.", "long_pause_offsets": [], "source_sha256": "7e27d85bcb57d6216985ded1129d44a2e6698093862c94ce43bc4648c119ba5e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, scope of requests, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/false_start/false_start_03_app/7323667685c20696/speech.wav", "wer": 0.2777777777777778, "wav_sha256": "f3a3d913e931d1c6224eb3307d9f9220f45ca9872ecdb04d602bc46e815b5e26", "seconds": 10.32, "spoken_utterances": [ "Accept all pending CashApp—um... I should say Venmo payment requests from my coworkers and let's see... friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'cash app' but immediately self-corrects to 'Venmo payment requests'. This is a false start, and the final corrected instruction matches the original reference. The speaker's intent is clear and fully aligned with the original task after the correction.", "heard_transcript": "Accept all pending cash app. I should say Venmo payment requests from my coworkers and let's see, friends." }, "candidate_id": "wav_163476b86470c0f5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip includes a false start with 'cash' and a self-correction to 'Venmo payment requests'. The speaker uses a natural-sounding 'I should say' to transition, and the correction is delivered with a slight rise in pitch and a brief pause, indicating a moment of realization. The flow is coherent and the delivery sounds like a real-time thought process." } ] }, "is_default": false }, { "variant_id": "false_start_02_request_status", "proposed_instruction": "Accept all unresolved—no, make that pending Venmo payment requests from my coworkers and friends.", "long_pause_offsets": [], "source_sha256": "af718e9ea2a9d1f29d7fcab094f321533a2d8577be555a413527e5722f5ff2ab", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'unresolved' but immediately corrects it to 'pending'. The final value is 'pending', which matches the original instruction. All other elements—action, app, recipient group, and scope—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/false_start/false_start_02_request_status/09d1c9c6277f87bf/speech.wav", "wer": 0.29411764705882354, "wav_sha256": "da0a68f7186798592e30d16ea1bc2026518dd11ff7c2afcb8caef1a86c14ea91", "seconds": 11.24, "spoken_utterances": [ "Accept all unresolved—uh... no, make that pending um... Venmo payment requests from my ... coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unresolved' but immediately self-corrects to 'pending' with 'no, make that pending'. This is a false start where the final value is the corrected one. The rest of the instruction matches the original exactly, including the target ('Venmo payment requests') and the source ('coworkers and friends'). The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Accept all unresolved uh no, make that pending um Venmo payment requests from my coworkers and friends." }, "candidate_id": "wav_b8ef92388c807679", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a false start with 'unresolved' and a correction to 'pending'. The speaker says 'no' before the correction, which is a common way to self-correct, but the transition feels slightly abrupt. The prosody is mostly natural, but the correction lacks the subtle intonation shift seen in the top two clips, making it feel a bit more mechanical." } ] }, "is_default": false }, { "variant_id": "false_start_01_request_action", "proposed_instruction": "Decline—sorry, I meant Accept all pending Venmo payment requests from my coworkers and friends.", "long_pause_offsets": [], "source_sha256": "7136adda8a36cb2e0d48a9a411623f7591472ffac4f4b4dfd26155f5e7557d1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Decline' to 'Accept' within the same utterance, preserving the app, action scope, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/false_start/false_start_01_request_action/17970dd828faa769/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "3a455bfec6b1eb113026195dd0a681eec3b541e12cc7028593f66f176f68b674", "seconds": 10.52, "spoken_utterances": [ "Decline—uh... sorry, I meant Accept all um... pending Venmo payment requests from my coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Decline' but immediately self-corrects to 'accept all'. According to the 'false_start' type criterion, the immediate self-correction should be applied, and the original mistaken value discarded. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Decline uh, sorry, I meant accept all um pending Venmo payment requests from my coworkers and friends." }, "candidate_id": "wav_485fb1d37d6198d0", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip begins with a false start ('decline') and a correction ('accept all'). However, the correction is delivered with a flat, slightly rushed tone, and the pause before 'pending Venmo payment requests' feels unnatural and drawn out. The overall delivery lacks the organic flow and intonation variation of the other clips, making it sound less spontaneous and more rehearsed." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Accept all pending payment requests from my coworkers and friends. Oh, and one more thing: do this on Venmo.", "long_pause_offsets": [ 67 ], "source_sha256": "bedeb3f4624e81757b9c61ba1d1980dbda8183a057026ae795fe17538991e27f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Venmo' without changing any other requirement; the initial instruction already specifies the action, scope, and recipients." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/addition/addition_02/90cc2e6c326ec652/speech.wav", "wer": 0.125, "wav_sha256": "ce998fb3007878dab7c5b237429a4ba47bd144a669140d63851a56b43129f9bf", "seconds": 17.28, "spoken_utterances": [ "Accept all pending payment requests from my hmmm... coworkers and uh, let's see... friends. ", "Oh, and one more thing: do this on um... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to accept all pending payment requests from coworkers and friends. The speaker initially omits the app name, then adds 'Do this on um Venmo' after a pause, which specifies the platform. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, with the addition of the app name.", "heard_transcript": "Accept all pending payment requests from my coworkers and uh let's see, friends. Oh, and one more thing. Do this on um Venmo." }, "candidate_id": "wav_f960bcb675e497c2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'uh', 'let's see') at plausible points during formulation. The addition of 'Oh and one more thing' followed by a new instruction ('Do this on Venmo') is a believable conversational addition with a clear shift in focus. The prosody is continuous and the pacing reflects spontaneous thought, including a slight hesitation before 'Venmo'." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Accept all Venmo payment requests from my coworkers and friends. I forgot to add, only accept pending payment requests.", "long_pause_offsets": [ 65 ], "source_sha256": "bbfe7a148e447f80edf679812877fe838ecccce012007e99a92623b28e1ed88e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the constraint 'only accept pending payment requests', which refines the initial broad request to match the original instruction's specificity. The recipient group and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/addition/addition_01/84c9ba98e9fbd378/speech.wav", "wer": 0.3181818181818182, "wav_sha256": "853a2033200ac4a1ab07dc82e6f9623dcc5a6011d9bb7823dfdba5735a09a8d1", "seconds": 14.56, "spoken_utterances": [ "Accept all ... Venmo payment requests from my um, let's see... coworkers and friends. ", "I forgot to add, only accept pending payment requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the original instruction to accept all Venmo payment requests from coworkers and friends. After a pause and self-correction, they add the constraint 'only accept pending payment requests'. This addition narrows the scope of the original request, and the final intent is to accept only pending requests. The addition is clear and modifies the initial instruction as expected for the 'addition' type. The final goal matches the reference instruction with the added constraint.", "heard_transcript": "Accept all Venmo payment requests from my um let's see coworkers and friends. I forgot to add, only accept pending payment requests." }, "candidate_id": "wav_c7102561918f1878", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a slightly robotic, flat delivery ('Accept all Venmo payment requests...'). The 'um' and 'let's see' are natural, but the self-correction ('I forgot to add, only accept pending payment requests') is abrupt and lacks the smooth intonation of a genuine self-repair. The correction feels tacked on rather than integrated into the flow, and the overall delivery is less fluid than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_request_status", "proposed_instruction": "Accept all completed Venmo payment requests from my coworkers and friends. Make that pending payment requests, not completed ones.", "long_pause_offsets": [ 75 ], "source_sha256": "5d5672f2091c9b78a60a02c0fa639b38e3d6969a5ab1986abdc5ccfcce30128d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial request by replacing 'completed' with 'pending' and explicitly excludes 'completed ones', restoring the original condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/revision/revision_02_request_status/1953ec9c5ec18c3c/speech.wav", "wer": 0.38095238095238093, "wav_sha256": "a9898a2f1bad7303272feda2e2db0e8768e832199d69087a9a49fc34fb67c5aa", "seconds": 13.16, "spoken_utterances": [ "Accept all completed uh... Venmo payment requests from my coworkers and um... friends. ", "Make that pending payment requests, not completed ones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'completed' but immediately corrects to 'pending' after a pause, which is the revision type. The final audible intent is to accept pending requests, matching the original instruction. The correction is clear and unambiguous.", "heard_transcript": "Accept all completed uh Venmo payment requests from my coworkers and um friends. Make that pending payment requests, not completed ones." }, "candidate_id": "wav_ec915012759dd689", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction. The initial 'completed' is followed by a smooth, plausible revision to 'pending' without awkward phrasing. The prosody is consistent, and the correction feels spontaneous rather than forced. The speaker maintains a steady pace and tone throughout, with only a brief, natural hesitation before the correction." } ] }, "is_default": true }, { "variant_id": "revision_04_requester_group", "proposed_instruction": "Accept all pending Venmo payment requests from my family. Change the requester group to coworkers and friends, not family.", "long_pause_offsets": [ 58 ], "source_sha256": "e292b5b2bf3874ef41c141cffce7819667fd59000b515a620b529ae32d477ea5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the requester group from 'family' to 'coworkers and friends', preserving the action and app. The original instruction's group is fully overridden by the correction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/revision/revision_04_requester_group/af5ad99e4f074af4/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "798bab8fed286c0e0e2f0189625086bce4efec4f293f185b29961969c038230c", "seconds": 15.44, "spoken_utterances": [ "Accept all pending um, let me think... Venmo payment requests from my family. ", "Change the requester group to um, let's see... coworkers and friends, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the requester group as 'my family' but immediately corrects this to 'coworkers and friends, not family' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the initial incorrect instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Accept all pending um let me think Venmo payment requests from my family. Change the requester group to um let's see coworkers and friends, not family." }, "candidate_id": "wav_43943c54ef353d2d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a revision where the speaker first says 'pending' and then corrects the group from 'family' to 'co-workers and friends'. The correction is clear and the speaker uses natural fillers like 'let me think' and 'let's see'. However, the flow is slightly less smooth than audio_01, with a more noticeable pause and a slightly more hesitant tone during the correction, making it feel a bit more deliberative than spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_01_request_action", "proposed_instruction": "Decline all pending Venmo payment requests from my coworkers and friends. Wait, the action should be Accept, not Decline.", "long_pause_offsets": [ 74 ], "source_sha256": "2ddfb98e9cbc292876306eb841079f25489c6d5819e36490d85130c52cd13615", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'Decline' to 'Accept', while retaining the app, request type, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/revision/revision_01_request_action/e4191c93bf3516fd/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "8503bd4ca02405f04964ece297afd433aadc60d476d4d3a85583d8180d3ed703", "seconds": 15.64, "spoken_utterances": [ "Decline all pending uh... Venmo payment requests from my ... coworkers and friends. ", "Wait, the action should be uh, let's see... Accept, not Decline." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Decline' but immediately corrects themselves to 'accept' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent is to accept, which matches the original reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Decline all pending Venmo payment requests from my coworkers and friends. Wait, the action should be uh let's see, accept, not decline." }, "candidate_id": "wav_9843e46dc5ac2bbd", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with 'decline' and then corrects to 'accept'. The correction is clear, but the phrasing 'wait, the action should be uh let's see accept' sounds slightly more rehearsed or less fluid than the others. The hesitation and self-correction are present, but the delivery feels a bit more mechanical, and the prosody is less consistent, with a slight rise in pitch during the correction that sounds less natural." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Accept all pending CashApp payment requests from my coworkers and friends. Let me correct the app: use Venmo, not CashApp.", "long_pause_offsets": [ 75 ], "source_sha256": "284f2a270273c5fc5a69e9ee320900f74605b9d8a4f2fbd03354558cd8dbf249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from CashApp to Venmo, while retaining the action, request status, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/revision/revision_03_app/5847badbce8df1cf/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "8ae9c98d39654fea4bfec6b1a62c0ca2321751fc3798e3f8944c5b8a82a00271", "seconds": 15.8, "spoken_utterances": [ "Accept all ... pending let me think... CashApp payment requests from my coworkers and friends. ", "Let me correct the app: use uh... Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects themselves to 'Venmo, not Cash App' after a pause. This is a clear revision where the final intent is to use Venmo. The original instruction matches the final corrected intent, so the task is executed as revised.", "heard_transcript": "Accept all pending, let me think, Cash App payment requests from my coworkers and friends. Let me correct the app, use uh, Venmo, not Cash App." }, "candidate_id": "wav_d24496c51070fd1a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a revision where the speaker initially says 'cash app' and then corrects to 'Venmo'. The correction is clear, but the phrasing 'let me correct the app use uh Venmo not cash app' sounds more like a literal correction than a spontaneous thought. The delivery is slightly robotic, and the prosody is less natural, with a flat intonation and a slightly unnatural pacing, especially in the correction phrase." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "35de8e3d5cc5974c750c485382476bd102329cd19027f9bc926105ed78505ac8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/cancellation/cancel_001/a5334f81194f1048/speech.wav", "wer": 0.2, "wav_sha256": "4db760db37334a41381e33d18d61419ba68838db227ea16e5db2786f9bbc2137", "seconds": 11.6, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends.", "Hang on, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then immediately follows with 'Hang on, cancel that,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends. Hang on, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "7e948cfbeefd4a087ec9e3f8f58d6870113c8df859e5fb375a7dce608bc74af1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/ambiguous_halt/halt_022/ada2d851d6b53707/speech.wav", "wer": 0.1875, "wav_sha256": "7350c5c43bf04ea8c14d8d96c612e981d1d023ad51b627f926e693654e58339c", "seconds": 11.84, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The initial instruction is not canceled, and the speaker does not provide a new command or correction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "6ea6792_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "84d5b49f935d18e923503c1030fd7564560eebbbebb1b3e62715aa60ed0ab1f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/6ea6792_3/audio_natural/filler/original/3109fbe8826ec23a/speech.wav", "wer": 0.2727272727272727, "wav_sha256": "c0b94bb29bd1a5bdc725b84b163a1b51f3ab82fdd659c83e076ed4ca1a256e3d", "seconds": 7.88, "spoken_utterances": [ "Accept all ... pending Venmo payment requests from my ... coworkers and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not alter the goal. Since the goal is fully and correctly stated, the task should be executed.", "heard_transcript": "Accept all pending Venmo payment requests from my coworkers and friends." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to watermelon—sorry, make that sunshine, all lowercase.", "long_pause_offsets": [], "source_sha256": "7365b608ba015b1046c446e94bebd08d1f28e37c7ac4136768643a2505709fca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'watermelon' with 'sunshine'; the app, action, and case requirement remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/false_start/false_start_02_password/751fcc92a014712b/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "e955a71fbef31f085529f0e1d06acc073a629e96da7c4c98daa81e13b9c1efd9", "seconds": 9.72, "spoken_utterances": [ "Change my um... venmo password to watermelon—hmmm... sorry, make that uh... sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'watermelon' but immediately self-corrects to 'sunshine' after a pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value 'sunshine' should be applied, and the original incorrect value 'watermelon' should be discarded. The final instruction matches the original reference instruction exactly, including the lowercase requirement.", "heard_transcript": "Change my um Venmo password to watermelon. Hmm. Sorry, make that uh sunshine, all lowercase." }, "candidate_id": "wav_977a96d76c9943f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'Venmo' followed by a self-correction to 'PayPal', then a pause and a clear, spontaneous correction to 'sunshine'. The intonation and rhythm are consistent with someone formulating a request in real time, and the correction flows smoothly without robotic repetition. The filler 'um' and the thoughtful 'hmm' are placed naturally, enhancing realism." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Change my PayPal—I should say venmo password to sunshine, all lowercase.", "long_pause_offsets": [], "source_sha256": "50eb9a3f217c4fa27939d44acd09b5ccf2e8c6dde3a3baee8d81cd0cf7f0bcb9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the password value and case requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/false_start/false_start_01_app/d06c3870ac3885ff/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "7a02809c62de78fccdb54fb624adfc46e1de28e1a6924d697a715fbdcc9d7dd4", "seconds": 7.12, "spoken_utterances": [ "Change my PayPal—um... I should say uh... venmo password to uh, let me think... sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo'. The final instruction matches the original reference exactly, with the correct app and password value. The false start is corrected, so the final intent is clear and executable.", "heard_transcript": "Change my PayPal, um, I should say uh Venmo password to uh let me think, sunshine, all lowercase." }, "candidate_id": "wav_8e1ed308f4137a3e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a false start ('PayPal') and a self-correction ('Venmo'), but the correction is less fluid and the pacing feels slightly rushed. The 'let me think' filler is placed awkwardly mid-sentence, and the overall delivery lacks the natural hesitation and breath control of the top-ranked clip. While still realistic, the flow is less convincing than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Change my password to sunshine, all lowercase. Oh, right, this is for my Venmo account.", "long_pause_offsets": [ 47 ], "source_sha256": "80a68bf2d3e8fb49259ee2e4d54d8f3c371af9c1cfcdb62a2de0d1b54c341a5d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the app context (Venmo) for the password change, which was missing in the initial utterance. The action, new password value, and case requirement are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my password to uh... sunshine, all lowercase. ", "Oh, right, this is for my um, let me think... Venmo account." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "7574325_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Change my PayPal password to sunshine, all lowercase. Let me correct the app: change your Venmo password to sunshine, all lowercase.", "long_pause_offsets": [ 54 ], "source_sha256": "9e6bc0bfe297ec24671c236b4c1085d63fa0a3764f4ca325e1bc8a4da0d0fa63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from PayPal to Venmo, while retaining the password value and case requirement. The final instruction matches the original instruction in all meaningful aspects." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/revision/revision_01_app/88fd55494750cd00/speech.wav", "wer": 0.04, "wav_sha256": "aa2d4e794a2cd5b05ec057803254d0dc3c0e354f0319c27526b39e7d03fc7687", "seconds": 13.8, "spoken_utterances": [ "Change my uh... PayPal password to sunshine, all lowercase. ", "Let me correct the app: change your uh, let's see... Venmo password to sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a self-initiated repair. The final instruction matches the original reference exactly, including the password and case requirement. The revision type criterion is satisfied as the correction is spoken after a pause and applies to the earlier instruction.", "heard_transcript": "Change my uh PayPal password to sunshine, all lowercase. Let me correct the app. Change your uh, let's see, Venmo password to sunshine, all lowercase." }, "candidate_id": "wav_13509096202aafe3", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_02_password", "proposed_instruction": "Change my venmo password to blueberry, all lowercase. Change the password to \"sunshine\", not \"blueberry\".", "long_pause_offsets": [ 54 ], "source_sha256": "fcbefcb553db3f90fb83556f453f1d702f68e68a45bf503d079c4a69d835a233", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the password value from 'blueberry' to 'sunshine' and confirms the lowercase requirement, preserving the app and action." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my venmo password to uh... blueberry, all lowercase. ", "Change the password to um... \"... sunshine\", not \"blueberry\"." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "52d7501bcfbbc7411e0e5d9ef16784e1649ef5a7c39459089ca78b061e0b3129", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/cancellation/cancel_032/9130d96425d7f51e/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "83e0384cc536bb851ff6dfc4d56c7950034bdf581656caa88a4f0b3b52b782db", "seconds": 11.92, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase.", "Wait, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially attempts to execute the task but clearly revokes it at the end by stating, 'Wait, I don't need this anymore. Please cancel the request.' This constitutes a direct cancellation of the entire task, satisfying the cancellation type criterion.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine, all lowercase. Wait, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "8d35d58ce041b194cb8db100aac9f166591c88a0320c1ee89eec342deae25dee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/ambiguous_halt/halt_017/99bea85cee5cfc84/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "bafdf8a600dcb8508140d21d03a8e4ec608aa58f040bb4ac9f9c7798782afb16", "seconds": 9.84, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to change the password is not canceled or revised, but the user explicitly requests a hold on execution. Therefore, the final control intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine all lowercase. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5d5a7526da40f1cf8b7307f66b52789090e9c5ab26a694c24770f7a452304fbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_1/audio_natural/filler/original/41e895cda44a2c80/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "f22b95d87ebe03501d608dd0bbe4dec23f2525f440bf3bde3decf4b0a16157bb", "seconds": 6.48, "spoken_utterances": [ "Change my um... Venmo password to uh, let's see... sunshine, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh', 'let's see') but does not alter the core request. The final instruction clearly matches the original task: changing the Venmo password to 'sunshine' in all lowercase. The fillers do not constitute a change in intent, so the execution criterion is satisfied.", "heard_transcript": "Change my um Venmo password to uh let's see sunshine, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to strawberry—sorry, I meant blueberry, all lowercase.", "long_pause_offsets": [], "source_sha256": "48d80305554af082b56f6506d11e66dbdb38f1bc2e6304435ca79cd7fc58a2df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'strawberry' but immediately corrects to 'blueberry' with 'I meant'. The final password value is blueberry, and the lowercase requirement is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/false_start/false_start_02_password/ac0ab046cc9aa661/speech.wav", "wer": 0.0, "wav_sha256": "afc9c6a36920706c924e63ec9d316ff29df7af17ad4d462712337f62d4e38453", "seconds": 8.12, "spoken_utterances": [ "Change my um... venmo password to strawberry—uh... sorry, I meant blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'strawberry' but immediately self-corrects to 'blueberry'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, including the lowercase requirement.", "heard_transcript": "Change my um Venmo password to strawberry uh sorry, I meant blueberry, all lowercase." }, "candidate_id": "wav_67e8bb0d6e3ff7b0", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Change my PayPal—I should say venmo password to blueberry, all lowercase.", "long_pause_offsets": [], "source_sha256": "a655032edaf630239d1796b44a7feed4b93dbfa096239920bf2cffec18fcb8d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the password value and case requirement." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my PayPal—um... I should say venmo password to uh... blueberry, all lowercase." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Change my venmo password to blueberry,. I left out a detail: make sure the password is all lowercase.", "long_pause_offsets": [ 40 ], "source_sha256": "4203fae4e04b8c9f0a2b7b19d562b10a46523b8615a04eaa6794fa2ba43f0c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the lowercase requirement, which is consistent with the original instruction's specification of 'all lowercase'. The password value 'blueberry' is preserved from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/addition/addition_02/c91ff68817d4ffdf/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "c692b300b8a4c11c41e5a13e5e1481a7647bb5818f19009804e70cabc2ac4609", "seconds": 14.36, "spoken_utterances": [ "Change my venmo password to uh, let's see... blueberry,. ", "I left out a detail: make sure the password is hmmm... all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the password change request, then adds a constraint ('all lowercase') after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final audible intent is to change the password to 'blueberry' and ensure it is all lowercase, which matches the original instruction.", "heard_transcript": "Change my Venmo password to uh let's see blueberry. I left out a detail. Make sure the password is um all lowercase." }, "candidate_id": "wav_0acff4d5c29a29a7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent initial request with a natural pause and filler ('uh, let's see') before stating the password. The follow-up addition ('I left out a detail') is smoothly integrated with a believable self-correction and a pause before completing the thought. The prosody is consistent, and the flow reflects a person recalling and adding information in real time." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Change my password to blueberry, all lowercase. Oh, I forgot to mention, change the password for Venmo.", "long_pause_offsets": [ 48 ], "source_sha256": "faedc5e26220087b65ac43984f9d69c8f6f375c4b2877a3ce9bb87a1e17a380a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Venmo' to the initial password change request, preserving the specified password and case requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/addition/addition_01/7bb0841d39381773/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "1a27febb0531e1a4b0f7c75c09924ef60d714c26d035bba8bd0e2203fd315299", "seconds": 13.92, "spoken_utterances": [ "Change my password to uh... blueberry, all lowercase. ", "Oh, I forgot to mention, change the password for uh, let's see... Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to change 'my password' to 'blueberry', but this was followed by a longer pause and an addition specifying the target as 'Venmo'. According to the 'addition' type criterion, the final goal combines the original instruction with the added information. The final audible intent is to change the password for Venmo, which matches the original instruction's action and value, with the target now specified. Therefore, the final intent satisfies the addition type criterion.", "heard_transcript": "Change my password to uh blueberry, all lowercase. Oh, I forgot to mention, change the password for uh let's see, Venmo." }, "candidate_id": "wav_4e4a744372f1f8e6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a clear request but then adds a new detail ('change the password for... Venmo') after a longer, slightly awkward pause. The transition feels less fluid than audio_02, and the second part of the request is delivered with a more abrupt shift in focus, making the addition feel less spontaneous. The prosody is slightly less natural, and the overall delivery is less cohesive." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Change my PayPal password to blueberry, all lowercase. Wait, change my Venmo password to blueberry, all lowercase.", "long_pause_offsets": [ 55 ], "source_sha256": "8b48973fe17383927673107e60516f6d3fc39309b72b0c8e7b08f078b491f8b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the password value and case requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/revision/revision_01_app/571faa2ea19d7aec/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "4fdf8df1fe6c62a6276b24dd6fe2d3510bd5c846eb77728f377bfb478d8e7750", "seconds": 14.64, "spoken_utterances": [ "Change my um... PayPal password to ... blueberry, all lowercase. ", "Wait, change my hmmm... Venmo password to blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and the word 'Wait'. This is a clear revision where the final target app is 'Venmo'. All other constraints (password value 'blueberry', all lowercase) remain consistent. The revision criterion is satisfied.", "heard_transcript": "Change my um PayPal password to blueberry, all lowercase. Wait, change my Venmo password to blueberry, all lowercase." }, "candidate_id": "wav_9eaa2f5b018cf455", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request to change a PayPal password, then self-corrects to Venmo with a natural-sounding 'wait' and a brief pause. The correction is smoothly integrated, and the prosody reflects genuine thought processing. The second part of the utterance is delivered with consistent pacing and intonation, making the revision sound spontaneous and credible." } ] }, "is_default": true }, { "variant_id": "revision_02_password", "proposed_instruction": "Change my venmo password to strawberry, all lowercase. Change the password to \"blueberry\", not \"strawberry\".", "long_pause_offsets": [ 55 ], "source_sha256": "e5da7e1330377248c15383f758c00091bc755d7fd920e98a060c5b3f52a30186", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the password value from 'strawberry' to 'blueberry' and retains the lowercase requirement. The app and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/revision/revision_02_password/0617793edc2092f3/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "b581aba991034c6031cca7105a3adea07b1a7e39542cb1438f6aae739f93bb0d", "seconds": 13.56, "spoken_utterances": [ "Change my venmo password to uh... strawberry, all lowercase. ", "Change the password to um... \"... blueberry\", not \"strawberry\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong password ('strawberry') but immediately corrects it to 'blueberry' after a pause and self-correction. According to the revision type criterion, the final corrected value overrides the initial incorrect one. The final intent matches the original instruction exactly, including the lowercase requirement.", "heard_transcript": "Change my Venmo password to a strawberry, all lowercase. Change the password to um, blueberry, not strawberry." }, "candidate_id": "wav_141df347606218c4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request to change a Venmo password to 'strawberry', then revises it to 'blueberry'. The revision is delivered with a noticeable pause and filler ('um'), which is realistic, but the correction feels slightly more abrupt and less fluid than in audio_01. The prosody is less continuous, and the shift from 'strawberry' to 'blueberry' lacks the same natural hesitation and self-correction flow, making it sound slightly more rehearsed or less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "ff6d25673391d825d64cccbbe97405273b0cab6ae1e4c9ba08a6b132920b7d97", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/cancellation/cancel_025/ecd5590fcc600432/speech.wav", "wer": 0.0625, "wav_sha256": "fe93a588e91fc639791ecf3eefa68a7bf0572e2a35b40bd49d1c2079279cb965", "seconds": 11.96, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase.", "Wait, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to change the Venmo password, but immediately follows with 'Wait, forget that request. Don't do it,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Change my um Venmo password to blueberry, all lowercase. Wait, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_039", "hide_variant_original": true, "source_sha256": "2da7ae0a2ded091665aef4876c301f574c36e506e84371ad42e39d67350bfa42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/ambiguous_halt/halt_039/cab73c78a337d6eb/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "3442592b1a3b6e9df6e19e6571b1342442d203ca699e0d07cb4b389c9475edcd", "seconds": 11.16, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase.", "Wait, give me a second first." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, give me a second first,' which is a clear request to pause or hold. The initial instruction to change the password is not canceled, but the final interjection indicates a temporary halt. This matches the 'ambiguous_halt' type, where the final intent is to pause, not execute or cancel.", "heard_transcript": "Change my um Venmo password to blueberry, all lowercase. Wait, give me a second first." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c7340b1672dd590775e342451868dd6d1cd7bac146670a5f56d031e7934d0e56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_2/audio_natural/filler/original/9a0d08095e703b3b/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "9bd3c0ce552a6cc0fa126ad4e3cd008b8de17c42c07da7d163faadd9baa33900", "seconds": 7.2, "spoken_utterances": [ "Change my um... Venmo password to ... blueberry, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler ('um') but continues with the full original instruction without any correction or change in intent. The final audible task matches the original instruction exactly, including the password value and formatting requirement. The filler does not alter the goal, so the execution intent is preserved.", "heard_transcript": "Change my um, Venmo password to blueberry, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_password", "proposed_instruction": "Change my venmo password to watermelon—scratch that, make it pineapple, all lowercase.", "long_pause_offsets": [], "source_sha256": "bea1579250b80e72bb07203ad4a4181ea662df0527e0308606691652865e7128", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces watermelon with pineapple; the app, action, and case requirement remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/false_start/false_start_02_password/11d5e7e5600c37be/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "01317ed9c987afe1358c6286df5eddcafb196dcb9802112c3f643ad1eaef266b", "seconds": 7.76, "spoken_utterances": [ "Change my venmo password to watermelon—um... scratch that, make it ... pineapple, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a wrong password ('watermelon') but immediately self-corrects with 'scratch that' and provides the correct value ('pineapple, all lowercase'). According to the 'false_start' type criterion, the final corrected value must be executed, and the initial incorrect value is discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Change my Venmo password to watermelon. Um, scratch that, make it pineapple, all lowercase." }, "candidate_id": "wav_fff1a9dc81a7471b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a highly natural revision pattern. The speaker first makes a complete request ('Change my Venmo password to watermelon'), then uses a believable self-correction ('Scratch that, make it pineapple all lowercase'). The intonation and pacing of the correction sound spontaneous, with a clear shift in focus and a pragmatic back-reference to the initial request. The flow is continuous and the change-of-mind is expressed in a way that mimics real human hesitation and reformulation." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Change my PayPal—I should say venmo password to pineapple, all lowercase.", "long_pause_offsets": [], "source_sha256": "e37308e198fd823593ce1343de57100785523c8028613339c17656e991132859", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the password value and case requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/false_start/false_start_01_app/5415ad412537795f/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "0e199082a7dca3d87e4fb3987c66ab1da760b3019e00b256ed8e87445b167554", "seconds": 6.32, "spoken_utterances": [ "Change my PayPal—um... I should say ... venmo password to pineapple, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say'. The final instruction clearly specifies the correct app (Venmo), the new password ('pineapple'), and the required format ('all lowercase'). This matches the false_start type criterion, where the immediate self-correction overrides the initial mistaken value. The final intent is clear and executable.", "heard_transcript": "Change my PayPal um, I should say Venmo password to pineapple, all lowercase." }, "candidate_id": "wav_7e76aabed2d902cc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a false start where the speaker initially says 'PayPal' and then corrects to 'Venmo'. While the self-repair is present, the correction is very brief and the overall delivery is flatter and less prosodically varied than audio_02. The transition from 'PayPal' to 'Venmo' feels slightly abrupt, and the rest of the sentence lacks the natural rhythm and emphasis seen in the more fluent revision of audio_02. It is still realistic but less dynamically natural." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7574325_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Change my password to pineapple, all lowercase. Oh, I forgot to mention, change the password for Venmo.", "long_pause_offsets": [ 48 ], "source_sha256": "0b31c13d81eec2ed575fff6886bc5b30ef3df4b4227dfbf170ff719f4709488a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the app as Venmo, which was omitted in the initial utterance but is present in the original instruction. The password value and case requirement are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my password to uh... pineapple, all lowercase. ", "Oh, I forgot to mention, change the password for um... Venmo." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Change my venmo password to, all lowercase. I left out a detail: the new password should be pineapple.", "long_pause_offsets": [ 44 ], "source_sha256": "e5baf29b93d0603cab2dd8f3b33886617561aacb46340ebbecae0f4bf2abc7e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing password value 'pineapple' and retains the 'all lowercase' constraint from the initial utterance. The final goal matches the original instruction exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my uh... venmo password to, all lowercase. ", "I left out a detail: the new password should be uh, let me think... pineapple." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "7574325_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Change my PayPal password to pineapple, all lowercase. Wait, change my Venmo password to pineapple, all lowercase.", "long_pause_offsets": [ 55 ], "source_sha256": "03efe38b6eb0fcca12be7bf82e354c332b45468cfa766ee805a00061a90e40e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the password value and case requirement." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my PayPal password to um, let me think... pineapple, all lowercase. ", "Wait, change my ... Venmo password to pineapple, all lowercase." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_password", "proposed_instruction": "Change my venmo password to watermelon, all lowercase. Change the password to \"pineapple\", not \"watermelon\".", "long_pause_offsets": [ 55 ], "source_sha256": "8375acc85786ce1257ff66919a89e92b4fe823519e35686b89da449b12d987dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the password value from 'watermelon' to 'pineapple' and maintains the lowercase requirement." }, "audio": { "status": "pending", "spoken_utterances": [ "Change my uh... venmo password to watermelon, all lowercase. ", "Change the password to \"... pineapple\", not \"watermelon\"." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "7574325_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "2d951c8fc625df19754ccc5dfd1fce3e3f03e3e5ceae04896e503ccf439b8abb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/cancellation/cancel_029/6533e910e9037beb/speech.wav", "wer": 0.0, "wav_sha256": "4a9fe401f702d4be1373a1f7a7f969f4c085f84f04beeaa9390a4a491832d7e1", "seconds": 10.96, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to change the Venmo password, but immediately follows with 'Hang on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection confirms the cancellation intent, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "49be5d08a04b9dc91de19558bb2addbc494916cff81086dbf6d265574c433a16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/ambiguous_halt/halt_031/67f907be70e7ead3/speech.wav", "wer": 0.0, "wav_sha256": "91070b7d6002d4ae36095428914a4eea5aa16b28a3517df0c27771b5e5568c9c", "seconds": 10.04, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. This constitutes an ambiguous_halt as per the rubric, where the speaker indicates a need to stop and think without canceling the task. The final interjection is a pause request, not a cancellation or a new instruction.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7574325_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "58c8eee4eddc6b55e5cb90928a50c4dca4a19fb88748b8063b8c31c5616e2694", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7574325_3/audio_natural/filler/original/1cccb2c89ab55ebf/speech.wav", "wer": 0.0, "wav_sha256": "ba9500df1e7a7fd06384ac32e22ae69a9a919aa71f3a88a777dded8c5ce505f6", "seconds": 6.4, "spoken_utterances": [ "Change my ... Venmo password to ... pineapple, all lowercase." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker clearly states the full instruction to change the Venmo password to 'pineapple' in all lowercase. The 'filler' type allows for minor hesitations or fillers, but the final intent is clear and matches the original instruction. No errors or omissions are present in the final utterance.", "heard_transcript": "Change my Venmo password to pineapple, all lowercase." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 3654328627—sorry, that should be 3654328626 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "cd03732f16fcc735128f9e99d0949fd7ab58b3275d637d5742776d6263569523", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in seven, then corrects it to end in six. The final number matches the original instruction, and all other conditions (message types, action, and intent) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/false_start/false_start_02_phone_number/0cde7f08af45041c/speech.wav", "wer": 0.10256410256410256, "wav_sha256": "1a3ebe0d5b611d005a92cc2b59e62e057b4b76b4aaac48be57f5a5575e1214ca", "seconds": 16.88, "spoken_utterances": [ "All phone text messages and voice messages from uh... three six five, four three two, eight six two seven—ah... sorry, that should be three six five, four three two, eight six two six uh... are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the number as 3654328627, then immediately self-corrects to 3654328626. According to the 'false_start' type criterion, the final intent is the corrected version. The action 'delete them' remains unchanged. The final audible task matches the original instruction after the correction.", "heard_transcript": "All phone text messages and voice messages from uh three six five four three two eight six two seven uh sorry, that should be three six five four three two eight six two six uh are spam, delete them." }, "candidate_id": "wav_852a6be7d6b79118", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects the phone number from '8627' to '8626' with a natural-sounding 'uh, sorry, that should be...' followed by a smooth continuation. The prosody includes a believable pause and filler, and the correction is integrated seamlessly into the flow. The delivery sounds like a real person recalling and fixing a detail mid-sentence." } ] }, "is_default": true }, { "variant_id": "false_start_01_message_channel_scope", "proposed_instruction": "All phone text messages—or rather, phone text messages and voice messages from 3654328626 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "e8beb2ba51f879a66b7896aa717a226c9aa3669e07ff53704f87398c80782628", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' and then corrects to include 'and voice messages', which matches the original instruction. The phone number is correctly spoken as 'three six five, four three two, eight six two six', which is equivalent to 3654328626. The action and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/false_start/false_start_01_message_channel_scope/15eab67a16b508b8/speech.wav", "wer": 0.27586206896551724, "wav_sha256": "9471b48a35f6bf37c8a993110b062d8adc20f87fa86bdb708d7acac48432c402", "seconds": 12.08, "spoken_utterances": [ "All phone text messages—um... or rather, phone text messages and voice messages from three six five, four three two, eight six two six hmmm... are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' but immediately corrects to 'phone text messages and voice messages'. The phone number is correctly stated as 'three six five four three two eight six two six', matching the original. The final instruction to delete spam messages from that number is clear and complete. The correction is immediate and does not change the scope of the task beyond adding voice messages, which is a reasonable clarification. The control intent is to execute the corrected task.", "heard_transcript": "All phone text messages um or rather phone text messages and voice messages from three six five four three two eight six two six are spam, delete them." }, "candidate_id": "wav_63861d6ccf0efb6f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with 'All phone text messages um...' and then corrects to 'or rather, phone text messages and voice messages...', which is a plausible self-correction. However, the correction feels slightly more abrupt and less integrated than in audio_01. The 'um' and 'or rather' are natural, but the overall flow is less smooth, and the second part of the sentence lacks the same level of spontaneous integration seen in the top-ranked clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "All phone text messages and voice messages from are spam, delete them. One more detail: these are from the number 3654328626.", "long_pause_offsets": [ 71 ], "source_sha256": "eaae1336a73cf63052f5b11a96821cb08605b54ee9d6206947f952b93534e4a8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing phone number, which matches the original instruction's number. The initial utterance omits the number but otherwise matches the final goal. The addition fills the missing value without altering other requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/addition/addition_02/4840821def2a04e9/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "bcc343e83180bdcd99fad618d985a788bded548051db337b630400e8d64e4cad", "seconds": 16.12, "spoken_utterances": [ "All ... phone text messages and uh... voice messages from ... are spam, delete them. ", "One more detail: these are from the number three six five, four three two, eight six two six." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to delete spam messages from a specific number. The speaker initially omits the number but then adds it clearly after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added detail. The number is correctly stated as 'three six five, four three two, eight six two six', which matches the original reference number. No other constraints are changed or removed.", "heard_transcript": "All phone text messages and uh voice messages from are spam, delete them. One more detail, these are from the number three six five, four three two, eight six two six." }, "candidate_id": "wav_9c6c78789c1bd92a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to delete spam messages, then adds a specific detail about the sender's number. The addition is delivered with a slight pause and a conversational tone, sounding like a spontaneous recollection rather than a rehearsed or robotic insertion. The prosody is consistent, and the number is pronounced clearly without hesitation." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "All from 3654328626 are spam, delete them. Oh, I forgot to mention, this should apply to both phone text messages and voice messages.", "long_pause_offsets": [ 43 ], "source_sha256": "af2e356729a88e09a9306084eedbaf9ce48ade1eca75873a71783f1dded39f73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing message type 'voice messages' to the initial instruction, which already included 'phone text messages'. The original instruction's phone number and action are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/addition/addition_01/a6f9650c2013a38c/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "2203be338b99caebca5394250b81f4569980f86692a42b09c0fa5d3c0cca328a", "seconds": 19.08, "spoken_utterances": [ "All from um... three six five, four three two, eight six two six ... are spam, delete them. ", "Oh, I forgot to mention, this should apply to both let's see... phone text messages and voice messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'All from 3654328626 are spam, delete them.' This is followed by an addition: 'this should apply to both, phone text messages and voice messages.' Since the type is 'addition', the final intent combines the original instruction with the added constraint. The final goal is to delete spam messages from the specified number, now explicitly including both text and voice messages. The addition clarifies the scope without contradicting the original, so the combined task matches the reference instruction.", "heard_transcript": "All from um, 3654328626 are spam, delete them. Oh, I forgot to mention, this should apply to both, let's see, phone text messages and voice messages." }, "candidate_id": "wav_d1039ef6f2185bce", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant 'um' and delivers the core request clearly. The addition is introduced with 'Oh, I forgot to mention,' which is a common and realistic conversational repair. However, the phrase 'this should apply to both' followed by 'let's see' introduces a slightly awkward, less fluent transition that sounds like a minor stumble or a moment of formulating the thought, making it less smooth than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_message_channel_scope", "proposed_instruction": "All phone text messages from 3654328626 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages.", "long_pause_offsets": [ 63 ], "source_sha256": "60989f3298609ddc99d7f40497755150db69d2d469962043f991205d8395d167", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the message type scope from only phone text messages to include both phone text and voice messages, while retaining the sender number and the delete action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/revision/revision_01_message_channel_scope/ecbe980b1a2b7993/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "2eb9ac09f065563b952db0e00350918739b0f94276f48d51b7491f59f445d046", "seconds": 19.36, "spoken_utterances": [ "All phone text messages from um... three six five, four three two, eight six two six uh... are spam, delete them. ", "Wait, the messages should be um... phone text messages and voice messages, not just phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests deletion of only 'phone text messages' from the number, but then corrects this to include 'phone text messages and voice messages'. This is a revision of the message type to be deleted. The final intent is to delete both types of messages, which matches the original instruction. The correction is clear and applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "All phone text messages from um, 3654328626 uh, are spam, delete them. Wait, the messages should be um, phone text messages and voice messages, not just phone text messages." }, "candidate_id": "wav_88f94dab86b60ffb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'wait' before revising the message type from 'phone text messages' to include 'voice messages'. The correction is smoothly integrated into the same utterance with a believable self-repair ('should be... not just...'), and the prosody reflects a genuine moment of reconsideration. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 3654328627 are spam, delete them. Change the number to 3654328626 instead of 3654328627.", "long_pause_offsets": [ 82 ], "source_sha256": "9217a331c7a55b3d38ed6b82a03be591482edf24713a59df8717425293f98aaf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the phone number from 3654328627 to 3654328626, which differs in the last digit." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/revision/revision_02_phone_number/6a9715e928c21aac/speech.wav", "wer": 0.0, "wav_sha256": "fa37624384de3d70e52ab7d62e97bb6f172d5f381586806e51605f85859fb8ec", "seconds": 23.4, "spoken_utterances": [ "All uh, let's see... phone text messages and voice messages from three six five, four three two, eight six two seven are spam, um... delete them. ", "Change the number to three six five, four three two, eight six two six instead of three six five, four three two, eight six two seven." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a mistaken number (8627) but immediately corrects it to 8626 in the revision clause. The final intent is to delete spam messages from the corrected number, matching the original instruction. The revision type criterion is satisfied as the correction is spoken after a pause and applies to the earlier instruction.", "heard_transcript": "All uh let's see, phone text messages and voice messages from three six five four three two eight six two seven are spam, um delete them. Change the number to three six five four three two eight six two six instead of three six five four three two eight six two seven." }, "candidate_id": "wav_cfa31f19d4912d38", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a filler ('uh') and a hesitation ('let's see'), which is realistic, but the revision is delivered as a separate, post-hoc instruction rather than a self-corrected clause. The change of number is stated as a new command ('Change the number to...'), which sounds more like a task instruction than a natural revision. The flow feels less integrated and more like a two-part command, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "9ceef6b0db06463cfd8004dad3cb82795543d50d4a5f37cdfd813ef62ebd18a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/cancellation/cancel_009/c338d42ab1b1862e/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "0300bc2671e163fac5bd8a69b920d37cf36809895c0997fe96429eb131516dde", "seconds": 16.04, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a deletion task, then immediately follows with 'Hang on, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "All phone text messages and voice messages from 3654328626 are spam, um, delete them. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_001", "hide_variant_original": true, "source_sha256": "a43005343635c827df2f7517d0c95c72717778c12c0911af1dae25f3f5722881", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/ambiguous_halt/halt_001/8d2ff83133b3069c/speech.wav", "wer": 0.12, "wav_sha256": "936ba5f35679181d47be590be6b5c8e941efea9e1b41447ca839e1366f031777", "seconds": 15.08, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them.", "Wait..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait,' after completing the instruction to delete messages. This constitutes a clear request to pause or hold the current action, consistent with the 'ambiguous_halt' type. The final interjection is not a cancellation or a correction, but a directive to stop and reconsider, which matches the expected behavior for this type.", "heard_transcript": "All phone text messages and voice messages from 3654328626 uh are spam, um delete them. Wait." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ac81a0cb1a1462fcdb8eea20da8b48aa1a8a15b805453f5bd46672b150765c1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_1/audio_natural/filler/original/b9cfd095475a6bd7/speech.wav", "wer": 0.125, "wav_sha256": "8bb675f766eaf46e946efd0719df2228192fe58906610e0f5725e986619fa639", "seconds": 12.32, "spoken_utterances": [ "All ... phone text messages and voice messages from three six five, four three two, eight six two six uh... are spam, um... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a brief hesitation ('um') but does not alter the core instruction. The final control request is 'delete them', which matches the original task. The hesitation is a filler and does not change the goal. The final intent is to execute the deletion of messages from the specified number.", "heard_transcript": "All phone text messages and voice messages from 3654328626 are spam, um, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_message_channel_scope", "proposed_instruction": "All phone calls and voice messages—sorry, I meant phone text messages and voice messages from 9294880327 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "96b024a94691ad302cf6338b409e082e5ab469691fc80db6f9f021db223e3a27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone calls' but immediately corrects to 'phone text messages' within the same utterance. The final intended message type is 'phone text messages and voice messages', and the phone number is correctly specified as 9294880327. The action and target remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/false_start/false_start_01_message_channel_scope/480e66cb753ce259/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "38298e3b693177b877a01f57618fbae269835cbb1723ab7919ebf4c4e5cacd50", "seconds": 12.2, "spoken_utterances": [ "All um... phone calls and voice messages—uh... sorry, I meant phone text messages and voice messages from nine two nine, four eight eight, zero three two seven are spam, hmmm... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone calls' but immediately self-corrects to 'phone text messages'. The final intent is clear and matches the original instruction with the correction applied. The type is 'false_start', so the immediate self-correction is preserved and the mistaken value is discarded. All other constraints, including the phone number and action, are consistent.", "heard_transcript": "All um phone calls and voice messages, uh sorry, I meant phone text messages and voice messages from nine two nine four eight eight zero three two seven are spam, delete them." }, "candidate_id": "wav_6f35c977d97706e5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural false start ('All um phone calls...'), immediately self-corrects with 'sorry, I meant...', and smoothly continues with the correct request. The correction is integrated into the flow without awkwardness, and the delivery includes a believable hesitation and a natural pause before the final command. The prosody and rhythm are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 9294880328—no, that should be 9294880327 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "3a9dd2dddea50ae5fb84a3b01237c4f7f2cec9d6a365911df264f857fbf068e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in eight, then immediately corrects it to seven. The final number matches the original instruction, and all other conditions (message types, action, and intent) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/false_start/false_start_02_phone_number/e4ab4f6391c558e1/speech.wav", "wer": 0.24324324324324326, "wav_sha256": "61a668e8a80219f6051df1548d81d8a16c75c80b0a741036419aae85c5e04fd0", "seconds": 16.68, "spoken_utterances": [ "All phone text messages and voice messages from ... nine two nine, four eight eight, zero three two eight—um... no, that should be nine two nine, four eight eight, zero three two seven ... are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the wrong phone number (8 instead of 7) but immediately self-corrects to the correct one. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the reference exactly.", "heard_transcript": "All phone text messages and voice messages from 9294880328 are spam, delete them. No, that should be 9294880327." }, "candidate_id": "wav_d0be88989d64f34f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear initial request, then pauses and corrects the phone number. While the correction is plausible, the delivery feels slightly more rehearsed or less fluid than audio_01. The pause before the correction is longer and more abrupt, and the correction itself is delivered with a flatter intonation, making it sound less like a spontaneous thought process and more like a deliberate edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "All phone text messages and from 9294880327 are spam, delete them. And make sure to include voice messages too.", "long_pause_offsets": [ 67 ], "source_sha256": "0550fcf8b52ba373e57f2ffa2b283ea2b14554a7d0a3a71c8b78968b476bbc1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include voice messages, which was not in the initial utterance but is consistent with the original instruction. The phone number and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/addition/addition_02/d986c6fb8d92e6a1/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "a68c2f43b5e5fb61f4da45ef9a84701ee2bcff47425c1a902cc3b01d72bfaf98", "seconds": 17.36, "spoken_utterances": [ "All um... phone text messages and um, let me think... from nine two nine, four eight eight, zero three two seven are spam, delete them. ", "And make sure to include ... voice messages too." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the reference. The speaker then adds 'And make sure to include voice messages too' after a pause, which is a clear addition to the original request. The final intent combines both the original and added constraints, fully satisfying the 'addition' type criterion.", "heard_transcript": "All um phone text messages and um let me think from nine two nine four eight eight zero three two seven are spam, delete them. And make sure to include voice messages too." }, "candidate_id": "wav_1954a690f4a8eafa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural, spontaneous fillers ('um', 'let me think') that are well-placed and not overused. The delivery has a realistic, hesitant rhythm, and the addition of 'voice messages too' at the end sounds like a remembered detail rather than a rehearsed command. The prosody is varied and the pacing reflects genuine thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "All and voice messages from 9294880327 are spam, delete them. Also, include phone text messages.", "long_pause_offsets": [ 62 ], "source_sha256": "2d2a1f0db043c44dc3f04d481c721f58e2ca57afebf08093a7928abb2149b7d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'phone text messages' category, completing the original instruction which already included voice messages and the sender number. The final goal matches the original in scope and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/addition/addition_01/b30d0182333493d7/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "0d7da93664111678abd8b7471695d61b277ab19f9d14c28e40fd08bd6f06f9c3", "seconds": 16.36, "spoken_utterances": [ "All and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, um... delete them. ", "Also, include phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is to delete spam messages from a specific number. The speaker initially says 'All and uh voice messages' and then adds 'Also include phone text messages' after a pause. This is an addition that expands the scope of the original instruction to include text messages. The final intent is to delete both voice and text messages from the specified number, which matches the original instruction plus the added constraint. The addition is clear and does not contradict the original request.", "heard_transcript": "All and uh voice messages from nine two nine four eight eight zero three two seven are spam. Um, delete them. Also include phone text messages." }, "candidate_id": "wav_8d38891906deb2eb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the content is similar, the delivery is slightly more robotic and less fluid. The initial 'all and uh' sounds like a stumble, but the rest of the sentence is delivered in a more read-aloud manner. The addition of 'also include phone text messages' at the end feels tacked on rather than naturally recalled, and the overall prosody is flatter compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_message_channel_scope", "proposed_instruction": "All phone text messages from 9294880327 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages.", "long_pause_offsets": [ 63 ], "source_sha256": "5a6d0470d14c9c0be2e34b924bfd5a8bd96f3e094f5413d87bd5e42b7b28d9ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the message type from 'phone text messages' to include 'voice messages', matching the original instruction's scope. The phone number and action (delete spam) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/revision/revision_01_message_channel_scope/19ef18dd0b008b45/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "23367f4f8b9304c035ce4369e07af048980520cfea81a08c10dcfc2af9ee03f1", "seconds": 19.92, "spoken_utterances": [ "All phone text messages uh, let me think... from nine two nine, four eight eight, zero three two seven are spam, ... delete them. ", "Wait, the messages should be ... phone text messages and voice messages, not just phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' but immediately corrects to 'phone text messages and voice messages' after a pause. This is a revision of the message type. The final intent includes both message types, which matches the original instruction. The phone number and action (delete spam) remain unchanged. Therefore, the revision is correctly applied and the task is fully executable.", "heard_transcript": "All phone text messages, let me think, from 9294880327 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages." }, "candidate_id": "wav_ddaf3518202f85ed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), delivers a clear initial request, then self-corrects with a plausible revision ('wait, the messages should be...') that smoothly integrates the missing 'voice messages' and corrects the phone number. The prosody is consistent, the correction is embedded within the same utterance, and the pacing reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "revision_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 9294880328 are spam, delete them. Change the number to 9294880327 instead of 9294880328.", "long_pause_offsets": [ 82 ], "source_sha256": "9a2893621050ca4e767bd5a3c56f0760f62d36d50a3a3b1066d16a86424cc5c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source number from 9294880328 to 9294880327, which matches the original instruction. The action and message type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/revision/revision_02_phone_number/ac012ccdf3fe04bc/speech.wav", "wer": 0.09259259259259259, "wav_sha256": "ccf17a17975a5d6095aa87ce7010ae6776989604125293c14205e0e20863edbd", "seconds": 22.52, "spoken_utterances": [ "All ... phone text messages and let's see... voice messages from nine two nine, four eight eight, zero three two eight um, let me think... are spam, delete them. ", "Change the number to nine two nine, four eight eight, zero three two seven instead of nine two nine, four eight eight, zero three two eight." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task for number 9294880328, then immediately corrects it to 9294880327. This is a revision where the final intent is the corrected version. The final instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "All uh phone text messages and let's see, voice messages from 9294880328 um let me think are spam, delete them. Change the number to 9294880327 instead of 9294880328." }, "candidate_id": "wav_ba4b676b8ea5efcf", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker also uses natural fillers ('uh', 'let's see', 'um', 'let me think') and delivers a revision. However, the correction is delivered as a separate, post-hoc statement ('Change the number to...') rather than being integrated into the original request. This makes the revision feel more like a task instruction than a spontaneous self-correction, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "8668665d70f7b6c4e8f1fd9bf0e72c0afc51361eabf54289945f6c8553fd11d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/cancellation/cancel_006/2d6ec7cf264ca9a4/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "c0d5206bd280a5f2960c32f4a136dc6fa8fe9d07b2f8eb529cf1e8c47cacf7b4", "seconds": 16.0, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them.", "Oh, wait, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a complete instruction to delete spam messages from a specific number. However, the speaker immediately follows with 'Oh, wait, please cancel the entire task,' which is a clear and unambiguous cancellation of the prior request. According to the cancellation criterion, the recording must clearly revoke/cancel the task, which it does. The final interjection is a direct cancellation command, not a pause or a request to wait. Therefore, the final intent is cancellation, and the task should not be executed.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam. Delete them. Oh, wait, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "24de9fc93801299518a2ed0f030e69d80648db632370c572ffd73038f0f971f6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/ambiguous_halt/halt_037/590d99201e7aa1d4/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "e349ff664a9ffe1a2f86271c7e392113724581b6a855731aa8c28f739ff39e0d", "seconds": 14.72, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the original instruction to delete spam messages from 9294880327, followed by the explicit final interjection 'Pause for a moment.' This constitutes a clear request to hold or wait, with no indication of cancellation or resumption. According to the rubric, such a final interjection qualifies as an 'ambiguous_halt' where the agent should pause. The control intent is correctly identified as 'pause'.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam. Delete them. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a5da11abfbd8f94d05140473405402224aa9881d640a6e5a9ee87d8b4fbed6a0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_2/audio_natural/filler/original/f824dab3c492049a/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "7b0114b68d3f599bbe7bf225dddf5a3faf54bd8a883d53395371af5ece6e8624", "seconds": 11.08, "spoken_utterances": [ "All phone text messages and uh... voice messages from nine two nine, four eight eight, zero three two seven are spam, ... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'uh' but continues the original instruction without changing the action, target, or value. The phone number is clearly stated and matches the reference. The final control intent is to execute the deletion task as originally requested.", "heard_transcript": "All phone text messages and uh voice messages from nine two nine four eight eight zero three two seven are spam, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 5708520673—sorry, that should be 5708520672 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "705b7ea6ec572785efa2d250bfa55e093a9cba4bbf304650e8bc78c592536c99", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in three, then immediately corrects it to end in two. The final number matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/false_start/false_start_02_phone_number/2674626f4c44cd0c/speech.wav", "wer": 0.0, "wav_sha256": "0b662333ecd0a75957c6e9aebd8e70a01a34a121827a94e0ce9d9ebce9724df2", "seconds": 18.36, "spoken_utterances": [ "All phone text messages and voice messages from um, let me think... five seven zero, eight five two, zero six seven three—um... sorry, that should be five seven zero, eight five two, zero six seven two are spam, delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the phone number as ending in 'three' but immediately self-corrects to 'two'. The final number spoken is 5708520672, which matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion. The task intent remains consistent and executable.", "heard_transcript": "All phone text messages and voice messages from um, let me think, five seven zero eight five two zero six seven three um, sorry, that should be five seven zero eight five two zero six seven two are spam, delete them." }, "candidate_id": "wav_1c6cf6eef3cdb699", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'um' followed by a plausible hesitation ('let me think'), then delivers a clear initial number. The self-correction ('sorry, that should be...') is spoken with a slightly lower pitch and faster pace, typical of a spontaneous correction. The number is corrected from 0673 to 0672, and the sentence flows smoothly into the final command. The prosody and rhythm are consistent with a person recalling and correcting a detail in real time." } ] }, "is_default": true }, { "variant_id": "false_start_01_message_channel_scope", "proposed_instruction": "All phone text messages—or rather, phone text messages and voice messages from 5708520672 are spam, delete them.", "long_pause_offsets": [], "source_sha256": "66de4839b3daa1dcfef07690d20749c9bd177bc43eea5aa4bf27d7a19a47337e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' but immediately corrects to include 'and voice messages'. The phone number is correctly spoken in full and matches the reference. The final goal includes both message types and the correct number." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/false_start/false_start_01_message_channel_scope/dfdb9ecae614351d/speech.wav", "wer": 0.0, "wav_sha256": "0e7411fb3544241d7379338b9b043a4593bf8144dfeb1132a8846c169b6d4511", "seconds": 14.56, "spoken_utterances": [ "All phone text messages—um... or rather, phone text messages and voice messages from five seven zero, eight five two, zero six seven two are spam, let's see... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'All phone text messages um', then self-corrects with 'or rather, phone text messages and voice messages'. This is a false start where the speaker corrects their own phrasing mid-sentence. The final, corrected instruction matches the original reference exactly in action, target, and value. The control intent is to execute the deletion task.", "heard_transcript": "All phone text messages um, or rather, phone text messages and voice messages from five seven zero eight five two zero six seven two are spam. Let's see. Delete them." }, "candidate_id": "wav_dcb4d7fc6db9c841", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with 'um' and then corrects 'phone text messages' to 'phone text messages and voice messages,' which is a plausible self-repair. However, the correction is less fluid than in audio_01, with a slight pause and a more hesitant tone. The phrase 'let's see' at the end sounds like a filler or a moment of hesitation, but it's not clearly connected to a new thought or correction. The overall delivery is slightly less natural due to the less seamless flow of the self-correction and the ambiguous ending." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "All from 5708520672 are spam, delete them. I left out a detail: only the phone text messages and voice messages.", "long_pause_offsets": [ 43 ], "source_sha256": "bd6d000e7dc742a66a94e1f94208b537d304eb91b7d03030cb8459da33a50b21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that limits the deletion to only phone text and voice messages, which is consistent with the original instruction. The original instruction already specified the sender and action, and the followup refines the message type without contradicting any other part." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/addition/addition_02/64d258bcd31093f7/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "f60749cae03a323b157b341c68e380b365b9228b7fb767b88a5c04a0a178315b", "seconds": 17.16, "spoken_utterances": [ "All ... from five seven zero, eight five two, zero six seven two um... are spam, delete them. ", "I left out a detail: only the uh... phone text messages and voice messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the full instruction to delete spam messages from the number. After a pause, they add a clarifying detail that the action should apply only to 'phone text messages and voice messages', which is a narrowing of the original instruction. Since the original instruction already specified 'phone text messages and voice messages', the addition does not change the scope. The final intent remains to delete spam messages from the specified number, and the addition is consistent with the original. Therefore, the task is executed as intended.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, delete them. I left out a detail, only the uh phone text messages and voice messages." }, "candidate_id": "wav_0bd870ac1e8e01c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to delete spam messages from a specific number. After a brief pause and filler ('um'), they add a detail about only deleting phone, text, and voice messages. The addition is smoothly integrated with a natural-sounding 'I left out a detail' self-correction, and the prosody and pacing are consistent with spontaneous speech. The flow feels like a real-time recollection and refinement of a task." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "All phone text messages and voice messages are spam, delete them. Oh, I forgot to mention, these are from the number 5708520672.", "long_pause_offsets": [ 66 ], "source_sha256": "06b7ae8dca218dbd33f676c61fec4fcfe5608de9a9bef50a8b5e259d1132eaee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing sender number, which matches the original instruction. The action and message type remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/addition/addition_01/41a0d9a581fe8508/speech.wav", "wer": 0.09375, "wav_sha256": "f12a86f40ee0ac50f4b6d3c799085a5d4d60f3be50ca2ec6d389c26ce225502a", "seconds": 17.92, "spoken_utterances": [ "All phone text messages and voice messages hmmm... are spam, delete them. ", "Oh, I forgot to mention, these are from the number ... five seven zero, eight five two, zero six seven two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'All phone text messages and voice messages are spam, delete them.' This is a broad request without a source number. After a pause and an interjection, the speaker adds 'these are from the number 5708520672,' which narrows the scope of the original instruction. The addition specifies the sender of the messages, completing the task as a combined request. The final intent matches the original instruction with the added constraint, satisfying the 'addition' type criterion.", "heard_transcript": "All phone text messages and voice messages, um, are spam, delete them. Oh, I forgot to mention, these are from the number 5708520672." }, "candidate_id": "wav_d381c1e4e23aac62", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a complete request to delete spam messages from a specific number, but the phrasing is slightly more formal and less conversational than audio_01. After a pause and filler ('um'), they add a detail about the message types. The addition is less smoothly integrated and sounds more like a separate thought than a spontaneous correction. The prosody is flatter, and the overall delivery is less dynamic, making it sound less like a natural, in-the-moment thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_phone_number", "proposed_instruction": "All phone text messages and voice messages from 5708520673 are spam, delete them. Let me correct the number: messages from 5708520672, not 5708520673.", "long_pause_offsets": [ 82 ], "source_sha256": "0ef64fd4d8dea5d74550582487cdffc2a5cca8e5865ba048f5ed2c71cb0482b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up revision corrects the phone number from 5708520673 to 5708520672, which matches the original instruction. The action, message type, and deletion directive remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/revision/revision_02_phone_number/c39996c591e87e24/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "67105c80d782d4522201b41082e168d6db4444afd9d0a5b2d9d9b48789c6d606", "seconds": 24.24, "spoken_utterances": [ "All uh... phone text messages and voice messages from five seven zero, eight five two, zero six seven three let me think... are spam, delete them. ", "Let me correct the number: messages from five seven zero, eight five two, zero six seven two, not five seven zero, eight five two, zero six seven three." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the wrong number (5708520673), then immediately corrects it to 5708520672 after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "All uh phone text messages and voice messages from 5708520673, let me think. are spam, delete them. Let me correct the number. Messages from 5708520672, not 5708520673." }, "candidate_id": "wav_5929b561183d0df9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural hesitation ('let me think') and a self-repair that flows smoothly within the same utterance. The correction of the phone number is immediate and integrated without restarting the sentence, showing a realistic thought process. The prosody is consistent, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "revision_01_message_channel_scope", "proposed_instruction": "All phone text messages from 5708520672 are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages.", "long_pause_offsets": [ 63 ], "source_sha256": "7176c0c20b520d53826598840636ac8ea16a469e6fd3484f0e18f9dcfd48abfd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the message type from 'phone text messages' to include 'voice messages', matching the original instruction's scope. The phone number and action (delete spam) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/revision/revision_01_message_channel_scope/5eca102854f075a3/speech.wav", "wer": 0.024390243902439025, "wav_sha256": "f7b065434116130bf90b166b846bd2d3064cee54c48c8ddd08caa05f7173d1f0", "seconds": 21.36, "spoken_utterances": [ "All um, let me think... phone text messages um... from five seven zero, eight five two, zero six seven two uh... are spam, delete them. ", "Wait, the messages should be phone text messages and voice messages, not just phone text messages." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'phone text messages' but immediately corrects themselves to 'phone text messages and voice messages' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to include both text and voice messages, which matches the original instruction. The phone number and action (delete spam) remain unchanged. Therefore, the final intent is fully aligned with the original instruction after the revision.", "heard_transcript": "All um, let me think. Phone text messages um, from 5708520672 uh, are spam, delete them. Wait, the messages should be phone text messages and voice messages, not just phone text messages." }, "candidate_id": "wav_c855da223376e447", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker includes natural fillers ('um', 'let me think') and pauses, but the revision occurs as a separate, completed sentence after a longer pause. While the correction is plausible, the shift from 'phone text messages' to 'phone text messages and voice messages' feels more like a post-hoc addition than a real-time self-correction. The delivery is slightly less fluid than audio_01, with a more noticeable break between the initial request and the revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "e2f7f48a2bdadeea4aa246bb5f18f616ecac858c57b045dbc1b240ef4dd277fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/cancellation/cancel_029/1e10a5ea04ecda8e/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "53dacb82d1fb2875ec6e8a0a9cc98ad4ed43d3ba6c557ee8fb76d3939a6329b8", "seconds": 15.56, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them.", "Hold on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a task to delete messages from a specific number, then immediately follows with 'Hold on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "All phone text messages and voice messages from five seven zero eight five two zero six seven two are spam, uh, delete them. Hold on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "9232d9ceb9214adb09b1b084b14a3f47bc25b0001bb33167179f095e018401ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/ambiguous_halt/halt_040/3b93d7c07de7b475/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "b29606eec566f92eafeae4737addf9a7ab1f16bd1d855321e08c918ee8f77f3e", "seconds": 14.84, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, just a minute,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or provide a new instruction. The initial request to delete spam messages is not revoked, but the speaker explicitly halts execution. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, uh, delete them. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "771d8fc_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ada2b046166947299ba8f7656c381f25f26c258666bd9513547d433cccd24a09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/771d8fc_3/audio_natural/filler/original/97b83febd2ea4a75/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "845850b4ac20f319abbd135e38e5d390367488cd5e3a820bbf1e168f9751d54c", "seconds": 11.32, "spoken_utterances": [ "All phone text messages and ... voice messages from five seven zero, eight five two, zero six seven two are spam, uh... delete them." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task with a filler word 'uh' but immediately proceeds to complete the original instruction without any correction or change in intent. The final audible command is to delete the messages, matching the original instruction. The filler does not alter the goal, so the task is executed as intended.", "heard_transcript": "All phone text messages and voice messages from 5708520672 are spam, uh, delete them." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Place an order for everything in my eBay—sorry, I meant amazon cart and wishlist for my home address.", "long_pause_offsets": [], "source_sha256": "760052ddfd6ec22b02c9ccea195900a6f1c0c46f55f13d66e507b63a16735f5f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'eBay' to 'Amazon' within the same utterance, preserving the action, scope (cart and wishlist), and destination (home address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/false_start/false_start_01_app/a94ba26e59fc2bba/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "cac51569bb73e603c610d42e56aacfb01fc0763ebc469a629c82251c091d1906", "seconds": 12.92, "spoken_utterances": [ "Place an order for everything in my eBay—ah... sorry, I meant um... amazon cart and ... wishlist for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' after a brief hesitation. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('eBay') is discarded. The final instruction matches the original reference instruction in both action and target, with no other changes.", "heard_transcript": "Place an order for everything in my eBay uh uh sorry, I meant um Amazon cart and wishlist for my home address." }, "candidate_id": "wav_75011ad82f143d17", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('eBay'), self-corrects smoothly to 'Amazon', and continues with a natural, flowing delivery. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause and filler 'uh uh' are placed at a believable point of hesitation, and the overall rhythm is that of someone thinking aloud while speaking." } ] }, "is_default": true }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart—or rather, cart and wishlist for my home address.", "long_pause_offsets": [], "source_sha256": "919edcb2642f3a812433ff2df6e1b0f27cc6f58b6430b332f183ba2f492cfa91", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'cart' to 'cart and wishlist', expanding the scope to include both collections. The app (Amazon), action (place an order), and destination (home address) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/false_start/false_start_02_source_collection/3bd49fa046115092/speech.wav", "wer": 0.1, "wav_sha256": "37a437a52130872f0842f7d9a5d6b35b4692934cdbc725fc3d38c5840035d832", "seconds": 10.88, "spoken_utterances": [ "Place an order for everything in my ... amazon cart—um... or rather, cart and wishlist uh... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon cart' and then self-corrects to 'cart and wishlist'. This is a false start where the final corrected version includes both the cart and wishlist, which matches the original instruction. The speaker does not introduce any new constraints or remove any existing ones. The final intent is clear and executable.", "heard_transcript": "Place an order for everything in my Amazon cart, um, or rather, cart and wishlist, uh, for my home address." }, "candidate_id": "wav_41fe48da226dd630", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a revision ('cart or rather cart and wishlist'), which is a plausible self-correction. However, the phrasing 'cart or rather cart and wishlist' is slightly awkward and less fluent than the correction in audio_01. The 'uh' filler is natural, but the overall delivery is less smooth, with a slightly more hesitant and less confident tone compared to audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address—no, that should be home address.", "long_pause_offsets": [], "source_sha256": "94a63c1af18e9a6296ced56d33b94697ff058bfa38ae3186889cfdb4fa26c850", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately corrects it to 'home address'. The final intended destination is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/false_start/false_start_03_destination_address/0ad13633f92358d0/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "08723ddcb558e2f65b86e09e40bb528805d7305a27ab59ef7ad18ceb22857eb5", "seconds": 9.68, "spoken_utterances": [ "Place an order for everything in my amazon cart and wishlist um... for my work address—uh... no, that should be home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately self-corrects to 'home address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, um, for my work address. Uh, no, that should be home address." }, "candidate_id": "wav_f280332160ae97d3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then makes a revision ('work address' to 'home address'). However, the correction is abrupt and lacks the natural flow of the other clips. The 'uh' filler before the correction is placed awkwardly, and the intonation on 'no, that should be home address' sounds more like a correction than a spontaneous thought. The delivery feels less organic and more like a rehearsed or forced edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Place an order for everything in my amazon for my home address. I left out a detail: include both my cart and wishlist.", "long_pause_offsets": [ 64 ], "source_sha256": "496c7bfa60580f243602f09a0a149101263acdc8f71bb04388e5dfb68654f848", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include both the cart and wishlist, completing the initial instruction which only mentioned Amazon. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/addition/addition_02/a269e270f41245cf/speech.wav", "wer": 0.04, "wav_sha256": "f970338095f6911aa3b3a6db483845d993ed2b583c6e8e4889226dea1c572605", "seconds": 14.08, "spoken_utterances": [ "Place an order for everything in my um... amazon for my home address. ", "I left out a detail: include both my hmmm... cart and wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'everything in my Amazon' but then corrects themselves by adding 'Include both my cart and wishlist.' This is a clear addition that narrows the initial broad request. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Place an order for everything in my um Amazon for my home address. I left out a detail. Include both my cart and wishlist." }, "candidate_id": "wav_e1a634d8bfcc4b36", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'um' before stating the core request. The addition of 'I left out a detail' followed by a clear, smoothly integrated follow-up about including both cart and wishlist sounds like a genuine, spontaneous recollection. The prosody is consistent, and the flow from initial request to addition feels organic and unforced." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist. Oh, I forgot to mention, it should be delivered to my home address.", "long_pause_offsets": [ 62 ], "source_sha256": "5c1d4c22c8b77dff434e852d7c5381210a04ae75bc07224e80307063ca21abee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address constraint, which was not in the initial utterance but is present in the original instruction. The final goal combines both the initial order request and the added delivery condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/addition/addition_01/4cdd5daca89966f4/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "e3a6fb618fe44fcf9257eb1a30c4398d7683a6dde897cd8d95bc26ec4dc10ad5", "seconds": 17.16, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um... wishlist. ", "Oh, I forgot to mention, it should be delivered to um, let me think... my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to place an order for everything in the Amazon cart and wishlist. After a pause and a self-correction ('Oh, I forgot to mention'), the speaker adds the delivery address as 'my home address'. This is a clear addition of a constraint to the original task. The final intent combines the original request with the added condition, matching the reference instruction. The filler words and hesitation do not alter the final goal.", "heard_transcript": "Place an order for everything in my Amazon cart and um wishlist. Oh, I forgot to mention, it should be delivered to um, let me think, my home address." }, "candidate_id": "wav_7c303bfdde0963f6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail, the initial request is delivered more fluently and without hesitation. The addition is introduced with 'Oh, I forgot to mention,' which is a plausible self-correction, but the subsequent pause and 'let me think' before stating the delivery address feels slightly more rehearsed or less spontaneous than the natural flow in audio_01. The overall delivery is still realistic but slightly less organic in its progression." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address. Change the destination to home address, not work address.", "long_pause_offsets": [ 82 ], "source_sha256": "690aa102c2bac723f78836c727b532ae53c8c1e844901b15bbd84f5dc4ffa90b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination from work address to home address, preserving the action, items, and scope of the order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/revision/revision_03_destination_address/592ee1ed9ea72311/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "0d20c73ca2297dda643067006cc037a10bbdb2cf7123894091ce599446339ec9", "seconds": 15.64, "spoken_utterances": [ "Place an order for everything in my amazon cart and wishlist um... for my work address. ", "Change the destination to hmmm... home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the task for the 'work address' but immediately corrects it by saying 'Change the destination to, um, home address, not work address.' This constitutes a revision where the final intent is to use the 'home address.' The correction is clear and unambiguous, and all other conditions (ordering everything from cart and wishlist) remain unchanged. Therefore, the final audible intent matches the original instruction.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, um, for my work address. Change the destination to, um, home address, not work address." }, "candidate_id": "wav_b194633dd8dd315d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a coherent, multi-part request with natural hesitation and self-correction. The initial request is clear, followed by a plausible pause and filler ('um'), then a revision of the destination. The correction ('change the destination to...') is smoothly integrated with a natural shift in intonation, and the final phrase is spoken with appropriate emphasis. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart for my home address. Make that cart and wishlist, not just cart.", "long_pause_offsets": [ 69 ], "source_sha256": "cd4bc5bd58c7ce6bc952cc2ab4dba1200f39097ce183fe8e4634b021ea570187", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the initial scope from 'cart' to include both 'cart and wishlist', preserving the app, action, and destination address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/revision/revision_02_source_collection/94733126f2146f84/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "6b4626f1028b035da68dd6e5f08c7c7d4ea0a475bd824933851841a7eeba5da6", "seconds": 13.04, "spoken_utterances": [ "Place an order for everything in my um... amazon cart for my home address. ", "Make that ... cart and wishlist, not just cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon cart' but immediately corrects to 'cart and wishlist, not just cart' after a pause. This is a revision where the final intent clearly expands the original request to include both the cart and the wishlist. The correction is clear and the final goal matches the reference instruction.", "heard_transcript": "Place an order for everything in my um Amazon cart for my home address. Make that cart and wishlist, not just cart." }, "candidate_id": "wav_c20cd6a820287fd0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision: the speaker starts with a partial request, then corrects it by adding a missing component ('cart and wishlist not just cart'). The self-correction is natural, with a slight pause and a change in intonation. The delivery is fluent and the prosody is consistent. However, the initial request is less complete than audio_01's, and the correction is a simple addition rather than a more complex revision, making it slightly less rich in pragmatic realism." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Place an order for everything in my ebay cart and wishlist for my home address. Wait, use Amazon instead of eBay.", "long_pause_offsets": [ 80 ], "source_sha256": "626f29e155ce1e06ca0a27f59d3200aebd71009e22b6724b81e8a403bc13f1fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the action, scope (cart and wishlist), and destination (home address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/revision/revision_01_app/ba13a09a81f793ef/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "88a3d7a648e4cc185c883a1d7ae7664e49d483ad036138f3eb21d1a02513e2f3", "seconds": 15.48, "spoken_utterances": [ "Place an order for everything in my uh... ebay cart and wishlist let's see... for my home address. ", "Wait, use ... Amazon instead of eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately corrects to 'Amazon' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction, so the task is considered passed.", "heard_transcript": "Place an order for everything in my uh eBay cart and wishlist. Let's see. For my home address. Wait, use Amazon instead of eBay." }, "candidate_id": "wav_102d3f6eb065dd87", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a plausible false start ('eBay cart'), then corrects it to 'Amazon'. However, the correction is abrupt and the intonation shift feels less natural. The filler 'let's see' is placed awkwardly, and the final correction ('wait, use Amazon instead of eBay') sounds more like a mechanical insertion than a spontaneous thought. The prosody is less fluid, and the overall delivery feels slightly rehearsed or less authentic compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "922259422f8c8848e68d306bafa03db71670c1580b564ab6fbd294bd112445d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/cancellation/cancel_030/7b5529475493e486/speech.wav", "wer": 0.0625, "wav_sha256": "ebca1a5f72293c0b0d9ec736d12667483a19d6962ad96227fef38a9aea6148bd", "seconds": 19.12, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address.", "Hang on, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to place an order, then after a pause and hesitation, explicitly says 'Hang on. You can drop that task. I no longer need it.' This is a clear and final cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist. Uh, let me think. For my home address. Hang on. You can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "8015757d4ca1f709eea85a472c276721cee5798c0d87b7175275de526d3bc0e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/ambiguous_halt/halt_032/cedec8499a544653/speech.wav", "wer": 0.08, "wav_sha256": "99caeb152b43493c6dfa31283160aeafb07e2cf0a236830947472c0b778e5532", "seconds": 16.8, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Wait, before you do that,' which is a clear request to pause or halt the current action. The final interjection is a pause command, not a cancellation or a new instruction. The initial request is not executed due to the halt, which is consistent with the 'ambiguous_halt' type. No issues are present as the final intent is a valid pause.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think, for my home address. Wait, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "857496a94eac34c46365edba9ea7d54a3498afbd4a7b7bc803ade9906d6e42bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_1/audio_natural/filler/original/9c26c61d18565efc/speech.wav", "wer": 0.1, "wav_sha256": "780a97c77fc42dda8a3557467bf01e75688aecc43e5f9a9a7eeefd6ed64facb3", "seconds": 12.8, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and hmmm... wishlist uh, let me think... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the correct instruction, hesitates with filler words ('uh', 'let me think'), but completes the original task without changing the action, target, or value. The hesitation is a filler, not a final directive to pause or cancel. The final intent matches the original instruction.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think, for my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my home address—no, that should be work address.", "long_pause_offsets": [], "source_sha256": "4892b9c8e0cad2fec99a8527fc08c834f76433fb50648487905ac2fc44144cd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects it to 'work address'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/false_start/false_start_03_destination_address/38a81afc2cd12661/speech.wav", "wer": 0.0, "wav_sha256": "84f91d23389ca52e64e45f53a1dc3f76cf240d092dec56a77efc8ef32b9911fc", "seconds": 11.48, "spoken_utterances": [ "Place an order for everything in my uh, let me think... amazon cart and wishlist uh... for my home address—um... no, that should be work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately self-corrects to 'work address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for everything in my uh let me think Amazon cart and wishlist uh for my home address um no, that should be work address." }, "candidate_id": "wav_fe9faaa115b48a47", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. It begins with a clear false start ('uh, let me think'), followed by a complete but incorrect initial request ('Amazon cart and wishlist for my home address'), then a believable change-of-mind with a self-repair ('No, that should be work address'). The intonation and pacing of the correction sound spontaneous and conversational, with a natural pause and shift in tone. The filler 'uh' and hesitation are well-placed and contribute to the realism." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Place an order for everything in my eBay—I should say amazon cart and wishlist for my work address.", "long_pause_offsets": [], "source_sha256": "83c6429236492cfc1aca75f34dc254d61de3c06e21f485a7a77d375c97570568", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'eBay' to 'Amazon' within the same utterance, preserving the action, scope (cart and wishlist), and destination (work address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/false_start/false_start_01_app/d35e8edc6d78e8f2/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "4c3d0e85f6cf9aece28f9f84ab31bf1338226106ab381ec4ced6fd68943fad7a", "seconds": 10.4, "spoken_utterances": [ "Place an order for everything in my um... eBay—ah... I should say uh... amazon cart and wishlist for my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'I should say'. This is a false start where the final, corrected value is 'Amazon'. The rest of the instruction, including 'cart and wishlist' and 'work address', remains unchanged and is executed as intended.", "heard_transcript": "Place an order for everything in my um, eBay uh, I should say uh, Amazon cart and wishlist for my work address." }, "candidate_id": "wav_c9ae4a8fc1321bfe", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start ('eBay, uh, I should say...') and a self-repair to the correct platform. The correction is immediate and the phrasing sounds like a genuine recollection. However, the initial error ('eBay') is more of a categorical mistake than a minor slip, and the repair feels slightly less fluid than in audio_02. The prosody is natural, but the overall flow is less seamless than the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart—or rather, cart and wishlist for my work address.", "long_pause_offsets": [], "source_sha256": "9e06259c8cf1d3cb290a631d184242974a553fa5f49b04f0f6e58c0b02c912f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'amazon cart' but immediately corrects to 'cart and wishlist', which expands the scope to include both. The recipient address and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/false_start/false_start_02_source_collection/a75465899accc131/speech.wav", "wer": 0.15, "wav_sha256": "9b779dc1bd10f66a8a53b34ce32c367065a5dab5456cbcd1f9c088dd4f38e1d1", "seconds": 8.64, "spoken_utterances": [ "Place an order for everything in my um... amazon cart—uh... or rather, cart and wishlist for my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon cart' but immediately self-corrects to 'cart and wishlist'. This is a false start, and the final, corrected instruction matches the original reference. The user's intent is to execute the corrected task.", "heard_transcript": "Place an order for everything in my um, Amazon cart uh, or rather, cart and wishlist for my work address." }, "candidate_id": "wav_3ad81dbad52130cc", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start ('um') and a self-repair ('or rather, cart and wishlist'), but the correction is less natural. The phrase 'cart and wishlist' is a partial correction that doesn't fully resolve the initial ambiguity. The delivery is somewhat flat and lacks the dynamic intonation of the other clips. The repair feels more like a simple substitution than a spontaneous recollection, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist. Oh, I forgot to mention, use my work address.", "long_pause_offsets": [ 62 ], "source_sha256": "451140102b09884eb858f5b9cd68fee2353d9fd8d65745112d407481b102f7b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing address constraint, which was omitted in the initial instruction but is present in the original. The action, scope (cart and wishlist), and app (Amazon) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/addition/addition_01/2cd02d4b2f9bbd97/speech.wav", "wer": 0.0, "wav_sha256": "36fbe36e12bbe0a32e0f8402ac6c1fcb26e96f0937bbd8259a6c1fb6676afd38", "seconds": 9.96, "spoken_utterances": [ "Place an order for everything in my um... amazon cart and ... wishlist. ", "Oh, I forgot to mention, use my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. The speaker then adds a new constraint, 'use my work address,' which is a valid addition. The final intent combines the original task with the added detail, fully satisfying the 'addition' type criterion.", "heard_transcript": "Place an order for everything in my um Amazon cart and wishlist. Oh, I forgot to mention, use my work address." }, "candidate_id": "wav_ff54d6c2134163ea", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural, spontaneous flow with a clear addition after an initial request. The speaker begins with a complete instruction, then pauses and adds a missing detail ('Oh, I forgot to mention...') with a believable intonation shift and pragmatic back-reference. The prosody is smooth, the pace is conversational, and the filler 'um' is used naturally. The addition feels like a genuine afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for everything in my amazon for my work address. I left out a detail: include both my cart and wishlist.", "long_pause_offsets": [ 64 ], "source_sha256": "8dcd4afd99eb6aa82f01a6c59b66120b85fdb49068a3b90657dfb2784f14e99e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include both the cart and wishlist, which were not specified in the initial utterance. The final goal combines the initial instruction with the added detail, matching the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/addition/addition_02/4fec10c80b389386/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "6e86873e1e4942d3904e967a6cac89611a432e4b32432a3f8ec5406482709412", "seconds": 11.56, "spoken_utterances": [ "Place an order for everything in my uh... amazon for my work address. ", "I left out a detail: include both my ... cart and wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'everything in my uh Amazon' which is ambiguous, but immediately corrects and adds the specific detail 'include both my cart and wishlist'. This addition clarifies and completes the original instruction, making the final goal fully aligned with the reference. The addition type criterion is satisfied.", "heard_transcript": "Place an order for everything in my uh Amazon for my work address. I left out a detail, include both my cart and wishlist." }, "candidate_id": "wav_3c1d62185fd2ee74", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio starts with a complete request but includes a noticeable pause and filler ('uh') before the final detail. The addition of 'include both my cart and wishlist' is somewhat abrupt and less smoothly integrated than in audio_02. The prosody is slightly flatter, and the transition into the added clause feels less organic, possibly due to a slight hesitation or rephrasing mid-sentence. While still realistic, it lacks the seamless conversational flow of the top-ranked clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my home address. Let me correct the destination: ship to my work address, not home address.", "long_pause_offsets": [ 82 ], "source_sha256": "9a522cbab2847c54768abf63fbcece3f46f84e4c12b47865d37cd6e2494e34a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the destination from 'home address' to 'work address', replacing the initial incorrect value while preserving the action, items, and the requirement to order everything from both cart and wishlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/revision/revision_03_destination_address/5ec52c8faf340563/speech.wav", "wer": 0.0625, "wav_sha256": "242a6c5e9deea3a94f74695b5ec4ea4ea7627eedf98893dd23dfc226301421a5", "seconds": 15.08, "spoken_utterances": [ "Place an order for everything in my um... amazon cart and let me think... wishlist for my home address. ", "Let me correct the destination: ship to my work address, not home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests shipping to the 'home address' but immediately corrects this by stating 'Let me correct the destination, ship to my work address not home address.' This constitutes a revision where the final, corrected instruction overrides the initial incorrect one. The final intent matches the original reference instruction, so the task is considered passed.", "heard_transcript": "Place an order for everything in my um, Amazon cart and let me think, wishlist for my home address. Let me correct the destination, ship to my work address not home address." }, "candidate_id": "wav_7f0db28b3a4894e1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, inserts a plausible 'let me think' pause, and then smoothly corrects the destination with a believable 'let me correct' phrase. The intonation and pacing are consistent with spontaneous speech, and the correction is integrated into the same utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart for my work address. Make that everything in my cart and wishlist, not just the cart.", "long_pause_offsets": [ 69 ], "source_sha256": "3053efd79cb91a6eed3066b063e7f6342986850037f7f4f275f76866968d9e20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly expands the original scope from 'cart only' to include both 'cart and wishlist', while preserving the app (Amazon), action (place an order), and destination (work address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/revision/revision_02_source_collection/18c15e9481871720/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "065ae5c9518813b742f2e41cfd589f2672749cc505c32ca5246517111137bf84", "seconds": 12.12, "spoken_utterances": [ "Place an order for everything in my uh, let me think... amazon cart ... for my work address. ", "Make that everything in my cart and wishlist, not just the cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'everything in my Amazon cart' but then revises the instruction to 'everything in my cart and wishlist, not just the cart' after a pause. This is a revision that corrects the scope of the order to include both the cart and the wishlist. The final intent matches the original instruction, so the task is executed as revised.", "heard_transcript": "Place an order for everything in my uh, let me think, Amazon cart. For my work address. Make that everything in my cart and wishlist, not just the cart." }, "candidate_id": "wav_a4701e753218c131", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic with a natural 'let me think' filler and a clear revision. However, the correction 'Make that everything in my cart and wishlist' is slightly abrupt and less smoothly integrated than in audio_02. The prosody is slightly less fluid, and the revision feels more like a separate sentence than a seamless self-correction." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Place an order for everything in my Etsy cart and wishlist for my work address. Wait, place the order on Amazon, not Etsy.", "long_pause_offsets": [ 80 ], "source_sha256": "78f562ee29266d1cb70da7f8737220d2b1d93be99a171a65608a0b2290213fd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Etsy to Amazon, while retaining the scope (everything in cart and wishlist) and destination (work address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/revision/revision_01_app/4fe31eaff16c323c/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "e11809b526c5159875c898c666e14c4f92be6ba05f27ca32f5030fb61fdf4d86", "seconds": 11.84, "spoken_utterances": [ "Place an order for everything in my Etsy cart and wishlist ... for my work address. ", "Wait, place the order on ... Amazon, not Etsy." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the platform as 'Etsy' but immediately corrects it to 'Amazon' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, with the correct platform and unchanged conditions (cart, wishlist, work address).", "heard_transcript": "Place an order for everything in my Etsy cart and wishlist for my work address. Wait, place the order on Amazon, not Etsy." }, "candidate_id": "wav_ab60794f37e557fa", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a clear false start ('Etsy') and a self-correction ('Amazon'), but the correction is delivered with a noticeable pause and a hesitant 'Wait' that sounds slightly unnatural. The overall delivery is less fluid, and the correction feels more like a disjointed afterthought than a spontaneous thought process, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "c0b1466c7a4c8edc0231dd29f38f5e58511f04284a29e48d6e7e0acf018f471c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/cancellation/cancel_010/869bf9937957beb9/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "d3b3e4c8b437a48b386d9ce4001cd0cdf4a5299da7c6aafd5b9e23879fdd8a3d", "seconds": 10.64, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address.", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a cancellation command ('cancel the request I just made') following the initial instruction. The final interjection is a direct revocation of the task, satisfying the cancellation type criterion. The original instruction is irrelevant after the explicit cancellation.", "heard_transcript": "Place an order for everything in my amazon cart and wishlist for my work address. Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "8d2cb7d5033cb033531f5ca85ddb52aa1e0c42e30c29f7ad4269e6161afb4eb4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/ambiguous_halt/halt_023/753158371f91b0d5/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "f5faa03240fe5fd2de120b5c55d3c397ea5f46ecf3aad041aee6e1fa207e32b5", "seconds": 9.68, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the evaluation passes.", "heard_transcript": "Place an order for everything in my amazon cart and wishlist for my work address. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ce38099781cceadc158824f9d8cd548c88736204d8cc942ce483321ff8931e88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_2/audio_natural/filler/original/2b187ae6851b22e9/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "dbcbe05521d11808432143dd6025aae1c9dfef232d3005f0107537c8519c7b3e", "seconds": 5.92, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and ... wishlist ... for my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without modification. The final intent matches the original instruction exactly, with no missing or extra constraints. The type is 'filler', which permits minor hesitations or fillers but requires the final task to be executed as stated. Since the final utterance is a complete and unmodified execution of the original request, the task passes.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist for my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address—sorry, make that home address.", "long_pause_offsets": [], "source_sha256": "550bca5ab554ad5c9c7879541045ae0309c89b0588d8bd1f956ee68ade5352b8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'work address' with 'home address'; the action, source items, and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/false_start/false_start_03_destination_address/6f74faf115fff669/speech.wav", "wer": 0.12, "wav_sha256": "5f0dcb2f5b397e5772db72c6bf392ba1162cb60ff15c01e22f3e72175263c3b5", "seconds": 10.6, "spoken_utterances": [ "Place an order for everything in my amazon cart and wishlist uh, let me think... for my work address—uh... sorry, make that home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately self-corrects to 'home address' after a pause. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Place an order for everything in my Amazon cart and wishlist, uh, let me think. For my work address, uh, sorry, make that home address." }, "candidate_id": "wav_f2f4357f98ff5a70", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with a clear request, then hesitates with 'uh let me think' before correcting the delivery address from 'work address' to 'home address' with a spontaneous-sounding 'sorry, make that home address.' The self-correction is integrated smoothly into the utterance, with natural intonation and pacing. The prosody is consistent, and the repair feels like a genuine moment of recollection rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Place an order for everything in my Etsy—I meant amazon cart and wishlist for my home address.", "long_pause_offsets": [], "source_sha256": "9f640aac4ba4d6a7cf14a2e7cd3cdb57c7e6e9d9507e76f56eb8dedb7d678836", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon' within the same utterance, preserving the action, items (cart and wishlist), and destination (home address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/false_start/false_start_01_app/48e3ebee4ef2c132/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "31c3b59595e5d7c8f45d595f9c91f5e932cc15d03b12efd97c3d9f64078e319f", "seconds": 10.28, "spoken_utterances": [ "Place an order for everything in my Etsy—uh... I meant amazon cart and um, let me think... wishlist for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon'. The phrase 'let me think' is a hesitation filler and does not alter the final goal. The final audible intent matches the original instruction with the corrected app name. The type is 'false_start', so the correction is applied and the original mistaken value is discarded. All other conditions are preserved.", "heard_transcript": "Place an order for everything in my Etsy, um, I meant Amazon cart and um, let me think, wishlist for my home address." }, "candidate_id": "wav_d3f8d2cf44eeb158", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a false start ('Etsy... I meant Amazon') and a filler ('let me think'), which are realistic elements. However, the correction is less integrated than in audio_03. The speaker says 'let me think' after the correction, which slightly disrupts the flow. The prosody is natural, but the repair feels a bit more abrupt and less conversational compared to audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart—or rather, cart and wishlist for my home address.", "long_pause_offsets": [], "source_sha256": "919edcb2642f3a812433ff2df6e1b0f27cc6f58b6430b332f183ba2f492cfa91", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'cart' to 'cart and wishlist', expanding the scope to include both collections. The app (Amazon), action (place an order), and destination (home address) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/false_start/false_start_02_source_collection/93d0ed0a4bfc279a/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "c2c8527bf6fc25ed247576ae6d760f84ab6ce72ce56b80e36279bfad95210293", "seconds": 13.84, "spoken_utterances": [ "Place an order for everything in my amazon let's see... cart—um... or rather, cart and wishlist uh, let's see... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon, let's see, cart' and then self-corrects to 'or rather, cart and wishlist'. This is a false start where the speaker corrects their own initial omission of 'wishlist'. The final intent clearly includes both the cart and the wishlist, matching the original instruction. The correction is immediate and does not introduce any new constraints or remove any existing ones. Therefore, the task should be executed as per the final corrected intent.", "heard_transcript": "Place an order for everything in my Amazon, let's see, cart, um, or rather, cart and wishlist, uh, let's see, for my home address." }, "candidate_id": "wav_19b1a5941067f6d8", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start ('Amazon, but see...') and a filler ('let's see'), but the phrasing is awkward and less natural. The speaker says 'Amazon, but see...' which sounds like a mispronunciation or stutter rather than a genuine self-correction. The 'let's see' filler is placed awkwardly mid-sentence, and the overall delivery feels less fluent and more hesitant than the others. The prosody is less consistent, and the repair lacks the smoothness of the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist. Oh, I forgot to mention, deliver it to my home address.", "long_pause_offsets": [ 62 ], "source_sha256": "571688afa8a20a01add8378a5701fc19fd69ff2f3b3f317f8407efe0bc38728a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement, which was not in the initial utterance but is present in the original instruction. The final goal combines both the initial order request and the added delivery condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/addition/addition_01/e668da694fcfb1e8/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "e08d9fb70108d868fcf04af907c679e1be63761c885fc38e7b14941c9594b80a", "seconds": 16.56, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and uh, let's see... wishlist. ", "Oh, I forgot to mention, deliver it to my um... home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request to place an order for everything in the Amazon cart and wishlist. After a pause and a self-correction ('Oh, I forgot to mention'), the speaker adds the delivery address as 'home address'. This is a clear addition of a constraint to the original task. The final intent combines the original order request with the new delivery address, satisfying the 'addition' type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and uh let's see, wishlist. Oh, I forgot to mention, deliver it to my um home address." }, "candidate_id": "wav_f45a97519589311a", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Place an order for everything in my amazon for my home address. One more detail: include both my cart and wishlist.", "long_pause_offsets": [ 64 ], "source_sha256": "e89eb788987808096f6be59e6edb718e36363b9010c22f29d150b0346415c17d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing requirement to include both the cart and wishlist, which were only partially mentioned in the initial utterance. The final goal combines the initial scope with the added detail." }, "audio": { "status": "pending", "spoken_utterances": [ "Place an order for everything in my amazon uh... for my home address. ", "One more detail: include both my let me think... cart and wishlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_address", "proposed_instruction": "Place an order for everything in my amazon cart and wishlist for my work address. Change the destination to home address instead of work address.", "long_pause_offsets": [ 82 ], "source_sha256": "53c4abb74d35addeedff34399f31a07dadc75512ca5ad93165053edc6cb60c0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination from work address to home address, preserving the action, scope (cart and wishlist), and the intent to place an order." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/revision/revision_03_destination_address/96cf9069bdd262b7/speech.wav", "wer": 0.0, "wav_sha256": "8ba1d0ac6385e892ec0e56566fde8d0da47b194f8b6cbfe518236496adf9f355", "seconds": 14.0, "spoken_utterances": [ "Place an order for everything in my amazon cart and uh... wishlist for my work address. ", "Change the destination to let me think... home address instead of work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the order for the 'work address' but immediately corrects it with 'Change the destination to... home address instead of work address.' This constitutes a revision where the final intent is to use the 'home address.' The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and uh wishlist for my work address. Change the destination to let me think, home address instead of work address." }, "candidate_id": "wav_f9f835fc9b1bab39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complete initial request with natural hesitation ('and uh') and a clear, self-contained revision. The correction ('Change the destination to...') is smoothly integrated, with a believable 'let me think' filler and a pragmatic shift from 'work address' to 'home address'. The prosody is consistent, and the flow reflects a realistic thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Place an order for everything in my amazon cart for my home address. Wait, it should be cart and wishlist, not just cart.", "long_pause_offsets": [ 69 ], "source_sha256": "c48773bf1c47378aae56ddf7a068a86dde410a5bf39a751dc4f90d8cff8b1d1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial omission of 'wishlist' by specifying 'cart and wishlist', preserving the original action, app, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/revision/revision_02_source_collection/8885f11fcf500ed7/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "716cbc59a91bc5be57b52dcc4b79a150282c284cc274f11f2f72bfc99d2dd284", "seconds": 10.2, "spoken_utterances": [ "Place an order for everything in my ... amazon cart um... for my home address. ", "Wait, it should be cart and wishlist, not just cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cart' but immediately corrects to 'cart and wishlist' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent includes both cart and wishlist, matching the original reference instruction. The correction is clear and unambiguous, and no other constraints are altered.", "heard_transcript": "Place an order for everything in my Amazon cart um for my home address. Wait, it should be cart and wishlist, not just cart." }, "candidate_id": "wav_8d4553250f1b043f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a revision after a brief pause ('wait'), correcting 'cart' to 'cart and wishlist'. The delivery is fluent and the correction is integrated naturally. However, the initial request is slightly less detailed than audio_03, and the revision is less of a full rephrasing and more of a simple addition, making it slightly less complex in cognitive flow." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Place an order for everything in my ebay cart and wishlist for my home address. Sorry, I meant Amazon, not eBay.", "long_pause_offsets": [ 80 ], "source_sha256": "bb48da5c5f7482c4bd74f89fdbb0643caa97f7a441292d6efd0664fbb7fac6ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from eBay to Amazon, preserving the action, scope (cart and wishlist), and destination (home address)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/revision/revision_01_app/2a57ca23c6573d0b/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "6a4c5abab505c0f3520a932f85d353038055627089cfad483be02e4ae0adf902", "seconds": 12.56, "spoken_utterances": [ "Place an order for everything in my ebay cart and um... wishlist for my home address. ", "Sorry, I meant um, let me think... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately corrects to 'Amazon' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to place an order on Amazon, which matches the original instruction's intent except for the platform. The correction is clear and final, so the task is to execute the corrected version.", "heard_transcript": "Place an order for everything in my eBay cart and um, wishlist for my home address. Sorry, I meant um, let me think. Amazon, not eBay." }, "candidate_id": "wav_86b856fbd968525b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a request, then self-corrects from 'eBay' to 'Amazon'. However, the correction is abrupt and the 'let me think' filler is placed awkwardly mid-sentence, interrupting the flow. The prosody is less natural, and the correction feels more like a mechanical insertion than a spontaneous thought process, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "ee0385db25c54ee2f8c0ef76e07007e4067651f52dd3af19a221b81b4fffa277", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/cancellation/cancel_015/383c8b9e4cac3143/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "d975109f604119c5abd9ed45d6e3cf1ac64e88b01fc0258a9aa767d6c81b1db7", "seconds": 14.12, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address.", "Oh, wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to place an order, but after a pause and hesitation, clearly revokes the entire request by saying 'cancel the whole thing.' This constitutes a final cancellation of the original task, satisfying the cancellation type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address. Oh wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "4acc1b65a01739aa084ef54e22e20845b0a4688af9026d6ae66993eb7267e43e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/ambiguous_halt/halt_036/f066d5c94e088009/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "1992bfeb0915b1ec8c9214dd0ac35d475f284930280753fa14448ae56900e59a", "seconds": 13.88, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Wait a second before you go ahead,' which is a clear request to pause or hold the action. This is a final interjection that does not cancel the task but explicitly asks for a delay. The initial request to place an order is not revoked, and the speaker does not say 'cancel' or 'never mind.' Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address. Wait a second before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "77bcb81_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f6e51dbf9a82beaf552221c7565e1d1a4a2b6feb4c0481a54525abf5673f5058", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/77bcb81_3/audio_natural/filler/original/3272d17a407e43eb/speech.wav", "wer": 0.25, "wav_sha256": "135ace661bc7e5f622a2d5e9e4fc32188c28e8d85d048f0867c8de246d9129fe", "seconds": 9.72, "spoken_utterances": [ "Place an order for everything in my ... amazon cart and um, let me think... wishlist hmmm... for my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let me think', 'hmm') but does not change the core request. The final intent remains to place an order for everything in the Amazon cart and wishlist for the home address, matching the original instruction. The hesitation is a filler and does not alter the task goal.", "heard_transcript": "Place an order for everything in my Amazon cart and um, let me think, wishlist, hmm, for my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_completion_state", "proposed_instruction": "How many activities are incomplete—sorry, I meant completed in my bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "0d365ac03c8e060027f850343f4aa23e5e2add4237e5235a27db91baffc768d2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'incomplete' to 'completed' within the same utterance, preserving the app (Simple Note), note title ('bucket list'), and the action of counting activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/false_start/false_start_01_completion_state/d0eb25e97fd427ec/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "26f2fdc121ea2525161a3983b68019d9e9a0fd27b83ac710198e7fdd1cb42d98", "seconds": 9.8, "spoken_utterances": [ "How many activities are incomplete—um... sorry, I meant completed in my bucket list note in ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'incomplete' but immediately self-corrects to 'completed'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is not executed. The final intent matches the original instruction exactly.", "heard_transcript": "How many activities are incomplete? Sorry, I meant completed in my bucket list note in Simple Note." }, "candidate_id": "wav_a0377a6fd8841585", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and credible false-start pattern. The speaker initially says 'incompletem' with a clear mispronunciation, then immediately self-corrects with 'sorry, I meant completed' in a spontaneous, conversational tone. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and recovery. The pacing and intonation are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "How many activities are completed in my bucket list note in Google Keep—I should say Simple Note?", "long_pause_offsets": [], "source_sha256": "f71cc53024199c2468afc4aa2428a24e4a42f7bea556ae18a383a2f5f830d5ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Google Keep' to 'Simple Note' within the same utterance, preserving the rest of the query including the note title and the action of counting completed activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/false_start/false_start_03_note_app/71c2dcf2dbdbd9c8/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "99f916453b6431db55689a2b3472adc66660efb01d074c6a523d6e9e67894aa8", "seconds": 11.52, "spoken_utterances": [ "How many activities are completed in ... my bucket list note in Google Keep—uh... I should say ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Keeper' but immediately self-corrects to 'Simple Note'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many activities are completed in my bucket list note in Google Keeper? I should say, Simple Note." }, "candidate_id": "wav_38786030213009a2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'I should say' after the initial phrase, which is a plausible self-correction. However, the correction is less integrated than in audio_02; the speaker pauses and restarts the entire phrase rather than smoothly revising it. The prosody is slightly more hesitant and less fluid, making the correction feel a bit more rehearsed or less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_02_note_description", "proposed_instruction": "How many activities are completed in my daily goals—or rather, bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "71777da81cf3536be40c866f9ab20743e10f491f0b66b6b35744f21238d25c74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily goals' but immediately corrects to 'bucket list note' using 'or rather', indicating the intended target is the bucket list note. The app 'Simple Note' and the action of counting completed activities remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/false_start/false_start_02_note_description/c91ec6ce4337300a/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "2aded05a1548d9e2393b6f53c4c5de97761deb0253df8bca10122473968dac75", "seconds": 9.8, "spoken_utterances": [ "How many activities are completed in um... my daily goals—uh... or rather, ... bucket list note in Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily goals' but immediately self-corrects to 'bucket list note'. The final utterance clearly specifies the correct target note. The 'false_start' type requires applying the correction, which is done here. The rest of the instruction matches the original.", "heard_transcript": "How many activities are completed in um my daily goals uh or rather bucket list note in Simple Note?" }, "candidate_id": "wav_7f0cf28ec7d5e4b1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a more complex revision with 'my daily goals uh, or rather bucket list note in simple note'. While it includes a filler ('uh') and a revision, the flow is less natural. The speaker seems to be trying to correct two elements at once, and the transition from 'daily goals' to 'bucket list note' feels abrupt and less like a spontaneous thought process. The prosody is slightly rushed, and the correction lacks the organic rhythm of the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many activities are in my bucket list note in Simple Note? Oh, I forgot to mention, only count the activities that are completed.", "long_pause_offsets": [ 63 ], "source_sha256": "1bd95b8f68a6ba416adba26220475fb0922c92ee8007d8ee6f5d3401b8cc98be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a constraint to only count completed activities, which refines the initial broad request to count all activities. The final goal is fully specified and matches the intent of the original instruction with a clarifying restriction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/addition/addition_01/a56aaf659c9c6d52/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "82ba6bf692d21eb04aa08d9a206e41c15e16392eccc6579ca584a59c69b35f02", "seconds": 13.84, "spoken_utterances": [ "How many activities are in ah... my bucket list note in let's see... Simple Note? ", "Oh, I forgot to mention, only count the activities that are ... completed." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many activities are in my bucket list note in Simple Note?'. After a pause and a self-correction, the speaker adds the constraint 'only count the activities that are completed'. This is a clear addition of a new condition to the original request. The final intent combines the original task with the added constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "How many activities are in uh my bucket list note in let's see, Simple Note. Oh, I forgot to mention, only count the activities that are completed." }, "candidate_id": "wav_374b960f0247e956", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, uses a plausible filler ('uh') and hesitation ('let's see'), and then adds a relevant restriction ('only count the activities that are completed') in a way that sounds like a spontaneous afterthought. The prosody is consistent, and the addition flows naturally without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many activities are completed in my note in Simple Note? One more detail: it's specifically my bucket list note.", "long_pause_offsets": [ 61 ], "source_sha256": "29ba59dcb67ce050a44e1a55ac56efb4dc11a99b1f71aba710b579e0e40313db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specificity that the note is the 'bucket list note', which was implied but not stated in the original instruction. The action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/addition/addition_02/300cfd0ee7c69ee7/speech.wav", "wer": 0.16, "wav_sha256": "eaad94b3fd3652c89404642f04851de88f82a661dbdc6b86039cc8321201d6bb", "seconds": 14.64, "spoken_utterances": [ "How many activities are completed in um... my note in uh, let's see... Simple Note? ", "One more detail: it's specifically my um... bucket list note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many activities are completed in my note in Simple Note?'. After a pause and a self-correction, the speaker adds the detail 'it's specifically my bucket list note'. This addition narrows the scope of the original request, and the final intent combines the original request with the added specificity. The final goal matches the reference instruction, so the task passes.", "heard_transcript": "How many activities are completed in um my note in uh let's see, Simple Note. One more detail, it's specifically my um bucket list note." }, "candidate_id": "wav_ac818710eb745e07", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker's delivery is mostly natural, but the addition ('one more detail... specifically my bucket list note') feels slightly more forced and less integrated than in audio_01. The phrase 'one more detail' is a bit of a formulaic transition, and the correction is delivered with a slightly stiffer intonation. The overall prosody is good, but the addition lacks the seamless spontaneity of the first clip." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_note_app", "proposed_instruction": "How many activities are completed in my bucket list note in Evernote? Let me correct the app: look for the note in Simple Note, not Evernote.", "long_pause_offsets": [ 70 ], "source_sha256": "7c83452b4852caf47b62c81b5670dce296ed79daeb81f668349fbf0897210986", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Evernote to Simple Note, preserving the note title and the query about completed activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/revision/revision_03_note_app/8d1f9d28b7300ae8/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "61c4e7626c857f9833604bf18172a6c3491044ab0ad8f360763cd4204c67d896", "seconds": 17.84, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in um, let's see... Evernote? ", "Let me correct the app: look for the note in um... Simple Note, not Evernote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Evernote but immediately self-corrects, specifying 'Simple Note, not Evernote' as the correct app. This constitutes a revision where the final, corrected target app is Simple Note, matching the original instruction. The rest of the task (counting completed activities in the 'bucket list' note) remains unchanged and is executed as intended.", "heard_transcript": "How many activities are completed in my bucket list note in um, let me see, Evernote? Let me correct the app. Look for the note in um, Simple Note, not Evernote." }, "candidate_id": "wav_76f8bdb29ac5a720", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a complex, multi-part request with natural hesitation ('uh', 'um'), a clear false start ('Evernote'), and a self-correction ('Let me correct the app') that flows into a believable revision ('SimpleNote'). The prosody is varied, with appropriate stress and pacing, and the correction is integrated smoothly into the ongoing utterance rather than sounding like a separate sentence. The speaker sounds like they are actively thinking through the task." } ] }, "is_default": true }, { "variant_id": "revision_01_completion_state", "proposed_instruction": "How many activities are incomplete in my bucket list note in Simple Note? Wait, I meant how many activities are completed, not incomplete.", "long_pause_offsets": [ 74 ], "source_sha256": "ea1e9fc0c6387379e46e0a8163824d30845f9ca2143f9f1d85e28cbafa5e13af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial request from 'incomplete' to 'completed', preserving the app, note, and the question about the count of activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/revision/revision_01_completion_state/ca85a63907005c87/speech.wav", "wer": 0.0, "wav_sha256": "ea2b33930953da75d40417b469b0bb6527042acec95c2f5f6e94ffc597e934ee", "seconds": 13.32, "spoken_utterances": [ "How many activities are incomplete in ... my bucket list note in ... Simple Note? ", "Wait, I meant how many activities are completed, not incomplete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'incomplete' activities but immediately self-corrects to 'completed' after a pause. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "How many activities are incomplete in my bucket list note in Simple Note? Wait, I meant how many activities are completed, not incomplete?" }, "candidate_id": "wav_b315e521670e2159", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a clear self-correction ('Wait, I meant...') after a false start, which is a realistic repair. However, the initial request is delivered in a flatter, more robotic tone with less natural prosody and fewer fillers. The correction is also slightly abrupt and less integrated into the flow compared to audio_02, making the overall delivery feel less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_note_description", "proposed_instruction": "How many activities are completed in my shopping list note in Simple Note? Change the note description to \"bucket list\" instead of \"shopping list\".", "long_pause_offsets": [ 75 ], "source_sha256": "c9877fe88b7c2b8bf714a7bf4cfcb3387ca23d22f7b1844f41bb9546c1afd837", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the note description from 'shopping list' to 'bucket list', which corrects the initial instruction to match the original goal." }, "audio": { "status": "pending", "spoken_utterances": [ "How many activities are completed in my shopping list note in uh... Simple Note? ", "Change the note description to um... \"bucket list\" instead of \"shopping list\"." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "84329a448869c4cd235eca03e1e3319962253a4430dabd8ddec3f22c0b5453b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/cancellation/cancel_031/087ddea75ee7b4e9/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "8f26a3a121fd2c9c5f8d050758351eedc443b8bcc802895d47245126dfff9efc", "seconds": 11.12, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?", "Wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the entire task by stating 'Wait, there's no need to do that anymore. Cancel it.' This constitutes a direct cancellation of the original request. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note? Wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "5dc14754cc5f601e61fd9efc209ee8bcaff037319f1b2ef27131eb72a1d82dfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/ambiguous_halt/halt_025/b8cd3df571e7a85a/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "ad12bcfbe0a7afc621e11d344e05274b37d91ebcdaf7e8e8de9cb5ea21aad55e", "seconds": 9.92, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I need to think for a second,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note? Wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ffbdcfda2c5e94d5c1f1fd0a6dd44581f5c086c480472860eaedbeab3b541bfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_1/audio_natural/filler/original/5fa03f6b0c78a9ff/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "768be97ba0850ab6f4519896fd6691bf1b3ef72e2e2685fe263ac2159b441dcf", "seconds": 5.68, "spoken_utterances": [ "How many activities are completed in uh... my bucket list note in ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'uh' as a filler, then proceeds to complete the original instruction without any correction, cancellation, or addition. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "How many activities are completed in uh my bucket list note in Simple Note?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_note_app", "proposed_instruction": "How many activities are left to do in my bucket list as per my Google Keep—I should say SimpleNote note?", "long_pause_offsets": [], "source_sha256": "48e3ae66eb605cf79d3ffe86d5e7fb0a46bcbf5d6330b0769e5d7e3fb7e06a73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Google Keep but immediately corrects to SimpleNote, which is the app specified in the original instruction. The rest of the query remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/false_start/false_start_03_note_app/c34fbbb3064afaeb/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "36acf89de8eb08c172620d63d24a3595eaa82e21cb04a508ddadb7535e5e44bd", "seconds": 8.24, "spoken_utterances": [ "How many activities are um... left to do in my ... bucket list as per my Google Keep—uh... I should say SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Keep' but immediately self-corrects to 'SimpleNote'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is preserved otherwise. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many activities are uh left to do in my bucket list as per my Google Keep uh, I should say SimpleNote." }, "candidate_id": "wav_a54916fcad529973", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('uhm') and smoothly self-corrects from 'Google Keep' to 'Simple Note' without any awkwardness. The prosody is consistent, the pace is conversational, and the correction feels spontaneous and integrated into the flow of the sentence." } ] }, "is_default": true }, { "variant_id": "false_start_02_note_description", "proposed_instruction": "How many activities are left to do in my daily tasks—or rather, bucket list as per my SimpleNote note?", "long_pause_offsets": [], "source_sha256": "1b4d3fe4f89c5e0c7a5acdc6de2f06afdcf1243941eacd900ac09158dbe0886e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily tasks' but immediately corrects to 'bucket list' using 'or rather', which clearly replaces the intended list. The app (SimpleNote) and the query type (count of remaining activities) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/false_start/false_start_02_note_description/ee3591e17469af55/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "b87c76b24a933b743c822965980d1bd105b170d5963cbbe28e7daa10f11bf00f", "seconds": 9.44, "spoken_utterances": [ "How many activities are um... left to do in my daily tasks—uh... or rather, bucket list as per my SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily tasks' but immediately self-corrects to 'bucket list'. According to the 'false_start' type criterion, the final corrected value must be applied. The correction is clear and immediate, and the rest of the instruction matches the original. Therefore, the final intent is to query the bucket list in SimpleNote, which matches the original instruction.", "heard_transcript": "How many activities are on left to do in my daily tasks uh or rather bucket list as per my SimpleNote note?" }, "candidate_id": "wav_84587003481d8f23", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a more complex self-correction ('daily tasks' to 'bucket list') with a filler ('uhm') and a 'rather' transition. While the correction is plausible, the 'uhm' and 'rather' create a slightly more hesitant, less fluid delivery compared to audio_01. The prosody is still natural, but the correction feels a bit more labored." } ] }, "is_default": false }, { "variant_id": "false_start_01_completion_state", "proposed_instruction": "How many activities are completed—sorry, I meant left to do in my bucket list as per my SimpleNote note?", "long_pause_offsets": [], "source_sha256": "d106ac7663af31349f88a9a38b9d66135a9c27d4f9f4ec5158e75d1cab9028f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'completed' to 'left to do' within the same utterance, preserving the rest of the query including the bucket list context and the reference to the SimpleNote note." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/false_start/false_start_01_completion_state/79737728f25b6811/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "29f87080c7805502e4facd94dc4000698e2244b12b0011d176569d9c87f14abd", "seconds": 10.04, "spoken_utterances": [ "How many activities are completed—um... sorry, I meant left to do in my ... bucket list as per my SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'completed' but immediately self-corrects to 'left to do' after a pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many activities are completed, um, sorry, I meant left to do in my bucket list as per my SimpleNote note?" }, "candidate_id": "wav_5eb7d567dd5b8f4f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a false start ('um') and then corrects 'completed' to 'left to do'. However, the correction is abrupt and the phrase 'left to do' is delivered with a flat, slightly robotic intonation. The overall delivery lacks the natural conversational flow of the other two clips, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many activities are bucket list as per my SimpleNote note? Oh, I forgot to mention, it's only the activities left to do in my bucket list.", "long_pause_offsets": [ 63 ], "source_sha256": "7cb5f38a5edfba7b84538648d0bf9a5006a59214a111d33944fc9b4116fcae61", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the initial ambiguous phrasing 'activities are bucket list' by specifying 'only the activities left to do in my bucket list', which matches the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/addition/addition_01/7c35f75797d685ae/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "3fdd651289b1ebff93b3d65934250effc0200eaad7a8315ac2cd76375b989c0f", "seconds": 13.24, "spoken_utterances": [ "How many activities are uh... bucket list as per my ... SimpleNote note? ", "Oh, I forgot to mention, it's only the activities um... left to do in my bucket list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('activities are a bucket list'), but the speaker adds the clarifying phrase 'only the activities left to do in my bucket list' after a pause. This addition completes the original request by specifying the subset of activities to count. The final intent matches the reference instruction, which is to count activities left to do in the bucket list as per the note.", "heard_transcript": "How many activities are a bucket list as per my SimpleNote? Oh, I forgot to mention it's only the activities um left to do in my bucket list." }, "candidate_id": "wav_1e54c0ceff52f5d3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then naturally adds a clarifying detail ('Oh, I forgot to mention...') with a believable self-correction. The prosody is smooth, the pause before the addition is realistic, and the correction flows naturally without sounding forced. The speaker's voice is consistent and the delivery sounds like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many activities are left to do in my as per my SimpleNote note? I left out a detail: it's specifically about my bucket list.", "long_pause_offsets": [ 68 ], "source_sha256": "599791a6ef069a24da7b4bcc2f7b0c87c83f90d9c9e42d3052b7fee55adaeb84", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing specification 'my bucket list' to the initial incomplete question, completing the intended scope. The app (SimpleNote) and the action (counting activities) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/addition/addition_02/4d16b4d6c009b043/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "9127b50db28615c1e99f0cc2f5895169118439d44c841226c17e0e35930455f3", "seconds": 12.92, "spoken_utterances": [ "How many activities are um... left to do in my as per my uh... SimpleNote note? ", "I left out a detail: it's specifically about my hmmm... bucket list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('in my as per my uh SimpleNote'), but the speaker immediately corrects and adds the missing context ('It's specifically about my bucket list'). This addition completes the original instruction, making the final goal match the reference. The 'addition' type is satisfied.", "heard_transcript": "How many activities are um left to do in my as per my uh SimpleNote? I left out a detail. It's specifically about my bucket list." }, "candidate_id": "wav_fe3248c417e429af", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly mumbled and disorganized phrasing ('How many activities are um... left to do in my as per my uh...'), indicating genuine hesitation. The addition ('I left out a detail...') is delivered with a natural pause and a believable self-aware correction. However, the initial phrasing is less fluent than audio_01, and the correction feels slightly more awkward, making it less realistic overall." } ] }, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "How many activities are left to do in my bucket list as per my note? I should also mention, the note is in SimpleNote.", "long_pause_offsets": [ 69 ], "source_sha256": "63ac8f3a9ba99ea8547aaf9a216282c1af26f35276614fcdf868008c4a828a17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the app where the note is located, completing the original instruction's missing context. The final goal combines the initial question with the added constraint about the note's location." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/addition/addition_03/5125435def2eb76a/speech.wav", "wer": 0.17857142857142858, "wav_sha256": "effaa9abcefcedbd2984381008dd0e4724fd3bce3fc26c5d416a4b17f28a0207", "seconds": 12.32, "spoken_utterances": [ "How many activities are uh... left to do in my let me think... bucket list as per my note? ", "I should also mention, the note is in SimpleNote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How many activities are left to do in my bucket list as per my SimpleNote note?'. The speaker initially says 'my note' but then adds 'the note is in SimpleNote' after a pause, which clarifies the location of the note. This is a valid addition that refines the original instruction without changing its core intent. The final audible task is to count activities in the bucket list based on the note in SimpleNote, matching the original instruction. Therefore, the addition is correctly applied and the task is executed.", "heard_transcript": "How many activities are uh left to do in my, let me think, bucket list as per my note. I should also mention, the note is in SimpleNote." }, "candidate_id": "wav_96f31907075d4a7c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker includes a filler ('let me think') and a revision ('I should also mention...'), but the phrasing is less natural. The initial request is somewhat disjointed, and the revision feels tacked on rather than a smooth continuation. The prosody is less fluid, and the overall delivery sounds more rehearsed or less spontaneous compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_completion_state", "proposed_instruction": "How many activities are completed in my bucket list as per my SimpleNote note? Wait, it should be 'left to do' in my bucket list, not 'completed'.", "long_pause_offsets": [ 79 ], "source_sha256": "6169b83fc488d3e797def0ede10ab8da952080f645004bedf4c69a9ce39d0ce8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial phrasing from 'completed' to 'left to do', preserving the app (SimpleNote), list (bucket list), and the intent to count remaining activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/revision/revision_01_completion_state/0bc599429ecb867e/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "14ae32aca74830d5471b31331aa6be96d26c8a8eb53a97733f64167ce4bf09aa", "seconds": 16.12, "spoken_utterances": [ "How many activities are uh... completed in my ah... bucket list as per my ... SimpleNote note? ", "Wait, it should be 'left to do' in my bucket list, not 'completed'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'completed' activities but immediately corrects themselves to 'left to do' after a pause. This is a clear revision where the final intent is the corrected version. The original instruction is revised to match the final audible request, so the task is executed as revised.", "heard_transcript": "How many activities are uh completed in my uh bucket list as per my SimpleNote note? Wait, it should be left to do in my bucket list, not completed." }, "candidate_id": "wav_aef6cc5d6d3940fc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('completed') and immediately self-corrects with a natural-sounding 'wait, it should be left to do'. The correction is integrated smoothly into the sentence, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit. The pacing and filler ('uh') are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_03_note_app", "proposed_instruction": "How many activities are left to do in my bucket list as per my Gmail note? Use SimpleNote instead of Gmail for the note.", "long_pause_offsets": [ 75 ], "source_sha256": "8b8fbd72ee8dc89d7574d7c8cd3a5acf3635c540793885313f942ddcf719f7d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the note source from Gmail to SimpleNote, preserving the original action and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/revision/revision_03_note_app/2269e69a181e7626/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "4587acf5c1a10e4d7dc2bf80d57590f8c041c7ba4773dac5d8c7685ffc931a4a", "seconds": 12.88, "spoken_utterances": [ "How many activities are left to do in my bucket list as per my um... Gmail note? ", "Use um, let me think... SimpleNote instead of Gmail for the note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Gmail note' but then corrects it to 'SimpleNote' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction exactly, with the target note app corrected from Gmail to SimpleNote.", "heard_transcript": "How many activities are left to do in my bucket list as per my um Gmail note? Use um let me think, SimpleNote instead of Gmail for the note." }, "candidate_id": "wav_9e11dc7de20ee535", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a natural hesitation ('um') and then self-corrects 'Gmail note' to 'Simple Note'. The correction is embedded within the sentence, and the 'let me think' filler is used plausibly. However, the correction is less integrated than in audio_03, and the overall delivery is slightly more hesitant, making it less fluid." } ] }, "is_default": false }, { "variant_id": "revision_02_note_description", "proposed_instruction": "How many activities are left to do in my daily goals as per my SimpleNote note? Change 'bucket list' to 'daily goals' as per my note.", "long_pause_offsets": [ 80 ], "source_sha256": "b9ef3d10b905f60a9af685afb911df893244a8185736b64bb2ea3a50160c4386", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference from 'bucket list' to 'daily goals', aligning the final goal with the revised target. The app (SimpleNote) and the action (querying remaining activities) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/revision/revision_02_note_description/deee950205de4e02/speech.wav", "wer": 0.08823529411764706, "wav_sha256": "4aac2661c9d44c64103cc7a64fa5fad0a06ca7ce023bea7ab994a10b624d8ae4", "seconds": 15.32, "spoken_utterances": [ "How many activities are left to do in my daily goals as per my SimpleNote note? ", "Change 'uh, let me think... bucket list' to 'um, let me think... daily goals' as per my note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'bucket list' but immediately self-corrects to 'daily goals' after a pause and repair. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to query 'daily goals', which matches the revised instruction. The original 'bucket list' is discarded, and the rest of the constraints (SimpleNote note) are preserved. Therefore, the final intent satisfies the revision criterion.", "heard_transcript": "How many activities are left to do in my daily goals as per my SimpleNote note? Change uh, let me think, bucket list to um, let me think, daily goals as per my note." }, "candidate_id": "wav_6b8e23777adf51d6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker delivers a revision of the original request, but the correction ('bucket list' to 'daily goals') is made in a separate, disjointed clause ('Change uh let me think bucket list to um let me think daily goals'). This sounds more like a meta-instruction than a spontaneous thought process, breaking the natural flow. The pauses and fillers feel less organic and more like a rehearsed edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "fcff27053f79295fda80e68b1fd6956cba512f95f120cade0e8a00a21aa9df0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/cancellation/cancel_031/e31c39c447d51200/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "305aa14a659cb2e2c92d7a14080a5a2a458e5c2cf17159bf8ecb8a642f092597", "seconds": 11.96, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?", "Oh, wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the entire task by stating 'Oh wait, there's no need to do that anymore. Cancel it.' This constitutes a final, explicit cancellation of the original request. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "How many activities are left to do in my bucket list as per my um, let's see, SimpleNote? Oh wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "feb6cd2e2bb698c7270e5afa7be03bdec428b2e1f67ede6630349a03901189f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/ambiguous_halt/halt_017/6c4ea418167a88d3/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "d5f07dd50649ac3f2c0bbf0368fde81bc868169a7a7fa17093541ac5a6ee2653", "seconds": 10.32, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'One moment please,' which is a clear request to pause or wait. The user is in the process of checking their SimpleNote note, as indicated by 'as per my um let's see. SimpleNote.' This constitutes an ambiguous_halt, where the user asks for a pause without canceling the task. The final interjection is a request to wait, not to cancel or execute, satisfying the ambiguous_halt criterion.", "heard_transcript": "How many activities are left to do in my bucket list as per my um let's see. SimpleNote. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e213d8856d287fb2c66ee7f89091c28541dbba0242322bb13c13425e4640dac7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_2/audio_natural/filler/original/721327583ae21e5d/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "02ad8b925d250d87fdd567320c3c81232f12d7ba318c019457c68dc4b6672009", "seconds": 7.2, "spoken_utterances": [ "How many activities are ... left to do in my bucket list as per my um, let's see... SimpleNote note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'let's see') and a brief hesitation but does not change the core request. The final intent remains to count bucket list activities from the SimpleNote note, matching the original instruction. The hesitation is a natural speech pattern and does not alter the task goal.", "heard_transcript": "How many activities are left to do in my bucket list as per my um let's see SimpleNote?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_note_description", "proposed_instruction": "How many activities are done in my daily goals—wait, I meant bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "668fa89bcfc47b8589e6c64e7843c8099b71c50a7aafc19f5f623bb630024e34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'daily goals' to 'bucket list note' within the same utterance, preserving the app (Simple Note) and the action (counting done activities)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/false_start/false_start_02_note_description/b16783e73f18cee6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "5954a4705c38bccf031125d85d8f1167f9cf8b69fc4f600818b96dbb1fbef58d", "seconds": 9.96, "spoken_utterances": [ "How many activities are done in my um... daily goals—uh... wait, I meant hmmm... bucket list note in Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'daily goals' but immediately self-corrects to 'bucket list note'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many activities are done in my um daily goals uh wait, I meant bucket list note in Simple Note?" }, "candidate_id": "wav_20ecf1f869bbae1f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false start and self-correction. The speaker begins with 'daily goals', hesitates with 'uh', then clearly self-corrects to 'bucket list note' with a believable 'wait, I meant' phrase. The prosody includes a natural upward inflection on 'uh' and a downward correction, and the pacing feels like a real-time thought process. The correction is integrated smoothly into the same utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "false_start_01_completion_state", "proposed_instruction": "How many activities are pending—sorry, make that done in my bucket list note in Simple Note?", "long_pause_offsets": [], "source_sha256": "b68eb49bcaebceb69b7169f9aa99020fca3fcc4bad7392b9e1afe0916efe59ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pending' but immediately corrects it to 'done' within the same utterance. The final intended query is about 'done' activities, which matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/false_start/false_start_01_completion_state/ba7ddf9344e17889/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "4ebbc9d030d39d9abaf8e5fb845dbb010525e82cedb81bf8c39070753cd12276", "seconds": 8.52, "spoken_utterances": [ "How many activities are pending—hmmm... sorry, make that ... done in my bucket list note in ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'pending' activities but immediately self-corrects to 'done' after a pause and 'sorry'. This is a false start where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to count 'done' activities in the specified note.", "heard_transcript": "How many activities are pending? Hmm. Sorry, make that done in my bucket list note in Simple Note." }, "candidate_id": "wav_5aa9114a7a5a74b2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'pending' followed by a self-correction to 'done'. The correction is clear and the speaker uses 'sorry, make that' which is a common conversational phrase. However, the correction is slightly less fluid than audio_02, with a brief pause and a slightly more abrupt shift. The prosody is natural, but the flow feels a bit more rehearsed or less spontaneous than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "How many activities are done in my bucket list note in Google Keep—no, that should be Simple Note?", "long_pause_offsets": [], "source_sha256": "920ad60c7f6215125bb8f429d8c3d3acbc08aacf9faa8c9f3c8320a7983d36d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the app from Google Keep to Simple Note within the same utterance. The final intended app is Simple Note, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/false_start/false_start_03_note_app/2b7eeff8d37d99e2/speech.wav", "wer": 0.1, "wav_sha256": "593556e445134278873483c8d65715d8bd7e3dc3c5d9d8f6ae95aa39728fadfa", "seconds": 9.44, "spoken_utterances": [ "How many activities are done in my bucket list note in Google Keep—um... no, that should be ... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Keep' but immediately self-corrects to 'Simple Note'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent is to perform the task in Simple Note, which matches the original instruction's intent after correction.", "heard_transcript": "How many activities are done in my bucket list note in Google Keep? No, that should be Simple Note." }, "candidate_id": "wav_74f8fc32ae250ce8", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'bucket list note' followed by 'no, that should be simple note'. While the correction is clear, the delivery feels less natural. The speaker's tone is more declarative and less hesitant, and the correction is delivered as a separate sentence rather than being integrated into the original thought. The prosody lacks the subtle hesitation and intonation shifts seen in the other clips, making it sound more like a read-aloud correction than a spontaneous self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many activities are done in Simple Note? Oh, I forgot to mention, it's in my bucket list note.", "long_pause_offsets": [ 45 ], "source_sha256": "644e27e46dc0b0f91a256af068951fe72f9bd578f4a22dc1f7eb2d2f8358b8ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context that the note is the 'bucket list note', which specifies the exact note to query. The app and the action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/addition/addition_01/9b19f0387bb9350c/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "1f580f7de7f08849965f56c944c7309c49cbe8d876fb77fad0043cf2e63540db", "seconds": 14.0, "spoken_utterances": [ "How many activities are done in um, let me think... Simple Note? ", "Oh, I forgot to mention, it's in my uh, let's see... bucket list note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many activities are done in Simple Note?'. After a pause and self-correction, the speaker adds the constraint 'it's in my bucket list note'. This is a clear addition that narrows the scope of the original request. The final intent combines both the original and added information, matching the reference instruction. The addition type criterion is satisfied.", "heard_transcript": "How many activities are done in um, let me think, Simple Note? Oh, I forgot to mention, it's in my uh, let's see, bucket list note." }, "candidate_id": "wav_ae6011a023cf19a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural, spontaneous fillers like 'um' and 'let me think' at the beginning, followed by a believable self-correction ('Oh, I forgot to mention...') and a clear addition of the missing detail ('bucket list note'). The prosody is conversational, with a natural rise and fall in intonation, and the pacing includes plausible pauses that suggest genuine thought processing. The addition feels integrated into the flow rather than tacked on." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many activities are done in my bucket list note in? I should also mention, it's in Simple Note.", "long_pause_offsets": [ 56 ], "source_sha256": "b4f023d71cfc200982cb258f46f530686cfe75f8ecd3701c5fe3ea6c9ad3b7e4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the app where the note is located, completing the initial incomplete question. The final goal combines the initial query with the app specification." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/addition/addition_02/7064edafcd0a1455/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "640ed6f12cd7b73c21ecfe1dcb0678bf777dccab60a74719ada8963de8d1cdfe", "seconds": 11.72, "spoken_utterances": [ "How many activities are done in hmmm... my bucket list note in? ", "I should also mention, it's in let's see... Simple Note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial part of the audio matches the original instruction. The speaker then adds 'I should also mention it's in let's see Simple Note,' which clarifies the app where the note is located. This is a clear addition that specifies the location of the note, refining the original request. The final intent is the original instruction with the added context of the app being Simple Note, which is consistent with the addition type criterion.", "heard_transcript": "How many activities are done in my bucket list note in I should also mention it's in let's see Simple Note." }, "candidate_id": "wav_a1aa6eae42fc4d81", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also uses fillers ('mm') and a self-correction ('I should also mention...'), the phrasing 'in my bucket list note in' is slightly awkward and the correction feels less fluid than in audio_01. The addition of 'simple note' comes after a longer, more hesitant pause, and the overall delivery is slightly more disjointed. The prosody is less smooth, and the correction seems more like a mechanical insertion than a natural thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_completion_state", "proposed_instruction": "How many activities are all in my bucket list note in Simple Note? Wait, I meant 'done' activities, not 'all'.", "long_pause_offsets": [ 67 ], "source_sha256": "1500d2ca547d43f344648a398b8ee6edf12dc89c0d5b7f6383d21b9b9c13421f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects 'all' to 'done', preserving the app (Simple Note), note (bucket list), and the action of counting activities. The final goal is to count only the 'done' activities." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/revision/revision_01_completion_state/c0d1293db0444cc9/speech.wav", "wer": 0.12, "wav_sha256": "657201d0ecaf63348dc299598bb92853172514a5f35856a8af67dbcaf367c1c0", "seconds": 16.32, "spoken_utterances": [ "How many activities are all in um... my bucket list note in uh, let's see... Simple Note? ", "Wait, I meant 'uh... done' activities, not 'all'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'all' activities but immediately self-corrects to 'done activities' after a pause and repair. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to count 'done activities', which matches the original instruction's structure and target ('my bucket list note in Simple Note') but with the correct value ('done' instead of 'all'). No other constraints were changed, so the revision is correctly applied.", "heard_transcript": "How many activities are all in um my bucket list note in uh let's see Simple Note? Wait, I meant uh done activities, not all." }, "candidate_id": "wav_35478cef19152d88", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, spontaneous flow with a clear false start ('all in um... my bucket list note') followed by a self-correction ('wait, I meant uh done activities not all'). The intonation and pacing reflect genuine hesitation and recollection, with a believable shift in focus. The pauses and filler words ('um', 'uh', 'let's see') are well-placed and contribute to the impression of someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_02_note_description", "proposed_instruction": "How many activities are done in my daily goals note in Simple Note? Change the note description to 'bucket list', not 'daily goals'.", "long_pause_offsets": [ 68 ], "source_sha256": "0c023769382df437efea6a67d5838ef07f8e8a3a04946ee767387b597eb22f0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the note description from 'daily goals' to 'bucket list', which aligns with the original instruction's target note. However, the initial utterance incorrectly references 'daily goals' instead of 'bucket list'. The final goal is equivalent to the original instruction because the revision corrects the note name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/revision/revision_02_note_description/5d291a2e75d328d8/speech.wav", "wer": 0.08, "wav_sha256": "ba7c9d015426852d6c3d3c6c4661c5d877c3c77b3f8feba5dddd5270397fdc9e", "seconds": 13.92, "spoken_utterances": [ "How many activities are done in ah... my daily goals note in uh... Simple Note? ", "Change the note description to 'bucket list', not 'daily goals'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a revision: the speaker first asks about 'daily goals note' but then corrects to 'bucket list, not daily goals'. According to the revision type criterion, the final instruction after the correction must be applied. The final intent is to change the note description, which matches the revision criterion. The original instruction is overridden by the correction.", "heard_transcript": "How many activities are done in uh my daily goals note in uh Simple Note? Change the note description to bucket list, not daily goals." }, "candidate_id": "wav_3efc7127db40288d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a revision, but the delivery is less natural. The speaker starts with a clear request, then immediately follows with a correction ('change the note description to bucket list not daily goals'). While the correction is grammatically sound, the transition feels abrupt and lacks the organic hesitation or self-repair seen in audio_01. The prosody is flatter, and the correction sounds more like a read-aloud instruction than a spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "revision_03_note_app", "proposed_instruction": "How many activities are done in my bucket list note in Evernote? Let me correct the app: look in Simple Note, not Evernote.", "long_pause_offsets": [ 65 ], "source_sha256": "2b1ed1ef4ad93629411e033601521b93b473b3660022d685af17861862b992d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Evernote to Simple Note, preserving the rest of the query including the note title and the action of counting done activities." }, "audio": { "status": "pending", "spoken_utterances": [ "How many activities are done in uh... my bucket list note in ... Evernote? ", "Let me correct the app: look in um... Simple Note, not Evernote." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "41fe8ab1e93dfa0c4b3f25ed2f60bcd763d1884f09d9771a73983f71e475a798", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/cancellation/cancel_024/826bf35b093e2dc3/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "9cc5941d4f781f01fbb0bf31c7fcf04232faff2910eab0b07ca0568f25ed803c", "seconds": 13.48, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?", "Hang on, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Hang on, no, never mind, cancel the whole thing.' This constitutes a final, unambiguous cancellation of the entire task. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "How many activities are done in um my bucket list note in let me think Simple Note? Hang on, no, never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "90bb7e10da93fe42c1706e6c8d7b9ca8f921b59f932e79dac32511445f346517", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/ambiguous_halt/halt_031/da2989f3a085fc8a/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "ae172a0673b59c6757de90ccba098a3a603997ce15e27ed72e246d3e86bee0c2", "seconds": 12.56, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present as the final intent is to pause, not execute or cancel.", "heard_transcript": "How many activities are done in um my bucket list note in let me think, Simple Note? Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7847649_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "46feb7a44c720d62c86f64010741f1b447fd93d1857696a0642e3944765d4309", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7847649_3/audio_natural/filler/original/8c263be8b55fac0e/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8e774c708934895bbc8a1ccc89f620b3f26ba73d7924115504b2085f6656982d", "seconds": 8.64, "spoken_utterances": [ "How many activities are done in um... my bucket list note in let me think... Simple Note?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and a pause but does not change the core request. The final audible intent matches the original instruction. The type is 'filler', so hesitation and backchannels are ignored as long as the remaining task meaning is preserved.", "heard_transcript": "How many activities are done in um my bucket list note in let me think Simple Note?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I mean as many things as possible with $400 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "32161cb5bdea80c1140c5538e6045b66758eacb64077771c3735c7dd5a303501", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and scope, with only the initial mistaken phrasing being replaced." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/false_start/false_start_01_selection_objective/a5cdb506e3221b3e/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "6d4226cfb156e4766c14980aafa98dcc1ac6dadf47280467002dd9fa1779c640", "seconds": 10.24, "spoken_utterances": [ "Make an order for uh... as few things as possible—ah... no, I mean as many things as possible with four hundred dollars um... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately corrects themselves to 'as many things as possible' with a self-repair. According to the 'false_start' type criterion, the final corrected intent should be executed. The monetary value and target (Amazon wishlist) remain consistent and correctly stated. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Make an order for uh as few things as possible. I no, I mean as many things as possible with four hundred dollars um for my Amazon wishlist." }, "candidate_id": "wav_cdf975761504df49", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'as few things as possible' immediately corrected to 'as many things as possible'. The self-repair is smooth, the intonation reflects genuine reconsideration, and the pause before the correction feels like a real moment of thought. The speaker's voice and prosody remain consistent throughout, with no signs of artificiality or disjointedness." } ] }, "is_default": true }, { "variant_id": "false_start_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $300—sorry, that should be $400 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "22f57de27cccc8f141800391c29b3900f17172eb7f15243e342c165c30ebe51f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces three hundred with four hundred; the action, scope, and source remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/false_start/false_start_02_budget_amount/cf5dd08c861be78e/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "ce4a0b65f28ab115280f12235e954a3a62c97c8f4c5938c4730c9bbe9d0ad2fe", "seconds": 12.36, "spoken_utterances": [ "Make an order for as many things as possible with uh, let me think... three hundred dollars—uh... sorry, that should be four hundred dollars ... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'three hundred dollars' but immediately self-corrects to 'four hundred dollars'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Make an order for as many things as possible with uh let me think three hundred dollars uh sorry, that should be four hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_3ae4f06f950efc0c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip includes a believable hesitation ('let me think') and a revision of the budget from $300 to $400. The correction is natural, with a slight shift in intonation indicating a change of mind. The flow is continuous and the speaker's delivery sounds spontaneous, though the initial hesitation is slightly less pronounced than in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Make an order for as many things as possible with $400 from my ebay—I should say amazon wishlist.", "long_pause_offsets": [], "source_sha256": "4b3c21ab831257e1e655ded1582f6fa5f1669ceac2d045ce62c161db268057cd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the app, monetary limit, and intent to order as many items as possible." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/false_start/false_start_03_app/3196d6dd8bb9d84a/speech.wav", "wer": 0.0, "wav_sha256": "aa687807d84428f67dded7d36d95a6681e5ddad59445b04174ce4fbed4e4e514", "seconds": 9.6, "spoken_utterances": [ "Make an order for as many things as possible with four hundred dollars ... from my ebay—uh... I should say amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon wishlist'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The corrected target 'Amazon wishlist' matches the reference, and all other constraints (order as many things as possible with $400) are preserved. Therefore, the final intent aligns with the original instruction.", "heard_transcript": "Make an order for as many things as possible with four hundred dollars from my eBay uh, I should say Amazon wishlist." }, "candidate_id": "wav_8a184822cb7d8526", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a self-correction from 'eBay' to 'Amazon wishlist'. The correction is clear and the speaker's tone suggests a genuine realization. However, the initial phrase 'Make it order' is slightly awkward and the correction feels a bit rushed, making the overall delivery less fluid than the top two." } ] }, "is_default": false }, { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $400 from my amazon shopping list—scratch that, make it wishlist.", "long_pause_offsets": [], "source_sha256": "2ef25954e38dcabfd464c5c40c0958b8d22443cd0c679bf6ce237a183e91d8cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist'. The correction is clear and unambiguous, preserving the original intent of sourcing from the wishlist. The monetary limit and goal of maximizing items are unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/false_start/false_start_04_source_collection/7abb422cd0854852/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "dae113596338df9933d2e604ae73530e21a0c22786c33b602a40debb36a36921", "seconds": 11.88, "spoken_utterances": [ "Make an order for as many things as possible with four hundred dollars um... from my amazon shopping list—hmmm... scratch that, make it um... wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wishlist' after a pause and 'scratch that'. According to the 'false_start' type criterion, the final corrected value ('wishlist') should be applied, and the original incorrect value ('shopping list') should be discarded. The final intent matches the original instruction with the corrected target.", "heard_transcript": "Make an order for as many things as possible with four hundred dollars, um, from my Amazon shopping list. Hmm. Scratch that. Make it, um, wishlist." }, "candidate_id": "wav_9e51b9d7c428325f", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start with 'scratch that' and a correction from 'shopping list' to 'wishlist'. While the correction is present, the initial phrase 'Make an order' is slightly off, and the overall delivery feels a bit more robotic. The pauses are longer and less naturally integrated, reducing the sense of spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $400 from my amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "ab8a6236890e542a04709018ee8af40662e9182097ee2fa10afe94694962c34d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial instruction, which already specified the amount and source. The final goal combines both requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/addition/addition_01/fa451e2c6b49479f/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "85563bd168be3cb185514896c6a4cfbccf702218369aa701a37a661ff1abfd5b", "seconds": 13.8, "spoken_utterances": [ "Make an order for with ... four hundred dollars uh... from my amazon wishlist. ", "Oh, I forgot to mention, make the order for ... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was incomplete ('Make an order for with four hundred dollars...'), but the speaker immediately corrected and completed it with 'make the order for as many things as possible.' This constitutes a clear addition that combines with the initial instruction to form the complete goal matching the reference. The final intent is fully aligned with the original instruction.", "heard_transcript": "Make an order for with four hundred dollars uh from my Amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible." }, "candidate_id": "wav_043c75e9f018b847", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Make an order for as many things as possible with from my amazon wishlist. One more detail: use $400 for the order.", "long_pause_offsets": [ 75 ], "source_sha256": "f632147c3601ba3a0c4123f7d82de5cb03261007e5c4df6b9c4f190b43cb66ff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing monetary constraint, replacing the incomplete 'with' clause with the explicit value 'four hundred dollars'. The initial instruction's intent to maximize quantity within a budget is preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Make an order for as many things as possible with um... from my amazon wishlist. ", "One more detail: use uh... four hundred dollars for the order." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Make an order for as many things as possible with $400 from my ebay wishlist. Let me correct the app: use 'amazon', not 'ebay'.", "long_pause_offsets": [ 78 ], "source_sha256": "3f8360902b51bc03702f0300842695f57d03e432b0622301e02f1b1a88d9234b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from 'ebay' to 'amazon' while preserving the amount ($400) and the goal of ordering as many items as possible from the wishlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/revision/revision_03_app/3407e8f5a569dc7a/speech.wav", "wer": 0.0, "wav_sha256": "d309f32185453f1d4e9195ac83010960a8a03e93658da14c49dad4f1e8ae159c", "seconds": 14.68, "spoken_utterances": [ "Make an order for as many things as possible with four hundred dollars let's see... from my ebay wishlist. ", "Let me correct the app: use 'ah... amazon', not 'ebay'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects it to 'Amazon, not eBay' after a short pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to use Amazon, which matches the original reference instruction. The monetary value and goal of ordering as many items as possible remain unchanged.", "heard_transcript": "Make an order for as many things as possible with four hundred dollars, let's see, from my eBay wishlist. Let me correct the app, use Amazon, not eBay." }, "candidate_id": "wav_26e3147e69032be1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a clear, confident request, then hesitates with 'let me see' before correcting the app from 'eBay' to 'Amazon'. The self-correction is smooth, with a natural pause and a clear, unforced intonation shift. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible with $400 from my amazon wishlist. Wait, the selection objective should be 'as many things as possible', not 'as few things as possible'.", "long_pause_offsets": [ 79 ], "source_sha256": "90ba33e4f26f2c4691af6d9f2c96320cc5b6b9d7c6d5e14979581bf6a7afb446", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the selection objective from 'as few things as possible' to 'as many things as possible', while retaining the budget and source. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/revision/revision_01_selection_objective/13725e6819d7eb25/speech.wav", "wer": 0.02702702702702703, "wav_sha256": "6832d7952b53c5bd051680831fc90e84024b5875121068ff76467e88e51b0f8d", "seconds": 17.72, "spoken_utterances": [ "Make an order for as few things as possible with four hundred dollars um... from my amazon wishlist. ", "Wait, the selection objective should be 'let's see... as many things as possible', not 'as few things as possible'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a goal to order 'as few things as possible' but immediately corrects this to 'as many things as possible' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, including the budget and source. Therefore, the task should be executed as revised.", "heard_transcript": "Make an order for as few things as possible with four hundred dollars um from my Amazon wishlist. Wait, the selection objective should be let's see, as many things as possible, not as few things as possible." }, "candidate_id": "wav_9399f093c551ab64", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision where the speaker corrects the selection objective from 'as few things as possible' to 'as many things as possible'. The initial request is slightly hesitant with 'um', and the correction is delivered with a clear change of mind. However, the phrasing 'the selection objective should be' sounds slightly more formal and less spontaneous than the other clips, making it feel a bit more like a structured thought process than a natural spoken correction." } ] }, "is_default": false }, { "variant_id": "revision_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $300 from my amazon wishlist. Make that $400, not $300.", "long_pause_offsets": [ 80 ], "source_sha256": "0573ee3862e9d3eb380783c3143817f9210def0f1665f1304b895dbeb2270066", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the budget amount from three hundred to four hundred dollars, while retaining the action, goal (maximize items), and source (amazon wishlist)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/revision/revision_02_budget_amount/67678dfef27d1bfa/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "445ecc08638ca54c570b6ad133450b52aa60be1868bc8b8e04968de440737f8e", "seconds": 12.6, "spoken_utterances": [ "Make an order for as many things as possible with three hundred dollars ... from my amazon wishlist. ", "Make that um... four hundred dollars, not three hundred dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'three hundred dollars' but immediately corrects it to 'four hundred dollars, not three hundred dollars'. This is a clear revision where the final value is 'four hundred dollars'. The final intent matches the original instruction, so the task passes.", "heard_transcript": "Make an order for as many things as possible with three hundred dollars from my Amazon wishlist. Make that um four hundred dollars, not three hundred dollars." }, "candidate_id": "wav_be57fdfaaddbe51d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision of the budget from $300 to $400. The speaker starts with a clear request, then corrects the amount. The correction is natural, but the phrase 'make that um' introduces a filler that feels slightly awkward and less fluent than the other revisions. The overall delivery is credible but less smooth, with a slightly less natural flow compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $400 from my amazon cart. Change the source collection to 'wishlist', not 'cart'.", "long_pause_offsets": [ 76 ], "source_sha256": "72db27c6d772510e4ac1cd8e90770650c62a2f33e941182ab21e6665c5fd88d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup correctly revises the source collection from 'cart' to 'wishlist' while preserving the budget and the goal of maximizing items. The monetary value is consistent in both forms." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/revision/revision_04_source_collection/a1e431ca50929716/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "466c7b48fcb4d14b643441b63aab544471e5102856496d90f39b91e7fb11568d", "seconds": 14.2, "spoken_utterances": [ "Make an order for as many things as possible with four hundred dollars uh, let me think... from my amazon cart. ", "Change the source collection to 'um... wishlist', not 'cart'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'from my Amazon cart' but then revises the source to 'wishlist, not cart' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with the source collection corrected from 'cart' to 'wishlist'. All other constraints (amount, goal, and action) remain consistent.", "heard_transcript": "Make an order for as many things as possible with four hundred dollars, uh, let me think, from my Amazon cart. Change the source collection to, um, wishlist, not cart." }, "candidate_id": "wav_907fadeba9881789", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision from 'cart' to 'wishlist'. The speaker begins with a clear request, then hesitates with 'uh let me think' before correcting the source. The correction is delivered with a filler ('um') and a slightly hesitant tone. The prosody is less consistent, and the correction feels more like a mechanical edit than a spontaneous recollection, making it the least realistic of the four." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "aadf2b9abcb07cd32a4a336856c628d11ffe58f345a90224f973fc0ffca223b4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/cancellation/cancel_037/1ede6da09ffd79e2/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "9b8d904dc1c12acde1a20489cbeb7ca9bfa3359718d32190da623ae89951dd5d", "seconds": 14.56, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist.", "Hang on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly and unambiguously cancels the entire task with 'Hang on, stop there, cancel the whole task.' This meets the criterion for a cancellation, as the final audible intent is to revoke the task, not to proceed with it.", "heard_transcript": "Make an order for as many things as possible with um four hundred dollars um from my amazon wishlist. Hang on, stop there, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "e8face98b0da38d3d3d57d3eda363bac36be42a249eede7584b4b120a31592fe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/ambiguous_halt/halt_030/4ee0763e655a1cc1/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "498e132326d3c5062685988e67870e1f773e0d5b808f9ff5b38b84206a0c2b9c", "seconds": 12.96, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist.", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause the current task. The speaker does not cancel the task or provide a new instruction. The final interjection indicates a temporary halt, consistent with the 'ambiguous_halt' type. No issues are present as the final intent is to pause, not execute or cancel.", "heard_transcript": "Make an order for as many things as possible with um four hundred dollars um from my amazon wishlist. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "36586c81255f71baf86bdb5b4b9f8ebc3fb143c3b0248693fc74c204a3807aa9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_1/audio_natural/filler/original/1445fef377cc34d2/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "d85baca40827329f4eda8ba315215093a5fcbb938e1c670e4ad538bd7cfd4acd", "seconds": 9.44, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... four hundred dollars um... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um') but does not change the core request. The final intent matches the original instruction, with 'four hundred dollars' being the audible equivalent of '$400'. The task is executed as requested.", "heard_transcript": "Make an order for as many things as possible with um, four hundred dollars um, from my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I meant as many things as possible with $500 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "e293ccadce001b47e4c46d6c4011e7ed303f71e86687850216cb34818e56ac25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/false_start/false_start_01_selection_objective/247efd372e790d12/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "5723bf0ecfb6637e97e0650dcd97650e412e95a1fc7ad653c70b63f9cf9aa56e", "seconds": 12.2, "spoken_utterances": [ "Make an order for as few things as possible—uh... no, I meant hmmm... as many things as possible with ... five hundred dollars from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately self-corrects to 'as many things as possible'. This is a false start, and the final intent is the corrected version. The monetary value and source are consistent with the original instruction. The final control intent is to execute the corrected task.", "heard_transcript": "Make an order for as few things as possible. Um, no, I meant, as many things as possible with five hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_df11239ca6ab605b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous flow. The speaker begins with a clear false start ('as few things as possible'), immediately self-corrects with a believable 'no, I meant' and a brief throat clear, then smoothly continues with the corrected request. The prosody is consistent, the correction is integrated into the utterance, and the pacing feels like genuine real-time thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Make an order for as many things as possible with $500 from my ebay—I should say amazon wishlist.", "long_pause_offsets": [], "source_sha256": "d2b1a1fe1a870af10cb8d6cf6e0380b3e3794e8a98931b1b4f1e40fe7fe6e8c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the original intent of sourcing from the Amazon wishlist. The monetary value and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/false_start/false_start_03_app/b783a57e1ccec8b1/speech.wav", "wer": 0.0, "wav_sha256": "6613ac6e33a78393559b8566d96f736b0eaf7a1bab951556821c9c81037fc9af", "seconds": 10.2, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars from my ebay—uh... I should say amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon wishlist'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The corrected target ('Amazon wishlist') and the monetary value ($500) are preserved, and the overall goal remains consistent with the original instruction. Therefore, the final intent matches the expected execution goal.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars from my eBay uh I should say Amazon wishlist." }, "candidate_id": "wav_c62351cf240cbdbf", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible revision. The speaker starts with 'Make an order for as many things as possible with um... $500 from my eBay uh...', then corrects 'I should say Amazon wishlist'. The hesitation and self-correction are natural, and the intonation suggests a genuine change of mind. The flow is slightly less smooth than audio_04 due to the longer pause and the 'uh' filler, but the overall delivery is realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $300—sorry, that should be $500 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "79e2d5252723595bdc026971b5b6d8778ebac7998ea0d9f714579055f726369a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces three hundred with five hundred; the action, scope, and source remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/false_start/false_start_02_budget_amount/86332513f56bc70b/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "004c1cb3b61b66b1fc1bc330a6fb3ac283c16ea84d5a36fc3c425e81deb0bf4e", "seconds": 10.04, "spoken_utterances": [ "Make an order for as many things as possible with uh... three hundred dollars—um... sorry, that should be five hundred dollars from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'three hundred dollars' but immediately self-corrects to 'five hundred dollars'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is executed with the corrected amount. The rest of the instruction matches the original.", "heard_transcript": "Make an order for as many things as possible with uh three hundred dollars. Um, sorry, that should be five hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_e5ee56126b325deb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker first says '$300' and then corrects to '$500'. The correction is clear, but the initial value is stated without hesitation, making the revision feel slightly less spontaneous than in audio_02 or audio_04. The 'um' filler is present, but the overall delivery is a bit more hesitant and less fluid in its flow." } ] }, "is_default": false }, { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $500 from my amazon cart—scratch that, make it wishlist.", "long_pause_offsets": [], "source_sha256": "10f4a728cc07b2eeca0a7f8f666d7551ac9cfff7fd5ad34aa78d92d03b047543", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'amazon cart' but immediately corrects it to 'wishlist'. The final target collection is the wishlist, and all other conditions (amount, action, scope) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/false_start/false_start_04_source_collection/1b6e2b8b97aa811b/speech.wav", "wer": 0.25925925925925924, "wav_sha256": "2ba5e30b17dd1b79a3bf28d1599d1a33e38b5778b8de98ea3389c1bd505157ff", "seconds": 14.6, "spoken_utterances": [ "Make an order for as many things as possible with five hundred dollars from my amazon um, let me think... cart—um... scratch that, make it ... wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'cart' but immediately self-corrects with 'scratch that, make it wishlist.' This is a clear false start where the final target is corrected to 'wishlist.' The monetary value and action remain consistent. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Make an order for as many things as possible with five hundred dollars from my Amazon um, let me think, cart um, scratch that, make it wishlist." }, "candidate_id": "wav_dfa9d4d791b59253", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start where the speaker says 'Make an order for as many things as possible with $500 from my Amazon um... let me think... cart um... scratch that, make it wishlist'. The 'scratch that' is a very clear cancellation, but the delivery feels slightly more disjointed. The 'let me think' and 'um' fillers are present, but the correction is abrupt and the flow is less continuous compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $500 from my amazon wishlist. Oh, I forgot to say, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "8782a0e50e31677804aec059c395e4e9a1c63c9f2c39b5a98028eb96e0bcd93e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial order request, which already included the amount and source. The final goal combines both utterances without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/addition/addition_01/37a48c39b0cfd443/speech.wav", "wer": 0.13793103448275862, "wav_sha256": "d4f58e1efa677448230be62a58703361f3a86091e5ba51bfca88258b8d5a19ca", "seconds": 13.48, "spoken_utterances": [ "Make an order for with five hundred dollars let's see... from my amazon wishlist. ", "Oh, I forgot to say, make the order for hmmm... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make an order for with five hundred dollars' and then adds 'as many things as possible' after a pause and self-correction. This matches the 'addition' type, where the final goal combines the initial instruction with the added constraint. The final audible intent is to make an order for as many items as possible within the $500 budget from the Amazon wishlist, which aligns with the original instruction.", "heard_transcript": "Make an order for with five hundred dollars, let's see. From my Amazon wishlist. Oh, I forgot to say, make the order for, hmm, as many things as possible." }, "candidate_id": "wav_cc54df14955377af", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' and a pause, which is a believable moment of thought. The correction 'Oh, I forgot to say' is a plausible self-aware interjection, followed by a smooth, additive continuation. The prosody is consistent, and the flow from initial request to addition feels spontaneous and human." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Make an order for as many things as possible with from my amazon wishlist. I left out a detail: use $500 for the order.", "long_pause_offsets": [ 75 ], "source_sha256": "2fc6e3aba2ec3d822fa7fc5f6287e0f0abd86fc0d3b86faeef499527a9ef89b1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing monetary constraint, replacing the incomplete 'with' clause with the explicit amount. The initial instruction's intent to maximize items within a budget is preserved, and the final value matches the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/addition/addition_02/91433816f7688bdf/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "2951d5038d0131c1b17a9a3a3410c4fb71a6490bb396bc6ff1ae66921d4e22dc", "seconds": 12.0, "spoken_utterances": [ "Make an order for as many things as possible with uh... from my amazon wishlist. ", "I left out a detail: use um... five hundred dollars for the order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially omits the amount, then adds 'five hundred dollars' after a pause and self-correction. This matches the addition type criterion: the original instruction is combined with the added information. The final goal is identical to the reference instruction, with no conflicting or missing constraints.", "heard_transcript": "Make an order for as many things as possible with a, from my Amazon wishlist. I left out a detail, use um, five hundred dollars for the order." }, "candidate_id": "wav_c4190243b0db57ca", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also makes a natural-sounding addition, the initial request is slightly awkward with 'with a' left hanging, suggesting a possible recording glitch or incomplete thought. The correction 'I left out a detail' is followed by a filler 'um' and then the value, but the overall delivery feels a bit more rehearsed or less fluid than audio_01. The prosody is less varied, and the addition lacks the same level of conversational spontaneity." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $300 from my amazon wishlist. Change the budget to $500, not $300.", "long_pause_offsets": [ 80 ], "source_sha256": "53b9079b63b1adbdcce661b84f624931b43a8f7758b2e6328341290e8d01a17e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the budget from three hundred to five hundred dollars, preserving the action, scope, and source from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/revision/revision_02_budget_amount/dc8a2e3c6a4aed95/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "2eca55d7e10f69a99aa579722f5beb05b1c28f463fb374664d5126d56e9636ed", "seconds": 13.48, "spoken_utterances": [ "Make an order for as many things as possible with um... three hundred dollars from my amazon wishlist. ", "Change the budget to uh... five hundred dollars, not three hundred dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a budget of $300, then clearly revises it to $500 after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, including the action, target, and budget value.", "heard_transcript": "Make an order for as many things as possible with um, three hundred dollars from my Amazon wishlist. Change the budget to uh, five hundred dollars, not three hundred dollars." }, "candidate_id": "wav_19242ea9f710fa2a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural revision flow. The speaker starts with a clear request, then self-corrects the budget amount from $300 to $500 with a smooth, spontaneous-sounding 'change the budget to uh $500 not $300'. The prosody is consistent, the correction is integrated into the same utterance, and the intonation reflects a genuine moment of reconsideration. The delivery is fluent and lacks any robotic or disjointed qualities." } ] }, "is_default": true }, { "variant_id": "revision_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible with $500 from my amazon wishlist. Wait, make that 'as many things as possible', not 'as few things as possible'.", "long_pause_offsets": [ 79 ], "source_sha256": "d12ef2a02858e7f872bfc0e81f43fa33845985ef0c8e9f05ac7aa91a633e8213", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial 'as few things as possible' to 'as many things as possible', while retaining the app (amazon wishlist), budget ($500), and action (make an order)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/revision/revision_01_selection_objective/ccf1489a381da844/speech.wav", "wer": 0.08823529411764706, "wav_sha256": "b75ba9b6a2cb582b8f8d3572de5ad5c71d22223556146d3ec93b334feec9586d", "seconds": 16.32, "spoken_utterances": [ "Make an order for as few things as possible with ... five hundred dollars uh, let's see... from my amazon wishlist. ", "Wait, make that 'as many things as possible', not 'as few things as possible'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately corrects themselves with 'Wait, make that as many things as possible, not as few things as possible.' This constitutes a revision where the final audible intent is the corrected version. The correction is clear and unambiguous, and the rest of the instruction (amount, source) remains consistent. Therefore, the final intent matches the original instruction, satisfying the revision type criterion.", "heard_transcript": "Make an order for as few things as possible with five hundred dollars, uh, let's see, from my Amazon wishlist. Wait, make that as many things as possible, not as few things as possible." }, "candidate_id": "wav_9801523c62b40787", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision where the speaker initially says 'as few things as possible' and then corrects themselves to 'as many things as possible'. The correction is made with a natural 'wait, make that...' and the intonation clearly signals a change of mind. The 'uh let's see' filler is placed appropriately, and the overall flow is continuous and realistic, though the initial request is less coherent than in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $500 from my amazon shopping list. The source should be 'wishlist', not 'shopping list'.", "long_pause_offsets": [ 85 ], "source_sha256": "08a24ebafb0c882593bc24ad4a0dd486456aeaf6ac49f69d37c80ed29c35c771", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'shopping list' to 'wishlist' while preserving the action, budget, and intent to maximize quantity. The monetary value is equivalent ('five hundred dollars' matches '$500')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/revision/revision_04_source_collection/2d62465bd8ca0b20/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "9f231c60c034aa0d26a740a34674897a459db7305367bc5917e1227e2dc9d7b3", "seconds": 15.52, "spoken_utterances": [ "Make an order for as many things as possible with let me think... five hundred dollars ... from my amazon shopping list. ", "The source should be '... wishlist', not 'shopping list'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'with let me think, five hundred dollars' but immediately corrects to 'five hundred dollars' without changing the value or intent. The final instruction matches the original request, and the correction is a self-repair within the same clause. The revision type requires applying the correction after a pause, but here the correction is immediate and does not alter the core task. The final audible intent is to execute the order with $500 from the wishlist, which aligns with the original instruction. No other constraints were changed.", "heard_transcript": "Make an order for as many things as possible with let me think, five hundred dollars from my Amazon wishlist. The source should be wishlist, not shopping list." }, "candidate_id": "wav_6944213478b19d3e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker corrects the source from 'shopping list' to 'wishlist'. The correction is made in a separate sentence, which is plausible, but the flow is less integrated than in audio_03. The 'let me think' filler is used, but the correction feels slightly more like a post-hoc addition than a spontaneous self-repair within the same thought process. The delivery is otherwise natural and clear." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Make an order for as many things as possible with $500 from my ebay wishlist. Use amazon instead of ebay.", "long_pause_offsets": [ 78 ], "source_sha256": "5319380c3b34a79318e8ea575cffc68dd014aa393063fd5dfd3571a6f8d9921a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from ebay to amazon, while retaining the original monetary limit and the goal of maximizing items ordered. The final instruction matches the original in all meaningful requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/revision/revision_03_app/04d45fb15637075b/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "16034644f261585c277c737562fed3c76ff42d4bb5e09e966e98830c7c6e9606", "seconds": 17.92, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... five hundred dollars um, let me think... from my ebay wishlist. ", "Use hmmm... amazon instead of ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects it to 'Amazon instead of eBay' after a pause, which constitutes a revision. The final audible intent is to use Amazon, aligning with the original instruction. The monetary value and the goal of ordering as many items as possible within the budget are preserved.", "heard_transcript": "Make an order for as many things as possible with, um, five hundred dollar, um, let me think, from my eBay wishlist. Use home, Amazon instead of eBay." }, "candidate_id": "wav_58c046c732ac8f95", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has multiple fillers ('mm', 'um') and a correction ('use home Amazon instead of eBay'), but the delivery feels less natural. The 'mm' and 'um' fillers are placed awkwardly, and the correction is made in a separate sentence without a clear connection to the initial request. The prosody is somewhat flat, and the overall impression is that of a rehearsed or less spontaneous utterance compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "0cbe445c6425498a2212e252cd4388cd0ea389f1aed7890215e247ff5e6c9e40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/cancellation/cancel_003/d9447eba04d00ef0/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "cdeb907a1669726e39037ef46c5c810b993488ef6dccf55799d93d895ef1a854", "seconds": 13.04, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist.", "Wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the task to make an order with $500 from the Amazon wishlist, then immediately corrects themselves by saying 'Wait, please cancel that.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist. Wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_025", "hide_variant_original": true, "source_sha256": "27844723f93dcbaeb077ec7f4ad56ba2517d7f22229f24fa7a5fe465a8952887", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/ambiguous_halt/halt_025/b29649b2c29e8b69/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "f57c74b097470281f384e90a2cd518f6be6433c238aeaf4aaa68b3cbdb01d9d6", "seconds": 13.04, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist.", "Wait, I need to think for a second." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I need to think for a second,' which is a clear request to pause the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial instruction is not canceled or revised, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist. Wait, I need to think for a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9d562cd28ab39be50346aac3eb2047cb55b1fba7cdb848956df5b495c7b77588", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_2/audio_natural/filler/original/b0594a5b26b6cdb0/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "07c0444c2535ef6194665e17415cf97c86e2a920bd8fa2cb100ea2d74c7731d9", "seconds": 9.12, "spoken_utterances": [ "Make an order for as many things as possible with um... five hundred dollars hmmm... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler ('um') and a pause ('Hmm'), but the final instruction clearly matches the original request. The speaker corrects the initial hesitation and completes the task with the same intent. The monetary value is stated as 'five hundred dollars', which is equivalent to '$500'. The target app ('Amazon wishlist') and the goal ('order as many things as possible') are preserved. No correction or change to the core task is made after the initial hesitation.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars. Hmm. For my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_selection_objective", "proposed_instruction": "Make an order for as few things as possible—no, I mean as many things as possible with $600 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "d7626b298220f06db32c34515f8461068ffc118f306e9cca38624db280f5505f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'as few things as possible' to 'as many things as possible' within the same utterance. The final intent matches the original instruction in both action and constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/false_start/false_start_01_selection_objective/8f8d4b3d12a66bc2/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "1bb0c160e770458fd583795cc471402b01ac16fffea481c2f5119bbf75a2b0aa", "seconds": 12.08, "spoken_utterances": [ "Make an order for as few things as possible—um... no, I mean hmmm... as many things as possible with ... six hundred dollars from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as few things as possible' but immediately self-corrects to 'as many things as possible'. According to the 'false_start' type criterion, the final corrected intent should be executed. The monetary value and source (Amazon wishlist) remain consistent. The correction is immediate and unambiguous, satisfying the type-specific requirement.", "heard_transcript": "Make an order for as few things as possible. Um, no, I mean, as many things as possible with six hundred dollars from my Amazon wishlist." }, "candidate_id": "wav_6497ed878cb0cd23", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and fluent self-correction. The speaker initially says 'as few things as possible' and immediately corrects to 'as many things as possible' with a clear, spontaneous 'no, I mean' transition. The prosody is consistent, the pace is conversational, and the correction feels like a genuine moment of reconsideration rather than a rehearsed edit. The sentence flows smoothly without awkward pauses or unnatural phrasing." } ] }, "is_default": true }, { "variant_id": "false_start_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $500—scratch that, make it $600 from my amazon wishlist.", "long_pause_offsets": [], "source_sha256": "e7de5f7c1deb755fc77e813a4c1e6ecde12303445e5e6fd578f9dc7c20afc543", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'five hundred dollars' but immediately corrects to 'six hundred dollars'. The final amount is consistent with the original instruction, and all other conditions (order scope, source, and intent) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/false_start/false_start_02_budget_amount/ed1fc10817169b70/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "b57ec10127c549306e015027b0ed75e3b9c05c15095d76193ecf9d7ca02888ba", "seconds": 14.56, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... five hundred dollars—uh... scratch that, make it six hundred dollars uh, let's see... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'five hundred dollars' but immediately corrects to 'six hundred dollars' with 'Scratch that'. This is a clear false start where the second value is the intended one. The final instruction includes the correct amount ($600) and all other conditions (order as many things as possible, from Amazon wishlist) match the original reference. The 'let's see' is a hesitation, not a change in intent.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars um. Scratch that, make it six hundred dollars um. Let's see. From my Amazon wishlist." }, "candidate_id": "wav_9ff06b7a930e4303", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable false start with 'five hundred dollars' followed by a self-correction to 'six hundred dollars'. The speaker uses 'scratch that' as a natural way to retract the previous value. The 'let's see' pause after the correction is a plausible moment of thought before continuing. The delivery is conversational and the repair is integrated into the flow, though the initial hesitation is slightly more pronounced than in audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Make an order for as many things as possible with $600 from my ebay—I should say amazon wishlist.", "long_pause_offsets": [], "source_sha256": "c70ad3587073d488ab2f95225b80d487eff5a522165591888a28590e99dfe631", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'ebay' but immediately corrects to 'amazon wishlist'. The monetary value and intent to order as many items as possible within the budget are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/false_start/false_start_03_app/490e967e9d13bbfe/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "54e4d483d050f5ea3787cca03048d36c41bf05d9b91112b9389d39e1fdc932af", "seconds": 11.64, "spoken_utterances": [ "Make an order for as many things as possible with uh... six hundred dollars from my ... ebay—um... I should say amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ebay' but immediately self-corrects to 'amazon wishlist'. According to the 'false_start' type criterion, the final corrected target should be applied. The monetary value and the intent to order as many items as possible from the wishlist remain consistent. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Make an order for as many things as possible with uh six hundred dollars from my ebay um I should say amazon wishlist." }, "candidate_id": "wav_110984fd80b6c211", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a self-correction from 'ebay' to 'Amazon wishlist', but the phrasing 'I should say' is less natural than 'no, I mean' and sounds slightly rehearsed. The initial 'ebay' is an odd misstatement that doesn't fit the context of the rest of the sentence. The prosody is somewhat flat, and the correction feels like a mechanical fix rather than a spontaneous realization, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "false_start_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $600 from my amazon cart—I should have said wishlist.", "long_pause_offsets": [], "source_sha256": "5f057c3fa794320db688d3996642472d75718e022f1869142e51ec7605de7b26", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'amazon cart' to 'amazon wishlist' within the same utterance, preserving the action, budget, and intent to maximize quantity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/false_start/false_start_04_source_collection/83c6a4049ab9e74f/speech.wav", "wer": 0.12, "wav_sha256": "f6b5419fe1ee940412d3f6aa5b5e88465a14a3ff6bd010fbdf2f621f3e7f693e", "seconds": 12.24, "spoken_utterances": [ "Make an order for as many things as possible with uh... six hundred dollars from my amazon cart—um... I should have said hmmm... wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon cartoon' but immediately self-corrects to 'wishlist'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('cartoon') is discarded. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Make an order for as many things as possible with uh six hundred dollars from my Amazon cartoon. I should have said hmm wishlist." }, "candidate_id": "wav_e8d214dce95f3341", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. The speaker incorrectly says 'Amazon cartoon' and then self-corrects with 'I should have said, hmm, wishlist'. The 'cartoon' error is highly implausible in this context, and the correction is awkwardly phrased. The 'hmm' after 'I should have said' is unnatural and breaks the flow. The overall delivery sounds disjointed and less like a spontaneous thought process, more like a forced or poorly constructed example." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Make an order for with $600 from my amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible.", "long_pause_offsets": [ 53 ], "source_sha256": "0027efb19a80eb1addc0c6f446873781c683f7d354f28eba3f6ae3369d92a175", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'as many things as possible' to the initial order amount, combining both utterances into a complete goal that matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/addition/addition_01/93a07eaf19a6281e/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "fc5c11bf8626271fe0242e095b406faa4647d06b915a67d36e17042d8346e601", "seconds": 15.0, "spoken_utterances": [ "Make an order for with ... six hundred dollars uh... from my amazon wishlist. ", "Oh, I forgot to mention, make the order for ... as many things as possible." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make an order for with six hundred dollars from my Amazon wishlist' and then adds 'make the order for as many things as possible' after a pause. This is a clear addition that narrows the original broad request to a specific constraint. The final intent combines the original instruction with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "Make an order for with six hundred dollars uh from my Amazon wishlist. Oh, I forgot to mention, make the order for as many things as possible." }, "candidate_id": "wav_90bece520d3ba353", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Make an order for...') and then adds a detail ('Oh, I forgot to mention...') with a believable hesitation and self-correction. The prosody is consistent, the pacing includes a realistic pause and filler ('uh'), and the addition flows smoothly without sounding forced or robotic. The speaker's intonation reflects genuine recollection, making the addition feel spontaneous and credible." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Make an order for as many things as possible with from my amazon wishlist. One more detail: use $600 for the order.", "long_pause_offsets": [ 75 ], "source_sha256": "91c3e5bbde0328d57f19eb9520a18fa4a5db08b5ac26493b273849fdd58a9c19", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The initial utterance is missing the monetary value, but the followup adds the constraint 'use six hundred dollars for the order', which fills the missing value. The action, source, and goal remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/addition/addition_02/caf0e8df4310e7f2/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "79c0a4511b5c3ba1ed8e100f726ef30bf9d03d2f9232ac834c7a8b87bf75242c", "seconds": 11.6, "spoken_utterances": [ "Make an order for as many things as possible with um... from my amazon wishlist. ", "One more detail: use uh... six hundred dollars for the order." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('with um'), but the speaker adds the missing constraint 'use uh six hundred dollars for the order' after a pause. This addition completes the original instruction by specifying the budget. The final goal matches the reference instruction, so the addition type is satisfied.", "heard_transcript": "Make an order for as many things as possible with um from my Amazon wishlist. One more detail, use uh six hundred dollars for the order." }, "candidate_id": "wav_3a93b57b38bccaa1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker delivers a coherent request, the addition ('One more detail...') is less natural in flow. The phrase 'One more detail' is delivered with a slightly abrupt transition, and the subsequent correction ('use uh $600') sounds more like a rehearsed insertion than a spontaneous recollection. The prosody is flatter, and the filler 'uh' is used in a way that feels less integrated with the thought process, reducing the overall realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_source_collection", "proposed_instruction": "Make an order for as many things as possible with $600 from my amazon shopping list. Change the source collection to 'wishlist', not 'shopping list'.", "long_pause_offsets": [ 85 ], "source_sha256": "56cb2e3694e154d62cd5905832495a43dcdc38649902bb891d0b79d2195eed34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source collection from 'shopping list' to 'wishlist' while preserving the monetary limit and the goal of maximizing items. The original instruction and final spoken goal match in all meaningful requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/revision/revision_04_source_collection/e34c657078b76c18/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "1a22d0738c3162d2067124c3aff71038fee22528c0e3fc41dc3dffaa0e33a6cb", "seconds": 17.64, "spoken_utterances": [ "Make an order for as many things as possible with let me think... six hundred dollars let's see... from my amazon shopping list. ", "Change the source collection to 'um... wishlist', not 'shopping list'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the amount and the target collection.", "heard_transcript": "Make an order for as many things as possible with, let me think, six hundred dollars, let's see, from my Amazon shopping list. Change the source collection to, um, wishlist, not shopping list." }, "candidate_id": "wav_aed8d73528b0fe89", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a realistic 'let me think' filler, then states a value ('six hundred dollars'), hesitates with 'let's see', and then clearly revises the source collection from 'shopping list' to 'wishlist'. The intonation and pacing are consistent with someone formulating a request in real time, and the revision is smoothly integrated into the ongoing utterance without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "revision_02_budget_amount", "proposed_instruction": "Make an order for as many things as possible with $500 from my amazon wishlist. Make that $600, not $500.", "long_pause_offsets": [ 80 ], "source_sha256": "1785afbfbdd3401cdad87d335e31ff08211232054f9e7ad39e9c2f25eab0e67b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the amount from five hundred to six hundred dollars, while retaining the action, goal (maximize items), and source (amazon wishlist)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/revision/revision_02_budget_amount/d548a011eda2b2d2/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "ac17da2bb31e5354f7a6147fb52c02aa099bae692d1e05f6305d69eb6bf15bee", "seconds": 14.08, "spoken_utterances": [ "Make an order for as many things as possible with hmmm... five hundred dollars from my amazon wishlist. ", "Make that uh... six hundred dollars, not five hundred dollars." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'five hundred dollars' but immediately corrects to 'six hundred dollars, not five hundred dollars'. This is a clear revision where the final value is 'six hundred dollars'. The final intent matches the original instruction, so the task passes.", "heard_transcript": "Make an order for as many things as possible with um five hundred dollars from my Amazon wishlist. Make that uh six hundred dollars, not five hundred dollars." }, "candidate_id": "wav_f39ca8fbee3a99e5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear revision from $500 to $600, but the phrasing 'Make that uh six hundred dollars not five hundred dollars' sounds slightly less spontaneous than audio_02. The 'uh' is a natural hesitation, but the correction is delivered in a more declarative, almost rehearsed tone. The prosody is slightly flat, and the revision feels more like a direct replacement than a fluid thought process, making it less realistic than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_01_selection_objective", "proposed_instruction": "Make an order for only the cheapest items with $600 from my amazon wishlist. Wait, the selection objective should be 'as many things as possible', not 'only the cheapest items'.", "long_pause_offsets": [ 77 ], "source_sha256": "26c1c13af1964a7cd702d861ca2b5ee00c6d49de37bdcfb60b0c549edb5b72f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the selection objective from 'only the cheapest items' to 'as many things as possible', while retaining the budget and source. The original instruction's intent is restored." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/revision/revision_01_selection_objective/fbcdd97dcb7a91b0/speech.wav", "wer": 0.08823529411764706, "wav_sha256": "b3bbf2f375452b57b6f29432def8e357fecc710ed5d5a47917de1c07e10d5e17", "seconds": 16.32, "spoken_utterances": [ "Make an order for only the cheapest items with ... six hundred dollars uh... from my amazon wishlist. ", "Wait, the selection objective should be 'ah... as many things as possible', not 'only the cheapest items'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only the cheapest items' but immediately corrects themselves with 'Wait, the selection objective should be uh as many things as possible, not only the cheapest items.' This is a clear revision where the final intent overrides the initial incorrect value. The final goal matches the original instruction, so the task passes.", "heard_transcript": "Make an order for only the cheapest items with six hundred dollars from my Amazon wishlist. Wait, the selection objective should be uh as many things as possible, not only the cheapest items." }, "candidate_id": "wav_9c1c29de7dd8e8be", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker initially says 'only the cheapest items' and then corrects to 'as many things as possible'. The 'wait' is a believable hesitation, and the correction is clear. However, the initial request is phrased in a more rigid, task-oriented way ('Make an order for only the cheapest items'), and the correction feels slightly abrupt. The overall delivery is less conversational and more like a structured command with a late correction, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Make an order for as many things as possible with $600 from my Walmart wishlist. Use Amazon instead of Walmart.", "long_pause_offsets": [ 81 ], "source_sha256": "c35da2c043173388f809b581cc4c40aaac0ad985e1e3428d38d52bee183b06d2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Walmart to Amazon, while retaining the budget and the goal of maximizing items. The original instruction's app was Amazon, and the spoken instruction's final app is Amazon, so the final goal matches the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/revision/revision_03_app/1e26a86fd58b35ef/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "11bdf57d98d337ca5b6fa45c722be9a60d61a75c8150b98b89c32f724652445d", "seconds": 11.64, "spoken_utterances": [ "Make an order for as many things as possible with uh... six hundred dollars ... from my Walmart wishlist. ", "Use Amazon instead of Walmart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Walmart wishlist' but immediately corrects it to 'Use Amazon instead of Walmart.' This is a revision where the final instruction overrides the initial incorrect target. The monetary value ($600) and the goal ('as many things as possible') remain consistent. The final intent matches the original instruction, so the task passes.", "heard_transcript": "Make an order for as many things as possible with uh six hundred dollars from my Walmart wishlist. Use Amazon instead of Walmart." }, "candidate_id": "wav_7e5aa02707486e33", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio presents a revision where the speaker first says 'Walmart wishlist' and then corrects to 'Amazon'. However, the phrasing 'Use Amazon instead of Walmart' is delivered in a flat, declarative tone that sounds more like a command than a spontaneous correction. The initial statement lacks hesitation or filler, and the correction is abrupt and lacks the natural flow of a self-correcting thought process. The prosody is monotonous, and the revision feels mechanically inserted rather than arising from real-time deliberation." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "e9facae1b90eda74c99dda02ad60949de2cf95deddf0d786b72ccf3a4836c953", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/cancellation/cancel_029/ea6b18694b3a4411/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "b37bac35077d68e2e2f3727d7fc995d28249e4d960f1124c5c00894103efed52", "seconds": 11.76, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a complete order request, then immediately follows with 'Hang on, I'm withdrawing that request,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct revocation, satisfying the cancellation type criterion.", "heard_transcript": "Make an order for as many things as possible with six hundred dollars uh from my Amazon wishlist. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "3faf3aa15755c66b65306093ccc88b583c74ad3d1a57e91c0d571ce61e7c4970", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/ambiguous_halt/halt_026/7da286241974a4fc/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "5dbb85745dd5fb0d221890faaf199bf2d54bce33160065b9c5ccb694a9a55437", "seconds": 12.0, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment, let me think about that,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation. The initial instruction is not executed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "Make an order for as many things as possible with six hundred dollars uh from my Amazon wishlist. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7b97e6d_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "68b646f5398daefe0c306b80f33bada73f496f0d7353a6f44b78b1447e1c0681", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7b97e6d_3/audio_natural/filler/original/f061025cc20f96aa/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "287232743dcbdc607b002a70340d8912f8856e824cebe0ab77159ac041e1a8a8", "seconds": 7.48, "spoken_utterances": [ "Make an order for as many things as possible with ... six hundred dollars uh... from my amazon wishlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh') but does not change the core request. The final intent matches the original instruction in action, target, and value. The filler nature of the pauses and hesitations does not alter the goal, so the task should be executed.", "heard_transcript": "Make an order for as many things as possible with uh six hundred dollars uh from my Amazon wishlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Buy me a kitchen scale—I meant food scale as David recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "61fe4fd5cd507fb232861f55b26ccb5d0bb146d287575ce0241707a143748af5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'kitchen scale' to 'food scale' within the same utterance, preserving the rest of the instruction including the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/false_start/false_start_01_product_category/08d5daf691ce2e27/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "1a718cf22b46cd4fefeeda69d6e0eedf63e70120373bb26c10fcc4694e8a13f4", "seconds": 7.52, "spoken_utterances": [ "Buy me a kitchen scale—um... I meant ... food scale as David recommended in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'kitchen scale' but immediately corrects to 'food scale' with 'I meant'. This is a false start where the final corrected value is the intended target. The rest of the instruction, including 'as David recommended in his phone message', is preserved and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "Buy me a kitchen scale. Um, I meant, food scale as David recommended in his phone message." }, "candidate_id": "wav_082a9d224959d078", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'kitchen scale', self-corrects to 'food scale' with a clear, spontaneous 'I meant' repair, and smoothly continues with the full context 'as David recommended in his phone message'. The prosody is consistent, the pause before the correction is brief and realistic, and the entire utterance flows as a single, coherent thought process without any awkwardness or unnatural phrasing." } ] }, "is_default": true }, { "variant_id": "false_start_02_person_name", "proposed_instruction": "Buy me a food scale as Emily—sorry, I mean David recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "74781712726e066ac8deee9afd94e6c7552fd0987b165ee128a40c4d4714d7b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misnames the recommender as Emily but immediately corrects to David. The rest of the instruction, including the item (food scale), the action (buy), and the context (recommended in his phone message), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/false_start/false_start_02_person_name/5e3e78bed9866f76/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5edc102d383e3441cec95b9351b79b1404d6b418468f5ff0891cdd066be19f54", "seconds": 10.32, "spoken_utterances": [ "Buy me a hmmm... food scale as Emily—uh... sorry, I mean David recommended in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Emily' but immediately self-corrects to 'David'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction with the correct person.", "heard_transcript": "Buy me a um food scale as Emily um sorry, I mean David recommended in his phone message." }, "candidate_id": "wav_4f3a67843e8aa940", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a false start with 'kitchen scale' replaced by 'food scale', but the speaker then hesitates with 'Emily um' before correcting to 'David'. The repair is less smooth than audio_01, with a longer and more awkward pause and filler ('um') that disrupts the flow. The correction feels slightly forced, and the overall delivery is less fluid, making it less realistic than audio_01 despite having a similar structure." } ] }, "is_default": false }, { "variant_id": "false_start_03_message_channel", "proposed_instruction": "Buy me a food scale as David recommended in his email—scratch that, make it phone message.", "long_pause_offsets": [], "source_sha256": "981f143219823e4e18dd1af53afd5c7a7bffa4622d164ab7020b7fb677557856", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately corrects it to 'phone message' within the same utterance. The final intended channel is phone message, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/false_start/false_start_03_message_channel/83ecaea332febc77/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "8fc3dda4e8b21bc3a8d1eb37e1e515e51e987fd5e7f9df49372cf819846feab8", "seconds": 8.04, "spoken_utterances": [ "Buy me a food scale as David recommended um... in his email—ah... scratch that, make it phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately self-corrects to 'phone message' after a pause and the phrase 'scratch that'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction with the correct target medium.", "heard_transcript": "Buy me a food scale as David recommended um in his email uh scratch that, make it phone message." }, "candidate_id": "wav_2ec5788259226633", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision rather than a simple false start. The speaker initially says 'as David recommended in his email', then immediately corrects to 'phone message'. The correction is abrupt and lacks the natural hesitation or self-repair flow seen in the other clips. The 'scratch that' interjection sounds more like a command than a spontaneous thought correction, and the overall delivery feels less organic and more like a rehearsed or edited response." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Buy me a food scale in his phone message. Oh, I forgot to say, it should be as David recommended.", "long_pause_offsets": [ 42 ], "source_sha256": "13184b8ae78a1b9aa6378fb399b4150a0c4727a6d570d42088a71d3ed07a4f0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'as David recommended', which was present in the original instruction. The initial utterance incorrectly placed 'in his phone message' as the location or context, but the correction clarifies the recommendation source. The final goal matches the original by specifying the item and the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/addition/addition_01/d8fba02856b57b51/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "613dceffa90e5176c7f1cfef71340270e131d9692eecc4a830eac28efc2029dc", "seconds": 10.0, "spoken_utterances": [ "Buy me a um... food scale ... in his phone message. ", "Oh, I forgot to say, it should be as David recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Buy me a food scale in his phone message.' The speaker then adds 'Oh, I forgot to say, it should be as David recommended.' This addition clarifies the condition under which the scale should be bought, modifying the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction. The addition is clear and directly relevant, satisfying the addition type criterion.", "heard_transcript": "Buy me a um, food scale in his phone message. Oh, I forgot to say, it should be as David recommended." }, "candidate_id": "wav_4bcf3c6c73de073d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a slightly off-the-cuff phrasing ('in his phone message'), which sounds like a spontaneous thought. The addition of the constraint ('it should be as David recommended') is delivered with a believable, slightly apologetic tone and a natural pause before the correction, fitting the 'addition' pattern well. The prosody is consistent, and the flow feels like a real-time recollection." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Buy me a food scale as David recommended. I forgot to add, it was in his phone message.", "long_pause_offsets": [ 42 ], "source_sha256": "882cb105e25cc3c363bdbb2e856e25fafb207704067069949c681c1998258450", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context that the recommendation was in David's phone message, which was already implied in the original instruction. The core action and subject remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/addition/addition_02/9bc7ebfb5195d6bd/speech.wav", "wer": 0.2777777777777778, "wav_sha256": "fc066282199faf15b1a7e818acbe2fe11bd10f6436b457911a40b1234b16670f", "seconds": 10.72, "spoken_utterances": [ "Buy me a ... food scale as David recommended. ", "I forgot to add, it was in his ... phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Buy me a food scale as David recommended' is followed by an addition clarifying that the recommendation was 'in his phone message'. This addition narrows the scope of the original instruction by specifying the context of the recommendation, which matches the original reference. The final intent combines the original request with the added detail, satisfying the addition type criterion.", "heard_transcript": "Buy me a food scale as David recommended. I forgot to add it was in his phone message." }, "candidate_id": "wav_8f0fea7a348de918", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the content is similar, the delivery is less natural. The speaker starts with a sigh ('ah') and a slightly more formal tone. The addition ('I forgot to add it was in his phone message') is delivered with a flat, almost rehearsed intonation, and the phrasing feels more like a correction than a spontaneous memory. The prosody lacks the organic variation of a real speaker, making it sound less realistic despite being semantically coherent." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_person_name", "proposed_instruction": "Buy me a food scale as Sarah recommended in his phone message. I misspoke about the person: it should be David, not Sarah.", "long_pause_offsets": [ 63 ], "source_sha256": "339a1bb2dac4d67aa35668bf8a49b488a51611d81390330fd4a416ba3affd6da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recommender from Sarah to David, preserving the action, item, and context of the phone message." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/revision/revision_02_person_name/904e4025dfab4251/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "6c65a4820e3398f959d580e744b4adbc49fed94cde325b57a1a53c69e785518f", "seconds": 14.08, "spoken_utterances": [ "Buy me a uh, let's see... food scale as ... Sarah recommended um... in his phone message. ", "I misspoke about the person: it should be David, not Sarah." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the recommender as 'Sarah' but immediately corrects themselves with 'I misspoke about the person, it should be David, not Sarah.' This constitutes a revision where the final intent is the corrected version. The original instruction is fully preserved except for the name, which is corrected. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Buy me a uh let's see food scale as Sarah recommended um in his phone message. I misspoke about the person, it should be David, not Sarah." }, "candidate_id": "wav_7388041a588b08ab", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous revision flow. The speaker begins with a filler ('let's see'), makes a clear mistake ('Sarah'), immediately self-corrects with 'I misspoke about the person, it should be David, not Sarah,' and maintains a consistent, conversational pace. The correction is integrated smoothly into the utterance, and the prosody reflects genuine recollection and correction rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Buy me a kitchen scale as David recommended in his phone message. Change the product to a food scale, not a kitchen scale.", "long_pause_offsets": [ 66 ], "source_sha256": "795156c94af1f1acef73dbbc979177a6bb702c832aff86653ac6ed1dd2ac24c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the product from 'kitchen scale' to 'food scale', while retaining the rest of the instruction including the recommendation source and the action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/revision/revision_01_product_category/a5306a3f3e9b848e/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "717b07cac9287c5c156f21b7d146aac9e721f4c5034c3b9802ab22d90b55902d", "seconds": 14.28, "spoken_utterances": [ "Buy me a ... kitchen scale as let me think... David recommended in his phone message. ", "Change the product to a ... food scale, not a kitchen scale." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'kitchen scale' but immediately corrects it to 'food scale' after a pause, which matches the revision type criterion. The final intent is to purchase a food scale, consistent with the original instruction. The correction is clear and unambiguous, and no other constraints are altered.", "heard_transcript": "Buy me a kitchen scale as David recommended in his phone message. Change the product to a food scale, not a kitchen scale." }, "candidate_id": "wav_b31ef406d1d25cc4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible revision with a filler ('let me think') and a clear correction ('change the product to a food scale, not a kitchen scale'). However, the correction is delivered as a separate, slightly abrupt sentence rather than being fully integrated into the original request. The prosody is natural, but the flow feels slightly more segmented compared to audio_03, making it less fluid overall." } ] }, "is_default": false }, { "variant_id": "revision_03_message_channel", "proposed_instruction": "Buy me a food scale as David recommended in his email. Make that a phone message, not an email.", "long_pause_offsets": [ 55 ], "source_sha256": "691ea2ff64a618caed1d9ca88c57b5087dc07e8864740fa2c163946179d9d3c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the communication channel from email to phone message, preserving the action, item, and recommender." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/revision/revision_03_message_channel/82eaf12216126005/speech.wav", "wer": 0.15, "wav_sha256": "1a9d151939bd7d16106544f0313bcc730347aaee3c020c39d28931c9836369f4", "seconds": 10.76, "spoken_utterances": [ "Buy me a uh... food scale as David recommended in his email. ", "Make that a ... phone message, not an email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately corrects it to 'phone message' after a pause. This is a revision where the final intent is to use 'phone message' instead of 'email'. The correction is clear and the rest of the instruction remains unchanged. The final control intent is to execute the revised task.", "heard_transcript": "Buy me a food scale as David recommended in his email. Make that a phone message, not an email." }, "candidate_id": "wav_ba088745835572d4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision but feels less natural due to its structure. The speaker first says 'buy me a food scale as David recommended in his email,' then corrects it to 'phone message.' The correction is a simple substitution without any self-awareness or integration into the original sentence. The delivery is flat and lacks the organic hesitation or self-correction seen in the other clips, making it sound more like a direct edit than a spontaneous spoken revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "ab5484cf22b4718c62a3069dd2cfbb7f450c5a4c1e9b8d0a4b6567ad7163c926", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/cancellation/cancel_026/74055ddf60016657/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "338a0fcbe3dd7166089132842d6afc8fd907dff32f7db94b79b33a9016b1a1b7", "seconds": 11.04, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message.", "Hang on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to buy a food scale, then immediately follows with 'Hang on, forget what I just asked. Cancel it.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and not followed by any resumption or new instruction, satisfying the criterion for a 'cancellation' type task.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message. Hang on, forget what I just asked. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "ea23d2961e9d86ccbda037ad14f084e20a56d73beb74156d3a988d6f70a31862", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/ambiguous_halt/halt_040/03fdef4a4a712f0d/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "842ca29cc69fa60d2b49fa4c2d60808068f4710f336866b404b2a82e85ed71f7", "seconds": 9.76, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly requests to hold or wait with 'Hang on, just a minute.' This constitutes a final, audible request to pause, which matches the 'ambiguous_halt' type criterion. No cancellation or further instruction is given, so the task is not executed or revoked.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6dfa76d0e5cab002da6232c82a0538fcaeb626bb53c23f471e99aa93df19e476", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_1/audio_natural/filler/original/c6315b836a9132f4/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "e0f16af90b0936a825878c203c5b25229beffc1c9e26c8c4a2353338933418c7", "seconds": 6.2, "spoken_utterances": [ "Buy me a hmmm... food scale as David recommended uh... in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker hesitated with 'uh' but did not change the core request. The final intent matches the original instruction, and the hesitation is a filler that does not alter the task goal.", "heard_transcript": "Buy me a food scale as David recommended uh in his phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Buy me a blender—sorry, I meant stand mixer as Connor recommended in their phone message.", "long_pause_offsets": [], "source_sha256": "3cdb2643100dd9ebd6fcbe38f8061f926654843a7cdd92187bbc014b9e3d46b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'blender' to 'stand mixer' within the same utterance, preserving the rest of the instruction including the recommendation source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/false_start/false_start_01_product_category/24dd5c9de74aa45d/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "911bcae7432aa850e93a7b6c8d175b47cf253a8264c013dccb0b126191891bcc", "seconds": 7.0, "spoken_utterances": [ "Buy me a ah... blender—um... sorry, I meant stand mixer as Connor recommended ... in their phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blender' but immediately self-corrects to 'stand mixer'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including the condition 'as Connor recommended in their phone message', is preserved. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Buy me a a blender um sorry I meant stand mixer as Connor recommended in their phone message." }, "candidate_id": "wav_7d27948170295caf", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "false_start_02_person_name", "proposed_instruction": "Buy me a stand mixer as Megan—no, I meant Connor recommended in their phone message.", "long_pause_offsets": [], "source_sha256": "25bc34e407db92b5739161a2bb7d498f09396aeafc637da354c974909f92c924", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says Megan but immediately corrects to Connor. The final intended recommendation source is Connor, matching the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy me a stand mixer as Megan—uh... no, I meant Connor recommended uh, let's see... in their phone message." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "false_start_03_message_channel", "proposed_instruction": "Buy me a stand mixer as Connor recommended in their email—scratch that, make it phone message.", "long_pause_offsets": [], "source_sha256": "98412752e331c637b924125d86f4fddb6dc67418d275a8f0acc6bd89234eba08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'email' but immediately corrects it to 'phone message'. The final intended source of the recommendation is 'phone message', matching the original instruction. The action, item, and recommender remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy me a stand mixer as Connor recommended uh, let's see... in their email—uh... scratch that, make it phone message." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Buy me a stand mixer in their phone message. Oh, I forgot to mention, it should be as Connor recommended.", "long_pause_offsets": [ 45 ], "source_sha256": "2c5843f98a2e66b0c2d30c596af1afa656abcda5cda140f48f69217d637ab523", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'as Connor recommended', which clarifies the source of the recommendation. The initial phrase 'in their phone message' is replaced by the more specific and correct reference to Connor's recommendation." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/addition/addition_01/d90bfeb2af8506c6/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "f9654dc7dcdcd782248e88a5b94b68a661d9fa4fd714816c8b4dc0278a784d69", "seconds": 11.76, "spoken_utterances": [ "Buy me a stand mixer um... in their phone message. ", "Oh, I forgot to mention, it should be as hmmm... Connor recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Buy me a stand mixer, um, in their phone message.' This is followed by a longer pause and an addition: 'Oh, I forgot to mention, it should be as, hmm, Connor recommended.' The addition clarifies the condition under which the mixer should be bought, completing the original request. The final intent combines the initial request with the added constraint, matching the original instruction. The hesitation and filler words do not change the final goal.", "heard_transcript": "Buy me a stand mixer, um, in their phone message. Oh, I forgot to mention, it should be as, hmm, Connor recommended." }, "candidate_id": "wav_46773ed6a166ff52", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Buy me a stand mixer as Connor recommended. I left out a detail: the recommendation was in their phone message.", "long_pause_offsets": [ 44 ], "source_sha256": "080b1ce2808e5cade33866f848371ba8586d56bab7d52d3b12739a421f2483ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context that the recommendation was in a phone message, which was already implied in the original instruction. The core action and subject remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy me a hmmm... stand mixer as Connor recommended. ", "I left out a detail: the recommendation was in their uh... phone message." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_product_category", "proposed_instruction": "Buy me a blender as Connor recommended in their phone message. Change the product to a stand mixer, not a blender.", "long_pause_offsets": [ 63 ], "source_sha256": "19a18b1dcb0fa44013cd6f46a3cc853ef22715477f5749c7f39e97ae00dda59a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the product from 'blender' to 'stand mixer', preserving the recommendation source and the action of buying." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/revision/revision_01_product_category/681da2ad5e89669c/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "15525e87a755203b01b90fc6fbc3096afd2e023513fb8ed329ff51aa74aa8865", "seconds": 11.64, "spoken_utterances": [ "Buy me a uh... blender as Connor recommended um... in their phone message. ", "Change the product to a stand mixer, not a blender." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blender' but immediately corrects it to 'stand mixer' in the revision clause. The final audible intent is to buy a stand mixer, which matches the original instruction. The revision is clear and applies to the product type, preserving all other conditions.", "heard_transcript": "Buy me a blender as Connor recommended in their phone message. Change the product to a stand mixer, not a blender." }, "candidate_id": "wav_15aecd697058b64f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant 'uh' and 'um' filler, then delivers a coherent initial request ('Buy me a blender as Connor recommended in their phone message'). The revision ('Change the product to a stand mixer not a blender') is a clear, self-contained correction that flows logically from the initial statement. The prosody is consistent, with natural stress and pacing, and the correction is delivered without awkwardness or repetition. The overall delivery sounds like a single, spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_03_message_channel", "proposed_instruction": "Buy me a stand mixer as Connor recommended in their email. Make that in their phone message, not email.", "long_pause_offsets": [ 59 ], "source_sha256": "063c7bc848f2702eba40de02d6f66ea881a0ddace4b2bd757aa84b46c6ca340f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the communication channel from 'email' to 'phone message', preserving the action, item, and recommender." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/revision/revision_03_message_channel/e2d869dde062042a/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "ce9cb8b6b8b4824a1361d50e10fbc2c48de3e49b35056742b9b106229833b4fd", "seconds": 11.0, "spoken_utterances": [ "Buy me a let me think... stand mixer as Connor recommended uh, let me think... in their email. ", "Make that in their phone message, not email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'in their email' but immediately corrects it to 'in their phone message, not email' after a pause. This is a revision where the final intent clearly overrides the initial incorrect value. The rest of the instruction matches the original, so the final audible task is to buy the stand mixer as recommended in the phone message.", "heard_transcript": "Buy me a let me think stand mixer as Connor recommended uh let me think in their email. Make that in their phone message, not email." }, "candidate_id": "wav_837083600580d67e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker uses 'let me think' as a filler, which is plausible, but repeats it unnaturally within the same sentence ('...stand mixers Connor recommended uh let me think...'). This repetition sounds forced and disrupts the natural flow. The correction ('Make that in their phone message not email') is grammatically awkward and delivered with a flat, slightly robotic intonation. The overall delivery feels less like a genuine thought process and more like a rehearsed or synthesized utterance, reducing its realism compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_02_person_name", "proposed_instruction": "Buy me a stand mixer as Megan recommended in their phone message. Wait, the recommendation was from Connor, not Megan.", "long_pause_offsets": [ 66 ], "source_sha256": "99d6c2ec8dd39ae24c5541cdf0c462ae45f6a5b059e5a72a7c8c267e5be0a85a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recommender from Megan to Connor, preserving the action, item, and context of the phone message." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy me a stand mixer as Megan recommended um... in their phone message. ", "Wait, the recommendation was from uh, let me think... Connor, not Megan." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "179101a78dbe2edad7818794a35d892f53fac81fa4a22c1958685382e7f51393", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/cancellation/cancel_019/6e1cd12512b8a2ba/speech.wav", "wer": 0.25, "wav_sha256": "59ccb459db5e0ea10bc1ac50a0c018c7e428abd66a8989f0e7406fb887085827", "seconds": 9.84, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message.", "Hold on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then clearly cancels it with 'Hold on, never mind, don't do that.' This constitutes a final cancellation of the original request. The initial part of the audio is a false start, but the final interjection is a direct and unambiguous cancellation. The control intent matches the expected 'cancel' for this type.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. kind of recommend in their phone message. Hold on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "14eba60e54bfd8673e21cc73d93d50864845fa03ac263847ce8f97c75b95a4ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/ambiguous_halt/halt_021/e90b024643f55008/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "76c657c4558ee8030d29d3d5d22bec0ce9a360906488c0f2a165426f583bd45f", "seconds": 9.32, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial request is not completed, but the speaker does not revoke or modify it; they only signal a need to pause.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. Connor recommended in their phone message. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f2d553e68c719a42b74892b4a99a01e2fca5a4928c9938d123aaed834dfd5a68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_2/audio_natural/filler/original/0dcee260f221c6bc/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "7529af948439861020bef263f874877fd7aca580c6ad486bd437973c3d6fcb2a", "seconds": 5.32, "spoken_utterances": [ "Buy me a uh... stand mixer as uh, let me think... Connor recommended in their phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh' and 'let me think' but does not change the core request. The final audible intent matches the original instruction. The hesitation and backchanneling are typical of a filler type, and the task is completed as stated.", "heard_transcript": "Buy me a uh stand mixer as uh let me think. Connor recommended in their phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_person_name", "proposed_instruction": "Buy me a backpack as Liam—sorry, I mean Mason recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "f22ce438ac44b0f76cd4dc7e39fcc8a9cc143c4b8668400ddf117a63172d2ee5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misnames the recommender as Liam but immediately corrects to Mason. The action, item, and context (phone message) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/false_start/false_start_02_person_name/b782bdfc13f2e359/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "f68879892c99e8d080d8f832fc7ca80dc824206d7f9f390b6769c0663e0f5e15", "seconds": 8.76, "spoken_utterances": [ "Buy me a backpack as Liam—uh... sorry, I mean Mason recommended um... in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Liam' but immediately self-corrects to 'Mason' with 'sorry, I mean Mason'. This is a clear false start where the final intended target is Mason. The rest of the instruction, including the action 'buy me a backpack' and the context 'recommended in his phone message', remains unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Buy me a backpack as uh Liam uh sorry, I mean Mason recommended um in his phone message." }, "candidate_id": "wav_071fd03514c061f9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural false-start pattern: the speaker begins with 'Liam', self-corrects to 'Mason' with a clear, spontaneous 'sorry, I mean', and continues smoothly. The filler 'uh' and 'um' are placed naturally, and the overall prosody reflects genuine hesitation and correction without robotic pauses or unnatural phrasing." } ] }, "is_default": true }, { "variant_id": "false_start_03_message_channel", "proposed_instruction": "Buy me a backpack as Mason recommended in his email—scratch that, make it phone message.", "long_pause_offsets": [], "source_sha256": "c5ded2f201c476d837181e3362e0340b03fff32e6a3e77ed4b18ca12c5035b1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'email' but immediately corrects it to 'phone message'. The final intended channel is phone message, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/false_start/false_start_03_message_channel/ab58ff148a311dc8/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "1fc73ae68bbb71f02663d0b9d355b3632ac995ebbe620522f7dee941814e97e8", "seconds": 9.4, "spoken_utterances": [ "Buy me a backpack as Mason recommended um... in his email—uh... scratch that, make it phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Buy me a backpack as Mason recommended' but then immediately self-corrects with 'scratch that, make a phone message.' This is a false start where the final audible intent is to execute the correction. The original instruction is abandoned in favor of the new one, which is to 'make a phone message.' The correction is immediate and clear, satisfying the false_start type criterion.", "heard_transcript": "Buy me a backpack as Mason recommended. Um, in his email. Uh, scratch that, make a phone message." }, "candidate_id": "wav_048c48c0f4286687", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision rather than a false start. The speaker starts with 'email', then corrects to 'phone message'. The correction is clear and the intonation reflects a change of mind. However, the initial request is slightly more abrupt, and the filler 'uh' after 'email' feels less integrated than in audio_03, making the flow slightly less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_01_product_category", "proposed_instruction": "Buy me a laptop—I meant backpack as Mason recommended in his phone message.", "long_pause_offsets": [], "source_sha256": "c786a93a33eb8742ca09320189bdb412592308da9c8bb0a69bff397941ee546d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'laptop' to 'backpack' within the same utterance, preserving the rest of the instruction including the recommendation source (Mason) and context (phone message)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/false_start/false_start_01_product_category/29c9c2a03ebf88a7/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "700fc50e375f221edfccbc9ad8e1859d2d66f94a0129b5e66ea1a27a705f9ced", "seconds": 6.68, "spoken_utterances": [ "Buy me a laptop—um... I meant backpack as Mason recommended ... in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'laptop' instead of 'backpack' but immediately corrects to 'backpack' in the same sentence. The final audible intent is to buy a backpack as recommended, matching the original instruction. The correction is immediate and self-contained, satisfying the false_start criterion.", "heard_transcript": "Buy me a laptop as Mason recommended in his phone message." }, "candidate_id": "wav_d8b64b22d256c536", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio begins with a false start ('laptop') but fails to correct it. The speaker continues with 'backpack' as if unaware of the initial error, creating a disjointed and unnatural flow. The lack of self-correction and the abrupt shift from 'laptop' to 'backpack' make this the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Buy me a backpack as Mason recommended. One more detail: Mason recommended it in his phone message.", "long_pause_offsets": [ 40 ], "source_sha256": "f85bac2ba4ecd89b13956f4aa6e085cf3c8281ba6b3815b154ffc8dbfb9f0e29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail that the recommendation came from a phone message, which was already implied in the original instruction. The core action and subject remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/addition/addition_02/3dd78e2a0cb440a3/speech.wav", "wer": 0.0, "wav_sha256": "0f5fa8a44d2ea28c5a365da0e1ef11eab3187c8068920a8a8cf1774a3aaa8c31", "seconds": 11.48, "spoken_utterances": [ "Buy me a ... backpack as Mason recommended. ", "One more detail: Mason recommended it in his ... phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to buy a backpack as recommended by Mason, then adds the detail that the recommendation came from a phone message. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and complete, matching the original instruction with the added constraint.", "heard_transcript": "Buy me a backpack as Mason recommended. One more detail, Mason recommended it in his phone message." }, "candidate_id": "wav_f2ca66ecde3b76be", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the initial request with natural prosody and a clear, spontaneous-sounding addition. The phrase 'one more detail' acts as a believable conversational cue before the speaker recalls and integrates the missing information ('Mason recommended it in his phone message') smoothly and cohesively. The pacing and intonation suggest genuine recall rather than a rehearsed or robotic insertion." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Buy me a backpack in his phone message. Oh, I forgot to mention, it should be as Mason recommended.", "long_pause_offsets": [ 40 ], "source_sha256": "524377d111160b9ab48a4744cf9cf6a57780a72aa676c94165d0a929de84f64d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'as Mason recommended', which clarifies the backpack's selection criterion. The initial utterance incorrectly placed the recommendation inside the message context, but the final instruction correctly attributes the recommendation to Mason." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/addition/addition_01/58f84e50a27d0238/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "50cc01e3cc90dd4dcfd33585e98f9e96b36e08c5fb6d612dc6625be69a22b682", "seconds": 9.36, "spoken_utterances": [ "Buy me a ... backpack let's see... in his phone message. ", "Oh, I forgot to mention, it should be as Mason recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Buy me a backpack, let's see, in his phone message,' then adds 'Oh, I forgot to mention, it should be as Mason recommended.' This is a clear addition that refines the original instruction by specifying the condition 'as Mason recommended.' The final intent combines the original request with the added constraint, matching the reference instruction exactly. The addition is made after a pause and is not a correction of a prior error, so it satisfies the 'addition' type criterion.", "heard_transcript": "Buy me a backpack, let's see, in his phone message. Oh, I forgot to mention, it should be as Mason recommended." }, "candidate_id": "wav_4a95677dd09fe673", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker begins naturally, the addition feels less integrated. The phrase 'let's see' introduces a filler that is slightly awkward in context, and the correction 'Oh, I forgot to mention it should be as Mason recommended' is delivered with a noticeable shift in intonation and pacing, sounding more like a post-hoc insertion than a fluid thought process. The overall delivery is less cohesive than audio_01, making it less realistic in terms of spontaneous speech flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_person_name", "proposed_instruction": "Buy me a backpack as Oliver recommended in his phone message. Sorry, I meant Mason recommended it, not Oliver.", "long_pause_offsets": [ 62 ], "source_sha256": "8e7fbdad633f70ca5223bf2ad2b330cac8236294bede6229762980ba7b4ab1fa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recommender from Oliver to Mason, preserving the action, item, and context of the phone message." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/revision/revision_02_person_name/9701d4869064ce13/speech.wav", "wer": 0.2, "wav_sha256": "57c278f38435d0c4546f8dbef8fec3e3faa8d90e552d6654eeff973d81897c05", "seconds": 10.52, "spoken_utterances": [ "Buy me a uh... backpack as ... Oliver recommended in his phone message. ", "Sorry, I meant Mason recommended it, not Oliver." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misattributes the recommendation to 'Oliver' but immediately corrects themselves to 'Mason' after a pause. The revision is clear and final, changing the recommender from Oliver to Mason. All other conditions (buying a backpack, based on a phone message) remain unchanged. The correction satisfies the revision type criterion.", "heard_transcript": "Buy me a uh backpack as cus Oliver recommended in his phone message. Sorry, I meant Mason recommended it, not Oliver." }, "candidate_id": "wav_27f882a3699de2eb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker starts with a plausible false start ('uh uh backpack as cust...'), self-corrects to 'Oliver recommended', then later corrects to 'Mason recommended'. The correction is natural, with a clear shift in intonation and a believable 'sorry' before the revision. The pacing and filler usage ('uh uh') are consistent with spontaneous speech, and the final sentence flows smoothly after the correction." } ] }, "is_default": true }, { "variant_id": "revision_03_message_channel", "proposed_instruction": "Buy me a backpack as Mason recommended in his email. Make that in his phone message, not email.", "long_pause_offsets": [ 53 ], "source_sha256": "3fa9b6c7b1c71036ce8f5779f93bd021f171ffe918128a4346d7f111cd37851a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the communication channel from email to phone message, preserving the action, item, and recommender." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/revision/revision_03_message_channel/e26167629ba6e0e5/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "2d58efe3ff262533e5eebb935608dcec1bcf9ecc70ff72a397a71bddb171074f", "seconds": 14.32, "spoken_utterances": [ "Buy me a um... backpack as Mason recommended let me think... in his email. ", "Make that in his uh... phone message, not email." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'in his email' but immediately corrects it to 'in his phone message, not email'. This is a clear revision where the final intent is to use the corrected information. The original instruction is revised to match the final spoken version, satisfying the revision type criterion.", "heard_transcript": "Buy me a um backpack as Mason recommended. Let me think. In his email. Make that in his uh phone message, not email." }, "candidate_id": "wav_3f33ab82e375e740", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a filler ('um') and a plausible false start ('backpack as Mason recommended'), then hesitates with 'let me think' before revising 'email' to 'phone message'. The revision is clear and the intonation shift is natural. However, the initial phrase 'backpack as Mason recommended' is slightly awkwardly phrased compared to audio_03, and the filler 'let me think' is a bit longer and less integrated than the more spontaneous 'uh uh' in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Buy me a laptop as Mason recommended in his phone message. Wait, I meant a backpack, not a laptop.", "long_pause_offsets": [ 59 ], "source_sha256": "a94c2d3ee09b1a8b5fafd6b4035fda52bdbe4c584ab10216279c46e94ef7b13a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the item from 'laptop' to 'backpack', while retaining the rest of the instruction including the source (Mason's recommendation) and the action (buying for the speaker)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/revision/revision_01_product_category/c400c7c15c7cdc5e/speech.wav", "wer": 0.1, "wav_sha256": "ee324cc8782076dcd8c724737965111886ceac97684930438a224565bc7e0000", "seconds": 10.76, "spoken_utterances": [ "Buy me a laptop as Mason recommended ... in his phone message. ", "Wait, I meant a uh... backpack, not a laptop." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'laptop' but immediately corrects to 'backpack' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction exactly, with only the erroneous initial value being replaced.", "heard_transcript": "Buy me a laptop as Mason recommended in his phone message. Wait, I meant a uh backpack, not a laptop." }, "candidate_id": "wav_b55274136a956842", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear false start ('laptop'), self-corrects to 'backpack', and then adds 'not a laptop'. While the correction is present, the delivery is flatter and less natural in prosody compared to the others. The pause before 'wait' is abrupt, and the correction lacks the subtle intonation shift seen in the other clips. The overall rhythm feels more like a rehearsed correction than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_016", "hide_variant_original": true, "source_sha256": "05673dfb1439f193ea05b07b082a9c214ea86daeeecf9cd2a2ee89e94b9edfd6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/cancellation/cancel_016/e3e13aa533939081/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ba2a6dd924107d05a1b382f3fa2fc1c860f300a742ad435bb70ded82e79473ca", "seconds": 11.52, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message.", "Oh, wait, you know what? Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Cancel that.' This constitutes a cancellation as per the type-specific criterion. The final audible intent is to cancel, not to execute or pause.", "heard_transcript": "Buy me uh let's see backpack as Mason recommended in his phone message. Oh wait, you know what? Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "e53b68734208eb1d2932837d9d3855c50b07f1be7422bbbe040def3dba1d82e9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/ambiguous_halt/halt_023/05e3b35ce04f315b/speech.wav", "wer": 0.1, "wav_sha256": "d0f5f9f0e8b0b2928ceffc6270e92e53d9302c967ad36b1128da528dbf581697", "seconds": 11.0, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or modified, and the speaker does not resume or correct the instruction. The final audible intent is to pause, not to execute or cancel.", "heard_transcript": "Buy me a uh, let's see, backpack as Mason recommended in his phone message. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "7e1be84_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "426369198e46a044807e7720b324b283b89b7195521a4ced54bd1e81a950a125", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/7e1be84_3/audio_natural/filler/original/da54d2d968a1ff69/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "61db4c8971ac65c4ec60ba5172e94f78adfd2a5aadb13f14cc09a7d026dcbb06", "seconds": 6.88, "spoken_utterances": [ "Buy me a uh, let's see... backpack as ... Mason recommended in his phone message." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh let's see' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "Buy me uh let's see backpack as Mason recommended in his phone message." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_source_collection", "proposed_instruction": "What is the title of the most-liked song in my Spotify library—scratch that, make it playlists.", "long_pause_offsets": [], "source_sha256": "3aa0a5044729ac2c9e7c27cd8f9496697c8aa3aca1d985e1a1ff9f1d2b9a41a3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'library' but immediately corrects it to 'playlists' within the same utterance. The final target is 'playlists', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/false_start/false_start_03_source_collection/1e0629ddedd94852/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "fcf4c1cd1d07e03fc27091a601c04d9545af7fc8885a98f8c8eed088b8d1a342", "seconds": 9.84, "spoken_utterances": [ "What is the title of the ... most-liked song in my Spotify library—uh... scratch that, make it playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify library' but immediately self-corrects to 'playlists' after a pause and the interjection 'scratch that'. According to the 'false_start' type criterion, the final corrected value ('playlists') should be applied, and the original incorrect value ('library') should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "What is the title of the most-liked song in my Spotify library? Uh, scratch that, make it playlists." }, "candidate_id": "wav_1195e4d349c4940b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a very natural false start: the speaker begins with 'What is the title of the most liked song in my Spotify library' and then immediately self-corrects with 'uh, scratch that, make it playlists' in a spontaneous, conversational tone. The correction is integrated smoothly into the same utterance, with a realistic hesitation ('uh') and a casual, self-directed correction ('scratch that'). The prosody and pacing are consistent with someone thinking aloud and adjusting their request on the fly." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "What is the title of the most-liked song in my Apple Music—I should say Spotify playlists.", "long_pause_offsets": [], "source_sha256": "e9ff846ccee4f6c584d212c962f85f8410418e17dac55363b8ff1e3fe6249656", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the scope of 'my playlists' and the intent to find the most-liked song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/false_start/false_start_02_app/c2300d5a33628320/speech.wav", "wer": 0.0, "wav_sha256": "009d4bccd3fa2b25b978f50e06f0a0013d078a1f413feae0a925c544222cf07d", "seconds": 10.32, "spoken_utterances": [ "What is the title of the most-liked song in my uh, let's see... Apple Music—uh... I should say Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "What is the title of the most-liked song in my uh, let's see, Apple Music uh, I should say Spotify playlists." }, "candidate_id": "wav_69993a1832b53985", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a more complex revision process with multiple fillers ('uh', 'let's see') and a clear change of mind from Apple Music to Spotify playlists. The speaker starts with a correct formulation, hesitates, then corrects themselves. The flow is natural, but the correction is less immediate than in audio_01, and the multiple fillers slightly disrupt the continuity. The self-correction is believable but feels slightly more deliberated than a spontaneous false start." } ] }, "is_default": false }, { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most-played—or rather, most-liked song in my Spotify playlists.", "long_pause_offsets": [], "source_sha256": "06d3b2addb2da598fae636fcd7e12d0b120c36ab105bb641ffe2bd440620cb76", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'most-played' to 'most-liked' within the same utterance, changing the metric but preserving the app, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/false_start/false_start_01_song_selection/39f40fba3cea5cbf/speech.wav", "wer": 0.0, "wav_sha256": "dd18100d2bdc643549cc3d8e60b76e715ab6ec71e7235f1d382479825fb09c78", "seconds": 10.68, "spoken_utterances": [ "What is the title of the most-played—um... or rather, most-liked song in my uh... Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'most liked' with 'or rather'. According to the 'false_start' type criterion, the final corrected value 'most liked' should be applied, and the original 'most played' is discarded. The rest of the instruction, including the target 'my Spotify playlists', remains unchanged. The final audible intent matches the original instruction after the correction.", "heard_transcript": "What is the title of the most played um, or rather, most liked song in my uh, Spotify playlists." }, "candidate_id": "wav_69737ef0ada2acce", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker initially says 'most played' and then corrects to 'most liked'. However, the correction is less integrated and feels more like a read-aloud or rehearsed change rather than a spontaneous thought process. The 'or rather' transition is slightly awkward, and the overall delivery is flatter and less conversational compared to the others. The prosody lacks the natural hesitation and flow of the other two, making it sound less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "What is the title of the most-liked song in my playlists. I should also mention, this is in my Spotify playlists.", "long_pause_offsets": [ 58 ], "source_sha256": "910f84570691b5b99a887786a613b2dfd69e99a724e88d09283a324188d5dac9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the context of 'playlists' as being on Spotify, which was already implied in the original instruction. The core request for the most-liked song's title remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/addition/addition_02/d4c86e75ccba9d3d/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "f4b711f103c579dba367b166985a147d3adfd32c3d357df7bb3c76a4695c77b5", "seconds": 14.28, "spoken_utterances": [ "What is the title of the uh... most-liked song in my playlists. ", "I should also mention, this is in my ... Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'What is the title of the most liked song in my playlist?' and is followed by an addition clarifying the context as 'this is in my Spotify playlists.' The addition provides a specific target for the initial request, combining the two into a single, coherent instruction. The final intent matches the original instruction with the added specificity, satisfying the addition type criterion.", "heard_transcript": "What is the title of the um most liked song in my playlist? I should also mention, this is in my Spotify playlists." }, "candidate_id": "wav_13c6f849e7126105", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural hesitation ('um') and a believable additive structure, first asking about the most liked song and then adding context about Spotify playlists. The pacing is conversational, with a slight pause before the addition that sounds like spontaneous recall. The intonation and rhythm are consistent with someone formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "What is the title of the most-liked song in my Spotify. And just to be specific, look in my playlists.", "long_pause_offsets": [ 56 ], "source_sha256": "9985c73731f520080b2d3f4ec274463ea27d41c2cee0789b231465b71414f996", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'in my playlists' to the initial request, which was otherwise identical in scope and intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/addition/addition_03/5ba0281377217c2b/speech.wav", "wer": 0.0, "wav_sha256": "323f6a0cd2c8126b587be7015591e546c11bf7bfadae9ac0d0e604187bcca56f", "seconds": 11.68, "spoken_utterances": [ "What is the title of the most-liked song in my ... Spotify. ", "And just to be specific, look in my um... playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the reference. The speaker then adds a clarifying constraint, 'and just to be specific, look in my um playlists,' which narrows the scope to playlists. This is a valid addition that refines the original request without contradicting it. The final intent is to execute the original task with the added constraint.", "heard_transcript": "What is the title of the most-liked song in my Spotify, and just to be specific, look in my um playlists." }, "candidate_id": "wav_8258101cb4f000c2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the request is clear, the delivery is slightly more robotic and less fluid than audio_01. The speaker uses 'um' and 'just to be specific' as a filler, but the transition into the clarification feels a bit more rehearsed. The overall prosody is flatter, and the addition lacks the natural, breathy hesitation of a truly spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song in my Spotify playlists. Oh, I forgot to mention, it should be the most-liked song.", "long_pause_offsets": [ 55 ], "source_sha256": "081a6dc4c6100396164c36fa723748e1697aef3325deebfdf48a65dfd8b4081f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most-liked' to the initial request for a song title, refining the scope without contradicting the original intent." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the song in my ... Spotify playlists. ", "Oh, I forgot to mention, it should be the ... most-liked song." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "What is the title of the most-liked song in my Spotify library. That should be 'playlists', not 'library'.", "long_pause_offsets": [ 64 ], "source_sha256": "9747ad2d2d2a6ff951cd29784fdb5a61cfe022308b207139c724456c463bcb16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'library' to 'playlists', preserving the action and the rest of the query." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/revision/revision_03_source_collection/ffac2f7e7008b3b4/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "1897e39e27a3acb98a02ac560c8375fc22d57b4d4e335f0d0410714c1ec6a734", "seconds": 11.72, "spoken_utterances": [ "What is the title of the let's see... most-liked song in my ... Spotify library. ", "That should be 'playlists', not 'library'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially says 'library' but immediately corrects themselves to 'playlists' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the most-liked song in 'playlists', which matches the original instruction. The correction is clear and unambiguous.", "heard_transcript": "What is the title of the let's see, most liked song in my uh Spotify library. That should be playlists, not library." }, "candidate_id": "wav_6c790149295ec64a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let's see'), then delivers the initial request fluently. The correction ('That should be playlist not library') is a clear, self-contained revision with a believable shift in intonation and pacing, indicating a spontaneous realization. The flow is continuous and the prosody sounds like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "What is the title of the most-liked song in my Apple Music playlists. Change the app to Spotify, not Apple Music.", "long_pause_offsets": [ 70 ], "source_sha256": "29e6a935f4cd3c0fcea3decd521a97cde3ee56575e7c02a61e346edb2a9a1a1a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the music app from Apple Music to Spotify, preserving the rest of the query including the scope of 'my playlists' and the condition of 'most-liked song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/revision/revision_02_app/28211eb15128d798/speech.wav", "wer": 0.0, "wav_sha256": "2dd8a6055ee4b054f80fb9e432978b9c6c538e299f826200c778f390792f76bf", "seconds": 15.04, "spoken_utterances": [ "What is the title of the let me think... most-liked song in my Apple Music playlists. ", "Change the app to um, let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request incorrectly specifies 'Apple Music', but the speaker immediately corrects this by stating 'Change the app to... Spotify, not Apple Music'. This constitutes a revision where the final, corrected target app is Spotify, matching the original instruction. The final intent is clear and executable.", "heard_transcript": "What is the title of the let me think most liked song in my Apple Music playlists. Change the app to um let me think Spotify, not Apple Music." }, "candidate_id": "wav_229917899608067c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker uses fillers ('let me think') and pauses that sound natural, but the revision is less smooth. The second 'let me think' is abrupt and seems to interrupt the flow, making the correction feel slightly disjointed. The prosody during the correction is less fluid, and the overall delivery has a slightly more rehearsed or segmented quality compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the newest song in my Spotify playlists. Wait, the song selection should be 'most-liked', not 'newest'.", "long_pause_offsets": [ 62 ], "source_sha256": "c5195ac00ef4dc263c6c299cc832a3f2046eb6fd573de42587b169eb0622830d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the selection criterion from 'newest' to 'most-liked', while retaining the app (Spotify), the scope (playlists), and the action (querying the title)." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the um... newest song in my let's see... Spotify playlists. ", "Wait, the song selection should be 'most-liked', not 'newest'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "69f4cfa84c85383581c1fa4f936a54ea44fc9d6c660afcbe00af3d5859fec1e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/cancellation/cancel_001/78c00c00a7cc6610/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "0ac7e1511325c768b7df10fef410962fcec41e034ed3753f51837f0ec9ee2501", "seconds": 11.32, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to ask for the title of the most-liked song in their Spotify playlist but then immediately says 'Oh wait, cancel that.' This constitutes a clear and final cancellation of the initial request. The control intent matches the 'cancellation' type criterion, as the recording ends with an explicit revocation of the task.", "heard_transcript": "What is the title of the most like song in my um Spotify playlist? Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_009", "hide_variant_original": true, "source_sha256": "f595741e77037f82d56016e57db06b55a2faf94fc031d6f04885f8f8d4f9b04f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/ambiguous_halt/halt_009/d2d4d54c09e7a250/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "8a571363445944157c4311c40fc11ff5ee2c50cf8081b5c493333424316e72fd", "seconds": 10.6, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists.", "Wait a minute..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait a minute,' which is a clear request to pause or hold. The initial request is not completed, and there is no indication of cancellation or a new instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "What is the title of the most-liked song in my um Spotify playlist? Wait a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "056af54a913006b9a8970143e3a6b5dd8ff54081dd41796878e398896c14e354", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_1/audio_natural/filler/original/ba18b2f70d034cdb/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "ab9c793532518b696410563966593ad03eceba7e75c7de74558c82be8927b65c", "seconds": 7.56, "spoken_utterances": [ "What is the title of the ... most-liked song in my um... Spotify playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most like song' but corrects to 'most-liked song' in the context of the original instruction. The use of 'um' is a filler, and the final intent is clear. The task is to find the title of the most-liked song in the user's Spotify playlist, which matches the original instruction.", "heard_transcript": "What is the title of the most like song in my um Spotify playlist?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most-played—no, I mean least-played song in my Spotify song library.", "long_pause_offsets": [], "source_sha256": "aeb6904eb6b65d52945c6f51aff940c7799681f7851a1a06344fa0977bee9838", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most-played' but immediately corrects to 'least-played'. The final intent matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/false_start/false_start_01_song_selection/c0689bc84c380a9a/speech.wav", "wer": 0.2, "wav_sha256": "dbd6e5a9a6d2d2bcca4ab8e95f3dd7af8ae1640adbaaeaee119ce45c779c244b", "seconds": 8.72, "spoken_utterances": [ "What is the title of the um, let me think... most-played—uh... no, I mean um... least-played song in my Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'least played'. According to the 'false_start' type criterion, the immediate self-correction should be applied, and the final intent is to find the least-played song. The correction is clear and unambiguous, so the task is executed as intended.", "heard_transcript": "What is the title of the um, let me think, most played uh, no, I mean um, least played song in my Spotify song library." }, "candidate_id": "wav_0bb7a02c2a69618b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible false-start flow. The initial request 'most played' is clearly a mistake, immediately self-corrected with 'no, I mean least played' in a way that sounds spontaneous and not rehearsed. The filler 'um' and 'let me think' are placed naturally, and the correction is integrated smoothly into the same utterance without awkward pauses or repetition. The prosody and rhythm are consistent with someone thinking aloud." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "What is the title of the least-played song in my Apple Music—I should say Spotify song library.", "long_pause_offsets": [], "source_sha256": "e7d42667055bdd9df1476cd345c1ad4612507d3179625306dd0e7fa85610415a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the action, scope, and condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/false_start/false_start_02_app/4ea7b7f66f2002e7/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "2fb22d79b1a494ad1897e7eecefeb66fdd0e4d87ac09365b6b8a1304b5e46914", "seconds": 8.96, "spoken_utterances": [ "What is the title of the um... least-played song in my ... Apple Music—uh... I should say Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music app' but immediately self-corrects to 'Spotify song library'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "What is the title of the um least played song in my Apple Music app. I should say Spotify song library." }, "candidate_id": "wav_cee54993068ade1e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable false start with 'Apple Music app' followed by a self-correction to 'Spotify song library'. The correction is natural, but the initial error is less plausible than in audio_02 because 'Apple Music' is a different platform than 'Spotify', making the switch feel more like a factual error than a momentary slip of the tongue. The filler 'um' and the phrase 'I should say' are appropriate, but the overall flow is slightly less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "What is the title of the least-played song in my Spotify playlist—scratch that, make it song library.", "long_pause_offsets": [], "source_sha256": "dbd4528692030fb765db695a38a634f1803fde0f300d763a773a720811e7bfdf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects it to 'song library' using 'scratch that, make it'. The final target is the song library, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/false_start/false_start_03_source_collection/fd526f95c5725559/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "9e296df1ae3863cc82e754e877380cda92f3437ceab4cc49f456678bf45de757", "seconds": 10.12, "spoken_utterances": [ "What is the title of the uh... least-played song in my Spotify playlist—um... scratch that, make it ... song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects to 'song library' after a pause and the phrase 'Scratch that'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "What is the title of the uh least played song in my Spotify playlist um. Scratch that, make it song library." }, "candidate_id": "wav_419deb7643f136e1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start with 'playlist' and corrects to 'song library', but the correction is abrupt and lacks the natural integration seen in the others. The initial 'playlist' is not a strong candidate for a slip—'song library' is a more accurate term for Spotify's interface, so the correction feels less like a spontaneous error and more like a deliberate rephrasing. The filler 'um' is used, but the overall delivery is flatter and less conversational than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "What is the title of the least-played song in my song library. I left out a detail: the song library is on Spotify.", "long_pause_offsets": [ 63 ], "source_sha256": "f3fbddaa2f22ab60c89954367b0b92be573a52152a9a1555e2cabf5e52643033", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify' to the initial request, which was already about the least-played song in a song library. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/addition/addition_02/0ea61f462abbc714/speech.wav", "wer": 0.03125, "wav_sha256": "0f4835d0d9f0e83e9118baec81c6a623a4b5635e24c85f58aadf46ee2c7379fc", "seconds": 14.2, "spoken_utterances": [ "What is the title of the um, let me think... least-played song in my song library. ", "I left out a detail: the song library is on uh, let me think... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was missing the source 'Spotify'. After a pause and self-correction, the speaker added the missing detail. The final intent combines the original request with the added information, matching the reference instruction. The 'addition' type is satisfied.", "heard_transcript": "What is the title of the um, let me think, least played song in my song library. I left out a detail, the song library is on uh, let me think, Spotify." }, "candidate_id": "wav_3ffb1c8dcb032485", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song in my Spotify song library. Oh, I forgot to mention, it should be the least-played song.", "long_pause_offsets": [ 58 ], "source_sha256": "fecc3afdb643dcd93119278843d67449d21240b9dbd22fd887dcc6ea3819e3f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least-played' to the initial broad request about a song in the library, refining the goal to match the original instruction exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the song in my uh... Spotify song library. ", "Oh, I forgot to mention, it should be the ... least-played song." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "What is the title of the least-played song in my Spotify. One more thing: look in my song library.", "long_pause_offsets": [ 58 ], "source_sha256": "906eda0b5d2a2c84751505987e5a8130b11773492c7b9821ccc58d2d4ffa7c90", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing scope 'song library' to the initial request, which was otherwise complete. The final goal combines the initial query with the added constraint." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the least-played song in my um... Spotify. ", "One more thing: look in my uh... song library." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the most-played song in my Spotify song library. Wait, the selection should be the least-played song, not the most-played.", "long_pause_offsets": [ 70 ], "source_sha256": "20c5cbd942c62eda1e04df37e88de88211622314e2bc72a9f96370b9a474e160", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the selection criterion from 'most-played' to 'least-played', while retaining the app (Spotify), the data source (song library), and the requested output (title)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/revision/revision_01_song_selection/5d2650a0f5ecda0d/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "25e973450004808cb401e0aa1168f02ea4e8e895c17ebb281fc741df65672704", "seconds": 14.0, "spoken_utterances": [ "What is the title of the um... most-played song in my uh... Spotify song library. ", "Wait, the selection should be the ... least-played song, not the most-played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially requested the most played song but immediately corrected themselves to the least played song. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to find the least-played song, which matches the original reference instruction.", "heard_transcript": "What is the title of the most played song in my uh Spotify song library. Wait, the selection should be the least played song, not the most played." }, "candidate_id": "wav_6248ae53d2fa593a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('What is the title of the most played song...') with a natural 'uh' filler and a pause. The revision ('Wait, the selection should be the least played song...') is a believable change-of-mind, with a clear shift in intonation and a smooth reference back to the initial request. The prosody is continuous and the pacing feels like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "What is the title of the least-played song in my Spotify playlist. Change the source to 'song library' instead of 'playlist'.", "long_pause_offsets": [ 67 ], "source_sha256": "784398c238e886532266d4336a42ab9bbd1d3084d8f301dcedca947038a5f733", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision changes the source from 'playlist' to 'song library', which matches the original instruction. The action and other conditions remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/revision/revision_03_source_collection/31b69b57e955b135/speech.wav", "wer": 0.0, "wav_sha256": "e4fa24578ffe690d1d7079ed239f0c2e2478864f3530b88e56ad74aa378ca2d3", "seconds": 10.96, "spoken_utterances": [ "What is the title of the ... least-played song in my ... Spotify playlist. ", "Change the source to 'song library' instead of 'playlist'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially asks for the least-played song in their 'Spotify playlist' but immediately corrects the source to 'song library' in the same utterance. This is a revision where the second clause corrects the first. The final intent is to query the song library, which matches the original instruction. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "What is the title of the least played song in my Spotify playlist. Change the source to song library instead of playlist." }, "candidate_id": "wav_c85798bef123c850", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request and then makes a revision ('Change the source to song library instead of playlist'). The correction is delivered as a follow-up statement rather than a self-repair within the same sentence, which is still plausible. The prosody is natural, but the revision feels slightly more like a post-hoc addition than a spontaneous correction mid-sentence, making it less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "What is the title of the least-played song in my Apple Music song library. Let me correct the app: use Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "7474785688c13e90504dd24fdc7077ca4bcd09d2d497f18b9327660a38ca5d07", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the original request to find the least-played song in the user's song library." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/revision/revision_02_app/eb501c7ffcfe7f88/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "f810a0d79aa218dddf6b7e049736be52e447411e88c480fe2ec8e1205e851a81", "seconds": 14.8, "spoken_utterances": [ "What is the title of the ... least-played song in my let me think... Apple Music song library. ", "Let me correct the app: use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause, which matches the revision type criterion. The final audible intent is to find the least-played song in the Spotify library, aligning with the original instruction.", "heard_transcript": "What is the title of the least played song in my let me think, Apple Music song library. Let me correct the app, use Spotify not Apple Music." }, "candidate_id": "wav_14204a41021ebb84", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but then makes a self-repair ('Let me correct the app use Spotify not Apple Music'). The correction is delivered as a separate sentence rather than being integrated into the original request, and the phrasing sounds slightly more robotic or like a scripted correction. The prosody is less continuous, and the overall flow feels less like a natural thought process compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "1ce513ab97f936f12f35e3a0279ffb6ac5d5e892315378d3812c2bce9569ff0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/cancellation/cancel_040/5ce7cfac50905958/speech.wav", "wer": 0.0, "wav_sha256": "0631fd390025cc1000d6935336493037b490bbd6f92d18cdaeed62fa323af5a3", "seconds": 13.12, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library.", "Hang on, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by asking for the least-played song in their Spotify library but then immediately follows with 'Hang on, I don't want to proceed with that task. Please cancel it.' This constitutes a clear and final cancellation of the original request. The final audible intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library. Hang on, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "c52f08068ef38ed8944995943e1279239601075bee34932560f9f3e8e91f8921", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/ambiguous_halt/halt_014/91292a2f8e5b924a/speech.wav", "wer": 0.0, "wav_sha256": "33965ef4554690bc3b1ef38e1ad9d9ca1226e9b98bb1fd0c38a0e52b93284fb5", "seconds": 10.92, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original request and then clearly says 'Can you hold on a second?', which is an audible request to pause. There is no cancellation or instruction to resume. The final interjection is a pause request, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bfbf78ec20e7aa55f47c46eae116b9c83d161aecb3b2d93485b42df0ca60238a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_2/audio_natural/filler/original/ffc786fb027487a9/speech.wav", "wer": 0.0, "wav_sha256": "febf5de92e444ec2d5ae03b50875c016550bb0836aa4f097641e9b9fb91c5d5a", "seconds": 7.52, "spoken_utterances": [ "What is the title of the ah... least-played song in my uh... Spotify song library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'What is the title of the least played song in my uh Spotify song library?' and the 'uh' is a filler word that does not change the task goal. The final intent is to retrieve the title of the least-played song in the Spotify library, which matches the original instruction. The filler word does not alter the meaning or intent of the task.", "heard_transcript": "What is the title of the least played song in my uh Spotify song library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most-listened—sorry, I meant most-played song in my Spotify album library.", "long_pause_offsets": [], "source_sha256": "ade29adbe2e59e211b9414d415d56170fd10f710a5bece1fd997d000f60f7cf7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most-listened' but immediately corrects to 'most-played'. The final intended action is to identify the most-played song, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/false_start/false_start_01_song_selection/424e1b0f33032153/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "f49e19ba8d2958efe86f25b901509ced0df600029d42da0bd31a4a15c3382c5f", "seconds": 8.0, "spoken_utterances": [ "What is the title of the most-listened—hmmm... sorry, I meant most-played song in my uh... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most listened' but immediately self-corrects to 'most played' with 'sorry, I meant'. This is a false start where the final intended action is the corrected version. The rest of the instruction, including 'song in my Spotify album library', remains unchanged and is preserved. The final audible intent matches the original instruction after the correction.", "heard_transcript": "What is the title of the most listened hmm, sorry, I meant most played song in my uh Spotify album library." }, "candidate_id": "wav_59f16e32fbcaec32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a believable self-correction. The initial 'most listened' is immediately followed by a hesitant 'huh' and a smooth, unforced correction to 'most played song'. The prosody is consistent, and the correction integrates seamlessly into the sentence without awkwardness or repetition. The pacing and intonation suggest genuine spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "What is the title of the most-played song in my Apple Music—no, that should be Spotify album library.", "long_pause_offsets": [], "source_sha256": "04025ed1c339aeebc5c3f91443d0b4e181b47d7e9be1794bdb56ac12a46daaac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the scope of 'album library' and the action of retrieving the most-played song's title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/false_start/false_start_02_app/9ba7238f6264901c/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "0a0c3d1945718cd02662de485db6e4d116177bee1a7d436780774d862c6a68e7", "seconds": 11.76, "spoken_utterances": [ "What is the title of the most-played song in my um... Apple Music—uh... no, that should be um... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify' after a pause and a self-correction phrase ('no, that should be'). According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final audible intent matches the original instruction.", "heard_transcript": "What is the title of the most played song in my um Apple Music uh no, that should be um Spotify album library." }, "candidate_id": "wav_2e80cff234ae78ed", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a false start with a more complex correction. The speaker begins with 'Apple Music' and then self-corrects to 'Spotify album library'. The correction is plausible, but the initial error is more substantial and the transition feels slightly less fluid than in audio_02. The 'no that should be' phrase is natural, but the overall flow is a bit more labored, possibly due to the more significant change in the requested platform." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "What is the title of the most-played song in my Spotify playlist collection—scratch that, make it album library.", "long_pause_offsets": [], "source_sha256": "57ef19d3513575ba7b1f5446c65a06b7d525fe54ccbeb55a1d06ede148a3a2ae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially refers to 'playlist collection' but immediately corrects it to 'album library', which matches the original instruction. The final goal is unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/false_start/false_start_03_source_collection/b3d3dbc0212198de/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "5cca8f0ed595849924a3373f7d66650fcd5c29bec8329ae8ffb0897c794f1ca0", "seconds": 10.76, "spoken_utterances": [ "What is the title of the um, let me think... most-played song in my uh... Spotify playlist collection—um... scratch that, make it album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the target as 'playlist collection' but immediately self-corrects with 'scratch that, make it album library.' According to the 'false_start' type criterion, the final corrected target must be applied. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "What is the title of the uh let me think most played song in my uh Spotify playlist collection. Um, scratch that, make it album library." }, "candidate_id": "wav_290301a0ec0d13d1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a false start ('uh') and a filler ('let me think'), which is realistic. However, the correction is not a self-repair within the same utterance but a complete rephrasing ('scratch that, make it album library'). This sounds more like a cancellation and restart than a natural self-correction. The prosody is slightly off, and the transition feels abrupt, reducing the overall realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song in my Spotify album library. Oh, I forgot to mention, it should be the most-played song.", "long_pause_offsets": [ 59 ], "source_sha256": "f97817a63d6dcdc50a0c58e15ebd75be3a5bdd3d8b423ded4e1f047f9aa819d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most-played' to the initial request for a song title, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/addition/addition_01/731dbf3e9b797525/speech.wav", "wer": 0.0, "wav_sha256": "311016679da296e1a2dc3224da0b7b2f598aff13aa4ac02e9db9ce0de6f8a819", "seconds": 11.0, "spoken_utterances": [ "What is the title of the song in my ... Spotify album library. ", "Oh, I forgot to mention, it should be the ... most-played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is to find the title of a song in the user's Spotify album library. After a pause, the user adds the constraint that it should be the 'most played song'. This is a clear addition of a new condition to the original request. The final intent combines both the original scope and the new constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "What is the title of the song in my Spotify album library? Oh, I forgot to mention it should be the most played song." }, "candidate_id": "wav_41fbb0f21dd07ac0", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "What is the title of the most-played song in my album library. One more detail: check only in Spotify.", "long_pause_offsets": [ 63 ], "source_sha256": "0a68255df110b4d72ad4487ae738bd0f2fb6e2332c4f0c48dbd48e064910554b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial request, narrowing the scope to Spotify, which is consistent with the original instruction's implied context and does not contradict it." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the ... most-played song in my album library. ", "One more detail: check only in um... Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "What is the title of the most-played song in my Spotify playlist collection. Change the source to album library instead of playlist collection.", "long_pause_offsets": [ 77 ], "source_sha256": "f30b8228515bd35804c895fc38c82d7e9af44344b95f58bcc1aa331621786249", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly changes the source from 'playlist collection' to 'album library', preserving the action and the app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/revision/revision_03_source_collection/3a43858820015cce/speech.wav", "wer": 0.0, "wav_sha256": "9e38747e0549aa6e3f7facb34321c63565525b76d5514f3536160b8de6477408", "seconds": 14.52, "spoken_utterances": [ "What is the title of the um... most-played song in my uh, let's see... Spotify playlist collection. ", "Change the source to album library instead of playlist collection." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the most-played song in the 'playlist collection' but then immediately corrects the source to 'album library'. This is a revision where the final instruction overrides the initial incorrect source. The final intent matches the original reference instruction, so the task should be executed as requested.", "heard_transcript": "What is the title of the um most played song in my uh let's see Spotify playlist collection. Change the source to album library instead of playlist collection." }, "candidate_id": "wav_a4fea937a4c98857", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, continuous flow with plausible hesitation ('uh', 'let's see') and a clear, self-contained revision. The correction ('change the source to album library instead of playlist collection') is smoothly integrated into the same utterance, with a believable shift in intonation and pacing that suggests spontaneous thought. The prosody is consistent, and the revision feels like a genuine in-the-moment adjustment rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the top-rated song in my Spotify album library. Wait, the selection should be the most-played song, not the top-rated song.", "long_pause_offsets": [ 69 ], "source_sha256": "559d3aae35c39d6918c197c3617bf9e841e79191d1fd519cf511cc219f2dcd6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the selection criterion from 'top-rated' to 'most-played', which matches the original instruction. The app and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/revision/revision_01_song_selection/6b7d1d9f19d28893/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "565c3edfb67c8fe3a88b4054bf4d6ed3dbe049efb4b9aae62d3657f189fbe651", "seconds": 13.08, "spoken_utterances": [ "What is the title of the uh, let me think... top-rated song in my um... Spotify album library. ", "Wait, the selection should be the most-played song, not the top-rated song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'top-rated song' but immediately corrects themselves after a pause, stating the selection should be the 'most played song'. This constitutes a revision where the final intent is the corrected version. The original instruction was 'most-played', and the final audible intent matches it exactly.", "heard_transcript": "What is the title of the uh let me think top-rated song in my um Spotify album library. Wait, the selection should be the most played song, not the top-rated song." }, "candidate_id": "wav_9faa1bb75307dbd5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While this speaker also uses natural fillers ('uh', 'let me think', 'um'), the revision ('Wait, the selection should be the most played song, not the top-rated song') is delivered as a separate, abrupt utterance after a noticeable pause. This creates a slightly disjointed feel, as if the speaker stopped and then restarted the task rather than revising mid-sentence. The prosody remains natural, but the flow is less continuous than audio_01, making it feel less like a single, spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "What is the title of the most-played song in my Apple Music album library. Let me correct the app: use Spotify, not Apple Music.", "long_pause_offsets": [ 75 ], "source_sha256": "a00579eecd9259640bd3e2eaf9743ed769ceace268c2a69705bd5ffb57d4b3c1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the rest of the query about the most-played song in the album library." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the ... most-played song in my uh... Apple Music album library. ", "Let me correct the app: use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "d8d5f81443ef2074798897cd523c6ddb181e0352327f9f587a96cb1d638ebabb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/cancellation/cancel_014/ebf539f495b8fdf0/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6ebcb8f597c65ff21fafa398b964d0ffbb22c92ab2481dfeafb951d9cbaa99fb", "seconds": 11.04, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library.", "Wait, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly and unambiguously cancels it with 'cancel that' as the final interjection. This meets the criterion for a cancellation, as the final audible intent is to revoke the task.", "heard_transcript": "What is the title of the uh most played song in my um Spotify album library. Wait, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "3f4d833e353417879440a939fdf0b0612d0bac8965b6237891fdc3b26d853d80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/ambiguous_halt/halt_028/5589277f56d66e74/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "8ad29a3422191e0827947a58fea827bfdd4cb241357b26deeaa755aab1b922db", "seconds": 10.8, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request is not completed, but the final intent is to pause, which is the expected behavior for this type.", "heard_transcript": "What is the title of the most played song in my um Spotify album library? Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "82e2fac_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9621770f4e1428eb269c4cff2dc4d4cd8be5f0655bda26b0d94de70117119fbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/82e2fac_3/audio_natural/filler/original/a3768b5e77c49e68/speech.wav", "wer": 0.125, "wav_sha256": "b7ebe4e1f03bede71760c477d86b111c63033b43fad67d34e926affc25ac86f1", "seconds": 6.6, "spoken_utterances": [ "What is the title of the uh... most-played song in my um... Spotify album library." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "What is the title of the uh most played song in my um Spotify album library?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks—no, make that recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "92ec7681a07b9f3610cb15e9a54b9564f52c80e7d24079d5a7fa9c49fb7f098f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'top tracks' with 'recommended songs'; all other actions (reset, shuffle, play) and the app (Spotify) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/false_start/false_start_02_recommendation_source/99d3cb94754efce4/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "8d230a134b7cadd42348b59bdf67a179a7c809d68394c7315af7a717ed1ea077", "seconds": 14.32, "spoken_utterances": [ "Reset my Spotify queue with all of its um... top tracks—hmmm... no, make that recommended songs, uh... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately self-corrects to 'recommended songs' after a pause and 'No, make that'. The final instruction matches the original reference. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Reset my Spotify queue with all of its um top tracks. No, make that recommended songs. Uh, shuffle it and play it." }, "candidate_id": "wav_98f796f67cb94fef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a highly natural and plausible false start. The initial phrase 'top tracks' is clearly a spontaneous misremembering, immediately corrected with 'no, make that recommended songs' in a way that sounds like genuine self-correction. The prosody is fluid, with natural pauses and filler words ('um') that fit the context of formulating a request. The delivery is continuous and the correction is integrated smoothly into the utterance, making it the most realistic false start." } ] }, "is_default": true }, { "variant_id": "false_start_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, sort it—sorry, I meant shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "ad13700741022863fe2d967b5f2bdb5938bcdada49fe1d7643e10a128cf3380b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it' but immediately corrects to 'shuffle it' within the same utterance. The final action is shuffle, which matches the original instruction. All other components—resetting the queue, including recommended songs, and playing—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/false_start/false_start_03_queue_ordering_action/048307f21f06f1c4/speech.wav", "wer": 0.0, "wav_sha256": "8da986c6dd3c518ac04fdb764e5915efada4c218ac051e53828514f70d3d3527", "seconds": 12.44, "spoken_utterances": [ "Reset my Spotify queue with all of its recommended songs, ... sort it—um... sorry, I meant ... shuffle it, and play it." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_117aa45afd372bc4", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This is a clear false start with a self-correction from 'sort it' to 'shuffle it'. The speaker's intonation and pacing are natural, and the correction is immediate and contextually appropriate. The 'um' filler is placed naturally, and the overall flow is believable. While the correction is less detailed than in audio_03, it still sounds like a spontaneous, real-time adjustment." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Reset my Apple Music—I should say Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "4a83d9f82f556153f23704001c47f6dd68ed1c751bb4e8b7165fe71a3c9125f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving all other actions and conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/false_start/false_start_01_app/22cfccddef6d5ffc/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "8f25e96aac7649b56e0b60f64dd15a0924e6b1ec64d3d47da6a6c0ee6136297d", "seconds": 11.44, "spoken_utterances": [ "Reset my Apple Music—uh... I should say Spotify queue with all of its um... recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify' with 'I should say'. This is a clear false start where the final, corrected target is Spotify. The rest of the instruction, including 'reset queue', 'all of its recommended songs', 'shuffle it', and 'play it', is preserved and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Reset my Apple Music uh, I should say Spotify queue with all of its um recommended songs, shuffle it and play it." }, "candidate_id": "wav_e385ebea6a9eede5", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a false start, correcting 'Apple Music' to 'Spotify queue'. However, the correction is less integrated and the phrasing 'I should say' sounds slightly more rehearsed or unnatural compared to the more spontaneous 'no, make that' in audio_03. The 'um' filler is used, but the overall delivery feels a bit more hesitant and less fluid, reducing its realism slightly." } ] }, "is_default": false }, { "variant_id": "false_start_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it—wait, I meant play it.", "long_pause_offsets": [], "source_sha256": "0ea902db4d7e1c1802d32faccb9c3eeb67ac0d8e370e994450cd8f4f5fdcee3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately corrects with 'I meant play it'. The final action is to play the queue after resetting and shuffling, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/false_start/false_start_04_playback_action/ba99325daa427791/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "331b055b05c43b60e090d9270e6a36d3816cc8a875a97a2e1ee8179ac71f4b1e", "seconds": 13.6, "spoken_utterances": [ "Reset my Spotify queue with all of its ... recommended songs, shuffle it, and um... pause it—hmmm... wait, I meant play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately self-corrects with 'Wait, I meant play it.' This is a false start where the final audible intent is the corrected version. The correction is immediate and unambiguous, so the final goal is to execute the original instruction with 'play it' replacing 'pause it'. All other conditions remain unchanged.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and um, pause it. Wait, I meant play it." }, "candidate_id": "wav_22545acde3d8cd37", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a false start where the speaker says 'pause it' but then corrects to 'play it'. However, the correction is awkwardly phrased as 'wait, I meant play it', which sounds more like a post-hoc explanation than a natural self-correction. The delivery is less fluid, and the correction feels tacked on rather than integrated into the original thought process. The prosody and pacing also seem less natural, making it the least realistic of the four." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and. One more thing: start playing the shuffled queue.", "long_pause_offsets": [ 75 ], "source_sha256": "e15a4da481b180a9838f5864bd1ffe182cbbec3c368ae99a52639a299f8dc198", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing action to start playing the shuffled queue, completing the original instruction which already included reset, shuffle, and the queue context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/addition/addition_02/162ee9a4f04a7c51/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "bfc76366da4e2f47d043dca69430c41d0c8154241c058974b85665a352b84283", "seconds": 15.64, "spoken_utterances": [ "Reset my Spotify queue with all of its um... recommended songs, uh... shuffle it, and. ", "One more thing: ah... start playing the shuffled queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction to reset the Spotify queue with recommended songs and shuffle it. After a pause and an interjection ('one more thing'), the speaker adds the final action to 'start playing the shuffled queue'. This matches the 'addition' type criterion, where the final goal combines the initial instruction with the added information. The final intent is complete and executable, matching the reference instruction.", "heard_transcript": "Reset my Spotify queue with all of its um recommended songs uh shuffle it and one more thing uh start playing the shuffled queue." }, "candidate_id": "wav_f585f07307780ae3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with a believable flow. The initial request is followed by a pause and a filler ('uh'), then a smooth addition of a new instruction ('one more thing') with a second filler ('uh') before the final command. The prosody is consistent, and the intonation reflects a speaker formulating and expanding on a thought in real time, which is highly realistic." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs,, and play it. Oh, right, make sure to shuffle the queue.", "long_pause_offsets": [ 72 ], "source_sha256": "dc952f11e0847d57ceafd5b174391618f00fe41660644c4e93919fe5e5a03d24", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing shuffle action, which was not in the initial utterance but is consistent with the original instruction. All other requirements (reset queue, include recommended songs, play) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/addition/addition_01/1eb12e939a9aa7ae/speech.wav", "wer": 0.04, "wav_sha256": "22eabbaa2825ee932961db41baa9abb8ac66dab55280940b560816eb04be6416", "seconds": 14.92, "spoken_utterances": [ "Reset my Spotify queue with all of its um, let me think... recommended songs,, and play it. ", "Oh, right, make sure to ... shuffle the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Reset my Spotify queue with all of its recommended songs, and play it.' The speaker then adds 'Oh right, make sure to shuffle the queue.' This is a clear addition of a new constraint to the original task. The final goal combines the original request with the added instruction to shuffle, matching the reference instruction exactly. The hesitation and self-correction do not alter the core task, and the addition is explicitly stated after a pause.", "heard_transcript": "Reset my Spotify queue with all of its um, let me think, recommended songs, and play it. Oh right, make sure to shuffle the queue." }, "candidate_id": "wav_4d6646c802382041", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a natural-sounding request with a filler ('let me think') and a follow-up instruction. However, the second part ('Oh right, make sure to shuffle the queue') feels slightly more abrupt and less integrated than the addition in audio_02. The transition from the filler to the correction is less fluid, and the overall pacing is slightly less organic, making it less realistic in comparison." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Reset my Apple Music queue with all of its recommended songs, shuffle it, and play it. Wait, reset my Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [ 87 ], "source_sha256": "15044bc5b145c1f521f717be79533cb347d407c860be66ab05bae1c4f5ccfd7d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining all other actions and conditions from the initial instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/revision/revision_01_app/fb058a45a61c2e82/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "37dada05ce53e1914b29e0a970e5aa5cc69423d9f5e4abfedc996fa127207c4d", "seconds": 19.76, "spoken_utterances": [ "Reset my Apple Music queue with all of its ... recommended songs, let's see... shuffle it, and play it. ", "Wait, reset my uh... Spotify queue with all of its recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final goal matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Reset my Apple Music queue with all of its recommended songs, let's see, shuffle it and play it. Wait, reset my uh Spotify queue with all of its recommended songs, shuffle it and play it." }, "candidate_id": "wav_12adfbef453658f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker starts with a clear request, then hesitates with 'let's see' before self-correcting to the correct platform ('Spotify') and smoothly continuing. The correction is integrated into the ongoing utterance without awkwardness, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "revision_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, sort it, and play it. Change shuffle it to sort it.", "long_pause_offsets": [ 80 ], "source_sha256": "7183e99253ebfca1f1676b914464c5521c4f9a3df2a2260496680b1e58ebb282", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly changes the action from 'shuffle it' to 'sort it', which is a different operation and contradicts the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/revision/revision_03_queue_ordering_action/c3f7f7de077ce21e/speech.wav", "wer": 0.0, "wav_sha256": "486533de1793231177397e9afad34111ede5e3373106df143c4c8a78cadaa5b7", "seconds": 12.32, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, uh... sort it, and play it. ", "Change shuffle it to sort it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shuffle it' but immediately corrects themselves to 'sort it' after a pause, which is a revision. The final audible intent is to sort the queue, not shuffle it. All other conditions from the original instruction are preserved, and the correction is applied as required by the revision type.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, uh, sort it and play it. Change shuffle it to sort it." }, "candidate_id": "wav_823b8d6da59954b8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a coherent request with a clear revision. The initial command is stated, followed by a self-correction of 'shuffle it' to 'sort it'. The pause and filler 'uh' before the correction sound natural, and the correction is smoothly integrated. The delivery is fluent and realistic, though slightly less spontaneous than audio_03 due to the more abrupt phrasing of the correction." } ] }, "is_default": false }, { "variant_id": "revision_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that recommended songs, not top tracks.", "long_pause_offsets": [ 76 ], "source_sha256": "eedb50e504eb33445dcb56fcbbc1d707b52e37c4e7e870bec0bf242766e5547a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial 'top tracks' to 'recommended songs', preserving the app, action sequence, and other requirements." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/revision/revision_02_recommendation_source/91a1149295478e00/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "560912d952e7351f5baf71225b0c981831e617c27eefe7045e8210e77bdd1190", "seconds": 14.72, "spoken_utterances": [ "Reset my um... Spotify queue with all of its top tracks, hmmm... shuffle it, and play it. ", "Make that recommended songs, not top tracks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects to 'recommended songs' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task is considered passed.", "heard_transcript": "Reset my um Spotify queue with all of its top tracks, shuffle it, and play it. Make that recommended songs, not top tracks." }, "candidate_id": "wav_374b6a6b9e55b0b2", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker changes 'top tracks' to 'recommended songs'. The delivery is clear, but the revision is less integrated into the flow. The pause and filler 'hmm' before the revision feel slightly more hesitant and less connected to the initial request, making the correction seem more like a separate thought than a natural self-correction." } ] }, "is_default": false }, { "variant_id": "revision_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it. That should be play it, not pause it.", "long_pause_offsets": [ 84 ], "source_sha256": "6db531446d38ae760c26fe468e0fe474377437daf609abad3e9f2f53093689c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup corrects the final action from 'pause it' to 'play it', reversing the intended playback control." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/revision/revision_04_playback_action/59501f951606dd1a/speech.wav", "wer": 0.24, "wav_sha256": "1a342bd53496bf3f70c88c5bcf6049a8acb97d04ac16f6e870117b700c210622", "seconds": 13.72, "spoken_utterances": [ "Reset my um... Spotify queue with all of its hmmm... recommended songs, shuffle it, and pause it. ", "That should be play it, not pause it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately corrects themselves with 'That should be play it, not pause it.' This constitutes a revision where the final intent is to play the queue, not pause it. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Reset my um Spotify queue with all of its um recommended songs, shuffle it and pause it. That should be play it, not pause it." }, "candidate_id": "wav_a6debefd2cf47552", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker makes a revision from 'pause it' to 'play it', but the correction is delivered as a separate sentence rather than being integrated into the original request. The phrase 'That should be play it, not pause it' sounds more like a post-hoc correction than a spontaneous self-repair, reducing the realism of the overall utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "f00d9bc204c0d3d1afb6d16f1345c79665e56f5c797ef59ba4eabc3b221ea0f4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/cancellation/cancel_018/41390f2570a7981f/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "65406c959033a8704e855649abbd8404352fd93e94b32a48d59f55089ce9cd3a", "seconds": 15.12, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it.", "Hold on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a complete task request, then immediately follows with 'Hold on, I'd rather not do that after all. Please cancel it.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or modification of the initial request.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it. Hold on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "77813149d291b1b444be9abac9697116c3a0068e2fb04d0c5feb1a3230a841c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/ambiguous_halt/halt_034/92afc6a40039e3af/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "22aba060f930064cf9a39912e9a20d142be9a783ed9f12ea7c4f070772e066bd", "seconds": 12.8, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, I was just thinking,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The initial task is not canceled, but the speaker indicates they need a moment to reconsider, so the agent should pause.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1ed9d24df49faeadc97e4869828509e3cfbaa894259a475c0e3f6f44c125c723", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_1/audio_natural/filler/original/d7cad1f892b7d06f/speech.wav", "wer": 0.0625, "wav_sha256": "355bffaaa84a7630a770d37e27a7a2603b6ba2543ef1c5ea8669631852d21b3e", "seconds": 8.56, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, hmmm... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Reset my Spotify queue with all of its recommended songs, shuffle it, and play it.' The 'um' is a filler sound that does not alter the task goal. The final intent is to execute the full original instruction, which matches the reference. No correction, cancellation, or pause is requested.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, sort it—sorry, I meant shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "ad13700741022863fe2d967b5f2bdb5938bcdada49fe1d7643e10a128cf3380b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it' but immediately corrects to 'shuffle it' within the same utterance. The final action is shuffle, which matches the original instruction. All other components—resetting the queue, including recommended songs, and playing—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/false_start/false_start_03_queue_ordering_action/d931a092b88b357b/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "315935329b4f5a7492ca983d2b4286f3c571964536bf138cb036678157283032", "seconds": 13.32, "spoken_utterances": [ "Reset my Spotify queue with all of its hmmm... recommended songs, sort it—uh... sorry, I meant um... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it out' but immediately self-corrects to 'shuffle it'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('sort it out') is discarded. The final instruction is to shuffle and play the queue, which is a valid and complete task. The correction is immediate and unambiguous, so the task is executed as intended.", "heard_transcript": "Reset my Spotify queue with all of its hmm recommended songs, sort it out. Sorry, I meant um shuffle it and play it." }, "candidate_id": "wav_6f5cba31c3798d70", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural false start: the speaker begins with a clear, slightly hesitant request, then self-corrects from 'sorted up' to 'shuffle it' with a believable 'sorry, I meant um' transition. The prosody includes a realistic pause and filler ('hmm'), and the correction flows smoothly without awkwardness. The rhythm and intonation suggest genuine spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its recently played—no, make that recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "d3f0ea699d497706d70e86720f2d5a8cd6c445b328f2c4db49019e3696fc8af5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'recently played' with 'recommended songs'; all other actions (reset, shuffle, play) and the app (Spotify) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/false_start/false_start_02_recommendation_source/a5108da890fa5214/speech.wav", "wer": 0.0, "wav_sha256": "a14d42862da71f58e072d4ae240b91f1bd754d9f3d48416a54c59d0b4349e093", "seconds": 14.52, "spoken_utterances": [ "Reset my Spotify queue with all of its recently played—uh... no, make that recommended songs, um, let's see... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'recently played' but immediately self-corrects to 'recommended songs'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference instruction in action, target, and sequence.", "heard_transcript": "Reset my Spotify queue with all of its recently played, uh, no. Make that recommended songs. Um, let's see. Shuffle it, and play it." }, "candidate_id": "wav_dd3095e755cc21b8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible self-correction from 'recently played' to 'recommended songs' with a natural 'uh, no' and 'make that' revision. The speaker also includes a filler ('um') and a moment of hesitation ('let's see'), which enhances realism. The correction is integrated into the flow, though the 'no' is slightly abrupt compared to audio_01's smoother transition." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Reset my Apple Music—I should say Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "4a83d9f82f556153f23704001c47f6dd68ed1c751bb4e8b7165fe71a3c9125f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving all other actions and conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/false_start/false_start_01_app/eced1791faefc7c8/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "880419e3cceec5795fe0a18d5b8413fb4412b73f5efe18e5a9ebbeaed95a0117", "seconds": 10.48, "spoken_utterances": [ "Reset my ah... Apple Music—uh... I should say Spotify queue with all of its recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction matches the original, and the final intent is clear and executable.", "heard_transcript": "Reset my uh Apple Music uh, I should say Spotify queue with all of its recommended songs, shuffle it and play it." }, "candidate_id": "wav_14852820bcef0187", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a false start where the speaker initially says 'Apple Music' before correcting to 'Spotify queue'. However, the correction is less natural; the phrase 'I should say' sounds slightly rehearsed or unnatural in spontaneous speech. The overall delivery is fluent but lacks the hesitations and fillers that make the other clips more realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it—scratch that, make it play it.", "long_pause_offsets": [], "source_sha256": "6d7642d14461e63de82162e76c8c613305c67e659a9599f0908af04a6b9b78e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately corrects with 'make it play it'. The final action is to play the queue after resetting and shuffling, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/false_start/false_start_04_playback_action/205611bf1524508a/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "3a5004872c80fbb8aa4e17e8e1648b9668207d6e92563c904ea274344aff434e", "seconds": 15.56, "spoken_utterances": [ "Reset my Spotify queue with all of its recommended songs, um, let me think... shuffle it, and pause it—uh... scratch that, make it play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shuffle it, and pause it' but immediately corrects with 'scratch that, make it play it'. This constitutes a false start where the final audible intent is to play the queue, not pause it. The correction is immediate and unambiguous, so the final control intent is to execute the original task with the corrected action.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, um, let me think, shuffle it, and pause it. Uh, scratch that, make it play it." }, "candidate_id": "wav_480fb3031d481de1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip includes multiple interruptions and a cancellation ('scratch that, make it play it'), but the transitions feel less organic. The 'uh' after 'pause it' and the 'scratch that' are somewhat abrupt, and the final command is delivered with a flat, robotic intonation. The speaker seems to be reading or rehearsing rather than thinking aloud, reducing the realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs,, and play it. Oh, I forgot to mention, shuffle the queue.", "long_pause_offsets": [ 72 ], "source_sha256": "95a58f6ec9fbd63fb42b953e48c18c51e16627d4aa7867b333980d66e1055b55", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing shuffle action, which was not in the initial utterance but is present in the original instruction. The final goal combines all three actions: reset, shuffle, and play." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/addition/addition_01/b12c3c5764fd583d/speech.wav", "wer": 0.0, "wav_sha256": "438190b70b24c95e32fd91d8dd8b478b890b62e41f2ca1c332d7cbb511e0bb7e", "seconds": 13.32, "spoken_utterances": [ "Reset my Spotify queue with all of its ... recommended songs,, and play it. ", "Oh, I forgot to mention, um... shuffle the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to reset the queue with recommended songs and play it. After a pause and an interjection, the speaker added 'shuffle the queue' as an additional instruction. Since the type is 'addition', the final goal is the combination of the initial and added instructions. The final audible intent matches the original reference instruction exactly, including the shuffle action.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs and play it. Oh, I forgot to mention, um, shuffle the queue." }, "candidate_id": "wav_26e6a374a81dcf39", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous initial request with natural prosody. The addition of 'Oh, I forgot to mention' followed by a brief pause and the insertion of 'shuffle the queue' feels like a spontaneous afterthought. The filler 'um' is placed naturally, and the overall flow is coherent and conversational, with no signs of artificiality or forced correction." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and. Oh, and one more thing: play the shuffled queue.", "long_pause_offsets": [ 75 ], "source_sha256": "4d35367de76e06f5c871208e9dbeb18d8d8f93d9320ec01ea1925938bd0f7166", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'play' action after the initial instruction to reset and shuffle, completing the full sequence without contradicting any prior steps." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/addition/addition_02/782beb0dc0ec3253/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "ffc21756362ff61d93302c19727270c9c25359f523f62ae4ffa5f009a9cac5b1", "seconds": 14.76, "spoken_utterances": [ "Reset my Spotify queue with all of its um... recommended songs, shuffle it, and. ", "Oh, and one more thing: ... play the shuffled queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Reset my Spotify queue with all of its um recommended songs, shuffle it, and oh, and one more thing, play the shuffled queue.' The 'um' is a filler, and the final interjection 'play the shuffled queue' is an addition to the initial request. The final intent combines the original instruction with the added action, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Reset my Spotify queue with all of its um recommended songs, shuffle it, and oh, and one more thing, play the shuffled queue." }, "candidate_id": "wav_db617dbbe9a1e099", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a similar addition ('Oh, and one more thing'), the initial request is less fluent, marked by a noticeable hesitation ('um') and a slightly awkward phrasing ('reset my Spotify queue with all of its um recommended songs'). The addition is also less smoothly integrated, with a longer pause and a slightly unnatural transition. The overall delivery feels more hesitant and less conversational than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it. That should be play it, not pause it.", "long_pause_offsets": [ 84 ], "source_sha256": "6db531446d38ae760c26fe468e0fe474377437daf609abad3e9f2f53093689c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup corrects the final action from 'pause it' to 'play it', reversing the intended playback control." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/revision/revision_04_playback_action/72b54c4188634694/speech.wav", "wer": 0.08, "wav_sha256": "490c594c39175a7e967c6e7ab579bc18b0a14c50e6b81b78fc8942f32b3d3b35", "seconds": 16.32, "spoken_utterances": [ "Reset my ... Spotify queue with all of its uh... recommended songs, hmmm... shuffle it, and pause it. ", "That should be play it, not pause it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately self-corrects with 'That should be play it, not pause it.' This constitutes a revision where the final audible intent is to 'play it'. The correction is clear and directly modifies the earlier instruction. The rest of the task (reset queue, shuffle) remains unchanged and is executed as intended.", "heard_transcript": "Reset my Spotify queue with all of its uh recommended songs, hmm, shuffle it and pause it. That should be play it, not pause it." }, "candidate_id": "wav_bdf451b94be84b63", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, multi-stage revision. They begin with a clear initial request, pause thoughtfully ('hmm'), then self-correct the action from 'pause it' to 'play it' with a believable, slightly hesitant tone. The correction is integrated smoothly into the same utterance, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery." } ] }, "is_default": false }, { "variant_id": "revision_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, clear it, and play it. I misspoke about the queue action: shuffle it, not clear it.", "long_pause_offsets": [ 81 ], "source_sha256": "827186ebfb0863e415bffcfafa254cbbd3757b2f3c3c5080b4a9116a16953755", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'clear it' to 'shuffle it', preserving the initial reset and play actions, as well as the scope of recommended songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/revision/revision_03_queue_ordering_action/3d71c794258d9ac3/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "fb0a00bf3fe17867f1935c680fbd5dc3aa6c2b99050155e3721d97442757da97", "seconds": 18.28, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, let's see... clear it, and play it. ", "I misspoke about the queue action: hmmm... shuffle it, not clear it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'clear it' but immediately corrects themselves to 'shuffle it, not clear it' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, let's see, clear it and play it. I misspoke about the queue action. Hmm. Shuffle it, not clear it." }, "candidate_id": "wav_bdab6d337cad18cd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a clear revision after a brief pause and filler ('let's see'), correcting 'clear it' to 'shuffle it'. The self-correction is natural and the pacing reflects a moment of reconsideration. However, the initial request is slightly less fluid than audio_03, with a minor stumble on 'recommended songs' and a slightly more abrupt correction." } ] }, "is_default": true }, { "variant_id": "revision_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that 'recommended songs', not 'top tracks'.", "long_pause_offsets": [ 76 ], "source_sha256": "e92c3389105240d2b844885db8692b4e585c50fa093b8007ff6276c533bcacc0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'top tracks' to 'recommended songs', preserving the app, action sequence, and other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/revision/revision_02_recommendation_source/87db03ef51f0ba62/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "9a9e719d39cea2e6de9b175c7da8340741d75039ce3fdfcd48ca5fe43ea57fe9", "seconds": 15.96, "spoken_utterances": [ "Reset my um... Spotify queue with all of its ... top tracks, ah... shuffle it, and play it. ", "Make that 'recommended songs', not 'top tracks'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects to 'recommended songs' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Reset my um Spotify queue with all of its top tracks uh shuffle it and play it. Make that recommended songs, not top tracks." }, "candidate_id": "wav_d8c1b5083ba9d6b9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with a clear request but makes a revision from 'top tracks' to 'recommended songs'. The correction is integrated into the same sentence, but the initial phrasing is slightly less fluent ('reset my um Spotify queue with all of its top tracks uh...'). The revision is natural, but the overall delivery feels slightly more hesitant and less fluid than the top two." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Reset my Apple Music queue with all of its recommended songs, shuffle it, and play it. Wait, reset my queue in Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "1f9f6242b8d5968ab396b6bddc5442a05d769f4b88142181cbfdbb24bf631438", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the queue reset, shuffle, and play actions, as well as the 'all of its recommended songs' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/revision/revision_01_app/d16842d27dbb45f7/speech.wav", "wer": 0.03125, "wav_sha256": "965f2292e24f7b67a9498e9a24149f4880689074daefee56c849efb96c64fc93", "seconds": 19.28, "spoken_utterances": [ "Reset my let me think... Apple Music queue with all of its recommended songs, uh, let me think... shuffle it, and play it. ", "Wait, reset my queue in ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the app as 'Apple Music' but immediately corrects it to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the app, action, and conditions.", "heard_transcript": "Reset my, let me think, Apple Music queue with all of its recommended songs, uh, let me think, shuffle it and play it. Wait, reset my queue in Spotify, not Apple Music." }, "candidate_id": "wav_dff5ad53f61a80b9", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker exhibits multiple, somewhat disjointed revisions. They start with 'Apple Music', then correct to 'Spotify', and also hesitate with 'let me think' twice. The corrections feel more like a series of disjointed thoughts rather than a single, coherent revision. The prosody is less continuous, and the multiple pauses and restarts reduce the overall realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "da49f61f39d09212faef89b5b2e44da1568a6c6ce6fef55b6ed3514fd7a429e4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/cancellation/cancel_032/9cede3f92c46e258/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "e3079af065eef4d73e4c129c71d7a90d8cc3627dfaf66b6a56685dc08786d743", "seconds": 15.2, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it.", "Hold on, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to reset and play the Spotify queue, then immediately follows with 'Hold on, I don't need this anymore. Please cancel the request.' This constitutes a clear and final cancellation of the initial task. The cancellation is explicit and not merely a pause or hesitation.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it and play it. Hold on, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "81bb7371093b47b1f60433777a28a277b21ec8516063400d1202cfe2a175bf16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/ambiguous_halt/halt_010/8685b07a0c13a31f/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "fcdd9dc8af7134b75a1b4cb8e68b20722367fdb8804f05c1f8760f9ac429b235", "seconds": 12.56, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then audibly requests to pause with 'Hold on a second.' This constitutes a final, clear request to pause, which matches the 'ambiguous_halt' type criterion. The initial hesitation ('let me think') is a filler and does not alter the final intent.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it, and play it. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2d4faff18608fad85c5aa7a9f878b220ae0f99852e7440f2a24f5d61d830403d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_2/audio_natural/filler/original/d51047e6cee1d7f1/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "9417664420b3feab7471568a098edbc666df3cef227f78e41e0b9858987fcb32", "seconds": 9.28, "spoken_utterances": [ "Reset my let me think... Spotify queue with all of its recommended songs, ... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial hesitation 'let me think' is a filler that does not alter the final task goal. The speaker proceeds to complete the original instruction without correction or change. The final intent matches the reference, and the filler is ignored per the type-specific criterion.", "heard_transcript": "Reset my let me think Spotify queue with all of its recommended songs, shuffle it, and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Reset my Apple Music—I should say Spotify queue with all of its recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "4a83d9f82f556153f23704001c47f6dd68ed1c751bb4e8b7165fe71a3c9125f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving all other actions and conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/false_start/false_start_01_app/866c9d2ccceae62e/speech.wav", "wer": 0.0, "wav_sha256": "410951e78b764d5d9e32048e8098749765e6241de9bfdb05c15d428b7785718d", "seconds": 11.4, "spoken_utterances": [ "Reset my Apple Music—um... I should say Spotify queue with all of its uh... recommended songs, shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction. The rest of the task constraints are preserved and correctly executed.", "heard_transcript": "Reset my Apple Music, um, I should say Spotify queue with all of its uh recommended songs, shuffle it and play it." }, "candidate_id": "wav_d02cc4604030f25f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction ('I should say Spotify queue') that flows seamlessly into the rest of the sentence. The prosody is consistent, the pace is conversational, and the correction feels like a spontaneous realization rather than a forced edit. The phrasing is slightly more fluent than others, with minimal hesitation." } ] }, "is_default": true }, { "variant_id": "false_start_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks—or rather, recommended songs, shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "f044d88c0725a1d8ba93ca1a042ce704676b55dcac638468617e84d1b0b6733f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects to 'recommended songs' within the same utterance. The final intended content is 'recommended songs', which matches the original instruction. The other actions—reset, shuffle, and play—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/false_start/false_start_02_recommendation_source/94273d1f93b23d24/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "c7d21c73b981cddf2e63bda944705f5d4ed85459f288ada7c98a8eac6a748d12", "seconds": 9.52, "spoken_utterances": [ "Reset my Spotify queue with all of its top tracks—uh... or rather, recommended songs, um... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately self-corrects to 'recommended songs'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction, and the correction is clear and immediate.", "heard_transcript": "Reset my Spotify queue with all of its top tracks, uh, or rather, recommended songs, um, shuffle it, and play it." }, "candidate_id": "wav_b7738aad16435590", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable revision ('top tracks' to 'recommended songs') with a natural 'or rather' transition. The speaker uses 'um' and 'uh' appropriately as fillers, and the correction is integrated smoothly. The overall delivery is fluent and realistic, though slightly less fluid than audio_02 due to the revision." } ] }, "is_default": false }, { "variant_id": "false_start_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, sort it—sorry, I meant shuffle it, and play it.", "long_pause_offsets": [], "source_sha256": "ad13700741022863fe2d967b5f2bdb5938bcdada49fe1d7643e10a128cf3380b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sort it' but immediately corrects to 'shuffle it' within the same utterance. The final action is shuffle, which matches the original instruction. All other components—resetting the queue, including recommended songs, and playing—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/false_start/false_start_03_queue_ordering_action/04876768c414d4f5/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "716621412152d3c6912fb9a461c5399f3bdc17b29ca8764f8ecb25ce9016d249", "seconds": 12.76, "spoken_utterances": [ "Reset my uh... Spotify queue with all of its let's see... recommended songs, sort it—ah... sorry, I meant shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sorted' but immediately self-corrects to 'shuffle it'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction. The correction is immediate and unambiguous, so the task is to execute the corrected command.", "heard_transcript": "Reset my uh Spotify queue with all of its uh let's see recommended songs sorted uh sorry, I meant shuffle it and play it." }, "candidate_id": "wav_8211a4be691f9f7c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start with 'let's see' and then corrects 'sort it' to 'shuffle it'. The correction is clear, but the initial hesitation and the phrase 'let's see' sound slightly more forced or rehearsed. The prosody is somewhat flat, and the correction feels a bit abrupt compared to the more organic flow of audio_02 and audio_04." } ] }, "is_default": false }, { "variant_id": "false_start_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it—no, make that play it.", "long_pause_offsets": [], "source_sha256": "9afceac555d94d02e25217ca1af068385b229c31b031cbd1f2b94ed6f24f3980", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately corrects to 'play it' within the same utterance. The final action is to play the queue after resetting and shuffling, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/false_start/false_start_04_playback_action/5122aea63cfc3885/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "08387d227aa4618ecdd2558413319693cce074375650d2767da81445c16a0647", "seconds": 9.08, "spoken_utterances": [ "Reset my Spotify queue with all of its recommended songs, uh... shuffle it, and pause it—um... no, make that play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pause it' but immediately corrects with 'No, make that play it.' This constitutes a false start where the final audible instruction is the corrected version. The final intent matches the original instruction, so the task should be executed as corrected.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, uh, shuffle it and pause it. No, make that play it." }, "candidate_id": "wav_c529f6e91c1c25d6", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a self-correction from 'pause it' to 'play it', but the delivery is less natural. The speaker says 'pause it um' with an awkward pause and filler, and the correction 'no, make that play it' sounds more like a mechanical edit than a spontaneous thought. The prosody is inconsistent, and the overall flow is disrupted by the unnatural phrasing and timing of the correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and. One more detail: start playing the queue.", "long_pause_offsets": [ 75 ], "source_sha256": "d12d23c1eda0a5e1285be9063d4c65f7614b1ada93d861fda5857aa373cfdaa6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing action of starting playback, which was implied but not stated in the original instruction. The original instruction ends with 'and', suggesting an incomplete action list. The follow-up completes the goal by specifying 'start playing the queue'. All other requirements (resetting with recommended songs, shuffling) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/addition/addition_02/659085d35cba2283/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "ae0c8b2d293099dc51e8e111bb66de32321c35592dc1f6ef199dae8baf6a233a", "seconds": 12.72, "spoken_utterances": [ "Reset my uh... Spotify queue with all of its recommended songs, ... shuffle it, and. ", "One more detail: ... start playing the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. After a pause and the interjection 'one more detail', the speaker adds 'start playing the queue', which is an addition to the original task. The final goal combines the original action with the new condition, resulting in a complete and coherent instruction that matches the reference when the addition is included.", "heard_transcript": "Reset my uh Spotify queue with all of its recommended songs, shuffle it, and one more detail, start playing the queue." }, "candidate_id": "wav_5d4dc639991c90f0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, multi-part request with a clear addition. The initial request is fluent, followed by a brief pause and the filler 'uh' before 'shuffle it'. The second addition, 'one more detail', is a plausible conversational cue, and the final instruction 'start playing the queue' is smoothly integrated with natural intonation and pacing. The flow feels like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs,, and play it. Oh, I forgot to mention, shuffle the queue.", "long_pause_offsets": [ 72 ], "source_sha256": "95a58f6ec9fbd63fb42b953e48c18c51e16627d4aa7867b333980d66e1055b55", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing shuffle action, which was not in the initial utterance but is present in the original instruction. The final goal combines all three actions: reset, shuffle, and play." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/addition/addition_01/761ddefb9fa2d60b/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "9fe0951e25f9cff89145c329bda2bca840da0e9499be84c43df6d8d94c2a51da", "seconds": 12.4, "spoken_utterances": [ "Reset my Spotify queue with all of its ... recommended songs,, and play it. ", "Oh, I forgot to mention, ... shuffle the queue." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to reset the Spotify queue with recommended songs, shuffle it, and play it. The user first executed the reset and play part, then added the shuffle instruction. Since the final audio includes the complete task with the addition of 'shuffle the queue', the final intent matches the original instruction. The addition type requires combining the initial instruction with the added information, which has been done correctly.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs and play it. Oh, I forgot to mention, shuffle the queue." }, "candidate_id": "wav_86baca9c634d9298", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a fluent request, but the addition 'oh, I forgot to mention' is delivered with a slightly abrupt, rehearsed quality. The pause before 'shuffle the queue' is long and the delivery of the final phrase is flat and clipped, lacking the natural prosody of a spontaneous thought. The overall impression is less organic than audio_02, despite being semantically complete." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_recommendation_source", "proposed_instruction": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that 'recommended songs', not 'top tracks'.", "long_pause_offsets": [ 76 ], "source_sha256": "e92c3389105240d2b844885db8692b4e585c50fa093b8007ff6276c533bcacc0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'top tracks' to 'recommended songs', preserving the app, action sequence, and other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/revision/revision_02_recommendation_source/65c463ecd86e45e0/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "0e976ed994f77ba749aaa2116b9e61ad508252164de3f465884d93025f613561", "seconds": 15.68, "spoken_utterances": [ "Reset my ... Spotify queue with all of its ... top tracks, shuffle it, and play it. ", "Make that 'uh... recommended songs', not 'top tracks'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects to 'recommended songs' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Reset my Spotify queue with all of its top tracks, shuffle it, and play it. Make that uh recommended songs, not top tracks." }, "candidate_id": "wav_6b221fc4c9d120ec", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural revision flow. The speaker starts with a clear, fluent request, then hesitates with 'make that uh' before correcting 'top tracks' to 'recommended songs'. The correction is integrated smoothly into the same utterance, with a natural-sounding pause and filler. The prosody is consistent, and the correction feels like a spontaneous realization rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Reset my Apple Music queue with all of its recommended songs, shuffle it, and play it. Sorry, I meant Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "e01e07e1ec585ceebe6a697407f92be0bee84ac2fc6b0da1aad75e5fb781a646", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining all other requirements: resetting the queue, including recommended songs, shuffling, and playing." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/revision/revision_01_app/c754b2d399ba11bc/speech.wav", "wer": 0.0, "wav_sha256": "3d7cb6c24db9ab2ca778a7705df4c486aad5834f6c52d8ca5a5e373a7b0e69f9", "seconds": 13.24, "spoken_utterances": [ "Reset my um, let's see... Apple Music queue with all of its um... recommended songs, shuffle it, and play it. ", "Sorry, I meant Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Reset my um, let's see, Apple Music queue with all of its um, recommended songs, shuffle it and play it. Sorry, I meant Spotify, not Apple Music." }, "candidate_id": "wav_02161f1e79fb7e08", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable false start and self-correction. The speaker begins with 'reset my um let's see' and then corrects 'Apple Music' to 'Spotify'. The initial hesitation and filler 'um' are natural, and the correction is immediate and integrated. The prosody is slightly more hesitant than audio_04, but the flow feels authentic and not overly rehearsed." } ] }, "is_default": false }, { "variant_id": "revision_04_playback_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it. That should be 'play it', not 'pause it'.", "long_pause_offsets": [ 84 ], "source_sha256": "1242ee4d4fac8d122282646e309372e7c9dfb975eca8f67ac6b3ba4011fcd2c1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the final action from 'pause it' to 'play it', while retaining all other requirements including the queue reset, inclusion of recommended songs, and shuffling." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/revision/revision_04_playback_action/0e65bb75295d78f6/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "d5accbeeb9b0e39df8ec57beaeaae213adc0ad95f6a9683c3b5088ac4b615f10", "seconds": 14.84, "spoken_utterances": [ "Reset my ... Spotify queue with all of its recommended songs, ... shuffle it, and pause it. ", "That should be 'uh, let's see... play it', not 'pause it'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to 'pause it' but immediately self-corrects to 'play it, not pause it' after a brief hesitation. This constitutes a revision where the final audible intent is to play the queue, which matches the original instruction. The correction is clear and immediate, satisfying the revision type criterion.", "heard_transcript": "Reset my Spotify queue with all of its recommended songs, shuffle it, and pause it. That should be uh let's see, play it, not pause it." }, "candidate_id": "wav_3b88c027c02c4f97", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a revision from 'pause it' to 'play it', but the correction is delivered as a separate sentence ('That should be uh let's see play it') rather than being integrated into the original request. This makes the flow feel slightly less natural, as if the speaker is restarting the sentence. The prosody is otherwise consistent, but the revision structure is less fluid than in audio_01 or audio_04." } ] }, "is_default": false }, { "variant_id": "revision_03_queue_ordering_action", "proposed_instruction": "Reset my Spotify queue with all of its recommended songs, clear it, and play it. Change 'shuffle it' to 'clear it'.", "long_pause_offsets": [ 81 ], "source_sha256": "6e9db3c4b1cf4a76476c2f2a8e66c43fb8fc054dd465d700028a99ceb0fb547b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'shuffle it' to 'clear it', replacing the action while preserving the other requirements: resetting the queue with recommended songs and playing it." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/revision/revision_03_queue_ordering_action/40e7950586fd0033/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "e378e97f352e23b4cdd7938023784f6e445b08259195bfa4b8764fe50a91595b", "seconds": 15.92, "spoken_utterances": [ "Reset my Spotify queue with all of its uh... recommended songs, uh, let me think... clear it, and play it. ", "Change 'uh... shuffle it' to 'clear it'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shuffle it' but immediately corrects themselves to 'clear it' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to clear the queue, not shuffle it, which is a revision of the original instruction. The rest of the task (reset queue with recommended songs, play it) remains unchanged and is executed as intended.", "heard_transcript": "Reset my Spotify queue with all of its uh recommended songs, uh let me think, clear it and play it. Change uh shuffle it to clear it." }, "candidate_id": "wav_df6a183e9312f07b", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains multiple self-corrections and hesitations ('uh', 'let me think') that disrupt the natural flow. The speaker first says 'clear it and play it', then immediately corrects 'change uh shuffle it to clear it'. The second correction is awkwardly phrased and sounds like a mechanical insertion rather than a spontaneous thought. The prosody is inconsistent, and the overall delivery feels less like a natural request and more like a rehearsed or fragmented attempt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "58c7a66088d339a81d246eb0b037f500018a05c8e2a3ca292c3b94daaf8add47", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/cancellation/cancel_010/edb79120d4be2c7d/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "e527b68a78f155db83af47569a6a9f859cf4a28e25b82f85fb2b20c8ba7bab8a", "seconds": 13.68, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it.", "Hold on, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a complete request to reset, shuffle, and play the Spotify queue. After a brief pause and the interjection 'Hold on', they explicitly state 'cancel the request I just made'. This constitutes a clear cancellation of the entire task. The final control intent is a cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it. Hold on, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "50beb43770e3c58cf1b067a44cc7367e9e12004f2fd846ab3ce19a42512f0f6f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/ambiguous_halt/halt_012/bf8f82e63da6237f/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "143610c7c5ae9e9cae0e5c562a25376ff913fd4aef3b409707610929f21b8226", "seconds": 11.88, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then explicitly requests a pause with 'Give me a second.' This constitutes a valid ambiguous_halt as the final interjection is a clear request to wait, with no cancellation or resumption permission. The initial instruction is fully audible and complete.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "8749218_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "07341e37555d4db43365f7e409a8d48885c73b8375f21d2eebf1a964cb10c6e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/8749218_3/audio_natural/filler/original/404e53bfbc6f4a00/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "527e1240876c45027357d2d0357b30d65d07f0f29c8182255efc2c89dcdb28a1", "seconds": 8.72, "spoken_utterances": [ "Reset my um... Spotify queue with all of its recommended songs, let me think... shuffle it, and play it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler phrase 'let me think' but does not change the core task. The final instruction matches the original request in action, target, and sequence. The hesitation is internal and does not alter the goal, so the task is executed as intended.", "heard_transcript": "Reset my um Spotify queue with all of its recommended songs, let me think, shuffle it and play it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_destination_city", "proposed_instruction": "I am going on a trip to Shanghai—I meant Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "e2913707d281479114bee04e841093c8a1f772bbf5c3972f54a1fe881826f093", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Shanghai' to 'Beijing' within the same utterance. The correction does not affect the core task of incorporating comments, leaving a specific final comment, and marking the task complete. The original instruction's goal is fully preserved after the correction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_01_destination_city/cc441d4b17c5d615/speech.wav", "wer": 0.06756756756756757, "wav_sha256": "2c68dc8b2eb84b405fcc4927fd0cb1cbbf6d09fcc0d0fc35d5e9228bd7f79b64", "seconds": 27.16, "spoken_utterances": [ "I am going on a trip to Shanghai—uh... I meant ... Beijing with some of my roommates. ... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Shanghai' but immediately self-corrects to 'Beijing'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The rest of the instruction is delivered clearly and matches the original request. The final control intent is to execute the task as corrected.", "heard_transcript": "I am going on a trip to Shanghai, uh, I meant, Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "candidate_id": "wav_13383d1dd3f66b20", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits a natural, spontaneous self-correction ('Shanghai... I meant Beijing') with a clear pause and intonation shift, followed by a fluent and continuous delivery of the rest of the request. The speaker's prosody is consistent and conversational, with no signs of robotic pacing or unnatural pauses. The correction is integrated smoothly, and the overall flow is highly realistic." } ] }, "is_default": true }, { "variant_id": "false_start_04_music_app", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—or rather, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "48bc07fa451c3fafdeffc9ccea6632aa55ae857cb06fd41acefa12149d4d351a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify'. The final goal remains unchanged: incorporate comments, add the specified final comment, and mark the task complete. The correction is internal to the utterance and does not alter any other requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_04_music_app/d4a5a3cacb29ff5d/speech.wav", "wer": 0.08860759493670886, "wav_sha256": "aa94d63dd5b128109943158bb57c65a93972d9808c53d18ef0e245cde802b708", "seconds": 28.52, "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. uh, let me think... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—uh... or rather, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. ... Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlist'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction is clear and matches the original. The correction is immediate and does not introduce new constraints or remove existing ones. Therefore, the final audible intent is to execute the task with the corrected app name.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. Uh, let me think. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music, uh, or rather Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "candidate_id": "wav_24ae674f636b8fed", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable hesitation ('uh, let me think') and a natural self-correction from 'Apple Music' to 'Spotify'. The speaker's intonation and pacing are consistent with someone formulating a request in real time. The correction is integrated within the sentence, and the delivery remains fluent and plausible, though slightly less fluid than audio_02 due to the mid-sentence correction." } ] }, "is_default": false }, { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a trip to Beijing with some of my friends—no, I mean roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "ab778394450f88a5d3d904b9a92e90a832ad28c0ef6534e3c75af78360ac5bd8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'friends' to 'roommates' but the original instruction already specifies 'roommates'. The correction does not change the final goal, which remains focused on incorporating comments, adding the specified final comment, and marking the task complete." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_02_companion_group/c059e3516f87a233/speech.wav", "wer": 0.06493506493506493, "wav_sha256": "a6f09b6ebfdca6b754c892b93c0bf46f2217dbc3b09c01fa493e4f222723b88d", "seconds": 27.0, "spoken_utterances": [ "I am going on a trip to Beijing with some of my um... friends—uh... no, I mean roommates. We are managing its planning on a Todoist project for it. um... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'roommates'. The final intent is clear and matches the original instruction. The correction is a false start, and the final goal is to execute the task as described in the reference.", "heard_transcript": "I am going on a trip to Beijing with some of my um, friends. No, I mean roommates. We are managing its planning on a Todoist project for it. Um, one of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "candidate_id": "wav_8c0cd8fabc7a125f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio includes a self-correction ('friends... no, I mean roommates') with a natural pause and intonation shift. The speaker's delivery is conversational, and the correction is integrated smoothly. However, the initial hesitation ('some of my um...') and the correction are slightly less fluid than in audio_02 and audio_03, making it slightly less realistic in flow." } ] }, "is_default": false }, { "variant_id": "false_start_03_task_app", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Trello—sorry, make that Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "02e7a5c5af8d65c57d4f335b73dceec0b7d5e47dda832f5cf0d74788d58568fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately corrects it to 'Todoist'. The rest of the instruction, including the action (incorporate suggestions, comment, mark complete), the context (playlist task in the project), and the comment content, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_03_task_app/bb34470042ec53a2/speech.wav", "wer": 0.05333333333333334, "wav_sha256": "78541eee4b108328f2e95dfe387592e377aca4bcdae3bac1ebb499ce72bb18ac", "seconds": 29.32, "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a ... Trello—um... sorry, make that ... Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately self-corrects to 'Todoist'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The corrected target application is 'Todoist', and all other conditions (incorporating suggestions, leaving a comment, marking complete) remain unchanged and are clearly stated. The final control intent is to execute the task as corrected.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Trello, um, sorry, make that Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "candidate_id": "wav_474854cfbf323dd2", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a self-correction from 'Trello' to 'Todoist', which is a plausible error. However, the correction is delivered with a slightly abrupt and rehearsed tone, and the overall pacing feels less spontaneous. The speaker's voice also has a slightly unnatural cadence, and the correction lacks the subtle hesitation or breathiness of a truly spontaneous correction, making it less realistic than the top three." } ] }, "is_default": false }, { "variant_id": "false_start_05_comment_text", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Changes incorporated.\"—scratch that, make it \"Incorporated changes.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "a7635f002b15d6d3396fda239cf61ab6490239c13b812e347bd6af3ad8d2f7f2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests the comment \"Changes incorporated.\" but immediately corrects it to \"Incorporated changes.\" The final instruction clearly specifies the correct comment text. All other requirements—incorporating suggestions and marking the task complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/false_start/false_start_05_comment_text/99c7dba74f08c98b/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "588459fbd39d520b72eaafea8d30031d86b409fb8b310178a7b5d23e3380d037", "seconds": 31.96, "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, \"Changes incorporated.\"—um... scratch that, make it \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Changes incorporated.' but immediately self-corrects to 'Incorporated changes.' and repeats the corrected version. This is a false start where the final, corrected value is the intended instruction. The final control request is to execute the task with the corrected comment text.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Changes incorporated.\" um, scratch that, make it \"Incorporated changes.\" and mark it complete. um, scratch that, make it \"Incorporated changes.\" and mark it complete." }, "candidate_id": "wav_decbe59dc10d8472", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This audio features a false start with a self-correction ('uh... scratch that, make it incorporated changes...') that is repeated verbatim. The repetition of the same phrase with no variation or natural variation in intonation makes it sound artificial or mechanically inserted rather than a genuine spontaneous correction. The delivery also has a slightly robotic, flat quality, and the repetition undermines the realism of the self-correction." } ] }, "is_default": false }, { "variant_id": "false_start_06_completion_action", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and complete it—I should say mark it complete.", "long_pause_offsets": [], "source_sha256": "60c3ca068f4706dc67893d5667bb934b3ac0c42496b2df17c43fd80dce64a2fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'complete it' but immediately corrects to 'mark it complete'. The final instruction clearly specifies marking the task as complete, which matches the original instruction's requirement. All other elements, including incorporating comments and leaving the specified final comment, are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. hmmm... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, ... \"Incorporated changes.\", and complete it—um... I should say mark it complete." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I am going on a trip to Beijing. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Oh, I forgot to mention, it's just with some of my roommates.", "long_pause_offsets": [ 364 ], "source_sha256": "6cb1bbf09fa1ae219196915364ba41dd06428bd5e79116a96f48e8f6556937ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds a clarification about the group ('just with some of my roommates') but does not change any of the core actions or conditions. The original instruction already specified the task context (Beijing trip, Todoist project, playlist comments), and the follow-up only adds a detail about the group, which is consistent with the original. The final goal remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/addition/addition_01/ac52d0dec5804a62/speech.wav", "wer": 0.0641025641025641, "wav_sha256": "174397d64065d81fc4ed5032fab01e27409aa79ba374ee16b08997d7f1d6dbc8", "seconds": 28.44, "spoken_utterances": [ "I am going on a trip to Beijing. ... We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. uh... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. ", "Oh, I forgot to mention, it's just with some of my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains an addition to the original instruction. The speaker first gives the core task (incorporate suggestions, comment 'Incorporated changes', mark complete) and then adds the detail 'it's just with some of my roommates' after a pause. This addition clarifies the group involved but does not change the core action. The final intent is to execute the original task with the added context, which matches the 'addition' type criterion.", "heard_transcript": "I am going on a trip to Beijing. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. Uh, I've made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes,\" and mark it complete. Oh, I forgot to mention, it's just with some of my roommates." }, "candidate_id": "wav_cd9c76d06ce30cfb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural pacing and a believable addition at the end. The initial statement is fluent, followed by a pause and a self-corrected phrase ('for it') that sounds like a spontaneous recovery. The final addition ('Oh, I forgot to mention...') is a pragmatic, believable afterthought with appropriate intonation and filler, fitting the 'addition' augmentation type well. The overall prosody is consistent and conversational." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates.. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. I left out a detail: we are managing the planning on a Todoist project.", "long_pause_offsets": [ 333 ], "source_sha256": "5d1695abeaafe81d299367a3c7ea9f4499eea323bf0bb5db5bbfbdb85e687fb9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing context that the planning is managed on a Todoist project. This clarifies the environment but does not change the core actions: incorporate comments, leave the specified comment, and mark complete. The original instruction already implied the task was in a project, and the addition specifies the app, which is consistent with the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/addition/addition_02/13323de659e7afaf/speech.wav", "wer": 0.06493506493506493, "wav_sha256": "71f4cd3fa6a06d343e7e5512fff6c932b90393732413d4e1e0919a64218c9631", "seconds": 32.2, "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates.. One of the tasks in it is about preparing a Spotify playlist. uh... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. ... Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. ", "I left out a detail: we are managing the planning on a uh, let's see... Todoist project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed. The addition of 'we are managing the planning on a Todoist project' is a clarification of the context and does not change the core task. The final control intent is to execute the original request with the added detail as a contextual note.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. One of the tasks in it is about preparing a Spotify playlist. Uh, I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. I left out a detail, we are managing the planning on a uh, let's see, Todoist project." }, "candidate_id": "wav_1d768d64c7f3f439", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip begins with a fluent, natural-sounding request but includes a noticeable pause and filler ('uh') before the main task. The speaker then adds a detail ('we are managing the planning on a uh let's see...') that is awkwardly phrased and seems to be a spontaneous attempt to recall a project name, but the delivery is less smooth than audio_02. The addition feels slightly forced, and the overall flow is less cohesive, making it less realistic in comparison." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_destination_city", "proposed_instruction": "I am going on a trip to Tokyo with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Wait, the trip is to Beijing, not Tokyo.", "long_pause_offsets": [ 388 ], "source_sha256": "22fd9c8b4fc31a99221ae019e57bc7e4bef55f7313d95953759c4db18dd5c7a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination city from Tokyo to Beijing, which matches the original instruction. All other conditions, including the app (Todoist), task context, actions (incorporate suggestions, comment, mark complete), and the comment content, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_01_destination_city/0b63cce9d18ea2b5/speech.wav", "wer": 0.075, "wav_sha256": "c802bfa1a0e4fbff6fcc481e61323608c3a20e746a3d037a0af6732cc8ad5fd8", "seconds": 27.16, "spoken_utterances": [ "I am going on a trip to Tokyo with some of my roommates. uh... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. um... Please incorporate them, leave a final comment, ... \"Incorporated changes.\", and mark it complete. ", "Wait, the trip is to Beijing, not Tokyo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the trip is to Tokyo but immediately corrects themselves to Beijing. Since this is a revision of the destination, the final intent is to perform the requested actions (incorporate suggestions, comment, mark complete) for the trip to Beijing. The revision is clear and the final control intent is to execute the task with the corrected information.", "heard_transcript": "I am going on a trip to Tokyo with some of my roommates. Uh, we are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Um, please incorporate them, leave a final comment, \"Incorporated changes,\" and mark it complete. Wait, the trip is to Beijing, not Tokyo." }, "candidate_id": "wav_f2c5d24a0771fb10", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous narrative with a believable revision. The initial mention of 'Tokyo' is followed by a clear, self-corrected realization ('Wait, the trip is to Beijing, not Tokyo') with appropriate intonation and pacing. The filler 'uh' and 'um' are used naturally, and the correction flows seamlessly within the context of the task, making it sound like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_06_completion_action", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it incomplete. That should be mark it complete, not mark it incomplete.", "long_pause_offsets": [ 392 ], "source_sha256": "f24525d01d8f4cf10776ad44fac977cf9edf318e217f388a8d5f9debbc909f8e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The follow-up revision explicitly corrects the completion status from 'incomplete' to 'complete', reversing the final action required by the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_06_completion_action/097467f5ac939ee3/speech.wav", "wer": 0.10588235294117647, "wav_sha256": "aa6ab0b4e1e27fa6ae32d522a49b4f97a5c0945c2b4f999a07de160ffdf2699b", "seconds": 32.96, "spoken_utterances": [ "I am going on a trip to Beijing ... with some of my roommates. We are managing its planning on a Todoist project for it. uh... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh, let me think... Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it incomplete. ", "That should be mark it complete, not mark it incomplete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark it incomplete' but immediately corrects themselves to 'mark it complete, not mark it incomplete'. This is a clear revision where the final audible intent is to mark the task as complete. The correction is immediate and unambiguous, satisfying the revision type criterion. The rest of the task (incorporating suggestions and leaving the comment) remains unchanged.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, let me think. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it incomplete. That should be mark it complete, not mark it incomplete." }, "candidate_id": "wav_0ccb771f57182f68", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a realistic, hesitant delivery with a self-repair. The speaker starts with 'uh' and 'let me think' before correcting 'marketing complete' to 'mark it complete'. The correction is integrated smoothly into the sentence, and the pauses and fillers are well-placed, suggesting genuine thought rather than a rehearsed or robotic delivery." } ] }, "is_default": false }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a trip to Beijing with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Sorry, I meant roommates, not friends.", "long_pause_offsets": [ 388 ], "source_sha256": "94511ffde4dbe362d526e9bdb507887448250f41399d1593f683083e30520d69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects 'friends' to 'roommates', which is the only change. All other conditions, including the task context, action (incorporate suggestions, comment, mark complete), and the comment content, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_02_companion_group/efd669029aab951f/speech.wav", "wer": 0.03896103896103896, "wav_sha256": "4cbf67b12087ab13fc66ea79a58bcd52ecd3590346614cef7c0a03203eb6526e", "seconds": 30.88, "spoken_utterances": [ "I am going on a trip to Beijing with some of my friends. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. um... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, ... \"Incorporated changes.\", and mark it complete. ", "Sorry, I meant roommates, not friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately corrects to 'roommates' with 'Sorry, I meant roommates, not friends.' This is a revision of the target group. The final intent is to execute the core task of incorporating comments, leaving the specified final comment, and marking the task complete, which aligns with the original instruction's intent. The revision is a minor correction to a non-critical detail and does not change the primary action or its conditions.", "heard_transcript": "I am going on a trip to Beijing with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. Um, I have made the playlist and shared it with others on the project, but they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Sorry, I meant roommates, not friends." }, "candidate_id": "wav_56fa0d5dbf074809", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a revision from 'friends' to 'roommates' with a clear 'sorry' and a natural-sounding correction. The 'um' at the beginning and the slight hesitation before the correction contribute to a realistic, conversational feel. The flow is coherent, and the correction is delivered as a genuine moment of recollection." } ] }, "is_default": false }, { "variant_id": "revision_03_task_app", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Make that Todoist, not Trello.", "long_pause_offsets": [ 389 ], "source_sha256": "50c1f9218fb791d3021e5eae88b63236510594c895fc0a38ee24b7a4bded15a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the project management app from Trello to Todoist, while all other requirements—trip context, playlist task, incorporating comments, final comment, and marking complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_03_task_app/33bf37b6502945ae/speech.wav", "wer": 0.0379746835443038, "wav_sha256": "9a16d2db6f16771bdd95878459f5693320062a87a2ffdd9c96c719123f3a8c9b", "seconds": 33.28, "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. hmmm... We are managing its planning on a Trello project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. ", "Make that uh, let's see... Todoist, not Trello." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but then immediately corrects themselves to 'Todoist'. This is a revision of the target application. The final intent is to perform the same actions (incorporate comments, leave a comment, mark complete) but on the correct project, Todoist. The revision is clear and final, so the task should be executed on Todoist.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Make that uh let's see, Todoist, not Trello." }, "candidate_id": "wav_9623fbeec5b278ed", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip includes a revision from 'Trello' to 'Todoist' with a self-correction ('Make that, uh, let's see, Todoist, not Trello'). The 'uh' and 'let's see' are natural fillers, and the correction is integrated into the sentence. However, the phrasing 'Make that' is slightly awkward and less fluid than the other revisions, making it feel a bit more like a scripted edit than a spontaneous thought." } ] }, "is_default": false }, { "variant_id": "revision_04_music_app", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Change Spotify to Apple Music.", "long_pause_offsets": [ 394 ], "source_sha256": "ce7c08902bf491256725d86fac7f50bf633620ff583628c96fa0e89ef9cd5776", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the music app from Spotify to Apple Music, which aligns with the final spoken version. All other requirements—context, action, comment content, and completion status—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/revision/revision_04_music_app/5dd841fd5722abfe/speech.wav", "wer": 0.06329113924050633, "wav_sha256": "ab2fdcbdd4e98d29592caeab3fca5307cbe1f2d82d52dfb010bcb05d38cc1b83", "seconds": 35.44, "spoken_utterances": [ "I am going on a trip to Beijing ... with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. let me think... Please incorporate them, leave a final comment, ... \"Incorporated changes.\", and mark it complete. ", "Change Spotify to Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Spotify playlist' but then immediately corrects it to 'Apple Music playlist' in a revision. The final instruction is to 'change Spotify to Apple Music', which is a clear revision of the original task. The rest of the instructions remain unchanged and are to be executed as stated. The revision is spoken after a pause and is a direct correction of the initial value, satisfying the revision type criterion.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Let me think. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Change Spotify to Apple Music." }, "candidate_id": "wav_15f75b9c15877f3b", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip contains a revision from 'Spotify' to 'Apple Music', but the correction is delivered as a separate, abrupt statement ('Change Spotify to Apple Music') rather than being integrated into the ongoing sentence. This makes it sound less like a natural self-correction and more like a post-hoc instruction, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_05_comment_text", "proposed_instruction": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Changes incorporated.\", and mark it complete. Use \"Incorporated changes.\" instead of \"Changes incorporated.\"", "long_pause_offsets": [ 390 ], "source_sha256": "4fd2b763f7e04cb5f41a49bd922a29eb59af34d0a28ff68496eca4178480a13c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the exact wording of the final comment from \"Changes incorporated.\" to \"Incorporated changes.\", which is the only change. All other requirements—incorporating suggestions, leaving a comment, and marking complete—are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to Beijing with some of my roommates. um, let me think... We are managing its planning on a Todoist project for it. um... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Changes incorporated.\", and mark it complete. ", "Use \"... Incorporated changes.\" instead of \"Changes incorporated.\"" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "84056d5313a32f379a3b3d2baf20ab46fbe469ce693fc15e2bd343a4ba02be09", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/cancellation/cancel_018/43798e97b7e0442f/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "199da543e0ae6a7f0b38b801dff63078d31d3505664485a3da1c66df1cc7b8f8", "seconds": 32.52, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "Wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a complete instruction to incorporate comments, leave a specific comment, and mark a task complete. After a pause and a self-correction ('Wait, I rather not do that after all'), they explicitly state 'Please cancel it.' This constitutes a clear cancellation of the entire prior task request. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Wait, I rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "a60a01bf370c388ecd8f6cbbe7b5a505731a21c9b5b1f53ba2200171902efd75", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/ambiguous_halt/halt_015/277ea9dc5088a505/speech.wav", "wer": 0.07792207792207792, "wav_sha256": "c17d81e467a38177c903427ccaefa6d9e10f26850288d407b04a3aa23791ef4d", "seconds": 30.2, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete.", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Please wait a moment,' which is a clear request to pause or hold. This is consistent with the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or revoked, and the speaker does not provide a new instruction or correction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8b360a6149793c45cf5b24a544ee975e944a8a6f00235f58eb43389c928383e1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_1/audio_natural/filler/original/d2111f43aba3c11e/speech.wav", "wer": 0.0821917808219178, "wav_sha256": "8e0d7d5778210bbac3184a189b1179f2120b488da8a3880c7c661185214ad44b", "seconds": 27.12, "spoken_utterances": [ "I am going on a trip to ... Beijing with some of my roommates. hmmm... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. let's see... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'Hmm' and a brief pause, but continues with the full original instruction without any correction, addition, or cancellation. The final interjection is the completion of the original task request. The filler does not alter the goal, so the final intent matches the original instruction.", "heard_transcript": "I am going on a trip to Beijing with some of my roommates. Hmm. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. Let's see. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Incorporated changes.\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_destination_city", "proposed_instruction": "I am going on a trip to Paris—sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "72878c2bc1848a083a044845adb7c64a8999506a41485538eaa5f096586fddcd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Paris' to 'Edinburgh' within the same utterance. The correction does not affect the core task, which is to incorporate comments, leave a 'Done!' comment, and mark the task complete. The original instruction's goal is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_01_destination_city/1474dfd57775f408/speech.wav", "wer": 0.0641025641025641, "wav_sha256": "20789c0e0c548e1262dd04ee3f6b1150df0912b38e566277e13c79218a114cae", "seconds": 27.84, "spoken_utterances": [ "I am going on a trip to Paris—uh... sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. let me think... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. hmmm... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Paris' but immediately self-corrects to 'Edinburgh'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The rest of the instruction is delivered without further changes, so the final goal is to incorporate the comments, leave a 'Done!' comment, and mark the task complete. The correction is clear and immediate, satisfying the false_start rule.", "heard_transcript": "I am going on a trip to Paris, uh, sorry, I meant Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. Let me think. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "candidate_id": "wav_39cd632516104282", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and varied prosody, with believable hesitations ('Let me think'), self-corrections ('Paris... sorry, I meant Edinburgh'), and filler sounds ('huh') that are well-timed and integrated into the flow. The speaker's intonation reflects genuine thought processes, and the correction is smooth and contextually appropriate, making it sound like a spontaneous, real-time request." } ] }, "is_default": true }, { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a trip to Edinburgh with some of my friends—no, make that siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "44d4e2146c1405893ffce36bc59724a547e648c5ad347dd2682f5aa43075acf1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately corrects to 'siblings'. The rest of the instruction, including the project context, task focus, and required actions (incorporate comments, comment 'Done!', mark complete), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_02_companion_group/632733b2c05836d6/speech.wav", "wer": 0.05333333333333334, "wav_sha256": "725030dbf1aa33b69e643914c389c3991b6b3a1991bee53d9b9faa375d977278", "seconds": 26.44, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my uh... friends—um... no, make that ... siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'siblings'. The final intent is clear and matches the original instruction after the correction. The task is a false_start type, so the immediate self-correction is applied, and the rest of the instruction is executed as stated.", "heard_transcript": "I am going on a trip to Edinburgh with some of my uh friends. Um, no, make that siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "candidate_id": "wav_32b5ea9397633101", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start ('friends... no, make that siblings') with a natural-sounding correction and a brief pause. The speaker's voice and pacing are consistent, and the hesitation feels authentic. The delivery is slightly more rehearsed than audio_03 but still highly realistic, with a believable shift in thought." } ] }, "is_default": false }, { "variant_id": "false_start_06_completion_action", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and check it off—no, I mean mark it complete.", "long_pause_offsets": [], "source_sha256": "cc9fe7526862b1255d28ad3e264dc92573c9e0d6fbb9dc4674d0124d4633ea3e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'check it off' but immediately corrects to 'mark it complete', which is the same action as per the original instruction. The rest of the goal remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_06_completion_action/385d702c2a738017/speech.wav", "wer": 0.0641025641025641, "wav_sha256": "f8a5356f29510d7aa68e6d6c7c5e326a2d7b9633724186cd5987518743fe4c9d", "seconds": 29.28, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. let's see... We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. ... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and check it off—uh... no, I mean mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'check it off' but immediately self-corrects to 'mark it complete'. This is a false start where the final intent is clear and correct. The correction is immediate and unambiguous, so the final instruction is to mark the task complete, which matches the original reference instruction. No other constraints are changed.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. Let's see. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and check it off. Uh, no, I mean mark it complete." }, "candidate_id": "wav_7ad16654d186b00f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio includes a self-correction ('check it off... no, I mean mark it complete') that is well-placed and sounds spontaneous. The speaker uses a natural 'uh' and 'no' to signal the correction, and the intonation reflects a genuine moment of reconsideration. The pacing is slightly faster, but the repair is credible and fits the context." } ] }, "is_default": false }, { "variant_id": "false_start_03_task_app", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Trello—I should say Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "6006c54160de6b934930a467abbedd8a1c9aae2e930fecf6432d19dbdf58f10b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately corrects to 'Todoist'. The final goal remains unchanged: the task is in the correct app (Todoist), and the requested actions (incorporate comments, comment 'Done!', mark complete) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_03_task_app/7270e4801360c716/speech.wav", "wer": 0.039473684210526314, "wav_sha256": "e8ca069f4319a5a4c4501b0b4c69c8e4e55ebf7d1e774058adb106cc2e6de7e1", "seconds": 29.48, "spoken_utterances": [ "I am going on a trip to um... Edinburgh with some of my siblings. uh... We are managing its planning on a Trello—hmmm... I should say Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately self-corrects to 'Todoist'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The final instruction matches the original reference exactly, with the correct app name and all other conditions preserved. Therefore, the task should be executed as stated.", "heard_transcript": "I am going on a trip to um Edinburgh with some of my siblings. uh We are managing its planning on a Trello. um I should say Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "candidate_id": "wav_770be97d70d798b0", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains multiple fillers ('um', 'uh') and a self-correction ('Trello... I should say Todoist project'), which are delivered in a way that suggests real-time thinking. The speaker's voice is consistent, and the correction is integrated smoothly. However, the delivery is slightly more hesitant and less fluid than the top three, making it feel a bit more like a rehearsed or slightly awkward attempt." } ] }, "is_default": false }, { "variant_id": "false_start_04_music_app", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—sorry, make that Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "c8dcd44190c03af77fe078237e9c2dcb7cb6bba78ec6e2d41f47ded944596a5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify'. The rest of the instruction, including the app, task context, and required actions (incorporate suggestions, comment 'Done!', mark complete), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_04_music_app/9ab5664b32f3e313/speech.wav", "wer": 0.07894736842105263, "wav_sha256": "bcc5d7687c0ade4aed615ca4bfed4441c18e0b9c2ada562b7986d0a4bed835f2", "seconds": 28.96, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a um... Apple Music—uh... sorry, make that Spotify playlist. I have made the playlist and shared it with others on the project. ... But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlist'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction is clear and matches the original. The correction is immediate and unambiguous, so the final goal is to act on the Spotify playlist.", "heard_transcript": "I am going in a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing um, Apple Music uh, sorry, make that Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "candidate_id": "wav_472772b9c8e650cd", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This audio includes a self-correction ('Apple Music... sorry, make that Spotify playlist') that is natural, but the delivery is slightly rushed and the correction feels a bit abrupt. The speaker's voice is consistent, but the overall pacing and intonation are less varied, making it sound slightly less spontaneous than the others." } ] }, "is_default": false }, { "variant_id": "false_start_05_comment_text", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Completed\"—scratch that, make it \"Done!\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "2afc6499b2601ad4a118ccd9d3dc024db47bcbdb7113e963f9f0b2179eab3ab7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests leaving the comment \"Completed\" but immediately corrects it to \"Done!\". The final instruction clearly specifies the correct comment text. All other requirements (incorporating suggestions, marking complete) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/false_start/false_start_05_comment_text/811b7de129b719c5/speech.wav", "wer": 0.17105263157894737, "wav_sha256": "9b9b4d0f94658197af2ff2f72dc0cbe5f15f2a55976bbb3c362e2d7e2e5b2ec9", "seconds": 29.8, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. ... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, \"Completed\"—um... scratch that, make it \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says to leave a comment saying \"Completed\" but immediately self-corrects to \"Done!\". This is a false start, where the final corrected value is what should be executed. The final instruction is to make the comment \"Done!\" and mark it complete, which matches the original intent of finalizing the task with a specific comment. The repetition of the correction does not change the final goal.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Completed\". Um, scratch that, make it \"Done!\" and mark it complete. Um, scratch that, make it \"Done!\" and mark it complete." }, "candidate_id": "wav_905d070875ac0657", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This clip has the least natural delivery. It features multiple self-corrections ('scratch that, make it done and mark it complete') that are repeated and sound mechanical rather than spontaneous. The speaker's voice is flat, and the pauses and fillers ('uh', 'um') are less integrated into the flow. The repetition of the same correction suggests a possible editing artifact or a less authentic delivery." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist.. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Oh, and one more thing: I've already made the playlist and shared it with others on the project.", "long_pause_offsets": [ 309 ], "source_sha256": "c753d42139a57ed30d53de98c0d1d7e3122b648f5c34cbe421ed0944d4cc0c15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds background context about the playlist already being made and shared, but does not change or contradict the core action: incorporate suggestions, comment \"Done!\", and mark complete. The original instruction already included these steps, and the addition only clarifies the state of the playlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/addition/addition_03/0f4c06d56ed45125/speech.wav", "wer": 0.05333333333333334, "wav_sha256": "a3fb4ad271e18bbe1966fd9dd14a86133d221c385ac61207f58169de2f4d958b", "seconds": 29.32, "spoken_utterances": [ "I am going on a trip to Edinburgh with um... some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist.. ... But they have made some suggestions in comments. Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete. ", "Oh, and one more thing: I've already made the playlist and shared it with others on the project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the original instruction followed by an addition that provides extra context ('I've already made the playlist and shared it with others on the project'). The addition does not change the core task of incorporating suggestions, leaving a 'Done!' comment, and marking the task complete. The final intent is to execute the original task with the added context, which matches the 'addition' type criterion.", "heard_transcript": "I am going on a trip to Edinburgh with um some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. But they have made some suggestions in comments. Please incorporate them, leave a final comment, done, and mark it complete. Oh, and one more thing. I've already made the playlist and shared it with others on the project." }, "candidate_id": "wav_c1d2be6b07cd85b2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural flow. The addition of 'Oh, and one more thing' after the main instruction is a highly plausible conversational device, and the speaker smoothly integrates the previously omitted detail without awkwardness. The pacing, intonation, and use of 'um' are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it.. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. I left out a detail: one of the tasks is about preparing a Spotify playlist.", "long_pause_offsets": [ 314 ], "source_sha256": "143e9e4801803069ec0cb756c262efc9f981c6c39b326f4482c8d094353c0b1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail that the task is about preparing a Spotify playlist. This clarifies the target task but does not change the required actions: incorporate comments, leave a 'Done!' comment, and mark complete. The original instruction already implied this task was part of the project, and the addition only specifies its subject." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/addition/addition_02/cf13350911e56d71/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "2422127e30c71ab04c65773bcf84a0048a4695fb418824b9dcee4d2b5c522228", "seconds": 28.8, "spoken_utterances": [ "I am going on a trip to Edinburgh with um, let's see... some of my siblings. We are managing its planning on a Todoist project for it.. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. let's see... Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "I left out a detail: one of the tasks is about preparing a Spotify playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the original instruction and then adds a clarifying detail ('One of the tasks is about preparing a Spotify playlist') after a pause. This addition narrows the scope of the initial request, specifying which task the comments pertain to. The final control intent is to execute the original action (incorporate comments, comment 'done', mark complete) on the specified Spotify playlist task. The addition is a clarification, not a revision or cancellation, and the final intent is clear and executable.", "heard_transcript": "I'm going on a trip to Edinburgh with um let's see some of my siblings. We are managing its planning on a Todoist project for it. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Let's see. Please incorporate them, leave a final comment, done and mark it complete. I left out a detail. One of the tasks is about preparing a Spotify playlist." }, "candidate_id": "wav_8756d63d87289910", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a similar setup but adds the missing detail as a separate, self-contained sentence ('I left out a detail...') rather than integrating it into the ongoing thought. This makes the addition feel more like a post-hoc correction than a natural thought process. The delivery is fluent but slightly less organic in its structure compared to audio_02." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings.. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Oh, I forgot to mention, we're managing the planning on a Todoist project for it.", "long_pause_offsets": [ 318 ], "source_sha256": "121a2cff50a745b47548036d6d874d5829b02f0447b7f549a0b1d87fea2adc6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing context that the planning is managed on a Todoist project. This clarifies the environment but does not change the core actions: incorporate suggestions, comment \"Done!\", and mark complete. The original instruction already implied the task was in a project, and the addition specifies the app, which is consistent with the original intent." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to ... Edinburgh with some of my siblings.. One of the tasks in it is about preparing a Spotify playlist. uh, let me think... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "Oh, I forgot to mention, we're managing the planning ... on a Todoist project for it." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_task_app", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Let me correct the app: we're using Todoist, not Trello.", "long_pause_offsets": [ 374 ], "source_sha256": "4ad6e8cc9c086baa36e39b31742cd9e7e9dc17e4c0485f1843f6203d2e66d5f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the project management app from Trello to Todoist, while retaining all other requirements: incorporating comments, leaving a 'Done!' comment, and marking the task complete. The initial mention of Trello is replaced by the corrected app name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_03_task_app/1cfe8679bd9342f1/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "dce31afa496c8d2df0315e99b7b4c0b23866891151c6e4fb7d38b2684b618b4a", "seconds": 34.52, "spoken_utterances": [ "I am going on a trip to Edinburgh with hmmm... some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. um... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. ... Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "Let me correct the app: we're using Todoist, not Trello." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Trello but immediately corrects themselves to Todoist. Since this is a revision, the final instruction is to perform the same actions (incorporate comments, leave a comment, mark complete) but on the correct app, Todoist. The core task remains unchanged and is fully specified.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Trello project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Let me correct the app. We're using Todoist, not Trello." }, "candidate_id": "wav_a37af5f03776689d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant setup, then smoothly corrects the project tool from 'Trello' to 'Todoist' after a brief pause and self-correction. The prosody is consistent, with believable fillers ('um') and a natural, conversational rhythm. The revision is integrated seamlessly into the narrative without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_04_music_app", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Change the music app to Spotify, not Apple Music.", "long_pause_offsets": [ 379 ], "source_sha256": "5d5520f25b25c2c671df715e8d3d959f766d4612d281b76e97a7c15d5b9b93b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the music app from Apple Music to Spotify, which aligns with the original instruction. All other requirements—task context, action (incorporate comments), final comment, and completion—remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_04_music_app/d34c455ee23ff3b1/speech.wav", "wer": 0.07317073170731707, "wav_sha256": "8a730ea90cdb5ef334489416df6921ec4b5b5608034f6588cb0aa9d0e4391162", "seconds": 31.88, "spoken_utterances": [ "I am going on a trip to ... Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh, let's see... Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "Change the music app to ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music playlist' but then corrects it by saying 'Change the music app to Spotify not Apple Music.' This is a revision where the final instruction overrides the initial incorrect value. The final goal is to update the task to use Spotify, which matches the reference instruction. All other conditions are preserved and correctly executed.", "heard_transcript": "I am going on a trip to um Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh let's see. Please incorporate them, leave a final comment, \"Done!\" and mark it complete. Change the music app to Spotify not Apple Music." }, "candidate_id": "wav_61d6c2d590aa3948", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is very realistic, with a natural, slightly rambling delivery. The speaker makes a self-correction from 'Apple Music' to 'Spotify' after a pause, which sounds spontaneous. The prosody is varied and the pacing is human-like, with a believable 'let's see' pause before the correction. The only minor issue is a slight robotic quality in the final sentence, but overall it's highly natural." } ] }, "is_default": false }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a trip to Edinburgh with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. I misspoke about the group; it's my siblings, not friends.", "long_pause_offsets": [ 374 ], "source_sha256": "b4cc0d885539de713b4af486a9e37dc8d77d69d88af40a8b117859877745cff5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up corrects the group from 'friends' to 'siblings', but the core task—processing comments, adding a 'Done!' comment, and marking the playlist task complete—remains unchanged. The correction is a clarification of the group context, not a change to the action or its conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_02_companion_group/4a06990ecec1c7c4/speech.wav", "wer": 0.09411764705882353, "wav_sha256": "fb5c584ccf07c6a90ee399dfb5fa0102cacc82903293d8bb8f719a581cee3cee", "seconds": 36.76, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my friends. We are managing its planning on a Todoist project for it. um... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. um, let me think... But they have made some suggestions in comments. Please incorporate them, leave a final comment, hmmm... \"Done!\", and mark it complete. ", "I misspoke about the group; it's my siblings, not friends." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_48097fe16d5355ff", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a strong sense of realism with a clear false start and self-correction. The speaker initially says 'friends' but then corrects to 'siblings' after a 'let me think' pause, which is a very plausible moment of recollection. The delivery is conversational, with natural fillers and a believable change of mind. The correction is integrated smoothly into the flow." } ] }, "is_default": false }, { "variant_id": "revision_06_completion_action", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it as done. That should be mark it complete, not mark it as done.", "long_pause_offsets": [ 374 ], "source_sha256": "8d4f9f41460729a5da45e12b63272d27846582e6d4e514f76fb301a7437f8208", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the phrasing from 'mark it as done' to 'mark it complete', which is functionally equivalent. The rest of the goal—incorporating comments and leaving the 'Done!' comment—is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_06_completion_action/dbe1197e6160becf/speech.wav", "wer": 0.12048192771084337, "wav_sha256": "116d86c14a6a045bd58c8d4988c292dc8de910cab80eac1e8b02d3ca00180baf", "seconds": 30.84, "spoken_utterances": [ "I am going on a trip to uh... Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. um... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it as done. ", "That should be mark it complete, not mark it as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark it as done' but immediately corrects themselves to 'mark it complete'. Since this is a revision with a clear self-correction, the final intent is to mark the task as complete. The correction is immediate and unambiguous, satisfying the revision type criterion. All other parts of the instruction are preserved.", "heard_transcript": "I am going on a trip to uh Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. Um. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it as done. That should be mark it complete, not mark it as done." }, "candidate_id": "wav_e2bc93b9d13595f3", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker delivers a revision with a clear self-correction from 'mark it as done' to 'mark it complete'. The prosody is natural, and the correction is made with a believable shift in intonation. However, the delivery is slightly more robotic than the top-ranked clips, and the correction feels a bit more like a read-aloud than a spontaneous thought." } ] }, "is_default": false }, { "variant_id": "revision_01_destination_city", "proposed_instruction": "I am going on a trip to Paris with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Done!\", and mark it complete. Wait, the trip is to Edinburgh, not Paris.", "long_pause_offsets": [ 371 ], "source_sha256": "8d1ea9669ebea88ac27f7d3931bde0fef8a602039e818c99375b883f770d94f6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination city from Paris to Edinburgh, which is the only change. All other conditions—trip with siblings, Todoist project, playlist task, incorporating comments, leaving 'Done!' as a comment, and marking complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_01_destination_city/1142d99a8a4e1ec1/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "e91463b10d8a9eb46efcd2c547133b91c706d715b4b078d71657eac8693f1a8b", "seconds": 35.88, "spoken_utterances": [ "I am going on a trip to Paris with ... some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. ... But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, \"Done!\", and mark it complete. ", "Wait, the trip is to Edinburgh, not Paris." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the trip is to Paris but immediately corrects themselves to Edinburgh. Since this is a revision of the destination, the final intent is to perform the requested actions (incorporate comments, leave 'Done!', mark complete) for the trip to Edinburgh. The revision applies to the earlier instruction, and all other conditions remain unchanged. Therefore, the final intent matches the expected control intent for a revision.", "heard_transcript": "I am going on a trip to Paris with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\", and mark it complete. Wait, the trip is to Edinburgh, not Paris." }, "candidate_id": "wav_fd3b5efd1e2ff733", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip features a revision where the speaker corrects 'Paris' to 'Edinburgh'. The delivery is natural, but the correction is made abruptly after a short pause, and the speaker's tone shifts slightly, making it feel a bit less fluid than the top-ranked examples. The initial part of the audio is clear, but the correction lacks the smooth integration seen in the best examples." } ] }, "is_default": false }, { "variant_id": "revision_05_comment_text", "proposed_instruction": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Completed\", and mark it complete. Use \"Done!\" instead of \"Completed\" as the final comment.", "long_pause_offsets": [ 379 ], "source_sha256": "6883bfbb5cfbd75800574dc946440c53f5316f7255b6b6640190f54192309c18", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up revision corrects the final comment from \"Completed\" to \"Done!\" as requested, while preserving all other requirements: incorporating suggestions, leaving a final comment, and marking the task complete. The original instruction specified \"Done!\" as the final comment, and the revision ensures that exact phrase is used." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/revision/revision_05_comment_text/65ad34106ea59f32/speech.wav", "wer": 0.2839506172839506, "wav_sha256": "a28b2108ac0a24e944fb1559f20911f8524f7d6fe2547b54a57a0666fe6acf31", "seconds": 38.12, "spoken_utterances": [ "I am going on a trip to Edinburgh with ah... some of my siblings. We are managing its planning on a Todoist project for it. uh... One of the tasks in it is about preparing a Spotify playlist. um... I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Completed\", and mark it complete. ", "Use \"Done!\" instead of \"Completed\" as the final comment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says to leave a final comment of \"Completed\" but then corrects themselves with a longer pause to use \"Done\" instead. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use \"Done\" as the comment, which matches the reference instruction. All other conditions are preserved.", "heard_transcript": "I am going on a trip to Edinburgh with uh some of my siblings. We are managing its planning on a Todoist project for it. Uh one of the tasks in it is about preparing a Spotify playlist. Um I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Completed\" and mark it complete. And mark it use \"Done\" instead of \"Completed\" as the final comment." }, "candidate_id": "wav_9d7e9e37bac936b8", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This audio has the least realistic delivery. The speaker repeats the phrase 'they have made the playlist and shared it with others on the project' twice, which sounds like a recording glitch or a very awkward stutter. The correction about using 'done' instead of 'completed' is made, but the overall flow is disrupted by the repetition, making it sound less natural and more like a mechanical error." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "fdc3500768575e63dae3bb1510f619ef2f65dd8ea3e9d92358991cc937fd62b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/cancellation/cancel_001/f73781106dd4ad35/speech.wav", "wer": 0.05405405405405406, "wav_sha256": "388d320ea1b8707c4d435dd817e8ee970cecfd9f991c3f5c65b5679c3a56e29f", "seconds": 30.84, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Oh wait, cancel that,' which is a clear and final directive to revoke the entire preceding task. This matches the 'cancellation' type criterion, where the recording must clearly revoke/cancel the task. The final interjection is not a pause or a revision but a direct cancellation of the prior instruction.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\", and mark it complete. Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "555c716c13e4555d00bcbace4cfa7be92ff9df2b5e615e7564bf7378e06e849b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/ambiguous_halt/halt_005/303c3bb868888a19/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "c306007c3570fe0d668452731f61de554b7077d694655433db03907b3fc4c18f", "seconds": 30.64, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete.", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'One second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or change its goal. The speaker's intent is to temporarily halt the process, not to execute or cancel the original instruction. Therefore, the control intent is 'pause', which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\" and mark it complete. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5e17377cdfcd00c3946b7fa8da03789d3f82cbcf41ce59543f9b25cfe7d25a60", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_2/audio_natural/filler/original/10ea413568be4891/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "9aa10aa2e0cab47be2e2eb1886e3d4ecbafad6b7d5be0e2078db4c1723d24238", "seconds": 27.2, "spoken_utterances": [ "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. ... One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. uh... Please incorporate them, leave a final comment, ... \"Done!\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'Uh' but immediately proceeds to state the full task request without any correction or change in intent. The final interjection is the complete instruction, which matches the original reference. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "I am going on a trip to Edinburgh with some of my siblings. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Uh, please incorporate them, leave a final comment, \"Done!\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_music_app", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—or rather, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "bd5c30f1146d24c07d2e51fd93a5dbbc2e4e591c9b1da43677fa28970f0eca25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' within the same utterance. The final intended app is Spotify, which matches the original instruction. All other requirements—task context, action (incorporate comments), final comment content, and completion status—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_04_music_app/862e0c28e6defae2/speech.wav", "wer": 0.07894736842105263, "wav_sha256": "31278c402f92373a950c664ad1ca1c3d7d31c448257386f3945650abbbeb5bd4", "seconds": 27.08, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music—um... or rather, ah... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"... Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify playlist'. According to the 'false_start' type criterion, the final corrected value ('Spotify playlist') should be executed, and the initial incorrect value ('Apple Music') should be discarded. The rest of the instruction is clear and matches the original. The final control intent is to execute the corrected task.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music um, or rather a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_919931a7be77320c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent delivery. The speaker makes a self-correction from 'Apple Music' to 'Spotify playlist' with a smooth, conversational 'or rather' transition. The prosody is consistent, with a natural hesitation and intonation. The correction is integrated seamlessly into the flow without any awkwardness, and the final command is delivered clearly and confidently." } ] }, "is_default": true }, { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a trip to Bangkok with some of my family—I should say friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "c55a0a616e554878e3cef252804b5cfa5bde4334c2958ea9b0505bb4cf829b48", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance, but the final goal remains unchanged. The correction only affects the introductory context, not the core task of incorporating comments, adding the specified final comment, and marking the task complete." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_02_companion_group/ee45f3280b89c887/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "d43125b03c3cfb8b41140b7ccf36570c36f4e2770b7ab02d3d20193b722ec3af", "seconds": 24.76, "spoken_utterances": [ "I am going on a trip to ah... Bangkok with some of my uh... family—hmmm... I should say friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. This is a false start, and the final, corrected target is 'friends'. The rest of the instruction is delivered without further changes. The final intent matches the original instruction, with only a minor, corrected self-correction. The task is to execute the full request as stated after the correction.", "heard_transcript": "I am going on a trip to uh Bangkok with some of my uh family. Hmm. I should say friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_a49c0a055fc097be", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable self-correction from 'family' to 'friends' with a natural 'hmm' hesitation and a self-aware 'I should say' repair. The speaker's voice and pacing are consistent, and the correction feels spontaneous. The delivery is slightly more hesitant than audio_02 but still very realistic and fluent." } ] }, "is_default": false }, { "variant_id": "false_start_05_comment_text", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated!\"—scratch that, make it \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "55d47def4100c4d0dcc80e707e90ae3e520043b660f43376a1d4a6d0b7fe8d22", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests leaving the comment \"Thanks! Updated!\" but immediately corrects it to \"Thanks! Updated.\" The final instruction clearly specifies the correct comment text, and all other requirements (incorporating suggestions, marking complete) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_05_comment_text/3d70bb6808d909af/speech.wav", "wer": 0.05194805194805195, "wav_sha256": "8be1c36b84988fdd6988dc27ad5c83e1dbdedaa9e12d203126ce279e7bb348e1", "seconds": 29.88, "spoken_utterances": [ "I am going on a trip to ... Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, ... \"Thanks! Updated!\"—hmmm... scratch that, make it \"Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests leaving the comment 'Thanks! Updated.' and marking the task complete. After a brief hesitation ('Hmm'), they self-correct with 'Scratch that. Make it \"Thanks! Updated.\" and mark it complete.' This is a clear false start where the final instruction is a corrected version of the initial one. The final intent matches the original instruction's goal, with only a minor, non-meaning-changing change in the comment's phrasing (adding 'and mark it complete' as a separate clause). The correction is immediate and does not introduce new constraints or remove existing ones. Therefore, the final audible intent is to execute the task as corrected.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Hmm. Scratch that. Make it \"Thanks! Updated.\" and mark it complete." }, "candidate_id": "wav_2763f437f84baabd", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a self-correction from 'Thanks, updated' to 'Thanks, updated and mark it complete' with a natural 'hmm' and 'scratch that' phrasing. The delivery is conversational and the correction is integrated smoothly. However, the 'scratch that' is slightly more abrupt than the 'or rather' in audio_02, making it feel marginally less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_03_task_app", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Trello—sorry, make that Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "6f9ce915c6a7c25c6fdc56d91b9b37792345be22b045286e3c99b542e8bb41f9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately corrects to 'Todoist'. The final goal is unchanged: perform the same actions (incorporate suggestions, comment, mark complete) within the correct project context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_03_task_app/3934e0e05574007a/speech.wav", "wer": 0.0379746835443038, "wav_sha256": "03d96cf36892c944937e5d03c9b2e4c07290059c8009e36ccae66438fb917a00", "seconds": 27.76, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a uh, let's see... Trello—um... sorry, make that Todoist project uh... for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Trello' but immediately self-corrects to 'Todoist'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The corrected target application is 'Todoist', and all other conditions (incorporating suggestions, leaving a comment, marking complete) remain unchanged. The final control intent is to execute the corrected task.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a uh, let's see, Trello um, sorry, make that Todoist project uh for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_b01a57e258413d63", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio contains a false start with 'Trello' followed by a self-correction to 'Todoist project'. The speaker uses 'uh' and 'let's see' as fillers, which are realistic. However, the correction is slightly more mechanical and the 'sorry' interjection feels a bit forced compared to the more natural 'or rather' in audio_02. The overall delivery is still credible but less smooth." } ] }, "is_default": false }, { "variant_id": "false_start_01_destination_city", "proposed_instruction": "I am going on a trip to Paris—I meant Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "long_pause_offsets": [], "source_sha256": "e3e03465e4ae3bb4d16fb244a96586079f30155a2025d9c3e76f6fabd5f74f31", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Paris' to 'Bangkok' within the same utterance. The correction does not affect the core task, which is to incorporate comments, leave a specific note, and mark the task complete. The original instruction's goal is fully preserved in the final intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_01_destination_city/74ffebbf4bb05fe4/speech.wav", "wer": 0.08974358974358974, "wav_sha256": "3ea2f1ba2e8bf42936199f3ed0f3a1f102ec4e4e28a41e7b7b67a6c38510ed8f", "seconds": 26.28, "spoken_utterances": [ "I am going on a trip to Paris—um... I meant Bangkok with some of my friends. We are managing its planning on a Todoist project uh, let me think... for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"... Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Paris' but immediately self-corrects to 'Bangkok'. This is a false start, and the final, corrected destination is 'Bangkok', which matches the original instruction. All other conditions, including the task in Todoist, the playlist, incorporating comments, leaving the specified comment, and marking it complete, are preserved and clearly stated. The final intent is fully aligned with the original instruction after the correction.", "heard_transcript": "I am going on a trip to Paris. I meant Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_17dfa89e6709014d", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "The speaker makes a location correction from 'Paris' to 'Bangkok' with a 'I meant' repair, which is realistic. However, the initial 'Paris' is spoken with a flat, almost robotic tone, and the 'let me think' pause feels slightly unnatural and overused. The correction is clear but the overall delivery lacks the conversational flow of the top-ranked clips." } ] }, "is_default": false }, { "variant_id": "false_start_06_completion_action", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it as done—I should have said mark it complete.", "long_pause_offsets": [], "source_sha256": "c1dd7231f9787516656c19e472bf275e7ff991381aa2ca96a5cf90e78fae5f89", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark it as done' but immediately corrects to 'mark it complete'. The final instruction clearly specifies marking the task as complete, which is equivalent in meaning to 'mark it as done' in this context. All other requirements—incorporating suggestions, leaving the specified comment—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/false_start/false_start_06_completion_action/11f9c00d87042aeb/speech.wav", "wer": 0.12345679012345678, "wav_sha256": "fed892e966625cac509e0a73f3109c396a2a6a54edea3db0bf5ec118b6611e89", "seconds": 30.52, "spoken_utterances": [ "I am going on a trip to um... Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it as done—ah... I should have said um... mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark it as done' but immediately self-corrects to 'mark it complete'. According to the 'false_start' type, the final corrected instruction must be applied. The final intent is to mark the task as complete, which is a valid and clear instruction. The correction is immediate and unambiguous, so the task should be executed with the final value.", "heard_transcript": "I am going on a trip to um Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it as done. I uh I should have said um mark it complete." }, "candidate_id": "wav_6c3498dc07253406", "naturalness": { "rank": 6, "mean_rank": 6.0, "singleton": false, "reasons": [ { "round": 1, "rank": 6, "reason": "This audio contains a self-correction from 'mark it as done' to 'mark it complete'. While the correction is present, the initial phrase 'mark it as done' is delivered with a flat, almost robotic intonation, and the correction is made with a long, unnatural pause and a drawn-out 'uh'. The overall delivery feels less spontaneous and more like a rehearsed or read-aloud task, reducing its realism compared to others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. I left out a detail: the task is about preparing a Spotify playlist.", "long_pause_offsets": [ 348 ], "source_sha256": "2389202e9fe8302b52f0f8591b4c9bb12640a5e46fb3b5f1c94a8478bd5b6eb0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The initial utterance specifies the full action sequence and the final comment. The followup adds the missing detail that the task is about preparing a Spotify playlist. This narrows the scope without contradicting or removing any prior requirement. All actions and conditions from the original instruction are preserved and augmented with the missing context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/addition/addition_03/84d45049616d8498/speech.wav", "wer": 0.0625, "wav_sha256": "f83a8300da14c7de411a36a4e5f6dbd2090aa9d8e77560e5582ef47170c1e25c", "seconds": 30.96, "spoken_utterances": [ "I am going on a trip to uh... Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"um... Thanks! Updated.\", and mark it complete. ", "I left out a detail: the task is about preparing a Spotify playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio starts with the original instruction and then adds a clarifying detail ('the task is about preparing a Spotify playlist') after a longer pause. This matches the 'addition' type, where the final goal combines the original task with the added information. The final intent is clear and executable, and the addition correctly narrows the scope of the task to Spotify playlist preparation, which was only implied in the original. No contradictions or errors remain.", "heard_transcript": "I am going on a trip to uh Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is, I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, um, \"Thanks! Updated.\", and mark it complete. I left out a detail, the task is about preparing a Spotify playlist." }, "candidate_id": "wav_1a4ae5e6d70b6307", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable self-correction. The initial mention of 'Bangkok' is slightly hesitant ('uh'), and the speaker later adds a missing detail ('I left out a detail...') in a way that sounds like a spontaneous realization. The pacing is conversational, with a natural pause and filler ('um') before the correction, and the correction itself is smoothly integrated into the ongoing thought. The prosody and rhythm are consistent with someone recalling and refining a task request in real time." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "I am going on a trip to Bangkok. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Oh, I forgot to mention, include the detail that it's with some of my friends.", "long_pause_offsets": [ 359 ], "source_sha256": "2e584ccad838f227f80503006b59f1c49698c2c4810160f28da66389405d6870", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds a missing detail about the trip being with friends. This is a supplementary constraint that does not contradict or remove any prior requirements. The core actions—incorporating suggestions, commenting, and marking complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/addition/addition_01/924f5f20b4e1268e/speech.wav", "wer": 0.024691358024691357, "wav_sha256": "a69a81470d526253256da7886cecf51541799eb2ad075058918ad9ab65ed2f7a", "seconds": 32.8, "spoken_utterances": [ "I am going on a trip to Bangkok. We are managing its planning on a uh... Todoist project ... for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"... Thanks! Updated.\", and mark it complete. ", "Oh, I forgot to mention, include the detail that it's with some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains the original instruction followed by an addition. The addition is a clarification to include the detail that the trip is with some of my friends. This is a valid addition that narrows the scope of the original task. The final intent is to execute the original task with the added detail, which matches the expected behavior for the 'addition' type.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Oh, I forgot to mention, include the detail that it's with some of my friends." }, "candidate_id": "wav_135f8a39635cc2ef", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a believable addition, but the delivery is slightly less natural. The speaker begins with a clear, fluent statement, then adds the missing detail ('Oh, I forgot to mention...') with a natural-sounding hesitation. However, the initial part of the utterance is more rehearsed and lacks the subtle hesitation seen in audio_02. The addition feels tacked on rather than seamlessly integrated, and the overall prosody is flatter, making it sound less spontaneous despite the correct pragmatic structure." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Oh, and one more thing: we're managing the planning on a Todoist project for it.", "long_pause_offsets": [ 355 ], "source_sha256": "15d5217aa2510b8ed24a81995b809e924055c27c2b2ba4bb6ace55fd3442f8ea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail that the planning is managed on a Todoist project, which was already implied in the original instruction. The core actions—incorporating suggestions, leaving a comment, and marking complete—are preserved. The additional information clarifies the context without changing the required actions." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to ... Bangkok with some of my friends. We are managing its planning. One of the tasks in it is about preparing a um... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. ", "Oh, and one more thing: we're managing the planning on a ... Todoist project for it." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_04", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project.. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Oh, I forgot to add, they have made some suggestions in comments.", "long_pause_offsets": [ 335 ], "source_sha256": "8cef282c9a777ba57f421246bf6b45c097f63a86335185930b1b6ef741424433", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The initial utterance already includes the full goal: incorporate suggestions, add the specified comment, and mark complete. The followup adds the missing context that the suggestions are in comments, which clarifies the source of the suggestions but does not change the required actions. The final goal remains the same." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to uh, let's see... Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project.. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. ", "Oh, I forgot to add, ... they have made some suggestions in comments." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_06_completion_action", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it incomplete. That should be mark it complete, not mark it incomplete.", "long_pause_offsets": [ 385 ], "source_sha256": "06f26b44b9122259987bb62a6cb6e9a3479ff6c6b922c60cb8f954f8fe5d59a0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The follow-up revision explicitly corrects the final action from 'mark it incomplete' to 'mark it complete', reversing the original instruction's requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_06_completion_action/e1c0f008f38be01e/speech.wav", "wer": 0.10975609756097561, "wav_sha256": "d52dd2602383e0cc41972573d37814d1029dbef3899bc0cf628d9bb987581189", "seconds": 34.88, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a um... Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"hmmm... Thanks! Updated.\", and mark it incomplete. ", "That should be mark it complete, not mark it incomplete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark it incomplete' but immediately corrects themselves with 'That should be mark it complete, not mark it incomplete.' This is a clear revision where the final audible intent is to mark the task as complete. The revision applies to the earlier instruction, and all other conditions (incorporating suggestions, leaving the comment) remain unchanged. Therefore, the final intent matches the original instruction.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a um, Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it incomplete. That should be mark it complete, not mark it incomplete." }, "candidate_id": "wav_2b51bbdf9b660f22", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent delivery. The speaker begins with a clear, continuous sentence, uses a plausible filler ('Hmm') before stating the comment, and then self-corrects the task action ('mark it incomplete' to 'mark it complete') with a believable, slightly hesitant tone. The correction is integrated smoothly into the utterance, and the overall prosody sounds like a real-time thought process without robotic pauses or unnatural word boundaries." } ] }, "is_default": false }, { "variant_id": "revision_05_comment_text", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks!\", and mark it complete. Change the comment to \"Thanks! Updated.\", not just \"Thanks!\".", "long_pause_offsets": [ 374 ], "source_sha256": "8629197f8a4d7bf506d51b3eb2e196284823b9046e50cd884998fc5c79613dc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revises the exact text of the comment to include \"Updated.\", replacing the initially requested \"Thanks!\". The rest of the task—incorporating suggestions and marking complete—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_05_comment_text/d171278b03bb9fc0/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "08046421eae804c4314094ef71deacbb4fc51e197e95968ed31ab36e3853ea2c", "seconds": 34.0, "spoken_utterances": [ "I am going on a trip to uh, let me think... Bangkok with some of my friends. We are managing its planning on a Todoist project uh... for it. One of the tasks in it is about preparing a um... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks!\", and mark it complete. ", "Change the comment to \"Thanks! Updated.\", not just \"Thanks!\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'let me think' but correctly states 'Bangkok' and continues the original instruction. The final instruction is a revision of the comment content from 'Thanks!' to 'Thanks! Updated.', which is clearly stated and preserved. The rest of the task (incorporate suggestions, mark complete) remains unchanged. The revision is applied to the earlier instruction, and all other conditions are preserved, satisfying the revision type criterion.", "heard_transcript": "I am going on a trip to uh, let me think, Bangkok with some of my friends. We are managing its planning on a Todoist project uh for it. One of the tasks in it is about preparing a um, Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "candidate_id": "wav_738370c0c45b85fc", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a very natural flow with a realistic 'let me think' pause and a clear self-correction of the comment text. The speaker's intonation and rhythm are consistent with spontaneous speech. The correction is immediate and integrated, and the overall delivery is credible, though slightly less fluid than audio_03 due to a minor stutter ('it's planning') and a slightly longer pause before the correction." } ] }, "is_default": true }, { "variant_id": "revision_03_task_app", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Google Tasks project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Make that Todoist, not Google Tasks.", "long_pause_offsets": [ 388 ], "source_sha256": "daa7d8db22b1c2309b9161b223cba37cadf2cb890f27780b53bc2dec2bbd1e5f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the project management app from Google Tasks to Todoist, while preserving all other requirements including the task context, playlist action, comment incorporation, final comment content, and completion marking." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_03_task_app/83610f589d766475/speech.wav", "wer": 0.13414634146341464, "wav_sha256": "d87042a3e1bbb51908e90db5cc889c72b21252bdca2f42709e794dff1dd75076", "seconds": 36.96, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a uh... Google Tasks project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"ah... Thanks! Updated.\", and mark it complete. ", "Make that uh, let's see... Todoist, not Google Tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Google Tasks' but immediately corrects it to 'Todoist' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final goal is to perform the same actions (incorporate comments, leave a specific comment, mark complete) but in the correct project, Todoist. The original instruction's intent is preserved and corrected, so the task is executed as revised.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a uh Google Tasks project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, uh, \"Thanks! Updated.\", and mark it complete. Make that uh, let's see, Todoist, not Google Tasks." }, "candidate_id": "wav_a81a2c3068c797ae", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker delivers a coherent request with a revision of the project name from 'Google Tasks' to 'Todoist'. The 'uh' and 'let's see' fillers are placed naturally. However, the correction is less integrated than in audio_01 and audio_03, as it's delivered as a separate sentence rather than a smooth self-repair within the original thought. The prosody is slightly less fluid, with a noticeable pause before the correction." } ] }, "is_default": false }, { "variant_id": "revision_04_music_app", "proposed_instruction": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Use Spotify, not Apple Music.", "long_pause_offsets": [ 387 ], "source_sha256": "1647825be2550cfb2419d358db6f5019f20a3d4c63955593f076b662b1663ea0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "unchecked", "reason": "현재 문장과 일치하는 semantic check 결과가 없습니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_04_music_app/9bb2ba63c3ead953/speech.wav", "wer": 0.2716049382716049, "wav_sha256": "8baf0bef0ad6402bf4388c1f6c83d5179f6efc37735f95c026c2f2ac61d37a24", "seconds": 36.92, "spoken_utterances": [ "I am going on a trip to hmmm... Bangkok with some of my friends. We are managing its planning on a Todoist project um, let me think... for it. One of the tasks in it is about preparing a ... Apple Music playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. ", "Use Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music playlist' but immediately corrects it to 'Use Spotify, not Apple Music.' This is a revision of the music platform. The final intent is to update the playlist using Spotify, incorporate comments, leave a final comment, and mark the task complete. All other conditions from the original instruction are preserved. The revision is clear and final, satisfying the revision type criterion.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Apple Music playlist. I've made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Use Spotify, not Apple Music." }, "candidate_id": "wav_c4df11af2301edde", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio contains a revision of the playlist platform from 'Apple Music' to 'Spotify'. The delivery includes natural fillers ('Hmm', 'let me think') and a clear correction. However, the speaker repeats the entire sentence about the playlist being made and shared, which sounds like a memory lapse or a stutter rather than a clean self-correction. This repetition slightly reduces the realism of the flow." } ] }, "is_default": false }, { "variant_id": "revision_01_destination_city", "proposed_instruction": "I am going on a trip to Paris with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Wait, the trip is to Bangkok, not Paris.", "long_pause_offsets": [ 381 ], "source_sha256": "8b36fb363be7527aedb6caa874a1a1cdafdc0d27df40ab88829d94ab045b6370", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the destination city from Paris to Bangkok, which is the only change. All other conditions—such as the app (Todoist), the task context (Spotify playlist), the actions (incorporate suggestions, leave comment, mark complete), and the comment content—are preserved as specified in the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/revision/revision_01_destination_city/28acc333adf410eb/speech.wav", "wer": 0.09302325581395349, "wav_sha256": "5df4804acfc551bab764da8476035e2ef255e558cd9506613f65cd5417716213", "seconds": 37.52, "spoken_utterances": [ "I am going on a trip to um, let's see... Paris with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"uh... Thanks! Updated.\", and mark it complete. ", "Wait, the trip is to uh, let me think... Bangkok, not Paris." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Paris but immediately corrects themselves to Bangkok after a pause and self-correction. According to the 'revision' type criterion, the final instruction after the correction must be applied. The final audible intent is to perform the same actions (incorporate comments, leave a comment, mark complete) but for the trip to Bangkok, which is the corrected destination. The core task remains unchanged except for the destination, which is now correctly specified as Bangkok.", "heard_transcript": "I am going on a trip to um, what's see, Paris with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, uh, \"Thanks! Updated.\", and mark it complete. Wait, the trip is to uh, let me think, Bangkok, not Paris." }, "candidate_id": "wav_8cb071bebba0646d", "naturalness": { "rank": 5, "mean_rank": 5.0, "singleton": false, "reasons": [ { "round": 1, "rank": 5, "reason": "This clip features a revision of the destination from 'Paris' to 'Bangkok'. The speaker uses fillers ('what's see', 'let me think') and a clear correction. However, the correction is delivered as a separate sentence with a noticeable pause, and the initial mention of 'Paris' is spoken with a flat, slightly robotic tone. The overall delivery feels less spontaneous and more like a rehearsed or read-aloud task, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a trip to Bangkok with some of my family. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Sorry, I meant some of my friends, not family.", "long_pause_offsets": [ 382 ], "source_sha256": "53d1bdb59fe78d5c8ce296f9b0dfc4185607fc19cdbea81a2b8510a070f78904", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'friends', which is the only change. All other conditions—trip location, app, task context, actions (incorporate suggestions, comment, mark complete)—are preserved. The correction is explicit and unambiguous." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a trip to Bangkok with some of my family. We are managing its planning on a Todoist project um, let's see... for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"... Thanks! Updated.\", and mark it complete. ", "Sorry, I meant some of my uh... friends, not family." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "0409946485d5f207a31c6702e890a6ed88a928ad814cd7ec0b2446880c8aaeeb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/cancellation/cancel_012/89bc6fa36114538b/speech.wav", "wer": 0.024691358024691357, "wav_sha256": "c84091e19f7ce2405bb7baa4e3bc4c28a9143480f4a61d70534c6078000f2cfb", "seconds": 32.2, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "Hold on, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with a clear cancellation: 'Hold on, actually, I've changed my mind. Don't do that.' This is a direct revocation of the prior instruction to incorporate comments, leave a message, and mark the task complete. The cancellation is unambiguous and final, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Hold on, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "c28d76b0fdf660b4354d52a42da03c14a5064e327218da9111fd7fe0ecdacea0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/ambiguous_halt/halt_016/872a858e911174e1/speech.wav", "wer": 0.02666666666666667, "wav_sha256": "3480dfdd6db7d18858264a74982e6fe9aad5c6ac69bba34f4de33041347102a6", "seconds": 30.12, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete.", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a second please,' which is a clear request to pause or wait. This is a final interjection that does not cancel or modify the task, nor does it execute it. The original instruction remains intact and unaltered, and the speaker is not requesting any change to the task goal. Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a hmm Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete. Just a second please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "986aa4e_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8dd202ab44ce34669f01b23e6a5c7a729adef536ed667ff52158798457abd8e2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/986aa4e_3/audio_natural/filler/original/de9db71a6c6f2972/speech.wav", "wer": 0.028169014084507043, "wav_sha256": "45a95dea1184ba6049cc2bf78d7522545e69eff31c5caab1d11bce663a58ddb6", "seconds": 26.88, "spoken_utterances": [ "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a ... Todoist project for it. One of the tasks in it is about preparing a hmmm... Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler word ('hmm') but immediately corrects and continues with the full original instruction. The final interjection is the complete task request, with no cancellation or pause. The filler does not alter the goal, so the execution intent matches the reference.", "heard_transcript": "I am going on a trip to Bangkok with some of my friends. We are managing its planning on a Todoist project for it. One of the tasks in it is about preparing a hmm Spotify playlist. I have made the playlist and shared it with others on the project. But they have made some suggestions in comments. Please incorporate them, leave a final comment, \"Thanks! Updated.\", and mark it complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address—sorry, I meant work address.", "long_pause_offsets": [], "source_sha256": "1eb4a28573ff6c30b91adb8b49162d8df788eaf2322260486fc0d0ced0337fda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects to 'work address' using 'sorry, I meant'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/false_start/false_start_03_destination_address/c8aa5de0a4aa9b2c/speech.wav", "wer": 0.05, "wav_sha256": "ec8cdaa445495de5fa60b77fc972ae6a6210403a1763a73c2089ab87e874bfe3", "seconds": 8.84, "spoken_utterances": [ "Buy everything on my amazon wishlist, and have it delivered to my home address—um... sorry, I meant ... work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task with 'home address' but immediately self-corrects to 'work address' after a brief hesitation. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Buy everything on my Amazon wishlist, and have it delivered to my home address. Um, sorry, I meant work address." }, "candidate_id": "wav_7787890ee206c708", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with a clear request, then after a brief pause and a soft 'um', smoothly corrects 'home address' to 'work address' with a clear 'sorry, I meant' phrase. The prosody is consistent, the correction is integrated into the flow without awkwardness, and the pacing feels like a genuine moment of recall." } ] }, "is_default": true }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list—no, make that wishlist, and have it delivered to my work address.", "long_pause_offsets": [], "source_sha256": "73d617eddaaf5f2de21907be290c9b4abbc99631d4761c8c726a2eef4bbcfee1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist'. The rest of the instruction, including the action, target list, and delivery address, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/false_start/false_start_02_source_collection/372ad93550d59015/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "e8d851d9c511655e51188d65ffaa0d996a0bbb2f599f35b43a451635f01d00ff", "seconds": 12.88, "spoken_utterances": [ "Buy everything on my amazon shopping list—uh... no, make that wishlist, ... and have it delivered to um, let me think... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wishlist'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The delivery address remains unchanged and correctly stated as 'my work address'. All constraints are preserved except the corrected target list, which now matches the reference.", "heard_transcript": "Buy everything on my Amazon shopping list, uh, no, make that wishlist, and have it delivered to, um, let me think, my work address." }, "candidate_id": "wav_302b61182faa6f1c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start with 'shopping list' corrected to 'wishlist', which is a plausible error. The speaker then pauses and says 'let me think' before stating the delivery address. While the correction is natural, the 'let me think' pause feels slightly longer and more deliberate than the spontaneous recall in audio_03, making it feel a bit less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Buy everything on my ebay—I should say amazon wishlist, and have it delivered to my work address.", "long_pause_offsets": [], "source_sha256": "5da27524b49c9e289e44f77e8cc776e1462d26310dbf27022a02afc9974dc2f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the action, target list, and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/false_start/false_start_01_app/193c65e70d4dc51c/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "f48ce53335a3aef87b92a4d65d589180542e191e5aa25d731f13d836135b2597", "seconds": 11.0, "spoken_utterances": [ "Buy everything on my uh... ebay—um... I should say amazon wishlist, and have it delivered to ah... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'I should say Amazon wishlist'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original mistaken value is discarded. The final intent matches the original instruction exactly, with only a harmless self-correction.", "heard_transcript": "Buy everything on my uh, eBay um, I should say Amazon wishlist, and have it delivered to uh, my work address." }, "candidate_id": "wav_00e2057a9cce8e99", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with 'my uh eBay', which is an unnatural and awkward phrasing for a wishlist. The correction to 'Amazon wishlist' is abrupt and lacks the smoothness of the other clips. The 'ah' pause before 'my work address' also sounds less like a natural hesitation and more like a filler, reducing the overall credibility of the spontaneous request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my work address.", "long_pause_offsets": [ 38 ], "source_sha256": "bf984e7b97174bc7d9866e13d47d6ed33ab6fc91ec3d1af2286d7c03769bbc20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement without changing any other aspect of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/addition/addition_01/0fd24a071748033b/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "57dc7788eca64e91308235c1831677e3a9e4063a42c1ccf9d2708129fa1d8632", "seconds": 11.0, "spoken_utterances": [ "Buy everything on my um... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to buy everything on the Amazon wishlist is followed by an addition specifying delivery to the work address. The addition completes the original request by adding a missing constraint. The final intent matches the original instruction exactly, so the addition type criterion is satisfied.", "heard_transcript": "Buy everything on my um Amazon wishlist. Oh, I forgot to mention, have it delivered to my work address." }, "candidate_id": "wav_847e691e85aca228", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um'), then delivers the core request fluently. The follow-up 'Oh, I forgot to mention...' is a highly plausible, spontaneous self-correction with a natural intonation shift and clear reference back to the initial request. The delivery is continuous and the repair feels genuinely integrated into the thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to. I left out a detail: deliver it to my work address.", "long_pause_offsets": [ 64 ], "source_sha256": "e95ea5d709ee83ba10cd885f20dd766784b20dbd18c3fb8bbaf7b59bd5e069b1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address, which was omitted in the initial utterance. The final goal combines the initial request with the added constraint, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/addition/addition_02/8cd65f1eca3f3bb0/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "728a49bc149386d63c96ac53733c21f1e909203f86325ee42926967111b0f405", "seconds": 13.12, "spoken_utterances": [ "Buy everything on my ... amazon wishlist, and have it delivered to. ", "I left out a detail: deliver it to my let's see... work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to buy everything on the Amazon wishlist and deliver it to the work address. They then pause and add a clarifying phrase, 'I left out a detail,' before reiterating the delivery address as 'work address.' This is a clear addition that specifies the delivery location, which was already part of the original instruction. The final intent is consistent with the original instruction, with the addition of a clarification that does not change the core task. Therefore, the addition type is satisfied.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to, I left out a detail, deliver it to my, let's see, work address." }, "candidate_id": "wav_b688bb25e63e9fe7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then inserts a long, unnatural pause and a hesitant filler ('Let's see...') before completing the delivery address. This pause feels more like a recording glitch or an awkward edit than a natural thought process. The self-correction is less integrated and the overall flow is less smooth compared to audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Buy everything on my ebay wishlist, and have it delivered to my work address. Sorry, I meant Amazon, not eBay.", "long_pause_offsets": [ 78 ], "source_sha256": "01d921799eecb3bc0c46e496c73f1b7a374832b2e9ee1000321fa90930311379", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the platform from eBay to Amazon, while retaining the action, scope (everything on the wishlist), and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/revision/revision_01_app/6bb5e10a8fa68305/speech.wav", "wer": 0.0, "wav_sha256": "8ea45b86ae2ab3c3ba073a92ddbd801ba8f5734c68ef3f8cf5b1b5b076816eb9", "seconds": 14.56, "spoken_utterances": [ "Buy everything on my um, let me think... ebay wishlist, and have it delivered to uh, let me think... my work address. ", "Sorry, I meant um... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects to 'Amazon not eBay' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with all constraints preserved.", "heard_transcript": "Buy everything on my um let me think eBay wishlist and have it delivered to uh let me think my work address. Sorry, I meant um Amazon not eBay." }, "candidate_id": "wav_3498733fa222a821", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker starts with a clear request, hesitates with 'let me think' before specifying 'ebay wishlist', then pauses again before correcting 'sorry, I meant Amazon not eBay'. The self-correction is immediate, the intonation reflects genuine recollection, and the filler 'let me think' is used at natural points. The prosody is consistent and the delivery sounds like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list, and have it delivered to my work address. Wait, the source should be wishlist, not shopping list.", "long_pause_offsets": [ 85 ], "source_sha256": "6cb54b78c00d4e670d1d240a6a86b90928f8de73d0b6df2b7f156c5a57b46908", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'shopping list' to 'wishlist' while retaining the action, app, and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/revision/revision_02_source_collection/0173efe60e65f81c/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "6372366bfd8e2bd3e8d8c7cf21640034c595687510fac55153196a2509632f69", "seconds": 13.72, "spoken_utterances": [ "Buy everything on my uh, let me think... amazon shopping list, ... and have it delivered to ... my work address. ", "Wait, the source should be wishlist, not shopping list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects themselves to 'wishlist' after a pause and the word 'Wait'. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to buy from the wishlist, which matches the original reference instruction. All other conditions, including delivery to the work address, remain unchanged.", "heard_transcript": "Buy everything on my uh let me think Amazon shopping list and have it delivered to my work address. Wait, the source should be wishlist, not shopping list." }, "candidate_id": "wav_18affc7a65daf013", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a self-correction from 'shopping list' to 'wishlist'. The speaker uses 'wait' as a natural cue for a change of mind, and the correction is integrated smoothly. The 'let me think' filler is used appropriately. However, the initial request is slightly less fluent than audio_02, and the correction feels a bit more abrupt, making it slightly less realistic in flow." } ] }, "is_default": false }, { "variant_id": "revision_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address. Make that delivered to my work address, not home address.", "long_pause_offsets": [ 80 ], "source_sha256": "649cd0d6a1c523cfbcb37693b4e80d243f2cd0b700b2f7cb4b73f9cf805f681e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the delivery address from 'home' to 'work', explicitly rejecting the previous value. The core action and target list remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/revision/revision_03_destination_address/4be0586795776d54/speech.wav", "wer": 0.1724137931034483, "wav_sha256": "aca9dc0a2aab6005f96a1039c8c82c7bb8d58d40a5a60fc22f9ee5b28707b580", "seconds": 13.4, "spoken_utterances": [ "Buy everything on my um... amazon wishlist, uh... and have it delivered to my home address. ", "Make that delivered to my let me think... work address, not home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests delivery to the 'home address' but then corrects this to 'work address, not home address' after a pause and self-interruption. This constitutes a revision where the final intent clearly overrides the initial value. The final instruction matches the original reference, so the revision is correctly applied.", "heard_transcript": "Buy everything on my um Amazon wishlist uh and have it delivered to my home address. Make that delivered to my let me think, work address, not home address." }, "candidate_id": "wav_8948ba2368feb176", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision but the flow is less natural. The speaker starts with 'buy everything on my Amazon wishlist' and then corrects 'make that delivered to my work address not home address'. The correction is a complete rephrasing rather than a self-repair of a specific word, and the transition feels more like a new instruction than a spontaneous correction. The 'let me think' filler is used awkwardly mid-sentence, disrupting the natural rhythm. The overall delivery is less fluid and more like a rehearsed or edited sequence." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "d53c2a9ce1099026366de4490886414d0b88a350a041c2b6dee7f1c7a13a3fdd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/cancellation/cancel_004/485f7daa8d515008/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c0e1a6199015856556272632d383c7ca2f620f95d39b55c307a512051f7486e5", "seconds": 11.68, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address.", "Wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete request to buy items from the Amazon wishlist and deliver them to the work address. Immediately after, they say 'Wait, cancel that request,' which is a clear and final directive to revoke the entire task. This matches the 'cancellation' type criterion, where the recording must clearly revoke the task. The final interjection is not ambiguous and explicitly cancels the prior instruction.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um let's see my work address. Wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "183e323ef009a503a6509926bd92dbcaf2476d152e302334a4e57d459786536f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/ambiguous_halt/halt_037/2a74e1633fb98fb3/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "9c426ad7923bd8bb7a98815baac22657869c0a471c63691c7f9e73aba4ea157c", "seconds": 12.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Pause for a moment' repeated twice, which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial instruction is not completed, but the final control request is audible and unambiguous.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um, let's see, my work address. Pause for a moment. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "142e3a3ae9c8f7eb2dd34da96027e92b3bc326ef0777a7b0f22c8c83c087790c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_1/audio_natural/filler/original/d8470961b05d977c/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "25e3683329caccbc31e9f3db45f11814df12b8cdb83ae337b1093d0a25e54a00", "seconds": 7.76, "spoken_utterances": [ "Buy everything on my amazon wishlist, ... and have it delivered to um, let's see... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um', 'let's see') and a pause but does not change the core request. The final intent remains to execute the original instruction. The filler does not alter the goal, so the execution criterion is satisfied.", "heard_transcript": "Buy everything on my Amazon wishlist and have it delivered to um, let's see, my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my work address—no, I mean home address.", "long_pause_offsets": [], "source_sha256": "b405bd651cc563e428aab1b3f55e81699e594b5e15ed4ba6aad0548c51b8a943", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately corrects it to 'home address'. The final intended destination is 'home address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/false_start/false_start_03_destination_address/48b6bd6895971994/speech.wav", "wer": 0.0, "wav_sha256": "9a3cd4d5998cfc34cbb5fcd99e56d4bfe52b88411681e6f734634c99ed7762d5", "seconds": 11.08, "spoken_utterances": [ "Buy everything on my let's see... amazon wishlist, and have it delivered to my work address—uh... no, I mean home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'work address' but immediately self-corrects to 'home address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Buy everything on my let's see, Amazon wishlist, and have it delivered to my work address. Uh, no, I mean home address." }, "candidate_id": "wav_798f618690404331", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'Let's see' as a filler, then states the initial request, pauses, and clearly corrects 'work address' to 'home address' with a believable 'uh, no, I mean' repair. The intonation and pacing of the correction sound spontaneous and integrated into the flow of thought, making it the most realistic false start." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Buy everything on my Etsy—I should say amazon wishlist, and have it delivered to my home address.", "long_pause_offsets": [], "source_sha256": "6517afbac0228e2495be9afb8f28ff46868d221bfacad442e18b48dbfe2ec030", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon wishlist' within the same utterance, preserving the action, target list, and delivery instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/false_start/false_start_01_app/0eb9e311693edf49/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "1a580d9caeb4d4e917baea1d306dc02b5ce0e5c877451e7f967fb8f51d147368", "seconds": 10.52, "spoken_utterances": [ "Buy everything on my Etsy—ah... I should say ... amazon wishlist, and have it delivered to um, let's see... my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon' with 'I should say'. The final intent is clear and matches the original instruction. The 'false_start' type requires applying the correction, which has been done. The delivery address remains unchanged and correct.", "heard_transcript": "Buy everything on my Etsy uh I should say Amazon wishlist and have it delivered to um let's see my home address." }, "candidate_id": "wav_5bdd02f9549c31e7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a false start with 'Etsy' corrected to 'Amazon Wish List'. The correction is clear and the speaker continues smoothly. However, the initial error is less plausible than 'work address' vs 'home address' because 'Etsy' and 'Amazon Wish List' are different platforms, making the correction feel slightly more forced. The 'let's see' pause is also less integrated than in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list—sorry, I meant wishlist, and have it delivered to my home address.", "long_pause_offsets": [], "source_sha256": "a913695151703b1b542cda7e9e4ac3acaca3ff63f18f1c31248dace9d8abee4b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'shopping list' to 'wishlist' within the same utterance, preserving the action, target list, and delivery instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/false_start/false_start_02_source_collection/3146f74873d76c4e/speech.wav", "wer": 0.16, "wav_sha256": "86a74d63fa7759d60706b386807065d02d840a85e45e0162e349aa33e99776a5", "seconds": 13.12, "spoken_utterances": [ "Buy everything on my um, let me think... amazon shopping list—ah... sorry, I meant wishlist, and have it delivered to um... my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon shopping list' but immediately self-corrects to 'wishlist'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the delivery address, remains unchanged and clear. Therefore, the final audible intent matches the original instruction after the correction.", "heard_transcript": "Buy everything on my um, let me think, Amazon shopping list uh, sorry, I meant wishlist, and have it delivered to um, my home address." }, "candidate_id": "wav_7f688c705626bb13", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'Amazon shopping list' corrected to 'Wish List'. The correction is clear, but the initial phrase 'Amazon shopping list' is less likely to be confused with 'Wish List' than the other options, making the error less believable. The 'let me think' filler is also more artificial-sounding than the 'let's see' in audio_03, and the overall delivery feels slightly more hesitant and less fluid." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my home address.", "long_pause_offsets": [ 38 ], "source_sha256": "11f5dcd6323bb4cb0194c885ffd7ad961f00f7dddac6281e9c4da534605cea0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery instruction, which combines with the initial request to form the complete goal. The original instruction already included both actions, so the spoken variant completes the intended task." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/addition/addition_01/7109bedf10c61675/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "e8afeb00091d2c70f63d4d02c2c8130bbd9d8a8934e63fed973ea4a90c851d7b", "seconds": 11.96, "spoken_utterances": [ "Buy everything on my uh, let's see... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to uh... my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction to buy everything on the Amazon wishlist is followed by an addition specifying delivery to the home address. The addition completes the original instruction by adding a missing delivery detail. The final intent matches the reference instruction, and the addition type criterion is satisfied.", "heard_transcript": "Buy everything on my uh let's see Amazon wishlist. Oh, I forgot to mention, have it delivered to uh my home address." }, "candidate_id": "wav_9b2e64c9adfc5659", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Buy everything, and have it delivered to my home address. I left out a detail: buy everything on my amazon wishlist.", "long_pause_offsets": [ 58 ], "source_sha256": "4a1ea65f8095c95dd4abd9bd7fbf4beaeec4cb0d9ce9b46d258f2534c160ecbe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'on my amazon wishlist' to the initial instruction, which already included the delivery address. The final goal combines both requirements without contradiction." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy everything, and have it delivered to uh... my home address. ", "I left out a detail: buy everything on my ... amazon wishlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my work address. Change the destination to home address, not work address.", "long_pause_offsets": [ 80 ], "source_sha256": "3e2e131d565686e8565f4022b4cd1ca5a410ecee04e979da1e00c44d7e8f64c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the destination from work address to home address, aligning with the original instruction. The action and item scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/revision/revision_03_destination_address/003957c77512bc12/speech.wav", "wer": 0.10714285714285714, "wav_sha256": "540397c2d9ae4904d751410f55eec1ef3b5ac4957a3dd7e452a10f62e18bf13e", "seconds": 17.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, hmmm... and have it delivered to uh... my work address. ", "Change the destination to um, let's see... home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests delivery to the 'work address' but immediately corrects this by saying 'Change the destination to... home address, not work address.' This constitutes a revision where the final intent is to deliver to the home address. The correction is clear and unambiguous, and the rest of the instruction (buying everything on the wishlist) remains unchanged. Therefore, the final audible intent matches the original reference instruction.", "heard_transcript": "Buy everything on my Amazon wishlist, and have it delivered to uh my work address. Change the destination to uh let's see, home address, not work address." }, "candidate_id": "wav_0e46887a4169c852", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, pauses thoughtfully with a believable 'hmm', then smoothly revises the delivery address from 'work address' to 'home address' with a self-correcting tone and intonation. The revision is integrated into the same utterance, and the prosody reflects genuine reconsideration rather than a mechanical edit." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Buy everything on my ebay wishlist, and have it delivered to my home address. Wait, use Amazon instead of eBay.", "long_pause_offsets": [ 78 ], "source_sha256": "42fc2eb27d79d09fac00bb2da13d5db7498eb9e828da5a55cfbf5ab9f379448b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the platform from eBay to Amazon, preserving the action, scope (everything on the wishlist), and delivery destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/revision/revision_01_app/d121639023bd0046/speech.wav", "wer": 0.0, "wav_sha256": "b61383ddd341701d8ff866863122c18898406ecf6154a0146d29f4d3ea17d330", "seconds": 10.96, "spoken_utterances": [ "Buy everything on my ebay wishlist, ... and have it delivered to my home address. ", "Wait, use ... Amazon instead of eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay wishlist' but immediately corrects to 'Amazon instead of eBay' after a short pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly, with all constraints preserved.", "heard_transcript": "Buy everything on my eBay wishlist, and have it delivered to my home address. Wait, use Amazon instead of eBay." }, "candidate_id": "wav_7aeb77f8dd8561c3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is fluent and clear, but the revision ('Wait, use Amazon instead of eBay') is delivered as a separate, abrupt utterance with a sharp, almost robotic shift in tone. The pause before the revision is too short and lacks the natural hesitation or intonation of a spontaneous change of mind, making it sound more like a post-hoc correction than a genuine in-the-moment revision." } ] }, "is_default": false }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list, and have it delivered to my home address. I misspoke about the source collection; it should be wishlist, not shopping list.", "long_pause_offsets": [ 85 ], "source_sha256": "f3aae803360e2a86aa03c9f55188449284ffb3b0dc9cc0115a1462a2df76b4f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source collection from 'shopping list' to 'wishlist', while retaining the action and delivery destination." }, "audio": { "status": "pending", "spoken_utterances": [ "Buy everything on my ... amazon shopping list, and have it delivered to my home address. ", "I misspoke about the source collection; it should be let's see... wishlist, not shopping list." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "44d86adb00f6cd0d99476de15faaefc8c5dd290175dc72c8ddcfa22279270d18", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/cancellation/cancel_004/b4add5a28431e5e5/speech.wav", "wer": 0.0, "wav_sha256": "834a18dc2887237257bd8eca666c4e0fdd36d2a2fdd6b369d02bcbaa3dadecae", "seconds": 13.28, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address.", "Oh, wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task, then after a pause and self-correction ('let me think'), proceeds to state the full request. Immediately after, they say 'Oh wait, cancel that request,' which is a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. Therefore, the final control intent is cancellation, matching the type-specific criterion.", "heard_transcript": "Buy everything on my um, let me think, Amazon wishlist, let's see, and have it delivered to my home address. Oh wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "df5d92035e66750d2992274a7d5abd682f6e4c953244b0c286f49db1aa117408", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/ambiguous_halt/halt_010/320d1f64fd69fb1a/speech.wav", "wer": 0.0, "wav_sha256": "092e35f8a2452a9f043a582d085878eac98defd9e753d5eecf2f51f04b32f398", "seconds": 12.32, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then says 'Hold on a second' as the final interjection, which is a clear request to pause. The initial hesitation and 'let me think' are fillers within the ongoing task and do not constitute a final control command. The final audible intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Buy everything on my um, let me think. Amazon wishlist, let's see. And have it delivered to my home address. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9eebc7777bf038dc036d893e98b4d465df1c9c900fbdd088281614e08e7d9ab2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_2/audio_natural/filler/original/ec203ccf3ed6a2c5/speech.wav", "wer": 0.0, "wav_sha256": "d9e678a5072b8a1fa70e9cb45819841d5336d41f9dbb6c823a49406474a1c10c", "seconds": 9.04, "spoken_utterances": [ "Buy everything on my um, let me think... amazon wishlist, let's see... and have it delivered to my home address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation markers ('um', 'let me think', 'let's see') but does not change the core request. The final intent remains consistent with the original instruction. The filler words do not alter the goal, so the task should be executed as stated.", "heard_transcript": "Buy everything on my um, let me think, Amazon wishlist, let's see, and have it delivered to my home address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address—sorry, I meant work address.", "long_pause_offsets": [], "source_sha256": "1eb4a28573ff6c30b91adb8b49162d8df788eaf2322260486fc0d0ced0337fda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately corrects to 'work address' using 'sorry, I meant'. The final intended destination is 'work address', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/false_start/false_start_03_destination_address/f2b03449280ca906/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "febf19dc50382026282b711cc75520776fd660148a82116938229c8b5fcffe3f", "seconds": 8.16, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to ... my home address—uh... sorry, I meant work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'home address' but immediately self-corrects to 'work address'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to my home address. Uh, sorry, I meant work address." }, "candidate_id": "wav_f754e953e938ed24", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker begins with a clear request, then after a brief pause and a soft 'uh', corrects 'home address' to 'work address' with a believable 'sorry, I meant' phrase. The prosody includes a slight pitch drop on 'sorry' and a quick, natural recovery, matching how a real person would correct a slip of the tongue mid-sentence. The pacing and intonation are consistent with genuine hesitation and recovery." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Buy everything on my Etsy—I should say amazon wishlist, and have it delivered to my work address.", "long_pause_offsets": [], "source_sha256": "22c21491949f5614502afdd65da7dce40c596a7cc69ac8e031803161733748bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Etsy' to 'Amazon wishlist' within the same utterance, preserving the action, target list, and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/false_start/false_start_01_app/52f095749d84db68/speech.wav", "wer": 0.0, "wav_sha256": "cd4b9fa6c3cb6dc650a7492826b9fb62c9ef843a6c2d1349673972a9a55952eb", "seconds": 9.04, "spoken_utterances": [ "Buy everything on my Etsy—um... I should say amazon wishlist, ... and have it delivered to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Etsy' but immediately self-corrects to 'Amazon wishlist'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('Etsy') is discarded. The rest of the instruction, including the delivery address, remains unchanged and clear. Therefore, the final audible intent matches the original instruction after the correction.", "heard_transcript": "Buy everything on my Etsy, um, I should say Amazon wishlist, and have it delivered to my work address." }, "candidate_id": "wav_d77b1d8116fd2da5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'Etsy' followed by a self-correction to 'Amazon wishlist'. The correction is immediate and the speaker continues smoothly. However, the initial error is more abrupt and the transition feels slightly less fluid than in audio_03. The 'um' filler is used, but the overall delivery is a bit more hesitant and less conversational than the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list—or rather, wishlist, and have it delivered to my work address.", "long_pause_offsets": [], "source_sha256": "1d8fcd1cae460d5c188cefee3c9d6f3c8db8e4759c26acfac97db2d85f1b4342", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist'. The correction is clear and replaces the incorrect term. The delivery address remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/false_start/false_start_02_source_collection/32688fa06bff6ce0/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "08d6415e35e0fb157185ddcd0ab7dea6250b232175e08f505f42e5cfdcb11886", "seconds": 9.92, "spoken_utterances": [ "Buy everything on my um... amazon shopping list—hmmm... or rather, wishlist, um... and have it delivered to my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wishlist'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction, including the delivery address, remains unchanged and clear. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Buy everything on my um Amazon shopping list, hmm or rather wishlist um and have it delivered to my work address." }, "candidate_id": "wav_d47840583daa6fbf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains multiple fillers ('um', 'hmm') and a self-correction from 'shopping list' to 'wishlist'. While it includes realistic elements, the delivery feels more hesitant and less fluent. The pauses are longer and more frequent, and the correction is less integrated into the flow. The speaker seems to be struggling more to formulate the request, making it sound less spontaneous and more like a rehearsed or uncertain attempt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Buy everything on my amazon wishlist. Oh, I forgot to mention, have it delivered to my work address.", "long_pause_offsets": [ 38 ], "source_sha256": "bf984e7b97174bc7d9866e13d47d6ed33ab6fc91ec3d1af2286d7c03769bbc20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address requirement without changing any other aspect of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/addition/addition_01/736713db9db2bd9b/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "b44eca61c8118c4aad03a2b07bdcada14b2dfbb0509c2058203151ebd8abe595", "seconds": 11.88, "spoken_utterances": [ "Buy everything on my uh, let me think... amazon wishlist. ", "Oh, I forgot to mention, have it delivered to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Buy everything on my Amazon wishlist' is followed by an addition 'have it delivered to my work address' after a longer pause and self-correction. This matches the 'addition' type criterion, where the final goal combines the original request with the added constraint. The final audible intent fully satisfies the original instruction and adds the delivery address, so the task is complete and correct.", "heard_transcript": "Buy everything on my uh let me think Amazon wishlist. Oh, I forgot to mention, have it delivered to my work address." }, "candidate_id": "wav_29ae69965f317008", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a plausible 'let me think' filler, indicating genuine thought. The addition of the delivery address is smoothly integrated with a believable 'Oh, I forgot to mention' phrase, and the prosody remains consistent and conversational throughout. The pacing and intonation suggest a spontaneous, real-time formulation of the request." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to. Oh, and one more thing: deliver it to my work address.", "long_pause_offsets": [ 64 ], "source_sha256": "0e8b3ea8cf52a44bb8ae010da6400c3b65512f484c827828eb42638b8ea70a02", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing delivery address, completing the initial incomplete instruction. The action, target list, and delivery intent are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/addition/addition_02/9106b207d52c5fae/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "bbef0b0209daf64d5df5986804983638648c94ca1973ab5125cdcf6a68344586", "seconds": 11.96, "spoken_utterances": [ "Buy everything on my uh... amazon wishlist, um... and have it delivered to. ", "Oh, and one more thing: deliver it to ... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to buy everything on the Amazon wishlist and deliver it to the work address. The speaker then adds 'oh and one more thing, deliver it to my work address,' which reinforces the delivery address. Since the addition is redundant and does not contradict or change the original instruction, the final intent matches the original instruction. The addition type is satisfied as the final goal is the combination of the initial and added information, which aligns with the original instruction.", "heard_transcript": "Buy everything on my uh Amazon wishlist um and have it delivered to oh and one more thing deliver it to my work address." }, "candidate_id": "wav_1be8e56939849553", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the initial request is delivered naturally with a 'uh' and 'um' filler, the addition of the delivery address is less integrated. The phrase 'Oh, and one more thing' is a common conversational device, but the delivery of the address feels slightly more abrupt and less fluid than in audio_01. The prosody is slightly more hesitant and less continuous, making the overall flow feel a bit more rehearsed or less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_destination_address", "proposed_instruction": "Buy everything on my amazon wishlist, and have it delivered to my home address. Wait, deliver it to my home address, not work address.", "long_pause_offsets": [ 80 ], "source_sha256": "7231433dfc4cca5500ae4a88aba7a1cc4a366299dddb15be62a478c716c1d343", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly confirms 'deliver it to my home address, not work address', which corrects the initial delivery location and preserves the original action and target list." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/revision/revision_03_destination_address/dcd9d259b293c1c0/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "a26d551e5a6f3544f7082048a47265d51b315234fe1e9495cd3403629f80af70", "seconds": 11.88, "spoken_utterances": [ "Buy everything on my um... amazon wishlist, uh... and have it delivered to um... my home address. ", "Wait, deliver it to my home address, not work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests delivery to the 'work address' but immediately corrects this to 'home address' after a pause and the word 'Wait'. This constitutes a revision where the final intent is to deliver to the home address. The correction is clear and unambiguous, and the rest of the instruction remains consistent. Therefore, the final intent matches the revised instruction and satisfies the revision type criterion.", "heard_transcript": "Buy everything on my um Amazon wishlist uh and have it delivered to um my home address. Wait, deliver it to my home address not work address." }, "candidate_id": "wav_e6b24463a8a5f6de", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous prosody with believable hesitation ('um') and a clear, spontaneous self-correction ('Wait, deliver it to my home address not work address'). The correction is integrated into the flow of the utterance, with a natural shift in intonation and emphasis, making it sound like a genuine moment of reconsideration during a real task request." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Buy everything on my amazon shopping list, and have it delivered to my work address. Make that wishlist, not shopping list.", "long_pause_offsets": [ 85 ], "source_sha256": "ddbd699ac878c91ec9aca9ecbf18124a0096718f5bd10cd1b55fa362e713c20c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the list type from 'shopping list' to 'wishlist', preserving the action, app, and delivery address." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/revision/revision_02_source_collection/437684d250d5bde7/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "92edd3cb594793c1b92089f657b2addd44b03a4c2a1d9465301631af55344137", "seconds": 14.16, "spoken_utterances": [ "Buy everything on my amazon shopping list, ... and have it delivered to hmmm... my work address. ", "Make that ... wishlist, not shopping list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately corrects it to 'wishlist' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to buy everything on the wishlist, not the shopping list, which matches the original reference instruction. The delivery address remains unchanged and correct.", "heard_transcript": "Buy everything on my Amazon shopping list, and have it delivered to, um, my work address. Make that, wishlist, not shopping list." }, "candidate_id": "wav_9e8e3e657d728b7e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision from 'shopping list' to 'wishlist', but the correction is delivered as a separate, slightly abrupt sentence ('Make that wishlist not shopping list'). While the initial request is fluent and the filler 'hmm' is natural, the revision lacks the smooth integration seen in audio_01, sounding more like a post-hoc correction than a spontaneous thought process." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Buy everything on my ebay wishlist, and have it delivered to my work address. Change amazon to ebay.", "long_pause_offsets": [ 78 ], "source_sha256": "4a377dba35a6ef4a240aa0c8c7aa0ca8482badd2e39b89b609a792d9f2eb1e58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'amazon' to 'ebay', aligning the final goal with the spoken instruction. The delivery address remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/revision/revision_01_app/c8008e99eace0866/speech.wav", "wer": 0.125, "wav_sha256": "b63a37c54a8c284daaf58c33c6de67043148cfeca226264fd4f00dabad65d47e", "seconds": 14.48, "spoken_utterances": [ "Buy everything on my um... ebay wishlist, and have it delivered to ah... my work address. ", "Change uh, let me think... amazon to ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon' but immediately corrects to 'eBay' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to execute the task on eBay, which matches the original instruction's structure but with the correct target app. No other constraints are changed.", "heard_transcript": "Buy everything on my um, eBay wishlist, and have it delivered to uh, my work address. Change, uh, let me think, Amazon to eBay." }, "candidate_id": "wav_899a0a7845e83564", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains multiple fillers ('um', 'uh', 'let me think') and a revision ('change uh let me think amazon to ebay'), but the revision is delivered in a disjointed, hesitant manner. The phrase 'let me think' is followed by a long pause and a non-fluent correction, making the self-correction sound less spontaneous and more like a forced or rehearsed edit, reducing its realism compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_031", "hide_variant_original": true, "source_sha256": "69a335cec2a465ac7546723087bb016e2fc4dc7e44feb50018991455f6297246", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/cancellation/cancel_031/7302a100c5e22522/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "a13991068f2eadd46491eeb665fbb7bb04a7c5db80f4aed30da0fc9e82cf9ebd", "seconds": 14.68, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address.", "Oh, wait, there's no need to do that anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task request, then immediately follows with 'Oh wait, there's no need to do that anymore. Cancel it.' This constitutes a clear and final cancellation of the entire task. The cancellation is not ambiguous and is not followed by any resumption or modification of the original request. The final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address. Oh wait, there's no need to do that anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "70565641b539cc3f7e587a762ec575d62c69b2eef3faa519cf9d7f71f3211cc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/ambiguous_halt/halt_027/aa0b45f1b6ebf651/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "70695cd2f2c7a8ccc1d02bd923218725550fc1e51ec8dc86b48118ccdedfd1b7", "seconds": 13.44, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then clearly requests a pause with 'Hang on, I need a moment.' This constitutes a final interjection that is a direct request to hold the task, matching the 'ambiguous_halt' type. The original instruction is fully audible and not canceled or revised.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9871968_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4c3629a904f6d58950119594d568a3814bb8c0d7dcb61ccb658dce73fd1de747", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9871968_3/audio_natural/filler/original/7366d33894d9e32d/speech.wav", "wer": 0.1875, "wav_sha256": "207b97a148b1b2814ecaea78337b4c4e48504740e9a65a0d9ff599cfbbd2ca4a", "seconds": 9.6, "spoken_utterances": [ "Buy everything on my amazon wishlist, um... and have it delivered to hmmm... my work address." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or cancellations. The filler type allows for hesitation and backchannels without altering the task goal.", "heard_transcript": "Buy everything on my Amazon wishlist, um, and have it delivered to, um, my work address." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family—no, I meant friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "d09673a3b2a9a93d503a12a84bfb53d96eb82dc50a75e33ac2d79761f3743f67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance. The final intended group is 'friends', matching the original instruction. All other elements, including the project creation and task completion request, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/false_start/false_start_02_companion_group/bb9e9e351a45fb24/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "468d2a33f3ca93d828b116e55334701438a2512e2776fd3b00d164238c7220e8", "seconds": 13.36, "spoken_utterances": [ "I am going on a camping trip uh... with some of my family—um... no, I meant friends. let me think... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. This is a false start, and the final, corrected intent is to create a Todoist project and mark its tasks as complete. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "I am going on a camping trip uh with some of my family. Um, no, I meant friends. Let me think. I've created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_e47bfc0bac02d184", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false start. The speaker begins with 'family', hesitates with 'um', self-corrects to 'friends' with a believable 'no, I meant friends' and a brief 'let me think' filler, all while maintaining a consistent, conversational pace and intonation. The correction feels like a genuine moment of recall, and the overall delivery is fluid and unforced." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks—sorry, I should have said Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "ea2abd1a5eafa413ab49e669cc7e139bd94e2596dff95bcce3cc308aedfa4ded", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Google Tasks' to 'Todoist' within the same utterance. The final intended app is Todoist, and all other requirements (creating a project, doing the tasks, marking them complete) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/false_start/false_start_03_app/52d9ed4bbc4e5a57/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "4afc739f48b925c0024c36e534f28b9fa207036e6e53343d000662f915db3d33", "seconds": 15.72, "spoken_utterances": [ "I am going on a camping trip with some of my friends. I have created a new Google Tasks—um... sorry, I should have said Todoist project let's see... for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks' but immediately self-corrects to 'Todoist project'. According to the 'false_start' type criterion, the final corrected value ('Todoist project') should be executed, and the initial incorrect value ('Google Tasks') should be discarded. The final instruction to 'do them for me and mark them all complete' is preserved. The final intent matches the original instruction with the corrected app name.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Google Tasks, um, sorry, I should have said Todoist project. Let's see. For managing its tasks, please do them for me and mark them all complete." }, "candidate_id": "wav_f601fa42e31334a6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear self-correction from 'Google Tasks' to 'Todoist project', which is a plausible error. The speaker uses 'sorry' and 'let's see' as natural repair markers. However, the initial error is a different application name, which is less likely than a simple word swap, and the 'let's see' pause feels slightly more like a scripted transition than a spontaneous thought process, making it slightly less realistic than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and check them all off—wait, make that mark them all complete.", "long_pause_offsets": [], "source_sha256": "998bb6700a21406f0ea68ac44feaed407ee92b113832ac517e80dd7cb0af2855", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately corrects it to 'mark them all complete'. Both phrases are functionally equivalent in the context of task completion, and the correction is made within the same utterance. The final intended action is to mark all tasks in the Todoist project as complete." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/false_start/false_start_04_completion_action/56a9d64f22e3509b/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "985103e3c5f72a14a4ba5d4e9f9e80a191844f4bb6a5df7e4205ac973db17800", "seconds": 17.36, "spoken_utterances": [ "I am going on a camping trip with some of my friends. hmmm... I have created a new Todoist project for managing its tasks. Please do them for me, and ... check them all off—um... wait, make that mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately self-corrects to 'mark them all complete'. Since this is a false_start type, the final corrected instruction 'mark them all complete' is the intended goal. The action and target remain consistent with the original instruction, and the correction is immediate and clear. Therefore, the final intent matches the expected execution.", "heard_transcript": "I am going on a camping trip with some of my friends. Hmm. I have created a new Todoist project for managing its tasks. Please do them for me, and check them all off. Wait, make that mark them all complete." }, "candidate_id": "wav_46dea2ca194ef32d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker says 'check them all off' and then corrects it to 'mark them all complete'. While this is a plausible change, the correction is abrupt and lacks a natural back-reference or hesitation. The 'wait' is a common repair cue, but the overall flow feels a bit too clean and rehearsed compared to the more organic hesitation in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_01_activity_description", "proposed_instruction": "I am going on a hiking trip—or rather, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "b0ec1566c4cd0851b419fcef2584b3b120c6796286dcf2302d52af0363f2a595", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects to 'camping trip'. The final intended activity is camping, and all other elements—project creation, app (Todoist), and the action of marking all tasks complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/false_start/false_start_01_activity_description/50102ded4a69072a/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "44e079e31a1190db20c150daa0ed8d45ab81cbab7a93ac3ea37a0e4b1f6c6d53", "seconds": 14.12, "spoken_utterances": [ "I am going on a hiking trip—um... or rather, camping trip with some of my friends. I have created a new Todoist project uh... for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately self-corrects to 'camping trip'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction, including creating a Todoist project and marking all tasks as complete, remains unchanged and is executed as intended.", "heard_transcript": "I am going on a hiking trip, um, or rather camping trip with some of my friends. I have created a new Todoist project, uh, for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_18f5927b90669af9", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has a false start with 'hiking trip' corrected to 'camping trip', which is a plausible error. However, the correction is very smooth and the 'um' pause is minimal. The delivery is relatively flat and lacks the natural, breathy hesitations and self-correction flow seen in the other clips, making it sound more like a rehearsed or read-aloud instruction than a spontaneous utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/addition/addition_01/94408e53790596c9/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "099322b636b0af79f52a13ceb224672dba0a827d6bf83b83cb4baf80be88b24f", "seconds": 20.44, "spoken_utterances": [ "I am going on a camping trip. uh, let me think... I have created a new Todoist project for managing its tasks. um... Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with hmmm... some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to mark all tasks as complete. The addition of 'it's with some of my friends' is a clarifying detail about the trip, not a change to the core action. The final intent remains to mark all tasks in the project as complete, which matches the original instruction. The addition does not alter the primary goal.", "heard_transcript": "I am going on a camping trip. Uh, let me think. I have created a new Todoist project for managing its tasks. Um, please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_cbe4f216f38d7798", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and realistic speech patterns. The speaker uses a believable 'uh' and 'um' as fillers, a clear 'let me think' hesitation, and a spontaneous-sounding 'oh, I forgot to mention' that introduces a new detail. The intonation and rhythm are consistent with someone formulating a request in real time, and the flow from initial statement to addition feels organic and unscripted." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project. Please do them for me, and mark them all complete. And just to be specific, it's for managing its tasks.", "long_pause_offsets": [ 143 ], "source_sha256": "b3db5dde9114e82f73668129871e7b8246bcf77afdb145139254dd72c53556d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the purpose of the project but does not change the core action or target. The initial instruction already specified doing the tasks and marking them complete; the addition only adds context about the project's purpose." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/addition/addition_03/354b24dae4540b18/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "238609aa190cd929145869a9b40cad5537fa0a96d92373f8a33c1ba4f7e01b3e", "seconds": 18.16, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. ... I have created a new Todoist project. ah... Please do them for me, and mark them all complete. ", "And just to be specific, it's for managing its tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to create a Todoist project for managing camping trip tasks and execute them. The speaker then adds specificity by clarifying that the project is for managing 'its tasks', which is a minor elaboration on the original goal. The addition does not contradict or remove any part of the original request but rather adds a clarifying detail. The final intent is to execute the original task with the added specificity, which aligns with the 'addition' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project. Uh, please do them for me and mark them all complete. And just to be specific, it's for managing its tasks." }, "candidate_id": "wav_b3f2cf4c97d45f08", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is realistic but slightly less so than audio_02. The speaker uses 'let's see' as a filler, which is natural, and adds a clarifying detail at the end ('and just to be specific...'). However, the addition is delivered more like a formal afterthought than a spontaneous recollection, and the overall pacing is slightly more rehearsed. The prosody is good, but the flow lacks the organic hesitation and backtracking seen in audio_02." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a camping trip with some of my friends. for managing its tasks. Please do them for me, and mark them all complete. I should also mention, I've created a new Todoist project.", "long_pause_offsets": [ 129 ], "source_sha256": "2b125e5337fdc51f587565e00ef9f5b1680122588fc505b5e6b0cc27713f3ed9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context that a new Todoist project was created, which clarifies the target of the action. The core request to perform and complete all tasks remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/addition/addition_02/bbcf07416f1bc7bd/speech.wav", "wer": 0.02702702702702703, "wav_sha256": "08012a1cd329e589f75a726edc29f5824df7bc86119ff104a6d2dee58f370adc", "seconds": 17.28, "spoken_utterances": [ "I am going on a ... camping trip ... with some of my friends. for managing its tasks. uh... Please do them for me, and mark them all complete. ", "I should also mention, I've created a new Todoist project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the context of a camping trip and a Todoist project for task management, then requests to 'do them for me and mark them all complete'. The addition of 'I should also mention I've created a new Todoist project' clarifies the project's existence and scope, refining the initial request. The final intent is to complete all tasks within that project, which aligns with the addition type criterion: the original instruction is combined with the added information to form the complete goal.", "heard_transcript": "I am going on a camping trip with some of my friends. For managing its tasks, uh, please do them for me and mark them all complete. I should also mention I've created a new Todoist project." }, "candidate_id": "wav_3fe98948d81f3544", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio sounds the least realistic. The speaker delivers the request in a flat, monotonous tone with no natural pauses or fillers. The addition of 'I should also mention' feels tacked on rather than spontaneous, and the delivery lacks the prosodic variation and hesitation expected in a real-time request. The speech is too smooth and lacks the authentic rhythm of someone thinking aloud." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_activity_description", "proposed_instruction": "I am going on a hiking trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. I misspoke about the activity: it should be a camping trip, not a hiking trip.", "long_pause_offsets": [ 165 ], "source_sha256": "0aa30377907f693e09e7dc936dc5b57cac9d5f93e3d309b0c3acb34d91b7bb40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the activity from 'hiking trip' to 'camping trip' but retains all other requirements: the group ('some of my friends'), the app ('Todoist'), the action ('do the tasks'), and the completion status ('mark them all complete')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/revision/revision_01_activity_description/d902c602284359ad/speech.wav", "wer": 0.03773584905660377, "wav_sha256": "cc947d10db34abdcdf5ba466fdf42dd4544bfd8fa08a646537c011da1f3a8bc8", "seconds": 21.12, "spoken_utterances": [ "I am going on a uh, let me think... hiking trip with some of my friends. uh... I have created a new Todoist project for managing its tasks. ... Please do them for me, and mark them all complete. ", "I misspoke about the activity: it should be a camping trip, not a hiking trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects it to 'camping trip' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark tasks complete for a camping trip, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints are changed.", "heard_transcript": "I am going on a uh let me think hiking trip with some of my friends. uh I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. I misspoke about the activity. It should be a camping trip, not a hiking trip." }, "candidate_id": "wav_59ca78b814880712", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear false start ('hiking trip'), self-corrects with a believable 'I misspoke about the activity' followed by a smooth, contextually appropriate correction to 'camping trip'. The prosody includes natural hesitation and a clear change-of-mind intonation. The initial pause and filler 'uh' are well-placed, and the overall rhythm mimics genuine spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Wait, I meant 'friends', not 'family'.", "long_pause_offsets": [ 165 ], "source_sha256": "f2d98fd67e77860f1db046a25f0ca2fffbcb92f7a407a258237f9c17a5f7ef5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'friends', preserving all other elements including the project creation, task execution, and completion. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/revision/revision_02_companion_group/d992c28fdece248b/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "8fb34d71a2643b98b00acd9ad345f262489f18dd094beb7153fac9a70c66c5f7", "seconds": 19.84, "spoken_utterances": [ "I am going on a uh... camping trip with some of my family. let me think... I have created a new Todoist project let's see... for managing its tasks. Please do them for me, and mark them all complete. ", "Wait, I meant 'friends', not 'family'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_df2ddf81fd7d4c16", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a realistic false start and self-correction ('family' to 'friends'), with a natural-sounding 'wait' and a plausible retraction. The speaker's prosody includes a believable hesitation ('let me think') and a slight shift in intonation during the correction. However, the correction is less smoothly integrated than in audio_03, and the overall delivery is slightly more robotic in pacing." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. Please do them for me, and mark them all complete. Use Todoist, not Google Tasks, for managing the tasks.", "long_pause_offsets": [ 171 ], "source_sha256": "a4ff84eb810f6cecf785aabc5adc5b56ae1f5febff26bedb0060803171872ddd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Google Tasks to Todoist, which changes the target application for managing the tasks." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/revision/revision_03_app/08fb2d2a9ac76724/speech.wav", "wer": 0.022727272727272728, "wav_sha256": "633d6a00b69ed717bc1a9ec97e6f59a045d1d959c4b857beba4d75ee6d06457f", "seconds": 18.36, "spoken_utterances": [ "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. ... Please do them for me, and mark them all complete. ", "Use um... Todoist, not Google Tasks, for managing the tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks' but immediately corrects to 'Todoist' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use Todoist, not Google Tasks, which matches the original reference instruction. All other conditions are preserved.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. Please do them for me, and mark them all complete. Use um, Todoist, not Google Tasks for managing the tasks." }, "candidate_id": "wav_c553a29e9b8f8026", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but feels less natural due to the speaker's phrasing. The correction ('use um todoist not google tasks') is delivered as a separate, almost command-like sentence rather than a fluid self-correction. The prosody is flatter, and the 'um' is used awkwardly. The revision lacks the organic flow of a speaker reconsidering their own words in real time." } ] }, "is_default": false }, { "variant_id": "revision_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them as pending. Change 'mark them as pending' to 'mark them all complete'.", "long_pause_offsets": [ 164 ], "source_sha256": "f8b0788acdba0e60fe5f0968bf09af75dfb5e75bcba40efd9364e8d658252c20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the action from 'mark them as pending' to 'mark them all complete', preserving the initial request to perform the tasks and the context of the Todoist project. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/revision/revision_04_completion_action/7355ea800fcb6005/speech.wav", "wer": 0.022727272727272728, "wav_sha256": "8ca7848b8bc72fd5bd0c3185f8f339f10f80a3cac535dbae8f6318032e208c09", "seconds": 16.56, "spoken_utterances": [ "I am going on a camping trip with some of my friends. ... I have created a new Todoist project um... for managing its tasks. Please do them for me, and mark them as pending. ", "Change 'mark them as pending' to '... mark them all complete'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark them as pending' but immediately corrects this to 'mark them all complete' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to mark all tasks as complete, which matches the original reference instruction. No other constraints are changed.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them as pending. Change mark them as pending to mark them all complete." }, "candidate_id": "wav_110833cef947a470", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. The speaker makes a revision but does so in a way that sounds like a post-hoc instruction rather than a spontaneous correction. The phrase 'change mark them as pending to mark them all complete' is unnatural and sounds like a mechanical command. The prosody is flat, and the delivery lacks the hesitation and intonation shifts of genuine self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "a13c155d13b53b6297c7b5d048a6c5ff21a03542a439ae76590b30429b733dfc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/cancellation/cancel_023/6d1357d1ce30be50/speech.wav", "wer": 0.020833333333333332, "wav_sha256": "3b926c36d3485b77288fb0e3b30c7245cdb3cddc98c8064d30eaaaf4e1425d81", "seconds": 18.44, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete.", "Oh, wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives a complete instruction to mark all tasks in a Todoist project as complete. After a pause and a self-correction ('Oh wait, never mind about that'), they explicitly state 'Don't carry out the request,' which constitutes a clear cancellation of the original task. The final intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Oh wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "e3ca08d6575ea37c19309b3df34c2b081b3cf656e4a646a19b45ad649e0bcd32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/ambiguous_halt/halt_033/2c999e08141e7412/speech.wav", "wer": 0.024390243902439025, "wav_sha256": "ce64a1792fbd63df0a1ef45b22c2079e6fa1bc6dc55e0fbda76d8b54bc5077fc", "seconds": 17.0, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, there's something,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request to mark tasks complete is not executed due to the halt, which is correct for this type.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Hold on, there's something." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "bc51b34e902458e83d403a3dde4286de9f8a1a7f222aef1015fe07c6f118b1dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_1/audio_natural/filler/original/faad954a72ff2f35/speech.wav", "wer": 0.02702702702702703, "wav_sha256": "cbc452e4229854e96e6050f0612c2abc9eb0a7a1bbf3a03e9ed045921456be70", "seconds": 13.08, "spoken_utterances": [ "I am going on a um, let me think... camping trip with some of my friends. I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and a pause but does not change the core request. The final instruction is identical to the original, with no correction or modification. The fillers are internal hesitations and do not constitute a final directive to pause or cancel. Therefore, the final intent matches the original instruction, and the control intent is to execute.", "heard_transcript": "I am going on a um, let me think, camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family—no, I mean friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "b2fb8d9d47aefefda25e8b7bfe18eacc84a9598fcf99b083eea9d25e8644de98", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'family' to 'friends' within the same utterance, replacing the group reference. The rest of the instruction—creating a Todoist project and marking all tasks complete—remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/false_start/false_start_02_companion_group/88a13d27b4e333f0/speech.wav", "wer": 0.0, "wav_sha256": "7d9a570736c1bc417f83a98b0071e1da17f0f0c407ca0bef2b5d0c26db66b2fa", "seconds": 17.72, "spoken_utterances": [ "I am going on a camping trip with some of my um... family—uh... no, I mean friends. I have created a new Todoist project for managing its tasks. ... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('family') is discarded. The rest of the instruction, including the creation of a Todoist project and the request to mark all tasks complete, remains unchanged and is fully audible. Therefore, the final intent matches the original instruction after the correction.", "heard_transcript": "I am going on a camping trip with some of my um family uh no, I mean friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_9fc265efae9f851e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a very natural false start with 'family' followed by a clear self-correction to 'friends'. The intonation and pacing are consistent with spontaneous speech, and the correction is immediate and fluent. The overall prosody is smooth and the task request is completed without further hesitation." } ] }, "is_default": true }, { "variant_id": "false_start_01_activity_description", "proposed_instruction": "I am going on a hiking trip—I meant camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "68d8285612c1e4e7d276db6c466c5c8067b510c821923e13f4afa2d150c430f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'hiking trip' to 'camping trip' within the same utterance. The final intended activity is camping, which matches the original instruction. All other elements, including the project creation and task completion request, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/false_start/false_start_01_activity_description/4103eb3618c8710b/speech.wav", "wer": 0.1282051282051282, "wav_sha256": "c215bfcb618219f5be0f82e358e18a3613e0d190e60fbb0d36320681936b3391", "seconds": 17.0, "spoken_utterances": [ "I am going on a hiking trip—hmmm... I meant camping trip um... with some of my friends. I have created a new Todoist project for managing its tasks. ... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately self-corrects to 'camping trip'. According to the 'false_start' type criterion, the final intent is the corrected version. The rest of the instruction, including the creation of a Todoist project and the request to mark all tasks as complete, remains unchanged and is executed as intended.", "heard_transcript": "I am going on a hiking trip. I meant camping trip, um, with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_f0d50ba991e08efe", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start with 'hiking trip' corrected to 'camping trip'. The speaker uses a thoughtful 'hmm' and a natural-sounding 'I meant' to transition. The correction is integrated smoothly into the sentence, and the delivery maintains a consistent, realistic rhythm." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks—sorry, I should have said Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "ea2abd1a5eafa413ab49e669cc7e139bd94e2596dff95bcce3cc308aedfa4ded", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Google Tasks' to 'Todoist' within the same utterance. The final intended app is Todoist, and all other requirements (creating a project, doing the tasks, marking them complete) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/false_start/false_start_03_app/acfc95ae24865111/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "d46ad9424ef26f3ddbe80372e9ee22fbcc22209d80e54710af0b52964cb7882f", "seconds": 18.48, "spoken_utterances": [ "I am going on a camping trip ... with some of my friends. I have created a new Google Tasks—um... sorry, I should have said Todoist project for managing its tasks. uh... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks sum' but immediately self-corrects to 'Todoist project'. According to the 'false_start' type criterion, the final corrected value ('Todoist project') should be applied, and the original incorrect value ('Google Tasks sum') should be discarded. The final instruction to 'do them for me and mark them all complete' is clear and matches the original intent. The correction is immediate and unambiguous, so the task is to execute the corrected instruction.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Google Tasks sum. Sorry, I should have said Todoist project for managing its tasks. Uh, please do them for me and mark them all complete." }, "candidate_id": "wav_b0023d5e9e11b9fb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start with 'Google Tasks' and corrects it to 'Todoist project'. The correction is clear, but the phrase 'some, sorry, I should have said' introduces a slightly awkward, self-conscious flow. The prosody is otherwise natural, but the correction feels a bit more rehearsed than the others." } ] }, "is_default": false }, { "variant_id": "false_start_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and check them all off—scratch that, make it mark them all complete.", "long_pause_offsets": [], "source_sha256": "8a2bac507978f0904d2afeff57c4260a9e668e2dcdb18a61ddcfe887cd90881c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately corrects it to 'mark them all complete'. The final action is 'mark them all complete', which is semantically equivalent to the original instruction's 'mark them all complete'. The context and other requirements remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/false_start/false_start_04_completion_action/a9e5f7a7314125e1/speech.wav", "wer": 0.10869565217391304, "wav_sha256": "32e6363f6a82a084a50531c4a12d8c4054a0400336408bdc8eb3763446107b52", "seconds": 19.44, "spoken_utterances": [ "I am going on a um, let's see... camping trip with some of my friends. I have created a new Todoist project uh... for managing its tasks. Please do them for me, and check them all off—ah... scratch that, make it mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately self-corrects with 'Scratch that, make it mark them all complete.' This is a clear false start where the final, corrected instruction is 'mark them all complete.' The original instruction 'do them for me and mark them all complete' is semantically preserved in the final intent, with the only change being the verb 'check off' to 'mark complete,' which is a synonym in this context. The task scope and target remain consistent.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and check them all off. Scratch that, make it mark them all complete." }, "candidate_id": "wav_b2ee881a4ef5c86a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip begins with a false start ('let's see') and a self-correction ('scratch that, make it mark them all complete'). However, the correction is abrupt and the phrase 'scratch that' is followed by a very short pause before the new instruction, making the transition feel less fluid. The overall delivery is slightly more robotic compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/addition/addition_01/ff39845b8636e8d3/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "6a05ddb7da9c04958b591d0d6d74456d25d78b48e2424c7d717057415e4af8e6", "seconds": 19.96, "spoken_utterances": [ "I am going on a camping trip. uh, let's see... I have created a new Todoist project um, let me think... for managing its tasks. Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is fully executed, and the addition 'it's with some of my friends' is a clarifying detail that is incorporated into the final goal. The addition does not contradict or remove any part of the original request; instead, it adds a relevant context. The final intent matches the original instruction with the added detail, satisfying the 'addition' type criterion.", "heard_transcript": "I am going on a camping trip. Uh, let's see. I've created a new Todoist project, um, let me think, for managing its tasks. Please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_6cc4e30208525c5b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and continuous flow of spontaneous speech. It begins with a clear initial statement, includes a believable 'let me think' filler, and smoothly integrates a later addition ('Oh, I forgot to mention...') that feels like a genuine recollection. The prosody is varied and conversational, with pauses and fillers placed in a way that mimics real-time thought. The speaker's voice remains consistent, and the overall delivery is cohesive and unforced." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project. Please do them for me, and mark them all complete. And just to be specific, it's for managing its tasks.", "long_pause_offsets": [ 143 ], "source_sha256": "b3db5dde9114e82f73668129871e7b8246bcf77afdb145139254dd72c53556d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies the purpose of the project but does not change the core action or target. The initial instruction already specified doing the tasks and marking them complete; the addition only adds context about the project's purpose." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/addition/addition_03/654920dd2ef81b60/speech.wav", "wer": 0.09302325581395349, "wav_sha256": "e31695478ef76cdb9e27628bbfe6d5325bc9d0966975b80e00ba4472faae6d66", "seconds": 19.84, "spoken_utterances": [ "I am going on a um... camping trip with some of my friends. ... I have created a new Todoist project. Please do them for me, and mark them all complete. ", "And just to be specific, it's for uh, let's see... managing its tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to 'do them for me and mark them all complete' for a new Todoist project. Later, they add specificity by clarifying the project is 'for managing its tasks'. This is a valid addition that narrows the scope of the original task. The final intent is to mark all tasks in the project 'managing its tasks' as complete, which combines the original instruction with the added detail. The addition is clear and does not contradict the initial request.", "heard_transcript": "I am going on a um camping trip with some of my friends. I have created a new Todoist project. Please do them for me and mark them all complete. And just to be specific, it's for let's see, managing its tasks." }, "candidate_id": "wav_2f68a6f8c1ac69a0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is realistic and conversational, with a natural 'um' and a pause before 'let's see'. The speaker adds a clarifying detail at the end, which is a plausible addition. However, the delivery is slightly more hesitant and less fluid than audio_03, with a more abrupt transition into the final clause. The prosody is less varied, and the overall pacing feels a bit more deliberate, making it slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a camping trip with some of my friends. for managing its tasks. Please do them for me, and mark them all complete. I didn't mention this, but I've created a new Todoist project.", "long_pause_offsets": [ 129 ], "source_sha256": "2319f498411fddf6615e0a8ae132b3571e1d723f71822f2e7e39e23ec15b2c31", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail that a new Todoist project was created for the camping trip. The initial instruction already specified doing the tasks and marking them complete. The final goal combines both parts: the action (do and mark complete) and the target (the new Todoist project for the camping trip)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/addition/addition_02/924dbaefe9fe5c84/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "fba5638355b2fdd11add803c5354e2d6582bc7aaf677f7a72940735ebf631885", "seconds": 19.64, "spoken_utterances": [ "I am going on a uh... camping trip with some of my friends. um... for managing its tasks. Please do them for me, and mark them all complete. ", "I didn't mention this, but I've uh, let me think... created a new Todoist project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests marking tasks as complete for a camping trip. After a pause and self-correction ('let me think'), they add the information that they have created a new Todoist project. This is an addition to the original instruction, and the final intent combines both the creation of the project and the completion of its tasks. The addition is clear and does not contradict the original request, so the task is executed as a combined goal.", "heard_transcript": "I am going on a uh camping trip with some of my friends. Um, for managing its tasks, please do them for me and mark them all complete. I didn't mention this, but I've uh let me think, created a new Todoist project." }, "candidate_id": "wav_b4bb2fd901242d32", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with a clear statement but then makes a revision ('I didn't mention this, but I've...') that feels abrupt and less integrated into the flow. The 'let me think' filler is placed awkwardly mid-sentence, and the overall prosody is flatter and more robotic compared to the others. The transitions between ideas are less natural, and the voice sounds less engaged, reducing the sense of spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them as done. Change 'mark them all complete' to 'mark them as done'.", "long_pause_offsets": [ 161 ], "source_sha256": "dd9c6daf4d991dfa0f7d3bb1e95352be4b29b569f28b54deecfdbd2d78bd5934", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the phrase 'mark them all complete' with 'mark them as done', which is a synonymous instruction for completing the tasks. The action, target, and intent remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/revision/revision_04_completion_action/cb51f73c358f9f34/speech.wav", "wer": 0.0, "wav_sha256": "1fe3014b75050f6fbcb2707953fc55d096100410e168320dd408a013d594722f", "seconds": 21.92, "spoken_utterances": [ "I am going on a camping trip um... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them as done. ", "Change 'mark them all complete' to 'mark them as done'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark them all complete' but then immediately corrects this to 'mark them as done' in a revision. The final audible instruction is 'mark them as done', which is the intended goal. The revision type criterion is satisfied as the correction is applied to the earlier instruction, and the final control intent is to execute the revised task.", "heard_transcript": "I am going on a camping trip um with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them as done. Change mark them all complete to mark them as done." }, "candidate_id": "wav_9e3b77cd883400af", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody. The speaker's intonation and rhythm are consistent with spontaneous speech, including a believable pause and filler ('um') at the start. The revision is a smooth, self-contained correction of the phrase 'mark them all complete' to 'mark them as done,' delivered with a natural change in intonation and pacing that suggests a genuine moment of reconsideration. The overall flow is coherent and unforced." } ] }, "is_default": true }, { "variant_id": "revision_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. Please do them for me, and mark them all complete. Let me correct the app: use Todoist, not Google Tasks.", "long_pause_offsets": [ 171 ], "source_sha256": "4b7f43753a9c0847885db0dcb3b1bd30d1e9eb54eab87434006ff44651296eb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Google Tasks to Todoist, while retaining all other requirements including the action (do and mark complete), the project context, and the group of tasks. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/revision/revision_03_app/308714481c59c5f9/speech.wav", "wer": 0.0, "wav_sha256": "f037313f9449b9e27d87b2f4f52ab91f1b8bc752887ac6d592ed6ddb78506647", "seconds": 22.24, "spoken_utterances": [ "I am going on a ... camping trip um... with some of my friends. uh... I have created a new Google Tasks project for managing its tasks. Please do them for me, and mark them all complete. ", "Let me correct the app: use Todoist, not Google Tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Google Tasks' but immediately corrects it to 'Todoist' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use Todoist, not Google Tasks, and the rest of the task (creating a project and marking tasks complete) remains unchanged. Therefore, the final intent matches the reference instruction with the corrected app name.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. Please do them for me and mark them all complete. Let me correct the app, use Todoist, not Google Tasks." }, "candidate_id": "wav_d3b4a770967190fb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is realistic but slightly less so than audio_02. The speaker uses natural fillers ('um', 'uh') and maintains a consistent rhythm. The revision is a self-correction of the app name from 'Google Tasks' to 'Todoist,' which is plausible. However, the correction is delivered with a slightly more abrupt shift in tone and pacing compared to audio_02, making it feel a bit more like a scripted edit than a spontaneous thought." } ] }, "is_default": false }, { "variant_id": "revision_01_activity_description", "proposed_instruction": "I am going on a hiking trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Wait, I meant a camping trip, not a hiking trip.", "long_pause_offsets": [ 165 ], "source_sha256": "0556723c88f8872966ed2562c55287ff9f55c06c20a8f76a2758f6c91a4141c1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the trip type from 'hiking' to 'camping', while retaining all other elements including the project creation, task execution, and completion. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/revision/revision_01_activity_description/90d9e032693b0723/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "40670a094f295852349837e6237a6a2e4c4d791c907f48d1a14d2311ed0d2198", "seconds": 21.0, "spoken_utterances": [ "I am going on a hiking trip um... with some of my friends. ... I have created a new Todoist project ... for managing its tasks. Please do them for me, and mark them all complete. ", "Wait, I meant a camping trip, not a hiking trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects themselves to 'camping trip' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark tasks as complete for a camping trip, which is a revision of the original hiking trip context. The rest of the instruction remains unchanged and is executed.", "heard_transcript": "I am going on a hiking trip, um, with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Wait, I meant a camping trip, not a hiking trip." }, "candidate_id": "wav_af563d56237e719e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker begins with 'hiking trip' and then corrects to 'camping trip,' but the correction is delivered with a flat, robotic intonation and lacks the natural prosodic shift seen in the other clips. The pause before the correction is also unusually long and seems artificial. The overall delivery feels more like a rehearsed or synthesized prompt than a spontaneous utterance." } ] }, "is_default": false }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Make that with some of my friends, not family.", "long_pause_offsets": [ 165 ], "source_sha256": "1184f6b3d2fc10e97bd194d6c3568835477cdde3f535d57a8c54880c77b5aae8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'friends', which matches the original instruction. All other conditions, including the action (doing and marking complete) and context (Todoist project for the trip), are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "I am going on a camping trip with some of my family. um... I have created a new Todoist project for managing its tasks. uh... Please do them for me, and mark them all complete. ", "Make that with some of my ... friends, not family." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "a56ddc41e4e94fdcd7399ec8f15a92df5eadf7a4e1bb4e72f020af07e197ea53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/cancellation/cancel_027/9cf9d0de95188eab/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "e4278ac5930781a7f775e6001e5f9dd07674528cf0a46f0f6b662b3ea76e3387", "seconds": 18.12, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete.", "Hold on, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a clear request to mark all tasks in a Todoist project as complete. Immediately after, they say 'Hold on, please disregard my last request,' which is a direct and unambiguous cancellation of the preceding instruction. The final control intent is clearly to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Hold on, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "82b87ea78b159df7e562331a18344f1c5ad08ae32601fad1debe9415d0ee186b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/ambiguous_halt/halt_036/76e6f782b6779de6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "498d85b098122d0cdc320018e28fa6f2f00c8bc6d542519f4855f543d566db75", "seconds": 17.0, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete.", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second, before you go ahead,' which is a clear request to pause or hold the current action. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request to 'do them for me and mark them all complete' is not executed due to the final pause command.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Wait a second, before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2947b09dfd9e2cd63821f16f4fd4a4403ef374cf05560e5eda708904341420c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_2/audio_natural/filler/original/099f8e576763f865/speech.wav", "wer": 0.11428571428571428, "wav_sha256": "4f4691f4ba766aadbc1e932772f2be99c3a62e7261db1c2a0085c1e2d03b37da", "seconds": 12.8, "spoken_utterances": [ "I am going on a camping trip let's see... with some of my friends. I have created a new Todoist project ... for managing its tasks. ... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'let's see,' which is a filler, and then proceeds to state the original instruction without any correction or change to the goal. The final interjection is the completion of the original request, so the control intent is to execute the task as stated.", "heard_transcript": "I am going on a camping trip, let's see, with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and check them all off—sorry, I mean mark them all complete.", "long_pause_offsets": [], "source_sha256": "b2d926aaa2b6e59465dd213dfe7780453239f3d00ac22f64672a486d990ac272", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately corrects to 'mark them all complete'. These phrases are functionally equivalent in the context of task management, both meaning to finalize all tasks in the project. The app (Todoist), the action (completing tasks), and the scope (all tasks in the project) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/false_start/false_start_04_completion_action/01ed5d2cebb557a9/speech.wav", "wer": 0.020833333333333332, "wav_sha256": "1ce93b8c484feee4e9b9a006d18a30e20570a79f9f2f5556a3f3a8b5a06f39ab", "seconds": 19.96, "spoken_utterances": [ "I am going on a camping trip with some of my friends. I have created a new Todoist project um, let me think... for managing its tasks. Please do them for me, and um, let's see... check them all off—ah... sorry, I mean mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'check them all off' but immediately self-corrects to 'mark them all complete'. According to the 'false_start' type criterion, the final corrected instruction should be applied. The final intent is to mark all tasks as complete, which matches the original instruction's goal. The correction is immediate and unambiguous, so the task should be executed as corrected.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project, um, let me think, for managing its tasks. Please do them for me, and um, let's see, check them all off. Uh, sorry, I mean mark them all complete." }, "candidate_id": "wav_feadcce02dedbc58", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant statement, uses a believable 'let me think' filler, and then smoothly self-corrects 'check them all off' to 'mark them all complete' with a natural, corrective intonation. The pauses and fillers are well-placed, and the overall prosody sounds like a person genuinely formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_01_activity_description", "proposed_instruction": "I am going on a hiking trip—I should say camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "a48d3ac2c06e46ea91c84b95103c2e5630d40b25a19dfacb6848746ec616013d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects to 'camping trip'. The final intended activity is camping, and all other requirements—creating a Todoist project and marking all tasks complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/false_start/false_start_01_activity_description/722cc33f5f6d9cbe/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "89af7bad1f8a5550770be4bb8a839de0b7e96e41b3fdf2cd6280162790b40cec", "seconds": 19.24, "spoken_utterances": [ "I am going on a hiking trip—uh... I should say camping trip with some of my friends. hmmm... I have created a new Todoist project for managing its tasks. um, let me think... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately self-corrects to 'camping trip'. This is a false start, and the final, corrected intent matches the original instruction. The task to mark all tasks in the Todoist project as complete is clearly stated at the end and is not modified.", "heard_transcript": "I am going on a hiking trip. uh, I should say camping trip with some of my friends. Hmm. I have created a new Todoist project for managing its tasks. Um, let me think. Please do them for me and mark them all complete." }, "candidate_id": "wav_b7e557bae782e5c0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic with a clear false start ('hiking trip' to 'camping trip') and a natural 'let me think' pause. However, the correction is less fluid than audio_02, and the final request is delivered more flatly. The prosody lacks the subtle intonation shifts of a fully spontaneous speaker, making it slightly less convincing than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family—no, that should be friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "cbc2ddc8b7efc5ad31501e2c2ae0a99842631da1f64ad4b570bf6c64d0162b91", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance. The final intended group is 'friends', matching the original instruction. All other elements, including the project creation and task completion, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/false_start/false_start_02_companion_group/4fd12d9f2754bbb8/speech.wav", "wer": 0.05, "wav_sha256": "1339c31673f399cf3a25563168fb0f16aab4a21d07b52e5d41b509fac7560e10", "seconds": 18.44, "spoken_utterances": [ "I am going on a camping trip with some of my um... family—uh... no, that should be friends. ... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends'. According to the 'false_start' type criterion, the final intent is determined by the immediate self-correction. The rest of the instruction remains unchanged and is executed as intended.", "heard_transcript": "I'm going on a camping trip with some of my um family uh no that should be friends. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_83c52bdf22e4d732", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a believable false start ('family' to 'friends') and a natural-sounding correction. However, the initial hesitation ('I'm going on a camping trip with some of my um... family uh...') is prolonged and the correction feels slightly rushed, making the overall flow less smooth. The prosody is less continuous compared to the top two." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks—scratch that, make it Todoist project for managing its tasks. Please do them for me, and mark them all complete.", "long_pause_offsets": [], "source_sha256": "937a9745f53fe9c08f3877e2fad70fcc9d7f19347e290d2a6a870a066138831e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks' but immediately corrects to 'Todoist' within the same utterance. The final app is Todoist, and all other requirements—project creation, context (camping trip with friends), and marking all tasks complete—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/false_start/false_start_03_app/58c946fe0fe1ebcd/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ccde4b0b3622abdba56137b06374eaea1815922ae622d8e1038aa3d92ab32147", "seconds": 17.4, "spoken_utterances": [ "I am going on a uh... camping trip with some of my friends. I have created a new Google Tasks—ah... scratch that, make it Todoist project uh, let's see... for managing its tasks. Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks' but immediately self-corrects to 'Todoist project'. According to the 'false_start' type criterion, the final corrected value ('Todoist') should be applied, and the original incorrect value ('Google Tasks') should be discarded. The rest of the instruction, including the action to 'mark them all complete', remains unchanged. The final audible intent matches the original instruction's goal, with only the app name corrected.", "heard_transcript": "I am going on a uh camping trip with some of my friends. I have created a new Google Tasks uh. Scratch that, make it Todoist project. Uh let's see. For managing its tasks. Please do them for me and mark them all complete." }, "candidate_id": "wav_0b468029f401d076", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the most awkward delivery. The speaker starts with a stutter ('I am going on a uh... camping trip'), then makes a more complex self-correction ('Google Tasks' to 'Todoist project'), which sounds less spontaneous and more like a rehearsed edit. The 'uh' fillers are overused and the pacing feels unnatural, reducing the overall realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "I am going on a camping trip. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Oh, I forgot to mention, it's with some of my friends.", "long_pause_offsets": [ 142 ], "source_sha256": "077337390ac8c6dee38846a2b88a6ea20c86dd2cc122af4fb94279be904291d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up adds the missing detail about the trip being with some friends, which is a constraint on the context of the project but does not alter the core action of completing all tasks in the project. The original instruction already specified the project and the action, and the addition only provides additional context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/addition/addition_01/fe7dab962476809f/speech.wav", "wer": 0.046511627906976744, "wav_sha256": "64c17f658f3c0c5e280a5f81fe4372273e678cc69b87df1e7bcec6c455b9448e", "seconds": 21.92, "spoken_utterances": [ "I am going on a um, let me think... camping trip. ... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. ", "Oh, I forgot to mention, it's with ... some of my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates but then clearly states the original task. After a pause, they add the detail 'it's with some of my friends' as an additional constraint. The final intent combines the original instruction with this new information, which is a valid addition. The core action of marking all tasks in the project as complete remains unchanged and is fully audible.", "heard_transcript": "I am going on a um, let me think, camping trip. I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. Oh, I forgot to mention, it's with some of my friends." }, "candidate_id": "wav_3c5bc10fcb69f7cf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous delivery. The speaker begins with a realistic 'um' and 'let me think' filler, indicating genuine hesitation. The addition of 'Oh, I forgot to mention' is a believable, pragmatic follow-up that integrates smoothly with the prior request. The prosody is varied and conversational, with appropriate pauses and intonation that suggest a real-time thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "I am going on a camping trip with some of my friends. for managing its tasks. Please do them for me, and mark them all complete. I left out a detail: I have created a new Todoist project.", "long_pause_offsets": [ 129 ], "source_sha256": "2c35b2ebd81daa3796566b076ed9e4ad13ccce6953f109e2268925218acd967b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing detail that the tasks are in a new Todoist project, which completes the original instruction's reference to 'its tasks'. The action (do and mark complete), the subject (tasks), and the context (camping trip with friends) are preserved. The project is now explicitly identified." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/addition/addition_02/3c64451cdd6ecaff/speech.wav", "wer": 0.05, "wav_sha256": "c0e373a8ac141cecbf275ba95e1eee5bfc223e66de966cf44dfd0886f49e141d", "seconds": 21.04, "spoken_utterances": [ "I am going on a camping trip with some of my friends. um... for managing its tasks. ... Please do them for me, and mark them all complete. ", "I left out a detail: I have uh... created a new Todoist project." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to execute and complete all tasks in a new Todoist project for a camping trip. After a pause, they add the detail that they have created the project. This is a clear addition that provides a missing but essential condition (the project already exists). The final goal combines the original task with the new information, resulting in a complete and executable instruction. The addition does not contradict or revise the original request, so the 'addition' type criterion is satisfied.", "heard_transcript": "I am going on a camping trip with some of my friends. Um, for managing its tasks. Please do them for me, and mark them all complete. I left out a detail. I have uh created a new Todoist project." }, "candidate_id": "wav_bf4245584f9a46eb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a natural addition with a 'um' and a later self-correction ('I left out a detail'), which is a credible repair. The speaker's intonation and pacing are generally natural, and the addition flows logically. However, the initial phrase 'for managing its tasks' is slightly awkwardly placed, and the correction feels a bit more like a post-hoc insertion than a fully integrated thought process, making it slightly less realistic than audio_02." } ] }, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project. Please do them for me, and mark them all complete. Oh, and one more thing: it's for managing its tasks.", "long_pause_offsets": [ 143 ], "source_sha256": "f5108c10dd073db1ada8c78cfed58764511b7345746d398b6cdc6fef11dace4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the purpose of the project ('for managing its tasks'), which clarifies the project's intent but does not change the core action or target. The original instruction already specified doing the tasks and marking them complete, and the followup does not contradict or alter that goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/addition/addition_03/28502268dfd0e810/speech.wav", "wer": 0.0, "wav_sha256": "2877006531d581d476d2bb086cab420142ce34b18e4da4e6d1d11a1cdf251f56", "seconds": 19.52, "spoken_utterances": [ "I am going on a ... camping trip with some of my friends. I have created a new Todoist project. ... Please do them for me, and mark them all complete. ", "Oh, and one more thing: it's for ... managing its tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to mark all tasks in a new Todoist project as complete. The addition 'it's for managing its tasks' clarifies the purpose of the project but does not change the core action. The final intent is to execute the original task with the added context, which is consistent with the 'addition' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. I have created a new Todoist project. Please do them for me and mark them all complete. Oh, and one more thing, it's for managing its tasks." }, "candidate_id": "wav_ff1961f50be299d4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic due to its flat, monotonous delivery and lack of natural hesitation or self-correction. The speaker proceeds in a continuous, rehearsed manner without any fillers, pauses, or signs of spontaneous thought. The addition of 'Oh, and one more thing' is abrupt and lacks the natural intonation or hesitation that would make it sound like a genuine afterthought, making the entire utterance sound more like a scripted instruction than a real person formulating a request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_activity_description", "proposed_instruction": "I am going on a hiking trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. I misspoke about the activity: it should be a camping trip, not a hiking trip.", "long_pause_offsets": [ 165 ], "source_sha256": "0aa30377907f693e09e7dc936dc5b57cac9d5f93e3d309b0c3acb34d91b7bb40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the activity from 'hiking trip' to 'camping trip' but retains all other requirements: the group ('some of my friends'), the app ('Todoist'), the action ('do the tasks'), and the completion status ('mark them all complete')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/revision/revision_01_activity_description/c84e28ad698c792e/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "03489cece9430e291fe1ff91808664c475be05b183a305667407b83c043ceac7", "seconds": 21.76, "spoken_utterances": [ "I am going on a um... hiking trip uh... with some of my friends. um... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. ", "I misspoke about the activity: it should be a camping trip, not a hiking trip." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking trip' but immediately corrects it to 'camping trip' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark tasks complete for a camping trip, which matches the original instruction's goal except for the activity type. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "I am going on a um hiking trip uh with some of my friends. um I have created a new Todoist project for managing its tasks. Please do them for me and mark them all complete. I misspoke about the activity, it should be a camping trip, not a hiking trip." }, "candidate_id": "wav_295114179c0ad68a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, coherent initial request with natural hesitation ('um') and a believable, smoothly integrated revision. The correction ('I misspoke about the activity...') is a complete, self-contained sentence that logically follows the initial statement, with a natural change-of-mind intonation. The prosody is consistent, and the pacing feels like a genuine moment of recollection." } ] }, "is_default": true }, { "variant_id": "revision_02_companion_group", "proposed_instruction": "I am going on a camping trip with some of my family. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Wait, I meant 'friends', not 'family'.", "long_pause_offsets": [ 165 ], "source_sha256": "f2d98fd67e77860f1db046a25f0ca2fffbcb92f7a407a258237f9c17a5f7ef5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'friends', preserving all other elements including the project creation, task execution, and completion. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/revision/revision_02_companion_group/1ca291828d07c0fc/speech.wav", "wer": 0.024390243902439025, "wav_sha256": "9cead367977f69fdbaeaffaec4596d5f70d7a58f0b489c48ff9a8a1542c29e4a", "seconds": 20.08, "spoken_utterances": [ "I am going on a um... camping trip with some of my family. hmmm... I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. ", "Wait, I meant 'friends', not 'family'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'friends' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark the tasks as complete for a trip with friends, which matches the original instruction except for the group. The correction is clear and final, so the task should be executed with the revised target.", "heard_transcript": "I am going on a uh, um, camping trip with some of my family. Hmm. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them all complete. Wait, I meant friends, not family." }, "candidate_id": "wav_80052a7c69184023", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a believable revision ('wait, I meant friends, not family') after a natural pause and hesitation ('hmm'). The correction is concise and fits the flow of a spontaneous thought process. The prosody is consistent, and the delivery sounds like a genuine moment of self-correction, though the initial 'family' mention is less plausible than 'hiking' in the context of a camping trip." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Google Tasks project for managing its tasks. Please do them for me, and mark them all complete. Use Todoist, not Google Tasks, for managing the tasks.", "long_pause_offsets": [ 171 ], "source_sha256": "a4ff84eb810f6cecf785aabc5adc5b56ae1f5febff26bedb0060803171872ddd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Google Tasks to Todoist, which changes the target application for managing the tasks." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/revision/revision_03_app/2b8f2165481ccf82/speech.wav", "wer": 0.061224489795918366, "wav_sha256": "b9ec8b3dd3462a9eb6b440c18d198fa28d6de7498a2ebc224827c03ec22d27f5", "seconds": 20.96, "spoken_utterances": [ "I am going on a camping trip let me think... with some of my friends. let's see... I have created a new Google Tasks project uh... for managing its tasks. Please do them for me, and mark them all complete. ", "Use Todoist, not Google Tasks, for managing the tasks." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Google Tasks' but immediately corrects to 'Todoist' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use Todoist, not Google Tasks, which matches the revision requirement. The rest of the task (creating a project, executing tasks, marking complete) remains consistent with the original instruction.", "heard_transcript": "I am going on a camping trip, let me think, with some of my friends. Let's see. I have created a new Google Tasks project, uh, for managing its tasks. Please do them for me and mark them all complete. Use Todoist, not Google Tasks, for managing the tasks." }, "candidate_id": "wav_cd9ef807859758ec", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a self-correction ('use Todoist, not Google Tasks'), but the initial mention of 'Google Tasks' is immediately followed by a correction within the same sentence. This creates a slightly less natural flow, as the speaker seems to correct themselves mid-utterance rather than after a pause. The 'let me think' and 'let's see' fillers are natural, but the correction feels more like a mechanical insertion than a spontaneous recollection." } ] }, "is_default": false }, { "variant_id": "revision_04_completion_action", "proposed_instruction": "I am going on a camping trip with some of my friends. I have created a new Todoist project for managing its tasks. Please do them for me, and mark them as incomplete. Change the completion action: mark them all complete, not as incomplete.", "long_pause_offsets": [ 167 ], "source_sha256": "e9ac55dda611a8646c43be017994c8be85f0409f2b085bdcf323c4e1fe0cc6eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The follow-up explicitly corrects the completion status from 'incomplete' to 'complete', reversing the initial instruction. The final goal is to mark all tasks as complete." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/revision/revision_04_completion_action/4ce4735382f9e8f5/speech.wav", "wer": 0.125, "wav_sha256": "c40d0556c991449adec93a71611261dfb562896915f1aa8fc28ab4eb8f4f0ce9", "seconds": 23.6, "spoken_utterances": [ "I am going on a ... camping trip um, let's see... with some of my friends. I have created a new Todoist project um... for managing its tasks. Please do them for me, and mark them as incomplete. ", "Change the completion action: mark them all complete, not as incomplete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'mark them as incomplete' but immediately corrects this with 'Change the completion action, mark them all complete, not as incomplete.' This constitutes a revision where the final, corrected instruction overrides the initial one. The final intent is to mark all tasks as complete, which matches the original instruction. The correction is clear and unambiguous.", "heard_transcript": "I am going on a camping trip, um, let's see, with some of my friends. I have created a new Todoist project, um, for managing its tasks. Please do them for me and um, mark them as incomplete. Change the completion action, mark them all complete, not as incomplete." }, "candidate_id": "wav_0f519590c1c238ee", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip features a revision, but the correction ('change the completion action...') is delivered as a separate, abrupt sentence after the initial request. The transition feels disjointed, and the correction is phrased as a new command rather than a natural follow-up to the initial request. The prosody is somewhat flat, and the overall delivery sounds less like a spontaneous thought process and more like a scripted correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_006", "hide_variant_original": true, "source_sha256": "c236b39b5dff725219f1a8791fa498216c67f1293cac991360c6a14eead4c587", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/cancellation/cancel_006/6854c92c7c880190/speech.wav", "wer": 0.0, "wav_sha256": "d7baf9e5c72d4efea0ffee035a210079ea4956ad2f59f2e592fc8bcdee411413", "seconds": 23.0, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete.", "Oh, wait, please cancel the entire task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests to mark all tasks as complete, but immediately corrects themselves with 'Oh wait, please cancel the entire task.' This constitutes a clear cancellation of the previously stated task. The final interjection is a direct command to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete. Oh wait, please cancel the entire task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "a787f124a621e87bb4fb402160c561fa362320bd1a2f95465b5fd5070ba73c8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/ambiguous_halt/halt_040/138227d19ec4fc91/speech.wav", "wer": 0.0, "wav_sha256": "cfd7ec1e9dd0e956e389d3b0c208356dff2566273dda0caed978ddc36fbd4a25", "seconds": 21.8, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes with 'Hang on, just a minute,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or request a rollback, but rather asks for a temporary halt. The initial request to 'do them for me and mark them all complete' is not executed due to this pause command. The audio ends with this pause request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "988af8e_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c533cb96a82bc2e3e650b2d4a7fca6712dd8d69406b9f6ce56005c3d4268244e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/988af8e_3/audio_natural/filler/original/301cb740fa1bc864/speech.wav", "wer": 0.0, "wav_sha256": "6cd1ae007f472518fa95ff6a21fd2004e27a99b5440ca10f4f3bb4c32c9a15d9", "seconds": 18.08, "spoken_utterances": [ "I am going on a camping trip with some of my friends. let's see... I have created a new Todoist project uh... for managing its tasks. um, let me think... Please do them for me, and mark them all complete." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let me think') and a pause ('Let's see') but does not change the core request. The final instruction 'Please do them for me and mark them all complete' clearly retains the original intent to mark all tasks in the project as complete. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "I am going on a camping trip with some of my friends. Let's see. I have created a new Todoist project uh for managing its tasks. Um, let me think. Please do them for me and mark them all complete." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_recipient_reference", "proposed_instruction": "Email the driving license found in my file system to my colleague—no, I mean partner.", "long_pause_offsets": [], "source_sha256": "bf3499dffd5a3995b57b3ebcb375efe3430bd81b0c7728a0184a792c32f0ff1d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'colleague' with 'partner'; the app, source, and content remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/false_start/false_start_03_recipient_reference/dedd3fcfda4825e2/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "59b13f15af749caaa5c43db1c8d98ce928d3f1be47a8b46177854b02e5f3c7da", "seconds": 9.8, "spoken_utterances": [ "Email the uh... driving license found in my file system to my ah... colleague—um... no, I mean partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'colleague' but immediately self-corrects to 'partner' with 'no, I mean partner.' This is a clear false start where the final intended target is 'partner'. The rest of the instruction matches the original, so the final intent is fully aligned with the reference.", "heard_transcript": "Email the uh driving license found in my file system to my uh colleague um no, I mean partner." }, "candidate_id": "wav_827fd6c90107bf41", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible false-start pattern. The speaker begins with 'email the uh driving license found in my file system to my uh colleague um no I mean partner.' The initial hesitation ('uh') and filler ('um') are placed naturally, and the self-correction ('no I mean partner') is a smooth, spontaneous-sounding revision of the recipient. The prosody includes a slight pitch drop on 'colleague' and a rise on 'partner,' indicating a genuine change of mind. The pacing and rhythm feel like a real person formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_02_attachment_source", "proposed_instruction": "Email the driving license found in my email inbox—sorry, make that my file system to my partner.", "long_pause_offsets": [], "source_sha256": "993638d6e44351d1fbe71371738493e0f5a46eb7bfc6f87a71a1187d80df3d8a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'email inbox' but immediately corrects it to 'file system'. The final source location is 'file system', matching the original instruction. The recipient and the item to be emailed remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/false_start/false_start_02_attachment_source/328f7495378e4ea6/speech.wav", "wer": 0.0, "wav_sha256": "f6b6d09d145877c7f5e74e28926cafc72d95821485a56a689a2d691df2ed021f", "seconds": 9.16, "spoken_utterances": [ "Email the driving license found in my email inbox—um... sorry, make that ... my file system to my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email inbox' but immediately self-corrects to 'file system'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Email the driving license found in my email inbox, um, sorry, make that my file system to my partner." }, "candidate_id": "wav_66d3ef985e9567a6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio contains a believable false start with 'email the driving license found in my email inbox um sorry make that my file system to my partner.' The speaker correctly identifies the error ('email inbox') and self-corrects to 'file system' with a natural-sounding 'sorry make that.' The pause and filler ('um') are appropriately placed, and the correction flows smoothly. However, the correction is slightly more abrupt than in audio_02, and the initial phrase 'email the driving license found in my email inbox' is less likely to be spontaneously said than the recipient error in audio_02, making it slightly less realistic overall." } ] }, "is_default": false }, { "variant_id": "false_start_01_attachment_description", "proposed_instruction": "Email the passport—I meant driving license found in my file system to my partner.", "long_pause_offsets": [], "source_sha256": "482554bad512ca2bec68ee32afd6eb69d0a428596b3a3efc6b2203f196a0b2b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'passport' to 'driving license' within the same utterance, preserving the recipient, source, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/false_start/false_start_01_attachment_description/c8cda936f26a85b2/speech.wav", "wer": 0.1875, "wav_sha256": "97f1409faf0dfa39ba2b063ccdc5c12ce6dfd6371a95ce5bdfe5db22c57cd639", "seconds": 7.04, "spoken_utterances": [ "Email the passport—uh... I meant driving license found in ... my file system to my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'passport' but immediately self-corrects to 'driving license'. This is a false start, and the final audible intent matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as stated.", "heard_transcript": "Email the passport, uh, I mean driving license found in my file system to my partner." }, "candidate_id": "wav_7cdda515c0ee8738", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start but sounds less realistic due to its abrupt and unnatural delivery. The speaker says 'email the passport uh I mean driving license found in my file system to my partner.' The correction 'I mean driving license' is spoken too quickly and without the natural hesitation or prosodic shift seen in the other clips. The transition from 'passport' to 'driving license' feels more like a mechanical substitution than a genuine moment of self-correction. Additionally, the initial word 'passport' is an odd choice to begin with, making the error less plausible in a real-world context." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Email the driving license to my partner. Oh, I forgot to mention, it's the one found in my file system.", "long_pause_offsets": [ 41 ], "source_sha256": "f26224cb6fb635691f9733d395076adda87e4271b6e41918af1b7955f30813da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint 'found in my file system' to the initial email request, preserving the recipient and item. The final goal combines both utterances." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/addition/addition_01/76b12dbd4b6c1c42/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "705542864cc1da53e1afecb475840034ead5db05529be4a5fc1825028a46f592", "seconds": 12.84, "spoken_utterances": [ "Email the uh... driving license ah... to my partner. ", "Oh, I forgot to mention, it's the one um... found in my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially begins the task and then adds a clarifying detail ('found in my file system') after a pause. According to the 'addition' type criterion, the final goal is the combination of the original instruction and the added information. The final audible intent matches the original instruction with the added constraint, so the task is correctly executed.", "heard_transcript": "Email the uh driving license uh to my partner. Oh, I forgot to mention, it's the one um found in my file system." }, "candidate_id": "wav_aa3615dafb55acfe", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('uh') and a self-repair ('uh' after 'the'), which is followed by a plausible, additive clarification ('Oh, I forgot to mention...') that smoothly integrates the missing detail. The prosody is consistent, and the pause before the addition feels like genuine recollection rather than a scripted insertion. The flow from initial request to addition is very natural." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Email the driving license found in my file system. One more detail: send it to my partner.", "long_pause_offsets": [ 51 ], "source_sha256": "b7766484916952973da27f534d5c92270d9e50646697dc120cdbe7e9f5f857d5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient, preserving the original action, object, and source. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/addition/addition_02/9000a5e703ecb413/speech.wav", "wer": 0.0, "wav_sha256": "52a39357896b9f4996abfc704962da6c0eedea84b4b93e3777fc764faaef52a3", "seconds": 11.84, "spoken_utterances": [ "Email the ... driving license found in uh, let's see... my file system. ", "One more detail: send it ... to my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Email the driving license found in my file system'. After a pause and the interjection 'One more detail', the speaker adds 'Send it to my partner'. This is a clear addition of a missing recipient. The final intent combines the original action with the added detail, matching the reference instruction. The addition is not a correction or revision but a completion of the original request.", "heard_transcript": "Email the driving license found in uh let's see my file system. One more detail. Send it to my partner." }, "candidate_id": "wav_b803ec3dfc25d0c2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a filler ('uh') and a plausible hesitation ('let's see'), the addition of 'one more detail' sounds slightly awkward and less spontaneous than a true memory recall. The phrase 'send it to my partner' is delivered as a separate, almost clipped command rather than a natural continuation of the prior sentence, making the overall flow feel slightly disjointed compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_attachment_source", "proposed_instruction": "Email the driving license found in my email inbox to my partner. Make that the driving license from my file system, not my email inbox.", "long_pause_offsets": [ 65 ], "source_sha256": "10c08b31933b40c29a40785930eaf34c41a7b540784d317cb310c387035d600d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the source of the driving license from 'email inbox' to 'file system', while retaining the recipient and the item to be sent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/revision/revision_02_attachment_source/5ee65ec888a5aa14/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "26421618520fcb15a34ac6489f68a5da95b73626aced8d95e43b7135865224e3", "seconds": 11.56, "spoken_utterances": [ "Email the driving license found in ah... my email inbox uh... to my partner. ", "Make that the driving license from my file system, not my email inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'email inbox' but immediately corrects it to 'file system' with a clear revision. The final intent matches the original instruction exactly, so the revision is correctly applied and the task is valid.", "heard_transcript": "Email the driving license found in my email inbox to my partner. Make that the driving license from my file system, not my email inbox." }, "candidate_id": "wav_cb3da53df96bfa43", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous prosody with a believable self-correction. The speaker starts with 'email the driving license found in my email inbox' and then smoothly revises to 'Make that the driving license from my file system, not my email inbox.' The correction is integrated into the same utterance with a natural shift in intonation and pacing, suggesting spontaneous thought. The pauses are brief and contextually appropriate, and the overall delivery sounds like a single, coherent thought process." } ] }, "is_default": true }, { "variant_id": "revision_03_recipient_reference", "proposed_instruction": "Email the driving license found in my file system to my my friend. Change the recipient to my partner, not my friend.", "long_pause_offsets": [ 67 ], "source_sha256": "1627c7c7fed1f1648893ae3926c993faa072e3fc8f470ce56f9b32367cf94fe7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recipient from 'my friend' to 'my partner', while retaining the file and action. The initial error is fully overridden." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/revision/revision_03_recipient_reference/6b4745be97cd777c/speech.wav", "wer": 0.125, "wav_sha256": "64a669abaa707aaffe2fa0ec5f61daa83ed394930f0875c44b26a460dc5ac209", "seconds": 12.64, "spoken_utterances": [ "Email the uh... driving license found in my file system to my my friend. ", "Change the recipient to my ah... partner, not my friend." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests sending the file to 'my friend', then immediately revises the recipient to 'my partner' with the explicit correction 'not my friend'. This matches the revision type criterion: the final intent is the corrected version, and the original incorrect value is discarded. The final goal aligns with the reference instruction.", "heard_transcript": "Email the uh driving license found in my file system to my my friend. Change the recipient to my uh partner, not my friend." }, "candidate_id": "wav_8ef131ce796c30ce", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with a clear change of mind, but the delivery is slightly less fluid than audio_01. The speaker says 'email the driving license found in my file system to my friend' and then corrects 'change the recipient to my partner not my friend.' The correction is a separate sentence, which is plausible, but the shift feels a bit more abrupt. The 'uh' filler and slight hesitation are natural, but the overall flow is less continuous than audio_01, making it slightly less realistic in terms of spontaneous speech." } ] }, "is_default": false }, { "variant_id": "revision_01_attachment_description", "proposed_instruction": "Email the passport found in my file system to my partner. Wait, the attachment should be the driving license, not the passport.", "long_pause_offsets": [ 58 ], "source_sha256": "b251d22dc4ce9bfab9c6a009177921ccedb20e0c1d6cae7f510ef49221fb307a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the attachment from 'passport' to 'driving license' while retaining the recipient and location; the final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/revision/revision_01_attachment_description/656180387a667f00/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "8e57427bb8c87c20357f63a623f04c6198d270f75f72c7755b036c47eb0d876e", "seconds": 12.48, "spoken_utterances": [ "Email the ... passport found in um, let me think... my file system uh... to my partner. ", "Wait, the attachment should be the driving license, not the passport." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'passport' but immediately corrects themselves after a pause and a self-correction ('Wait, the attachment should be the driving license, not the passport'). This constitutes a revision where the final intent is to send the driving license, matching the original instruction. The correction is clear and unambiguous, and the rest of the context (recipient, location) remains consistent.", "heard_transcript": "Email the passport found in um, let me think, my file system uh to my partner. Wait, the attachment should be the driving license, not the passport." }, "candidate_id": "wav_d4375888449bd2b4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains multiple elements that reduce its realism. The speaker starts with 'email the passport found in my file system to my partner' but then corrects to 'the attachment should be the driving license not the passport.' The initial mention of 'passport' is a significant semantic error, and the correction is delivered as a separate sentence with a noticeable pause and filler ('wait'). The delivery feels more like a mechanical correction than a natural thought process, and the shift from 'passport' to 'driving license' is abrupt and less integrated into the flow, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "d959eb36102dd4f7b3c9c1a89622b5773749b65b1347190c78db9f4e067ceed3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/cancellation/cancel_011/05ada78dc0c0853b/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "fd1ad12949261bf5378f23d150692e06bf761e7a95dcba3c07744cf97ae64319", "seconds": 13.88, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner.", "Hang on, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue the original instruction but then clearly revokes it with 'Hang on, I've changed my mind. Cancel that.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final interjection is unambiguous and complete, leaving no room for execution of the initial request.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner. Hang on, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "4c9a1a3e65dd73eb1ea27618de62b8aec30a12cf7a57e541840fcab716ed393e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/ambiguous_halt/halt_037/fac1b4b49c68acb6/speech.wav", "wer": 0.1, "wav_sha256": "07f9de549c46d91a1c5a8f6a607788de5c42a2c8310fd2ce91bcca026f6bbef1", "seconds": 12.36, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This aligns with the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request to email the document is not canceled, but the final audible intent is to pause, satisfying the type-specific criterion.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "32846bf8d67158d5c3df6d128d0b36ae3202d4b501c68b803a734cd70e043318", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_1/audio_natural/filler/original/f0032f033f17d681/speech.wav", "wer": 0.125, "wav_sha256": "3ce2a58a0cd46651cc8d6c64cbdb0daf206ff5c875dfc4f11398a34bd0aeddbc", "seconds": 9.08, "spoken_utterances": [ "Email the um... driving license found in ... my file system um, let's see... to my partner." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see') but does not change the core task. The final intent remains to email the driving license to the partner, matching the original instruction. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "Email the um driving license found in my file system um let's see to my partner." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_attachment_description", "proposed_instruction": "Email the resume—sorry, I meant headshot found in my file system to my manager.", "long_pause_offsets": [], "source_sha256": "d3b331d3056d8a9a2ff2986913a3607a3216695858e045d68aaa318797d6e2c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'resume' to 'headshot' within the same utterance, preserving the action, source, and recipient." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/false_start/false_start_01_attachment_description/6ab153fdbcefb98b/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "5e0711fb367d1187d43674b08aac25a9f1798c8e0566e81e16a98ffce1a512f4", "seconds": 9.04, "spoken_utterances": [ "Email the resume—um... sorry, I meant headshot found in my file system to let's see... my manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'resume' but immediately self-corrects to 'headshot'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the 'false_start' type criterion.", "heard_transcript": "Email the resume. Sorry, I meant headshot found in my file system to let's see, my manager." }, "candidate_id": "wav_e8cc112d3b213ad3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous false-start pattern. The speaker begins with 'resume', self-corrects to 'headshot' with a clear, audible hesitation ('um'), and then continues with a believable, slightly hesitant formulation of the recipient ('to let's see, my manager'). The pacing, filler usage, and repair are all highly realistic and mimic genuine human speech." } ] }, "is_default": true }, { "variant_id": "false_start_03_recipient_reference", "proposed_instruction": "Email the headshot found in my file system to my director—wait, I mean manager.", "long_pause_offsets": [], "source_sha256": "086121fa9741cf530278ca36cc216ec144d9b9d10d51090450787cf5a3612856", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'director' but immediately corrects to 'manager' within the same utterance. The final intended recipient is 'manager', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/false_start/false_start_03_recipient_reference/8fffd51b8a85b31f/speech.wav", "wer": 0.0625, "wav_sha256": "5ba5b4ecc8f2a9d133e5a5516e2a0f9bd1a82472774c0c38c053bc371fc27acb", "seconds": 10.28, "spoken_utterances": [ "Email the headshot found in my file system to ... my director—um... wait, I mean ... manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'director' but immediately self-corrects to 'manager' with 'Wait, I mean, manager.' This is a clear false start where the final intended target is 'manager'. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Email the headshot found in my file system to my director. Wait, I mean, manager." }, "candidate_id": "wav_69e7de3e897106f7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio has a clear false start with 'director' corrected to 'manager'. The repair is natural, but the initial part of the sentence ('email the headshot found in my file system to') is delivered very smoothly and without hesitation, making the correction feel slightly less spontaneous than in audio_03. The overall delivery is fluent but less varied in prosody." } ] }, "is_default": false }, { "variant_id": "false_start_02_attachment_source", "proposed_instruction": "Email the headshot found in my email inbox—no, that should be my file system to my manager.", "long_pause_offsets": [], "source_sha256": "45ef0086485d7d89032d1b8dfc7d895347f1b04e0433ee83d48ac592e73bf9e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email inbox' to 'file system' within the same utterance, preserving the recipient and the object to be sent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/false_start/false_start_02_attachment_source/f3688c890cb84cc8/speech.wav", "wer": 0.0, "wav_sha256": "29e9d102917ea26ec0e5a606d567a773af83df69fd705fb782aa5eaeba7a4a37", "seconds": 10.44, "spoken_utterances": [ "Email the headshot found in ... my email inbox—um... no, that should be ... my file system to my manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email inbox' but immediately self-corrects to 'file system'. The final utterance matches the original instruction exactly. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Email the headshot found in my email inbox. Um, no, that should be my file system to my manager." }, "candidate_id": "wav_d042b81c2bbc5a50", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'email inbox' corrected to 'file system'. However, the correction is abrupt and the speaker's intonation lacks the natural flow of a genuine self-repair. The phrase 'no, that should be' sounds more like a mechanical correction than a spontaneous thought process, and the overall delivery is flatter and less expressive than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Email the headshot to my manager. Oh, I forgot to mention, it's the headshot found in my file system.", "long_pause_offsets": [ 34 ], "source_sha256": "2733ce2549e96ae1516ad9dd565b5d098514d26160081560eedce7350dfab26b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source restriction 'found in my file system' to the initial email request, preserving the recipient and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/addition/addition_01/078c365e6a93e4d1/speech.wav", "wer": 0.0, "wav_sha256": "f7a8115a24af3aece19b409271400eecfab148633eed19e87818a71b010e5628", "seconds": 13.36, "spoken_utterances": [ "Email the um, let me think... headshot to my manager. ", "Oh, I forgot to mention, it's the headshot found in um... my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email the headshot to my manager' and then adds the clarifying detail 'it's the headshot found in my file system' after a longer pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is fully aligned with the original instruction, and the addition correctly specifies the source of the headshot.", "heard_transcript": "email the um, let me think. headshot to my manager. oh, I forgot to mention it's the headshot found in um, my file system." }, "candidate_id": "wav_b591f84560cb92d5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Email the headshot found in my file system to. I left out a detail: send it to my manager.", "long_pause_offsets": [ 47 ], "source_sha256": "8144a9609b8f69ee28996b6c6c09314b4bb8ca2c4fe314d3d158872ff632ae69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient, completing the initial incomplete instruction. The subject (headshot), source (file system), and action (email) are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Email the um... headshot found in my uh... file system to. ", "I left out a detail: send it to let me think... my manager." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_recipient_reference", "proposed_instruction": "Email the headshot found in my file system to my my assistant. Wait, the email should go to 'manager', not 'my assistant'.", "long_pause_offsets": [ 63 ], "source_sha256": "5c47cbd5e95fa00f89cb9247271d5fd8cfd481c00be38eb3f27cf4c1b1f1a2ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the recipient from 'my assistant' to 'manager', aligning with the original instruction. All other elements, including the subject (headshot), source (file system), and action (email), remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/revision/revision_03_recipient_reference/97f7ff3c482a264b/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "f069d93994cdfa86c5449a1d035b0dcb0b5f6af5158f0fd0450b2defb2563f11", "seconds": 17.2, "spoken_utterances": [ "Email the um, let me think... headshot found in my file system to hmmm... my my assistant. ", "Wait, the email should go to '... manager', not 'my assistant'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to my assistant' but immediately corrects themselves with 'Wait, the email should go to manager, not my assistant.' This constitutes a revision where the final intent is to send the email to the manager. The correction is clear and unambiguous, and the rest of the instruction remains consistent with the original. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Email the um, let me think, headshot found in my file system to my my assistant. Wait, the email should go to manager, not my assistant." }, "candidate_id": "wav_06ab41022843467c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear task, hesitates with 'let me think' and 'um', then self-corrects from 'my assistant' to 'manager' with a believable 'wait' and intonation shift. The prosody is continuous, the pauses are appropriately placed, and the correction feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "revision_02_attachment_source", "proposed_instruction": "Email the headshot found in my Google Drive to my manager. Use 'my file system' as the source instead of 'my Google Drive'.", "long_pause_offsets": [ 59 ], "source_sha256": "613de5e4b43dd0a61d975c2e7c46934d278f3e704a424da71280accf09d475b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the source from 'my Google Drive' to 'my file system', while retaining the action, target, and object." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/revision/revision_02_attachment_source/195b0afcb1c50ee8/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "271c92fb6b9f5d7756102e3844824e2d8db6f2348ea704b7bce1cf59be59703d", "seconds": 16.64, "spoken_utterances": [ "Email the ah... headshot found in uh, let me think... my Google Drive to my manager. ", "Use '... my file system' as the source instead of 'my Google Drive'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my Google Drive' but immediately corrects to 'my file system' after a pause, which is the final and intended source. The revision is clear and applies to the earlier instruction. The final goal matches the original instruction.", "heard_transcript": "Email the uh headshot found in uh let me think my Google Drive to my manager. Use my file system as the source instead of my Google Drive." }, "candidate_id": "wav_275148e68f6fc630", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a believable revision with a 'let me think' pause and a natural correction from 'my Google Drive' to 'my file system'. The prosody is generally smooth, and the speaker's hesitation and correction are contextually appropriate. However, the initial 'uh' and 'uh' fillers are slightly more abrupt and less integrated than in audio_01, making the delivery feel slightly less fluid." } ] }, "is_default": false }, { "variant_id": "revision_01_attachment_description", "proposed_instruction": "Email the portrait found in my file system to my manager. Change the attachment description to 'headshot' instead of 'portrait'.", "long_pause_offsets": [ 58 ], "source_sha256": "42d9c1f8d9ec2b0a111c78a3f18767e95b824a55805a3f7c04c6dd0d28a2cbd7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revises the attachment description from 'portrait' to 'headshot', while preserving the recipient, source, and action. The final goal includes the corrected description." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/revision/revision_01_attachment_description/2d0c8c3853d28c32/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "5847cb3c92d47f2aa69c689b945c5f6b35c915b2d16211214af1e438d0c54812", "seconds": 16.04, "spoken_utterances": [ "Email the let's see... portrait found in my file system to hmmm... my manager. ", "Change the attachment description to 'uh... headshot' instead of 'portrait'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially refers to the file as 'portrait' but then corrects it to 'headshot' in the revision clause. The final intent is to send the file with the updated description. The revision is clear and applies to the attachment description, preserving the rest of the instruction. The control intent is to execute the revised task.", "heard_transcript": "Email the let's see, portrait found in my file system to hmm, my manager. Change the attachment description to uh, headshot instead of portrait." }, "candidate_id": "wav_c5b5cc5310ea8e28", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but feels less realistic due to its structure. The speaker first states a complete request, then pauses with 'hmm', and only afterward adds a revision. The revision is delivered as a separate, almost disjointed sentence ('Change the attachment description...'), lacking the natural flow of a single, self-correcting utterance. The prosody is also flatter and less varied than in the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "c2361e37719efa765fbe18c14783779e295807740e333c817ee086e3b40a40b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/cancellation/cancel_002/920ad87051c51c2b/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "4de8be8887d35cbc63bdc980beb253f43427f3fb627b7ddfdb2c712eeb861aa8", "seconds": 14.48, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager.", "Oh, wait, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original email task but then clearly revokes it by saying 'cancel the whole task' as the final interjection. This meets the criterion for a cancellation, as the final audible intent is to cancel the task, not to proceed with it.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager. Oh wait, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "6385f84085e1a2151ed37a7978dbb3696a8de36dc52235d3793798ac239df53e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/ambiguous_halt/halt_004/41daca11dece2256/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "8ce7f641a5717c005d9b98baee7403ac0c7cdd067dd8c66fd04f089db0a20724", "seconds": 13.6, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager.", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second,' which is a clear request to pause or wait. The initial part of the audio is a hesitant but complete instruction to email a headshot, but the final spoken control intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager. Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5812330f99e9596a8ef44f1e9807deab7852352296d68571ad558bf63c03d5d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_2/audio_natural/filler/original/a9cf26c9520973cb/speech.wav", "wer": 0.0625, "wav_sha256": "ce5f55c1405a2531b481f328df28bb2e8dfccd37c6bfffafa8e48b2f97c2f8ae", "seconds": 10.08, "spoken_utterances": [ "Email the uh, let's see... headshot found in my ah... file system to uh... my manager." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see') but does not change the core task. The final intent matches the original instruction, and the fillers do not constitute a pause or cancellation. The task is executed as intended.", "heard_transcript": "Email the uh let's see headshot found in my uh file system to uh my manager." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_attachment_source", "proposed_instruction": "Email the birth certificate found in my email inbox—I should say my file system to my husband.", "long_pause_offsets": [], "source_sha256": "959b7d31c56f65961df8358ad36b25190900cd6f6928b55d53a4da873ede8f63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'email inbox' to 'file system' within the same utterance, preserving the recipient and the document identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/false_start/false_start_02_attachment_source/a606bd12cf30640f/speech.wav", "wer": 0.1, "wav_sha256": "cb382ae04dd3c7f4f579fc8c6264c48d91b55a0eb88fc78a8b8f9ef31433a788", "seconds": 10.36, "spoken_utterances": [ "Email the birth certificate found in um... my email inbox—hmmm... I should say my file system to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'email inbox' but immediately self-corrects to 'file system'. This is a false start, and the final, corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "Email the birth certificate found in um my email inbox. Hmm. I should say my file system to my husband." }, "candidate_id": "wav_de5d0968f286f585", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a plausible false start ('in um... my email inbox'), followed by a natural self-correction ('I should say my file system') and a smooth continuation. The intonation and pacing reflect genuine hesitation and correction, with a believable pause and filler ('hmm') before the correction. The flow is continuous and the repair is integrated into the utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "false_start_03_recipient_reference", "proposed_instruction": "Email the birth certificate found in my file system to my wife—no, I mean husband.", "long_pause_offsets": [], "source_sha256": "b8d8a4b8f21e8234f3e849628cd5c59378b9372549747271f044119cac02075e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'wife' with 'husband'; the app, document, and source remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/false_start/false_start_03_recipient_reference/4ca53e6ca41567cc/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "4113b8a824f99db8dc43ae862043c73d755933f4ad0ece021b269ce48ab2d698", "seconds": 9.16, "spoken_utterances": [ "Email the birth certificate found in my file system to my wife—ah... no, I mean um... husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to my wife' but immediately corrects to 'husband' after a self-correction ('No, I mean...'). According to the 'false_start' type criterion, the final corrected value ('husband') should be applied, and the original incorrect value ('wife') should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Email the birth certificate found in my file system to my wife. No, I mean, um, husband." }, "candidate_id": "wav_191a9dde38062fcb", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a correct phrase, then makes a false start ('to my wife at...'), followed by a self-correction ('No, I mean... husband'). While the correction is present, the initial phrase is fully formed before the error, and the correction is more abrupt and less integrated than in audio_01. The 'at' at the end of the first attempt sounds like a trailing thought rather than a natural hesitation, and the correction lacks the same organic flow." } ] }, "is_default": false }, { "variant_id": "false_start_01_attachment_description", "proposed_instruction": "Email the marriage certificate—sorry, I meant birth certificate found in my file system to my husband.", "long_pause_offsets": [], "source_sha256": "bfa1c8f1e86ebe4a1dd53e1eaec0c9140c6f97b2fafdb267982dd0d49227226b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'marriage certificate' to 'birth certificate' within the same utterance, preserving the recipient, source, and action." }, "audio": { "status": "pending", "spoken_utterances": [ "Email the hmmm... marriage certificate—ah... sorry, I meant ... birth certificate found in my file system to my husband." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Email the birth certificate found to my husband. Oh, I forgot to mention, the birth certificate is in my file system.", "long_pause_offsets": [ 49 ], "source_sha256": "80c8faceaa07904248bbdb698b6478d676ca535df2dae4cf374dd270fa2f86be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing location constraint 'in my file system' to the initial email request. The recipient and object remain unchanged, and the added detail clarifies the source without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/addition/addition_01/fedd3ed60a77778b/speech.wav", "wer": 0.12, "wav_sha256": "91d75615fa40ae9dfa0b4f50b9686a4f26d679d3d08036bfefa12a122fcba580", "seconds": 14.76, "spoken_utterances": [ "Email the uh... birth certificate found ... to my husband. ", "Oh, I forgot to mention, the birth certificate is uh, let's see... in my file system." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Email the birth certificate found to my husband' and then adds the location 'in my file system' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final intent is clear and complete, matching the original instruction with the added constraint.", "heard_transcript": "Email the uh birth certificate found to my husband. Oh, I forgot to mention, the birth certificate is uh, let's see, in my file system." }, "candidate_id": "wav_b3f53cd8d8017975", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural request, then pauses and adds a relevant detail ('the birth certificate is... let's see... in my file system') with a believable 'let's see' filler and a slight hesitation before completing the thought. The prosody is consistent, and the addition feels like a spontaneous memory rather than a rehearsed or awkward insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Email the birth certificate found in my file system. I left out a detail: email it to my husband.", "long_pause_offsets": [ 53 ], "source_sha256": "c234db2a15b087c6a15ffce4337f8fad73fd402160239c916596357cb1b272a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing recipient, preserving the original action, object, and source. The initial instruction omits the recipient, but the addition completes the full goal." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/addition/addition_02/ef85ff0442670406/speech.wav", "wer": 0.0, "wav_sha256": "70954755d9de156e3d80a80f9d5d5781ac8d69f8a4484ccaa1f4ad6a6e737b17", "seconds": 10.2, "spoken_utterances": [ "Email the ... birth certificate found ... in my file system. ", "I left out a detail: email it to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to email the birth certificate. After a pause, the speaker adds the detail 'Email it to my husband,' which specifies the recipient. This is a clear addition that completes the original instruction. The final intent matches the reference instruction, and the addition is not a correction or revision but a clarification of the recipient. Therefore, the task is fully specified and should be executed.", "heard_transcript": "Email the birth certificate found in my file system. I left out a detail. Email it to my husband." }, "candidate_id": "wav_39e1aba6d3dcada9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker first makes a complete request, then pauses and adds a correction ('I left out a detail. Email it to my husband'). While this is a plausible revision, the second part sounds slightly more rehearsed and less fluid than the first. The transition from the initial request to the correction is abrupt, and the phrase 'I left out a detail' is delivered with a flat intonation that lacks the natural hesitation or self-correction prosody seen in audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_attachment_source", "proposed_instruction": "Email the birth certificate found in my email inbox to my husband. Let me correct the source: find the file in 'my file system', not 'my email inbox'.", "long_pause_offsets": [ 67 ], "source_sha256": "99b13b1ad3eb47365b167d4321747ba8c27c6d66cbcdf14e3920b87bc95b0e4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'email inbox' to 'my file system', while retaining the recipient and the document identity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/revision/revision_02_attachment_source/a1c1ee6f7ad22e13/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "15be6f487487089c46f356f2a7385b54eadba35a9d11190cecbb6969d5e89d8c", "seconds": 16.88, "spoken_utterances": [ "Email the uh... birth certificate found in my email inbox ... to my husband. ", "Let me correct the source: find the file in 'um... my file system', not 'my email inbox'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'email inbox' but immediately corrects it to 'file system' after a pause. The revision is clear and final, and the rest of the instruction (action, target, and object) remains consistent. The final intent matches the original instruction, so the revision criterion is satisfied.", "heard_transcript": "Email the uh birth certificate found in my email inbox to my husband. Let me correct the source. Find the file in um my file system, not my email inbox." }, "candidate_id": "wav_e2658ea703f8c412", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural false start ('uh') and a self-corrected revision ('find the file in um my file system not my email inbox'). The correction is smoothly integrated into the same utterance, with a believable hesitation and a clear, audible change of mind. The prosody and pacing are consistent with spontaneous speech, and the repair is linguistically and pragmatically plausible." } ] }, "is_default": true }, { "variant_id": "revision_03_recipient_reference", "proposed_instruction": "Email the birth certificate found in my file system to my my wife. Wait, send the email to 'my husband', not 'my wife'.", "long_pause_offsets": [ 67 ], "source_sha256": "7f0dc8846978264787ed885c8164ce815c8e3b6f771b4cdc9f696cc545f3d37f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the recipient from 'my wife' to 'my husband', which matches the original instruction. The subject, source, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/revision/revision_03_recipient_reference/4171cc64d286e395/speech.wav", "wer": 0.03571428571428571, "wav_sha256": "0995f2dd2a0a0043f8c7a5b485099a0d89867c91f24be8cac0f8dfb7239e2e93", "seconds": 15.76, "spoken_utterances": [ "Email the birth certificate found in my file system uh... to my my wife. ", "Wait, send the email to 'um, let me think... my husband', not 'my wife'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to my wife' but immediately corrects themselves with 'Wait, send the email to... my husband, not my wife.' This constitutes a clear revision where the final target is 'my husband'. The original instruction was to send to 'my husband', so the final audible intent matches the reference instruction exactly.", "heard_transcript": "Email the birth certificate found in my file system uh to my my wife. Wait, send the email to um, let me think, my husband, not my wife." }, "candidate_id": "wav_378de83c315d05c0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker makes a revision by first stating the wrong recipient ('my wife'), then pausing and self-correcting to 'my husband'. The hesitation ('wait, send email to um let me think') is realistic and fits a moment of reconsideration. However, the correction is slightly less fluid than audio_01, with a more abrupt shift and a less natural-sounding 'let me think' filler. The prosody is slightly less consistent, with a noticeable drop in energy after the correction." } ] }, "is_default": false }, { "variant_id": "revision_01_attachment_description", "proposed_instruction": "Email the marriage certificate found in my file system to my husband. Change the attachment to 'birth certificate' instead of 'marriage certificate'.", "long_pause_offsets": [ 70 ], "source_sha256": "7cbda8c3bf415351b690e5673d3790d225ac4379d4c07a9845187f9429310712", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the attachment description from 'marriage certificate' to 'birth certificate', while retaining the recipient and source location." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/revision/revision_01_attachment_description/0870c7802b37f253/speech.wav", "wer": 0.0, "wav_sha256": "27191ac7c1cc13b2bbf4f3feffa9e33bd1efa43117713c485a80039aa2b4a92b", "seconds": 13.56, "spoken_utterances": [ "Email the marriage certificate found ... in my file system to my husband. ", "Change the attachment to '... birth certificate' instead of 'marriage certificate'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'marriage certificate' but immediately corrects it to 'birth certificate' in the same sentence. This is a revision where the final intent is to send the birth certificate. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Email the marriage certificate found in my file system to my husband. Change the attachment to birth certificate instead of marriage certificate." }, "candidate_id": "wav_6a5ece4c14772531", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a clear request but then adds a new, unrelated instruction ('change the attachment to birth certificate instead of marriage certificate') as a separate clause. This addition is not a self-correction of the initial request but a new command, making it less realistic as a single, spontaneous task formulation. The delivery is fluent but lacks the natural hesitation or repair flow seen in the other two, sounding more like a read-aloud or a composed instruction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_026", "hide_variant_original": true, "source_sha256": "b664a77b6f2ef53817e2f7feb74a3863261723f091210b574afccea13ae2eef2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/cancellation/cancel_026/400ca2ba6f6d36d8/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "86868e0e0f42b02869fb058dc0f76d944cb40646b89dd102e176d934d1880a7a", "seconds": 13.2, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband.", "Hang on, forget what I just asked. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by attempting to issue the email task but then explicitly states 'Hang on, forget what I just asked, cancel it.' This constitutes a clear and final cancellation of the original request. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Email the birth certificate found, let me think, in my file system, uh, let's see, to my husband. Hang on, forget what I just asked, cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "37e03bb8a121edac9c454576ae16b317153b2acf502aa0ce1e2e313d5b310096", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/ambiguous_halt/halt_024/fa21b5cdc3f61bd6/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "f8b2d0c1a965040e77739c00ccae09fd53ce5500f7cbd1a9c48508a107fddbc5", "seconds": 12.36, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but then repeatedly uses hesitation fillers ('let me think', 'uh, let's see', 'give me a moment to think') and does not complete the task. The final interjection is 'Give me a moment to think,' which is a clear request to pause. There is no cancellation or correction, and the speaker does not resume the task. This matches the 'ambiguous_halt' type, where the final intent is to pause.", "heard_transcript": "Email the birth certificate found, let me think. In my file system, uh, let's see, to my husband. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "98d2608_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2f8faad742c79c14584f8b43edf3b8b63a08216d7d664801d83c08f5bd36c235", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/98d2608_3/audio_natural/filler/original/7f49e1bad71506d9/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "2345e377afa7ade0d3b93287c5db31e1bef9c0cf5846143da7a6a1b676305ec8", "seconds": 8.4, "spoken_utterances": [ "Email the ... birth certificate found let me think... in my file system uh, let's see... to my husband." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation fillers ('let me think', 'uh', 'let's see') but does not change the core request. The final intent remains to email the birth certificate to the husband, matching the original instruction. The filler type allows ignoring such hesitations, so the task is executed as intended.", "heard_transcript": "Email the birth certificate found, let me think, in my file system, uh, let's see, to my husband." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all blenders—no, I meant food processors from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "1354b5fa6bd6a4d3cfac981a728a31f4be62ea54f4f683d0428794cc5e9aa40d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'blenders' to 'food processors' within the same utterance, preserving the action, source, destination, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/false_start/false_start_01_product_category/a2342ad636f0a7a8/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "e4eeb554589a50b4b1cde5008bf7f5473c87985887b1c7ad214334cc70ee5491", "seconds": 11.4, "spoken_utterances": [ "Move all uh... blenders—um... no, I meant hmmm... food processors from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blenders' but immediately self-corrects to 'food processors'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as stated.", "heard_transcript": "Move all uh blenders um. No, I meant um food processors from my Amazon cart to wish list." }, "candidate_id": "wav_fd4ecaf733aabe64", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible false start. The speaker initially says 'blenders', then clearly hesitates with 'um' and 'no', followed by a self-correction to 'food processors'. The correction is smooth, with a natural pause and intonation shift, and the final sentence flows without awkwardness. The prosody and pacing are consistent with someone thinking aloud and correcting a mistake in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all food processors from my amazon inbox—no, that should be cart to wish list.", "long_pause_offsets": [], "source_sha256": "bf4f9513d7dc7b3467af282e861ad7c726cc2b02f3f9bcb1b509a1cefaeb548c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'inbox' to 'cart' within the same utterance, preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/false_start/false_start_03_source_collection/3f62d5282367b88e/speech.wav", "wer": 0.2777777777777778, "wav_sha256": "7ae88ce363160d8d3fb44f2cce4f22deb4bacad0460a92add8a0d8d99dae8afd", "seconds": 11.36, "spoken_utterances": [ "Move all uh... food processors from my amazon inbox—um... no, that should be cart ... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately self-corrects to 'cart' with 'no, that should be cart'. According to the 'false_start' type criterion, the immediate self-correction must be applied and the mistaken value discarded. The final audible intent matches the original instruction with the correct target ('cart').", "heard_transcript": "Move all uh food processors from my Amazon inbox um no, that should be cart to wish list." }, "candidate_id": "wav_07ac545652ed693c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'inbox' corrected to 'cart'. The correction is immediate and the speaker continues smoothly. However, the initial word 'inbox' is less likely to be confused with 'cart' than 'blenders' with 'food processors', making the error slightly less plausible. The prosody is natural, but the correction feels a bit more abrupt than in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all food processors from my amazon cart to saved items—scratch that, make it wish list.", "long_pause_offsets": [], "source_sha256": "6623143706e3d45d9906d03946c368e8e840790e63185ededff0622dfe0f1fc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved items' but immediately corrects to 'wish list'. The final destination is correctly specified as 'wish list', matching the original instruction. All other elements (item type, source, action) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/false_start/false_start_04_destination_collection/1f43d3bec927ebb7/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "ce2f415033fe7d73e3ca7ae3560a57be83cc39fe4dcbf0e130b2498f13dccaba", "seconds": 10.84, "spoken_utterances": [ "Move all food processors ... from my amazon cart ... to saved items—uh... scratch that, make it wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved items' but immediately self-corrects with 'scratch that, make it wish list.' According to the 'false_start' type criterion, the final corrected value ('wish list') should be executed, and the initial incorrect value ('saved items') should be discarded. The final intent matches the original instruction, so the task is executed correctly.", "heard_transcript": "Move all food processors from my Amazon cart to saved items. Uh, scratch that, make it wish list." }, "candidate_id": "wav_01a15a115da49555", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip features a revision rather than a false start. The speaker first says 'saved items', then corrects to 'wishlist'. The correction is clear, but the phrasing 'scratch that' is a bit more formal and less conversational than the self-corrections in the other clips. The initial request is also slightly more complex, and the correction feels like a later decision rather than a spontaneous slip." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all food processors from my ebay—sorry, that should be amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "d7943f1deea334c522e0e596d02ef67cad616483467ecbe22028da6afe9e82b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon cart' within the same utterance, preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/false_start/false_start_02_app/bf2f882a0c96d5b3/speech.wav", "wer": 0.3888888888888889, "wav_sha256": "2dd622c346c733b2d401e44fdc18eb7037b39ec73517dfcc1b821377d44f4677", "seconds": 9.48, "spoken_utterances": [ "Move all food processors ... from my ebay—uh... sorry, that should be amazon cart um... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' and continues with the rest of the original instruction. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is executed with the correction applied. The final control intent is to execute the corrected task.", "heard_transcript": "Move all food processors from my eBay, um, sorry, that should be Amazon cart, um, to wish list." }, "candidate_id": "wav_bb1e8fb736c37e0c", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a false start with 'ebay' corrected to 'amazon cart'. However, the correction is less natural because 'ebay' is a different platform, and the speaker doesn't seem to realize the error until after saying 'amazon cart'. The flow is disrupted by the mention of two different platforms, making the correction feel less spontaneous. The prosody is also slightly more hesitant and less fluid than in the other clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Move food processors from my amazon cart to wish list. Oh, I forgot to say, move all food processors.", "long_pause_offsets": [ 55 ], "source_sha256": "d33402703f09acaf74256ba6aece5ce7feaf5397af7cd898490340ae0539577b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly adds the 'all' modifier to the initial request, which was missing in the first utterance. The final goal now includes the scope 'all' while preserving the app (Amazon), item type (food processors), and action (move from cart to wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/addition/addition_01/1e339a22d7de75c3/speech.wav", "wer": 0.12, "wav_sha256": "c7ec7bc3ee24e70dbcdb924159163417b77a54d61993e0ae2bfe38f33883664e", "seconds": 16.72, "spoken_utterances": [ "Move ... food processors from my amazon cart let's see... to wish list. ", "Oh, I forgot to say, move um, let me think... all food processors." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Move food processors from my Amazon cart to wish list' and then adds 'all food processors' after a pause and self-correction. According to the 'addition' type criterion, the final goal is the combination of the initial instruction and the added information. The final audible intent matches the original instruction exactly, with the addition of 'all' to clarify the scope. Therefore, the task is correctly executed.", "heard_transcript": "Move food processors from my Amazon cart, let's see, to wish list. Oh, I forgot to say, move, um, let me think, all food processors." }, "candidate_id": "wav_a899f5a82c59c02f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'let's see' and 'oh I forgot to say,' followed by a self-correction ('move um... let me think') before completing the full, coherent instruction. The pauses and fillers are well-placed and sound spontaneous, and the correction flows naturally within the same utterance. The prosody is consistent and the delivery is fluent, making it the most realistic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Move all food processors to wish list. I left out a detail: move them from my Amazon cart.", "long_pause_offsets": [ 39 ], "source_sha256": "e64d169ba180ce9175f2196edb5c5cb5b06e3c6ab3e51ba73b607707e01f94fc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source location 'from my Amazon cart', which was omitted in the initial utterance but present in the original instruction. The action, item, and destination remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/addition/addition_02/fcdf02027af98085/speech.wav", "wer": 0.25, "wav_sha256": "3b6b0582ff50124853df23f3f4e30a450990b1d97e70459d910d018840f0f652", "seconds": 12.2, "spoken_utterances": [ "Move all food processors hmmm... to wish list. ", "I left out a detail: move them um... from my Amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Move all food processors to wish list.' This was followed by a longer pause and an addition: 'I left out a detail, move them from my Amazon cart.' This addition clarifies the source of the items. The final intent combines the original action with the new source, matching the reference instruction. The addition type requires combining the earlier instruction with the added information, which has been done correctly.", "heard_transcript": "Move all food processors to wish list. I left out a detail, move them from my Amazon cart." }, "candidate_id": "wav_c09290588b938e7c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'hmm' before adding a correction. The correction is delivered as a follow-up statement rather than a self-repair within the same utterance, which is less natural. The prosody is slightly less fluid, and the correction feels more like a separate thought than a spontaneous recollection. While still realistic, it lacks the seamless integration of the first audio." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all food processors from my amazon inbox to wish list. Make that from 'cart', not 'inbox'.", "long_pause_offsets": [ 60 ], "source_sha256": "0c44f3601f8f3b2458e703d45d3ef078d55fa50532884c45d5f962b579a27bf4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'inbox' to 'cart', preserving the app (Amazon), item type (food processors), and destination (wish list)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/revision/revision_03_source_collection/fb955ca9000bc648/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "7704250f7e8312f0f5f8081d2ae96cee1bab2d89b28ddcf65e09a78836d99da4", "seconds": 15.08, "spoken_utterances": [ "Move all food processors ... from my amazon inbox uh, let's see... to wish list. ", "Make that from 'uh... cart', not 'inbox'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects to 'cart' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly.", "heard_transcript": "Move all food processors from my Amazon inbox, uh, let's see, to wish list. Make that from uh, cart, not inbox." }, "candidate_id": "wav_c4e3d9b65d9c0be9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and plausible revision flow. The speaker begins with a clear request, then hesitates with 'uh let's see'—a believable filler indicating momentary recall—before correcting 'inbox' to 'cart'. The correction is smoothly integrated into the same utterance with a natural back-reference ('Make that from...'), and the prosody maintains a consistent, conversational rhythm without robotic pauses or unnatural stress." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Move all food processors from my ebay cart to wish list. Let me correct the app: move from 'amazon' to 'wish list', not 'ebay'.", "long_pause_offsets": [ 57 ], "source_sha256": "87293ce463694414038a20863f7c35d6f541049ab683892c419c986e64742b26", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup corrects the app from 'ebay' to 'amazon', but incorrectly states the destination as 'wish list' instead of the original 'amazon wish list'. The original instruction specifies moving to the 'wish list' on Amazon, but the correction misrepresents the target as 'wish list' without specifying the app, which may imply a different platform or context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/revision/revision_02_app/6312fcab275c74c1/speech.wav", "wer": 0.21428571428571427, "wav_sha256": "66404be0a6a60731336d052eef77d63836dffd09b3806f6855ccf804fa14e4a1", "seconds": 15.24, "spoken_utterances": [ "Move all food processors uh, let me think... from my ebay cart ... to wish list. ", "Let me correct the app: move from 'amazon' to 'wish list', not 'ebay'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay cart' but immediately corrects it to 'Amazon' after a pause, stating 'Let me correct the app, move from Amazon to wish list, not eBay.' This constitutes a revision where the final intent clearly overrides the initial incorrect value. The final instruction matches the original reference.", "heard_transcript": "Move all food processors, uh, let me think, from my eBay cart to wish list. Let me correct the app, move from Amazon to wish list, not eBay." }, "candidate_id": "wav_e223e95e153f1284", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a self-correction. The speaker starts with 'food processors', hesitates with 'uh let me think', then makes a clear correction from 'eBay' to 'Amazon'. The phrasing 'Let me correct the app' is a pragmatic and plausible way to frame a revision, and the intonation supports the speaker's self-awareness. The only minor issue is a slightly abrupt transition after the correction, but overall, the flow and hesitation feel authentic." } ] }, "is_default": false }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Move all blenders from my amazon cart to wish list. Change the product category to 'food processors' instead of 'blenders'.", "long_pause_offsets": [ 52 ], "source_sha256": "0e020a38c4763d97406f5950cd8381f871ba71eb3d6b57dc9314036d4050f7a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the product category from 'blenders' to 'food processors', preserving the action, source, destination, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/revision/revision_01_product_category/836597701c982061/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "7d9fe7fa5081874b81f405af30f465dbd57c363150136612bb1531cd29575d2a", "seconds": 16.56, "spoken_utterances": [ "Move all blenders um... from my amazon cart uh... to wish list. ", "Change the product category to 'um... food processors' instead of 'blenders'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'blenders' but then corrects to 'food processors' after a pause, which matches the revision type criterion. The final audible intent is to move food processors, not blenders, so the revision is correctly applied and the task is executed as intended.", "heard_transcript": "Move all blenders um from my Amazon cart uh to wish list. Change the product category to um food processors instead of blenders." }, "candidate_id": "wav_a5cb2e3e7a6fdbdc", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision, but the correction is less smoothly integrated. The speaker starts with 'blenders', then corrects to 'food processors' mid-sentence. However, the pause before 'change the product category' feels slightly too long and artificial, and the correction is delivered with a flat, almost robotic intonation. The overall delivery is less conversational and more like a rehearsed or read instruction, reducing its realism despite a correct revision structure." } ] }, "is_default": false }, { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all food processors from my amazon cart to saved items. Use 'wish list' instead of 'saved items' as the destination.", "long_pause_offsets": [ 61 ], "source_sha256": "25e6bb99ae17e6e57f451a454d27fb70c0e14d43f518282ba45b2f29de199706", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination from 'saved items' to 'wish list' while preserving the item type and source cart." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/revision/revision_04_destination_collection/188f62e219a63a52/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "b29a322c04d1fe16ee1c44af8a966374dd341b5bd1530f2278ef2aa284a72e4e", "seconds": 15.96, "spoken_utterances": [ "Move all hmmm... food processors from my amazon cart ... to saved items. ", "Use '... wish list' instead of 'saved items' as the destination." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved items' as the destination but immediately corrects it to 'wish list' with the phrase 'Use wish list instead of saved items as the destination.' This constitutes a revision where the final intent is to use the wish list. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Move all food processors from my Amazon cart to saved items. Use wish list instead of saved items as the destination." }, "candidate_id": "wav_ac847242a65af4cb", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. It begins with a filler 'mhm' and then states a request that is immediately followed by a revision. However, the revision is phrased as a separate, declarative sentence ('Use wishlist instead of saved items') rather than a natural continuation or self-correction. The prosody is flat, and the transition between the two parts feels abrupt and disjointed, as if two separate thoughts were stitched together. The lack of natural hesitation or back-reference makes it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "0abaf1978c9bd8e3c86c1a111056d4797f44e468e75fde9eb96e4f8961c24cbc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/cancellation/cancel_029/231d4b1ae47b1802/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "63d2a5386215f3e1ca9d9f7896d1900bf8dad234aaf5da3e8488339362a0e111", "seconds": 9.76, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task, then clearly states 'Hang on, I'm withdrawing that request,' which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_031", "hide_variant_original": true, "source_sha256": "f2b7301e749e625f4ae686bdd85a42538c4c29ff5c0bc7ab65027768511ff086", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/ambiguous_halt/halt_031/e4aabbe4b0e3edc7/speech.wav", "wer": 0.1875, "wav_sha256": "dff00bb604dce449bf246ff9f513b2389fbe828e2bf0dfe549c1423b4eeb592d", "seconds": 9.44, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list.", "Hang on, I just..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, I just', which is a clear request to pause or hold. There is no cancellation of the task, nor is there a new instruction. The speaker is halting the process mid-sentence, indicating a temporary stop. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list. Hang on, I just" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "c3bf83013b3e12a431839ecb5f759eff63209ce363e5e96a0a6ddf447c079de3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_1/audio_natural/filler/original/1067a65260b4b848/speech.wav", "wer": 0.25, "wav_sha256": "69b227ab94606f5ae9175dda4d418c6f22e26c54bef3cafa5ee20e6ac4537b19", "seconds": 5.4, "spoken_utterances": [ "Move all food processors uh... from my amazon cart ... to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'uh' but does not change the core request. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Move all food processors uh from my Amazon cart to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all shoes—sorry, I mean watches from my amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "ea6c86358b7f0dfebece744cbe440f440065ad9bed3b6531ae75473660a528e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shoes' to 'watches' within the same utterance, preserving the app (Amazon), source (wish list), destination (cart), and action (move all)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/false_start/false_start_01_product_category/dbac95b97f453eb1/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "8061ee1afffad11df493a925ed7ae8b495dd25e513ebc315c841a4f8788e7655", "seconds": 8.48, "spoken_utterances": [ "Move all hmmm... shoes—um... sorry, I mean watches from uh... my amazon wish list to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shoes' but immediately self-corrects to 'watches'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, with only a minor grammatical repair ('I mean watches') replacing the mistaken 'shoes'. All other constraints (from wishlist to cart, on Amazon) are preserved.", "heard_transcript": "Move all um, shoes um, sorry, I mean watches from uh my Amazon wish list to cart." }, "candidate_id": "wav_f4f863e2aaf78c1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a believable 'uh' and 'um' hesitation before correcting 'shoes' to 'watches'. The correction is smooth, with a natural pause and intonation shift. The final request is fluent and complete, with no awkward phrasing or unnatural pauses. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all watches from my ebay—no, make that amazon wish list to cart.", "long_pause_offsets": [], "source_sha256": "18080d348866505b34d71ee5d93059bcedc348b7384484f200c84529a0e54e66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'ebay' with 'amazon wish list'; the action, item, and destination remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/false_start/false_start_02_app/f73c576ac7fe61df/speech.wav", "wer": 0.2, "wav_sha256": "63c4d23e4a279f73e28bb2647dc1720eeafe170c6b3bed3a7a7ef8260e075d3c", "seconds": 7.4, "spoken_utterances": [ "Move all ... watches from my ebay—uh... no, make that amazon wish list to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon wish list'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Move all watches from my eBay uh, no, make that Amazon wish list to cart." }, "candidate_id": "wav_e767315bb24a00c8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable self-correction from 'eBay' to 'Amazon wishlist'. The speaker uses a natural 'uh' and 'no' to signal the change, and the correction is integrated smoothly into the sentence. The prosody is consistent, and the pacing feels like a real-time thought process. The only minor issue is a slightly abrupt transition after 'no', but overall it sounds very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all watches from my amazon wish list to wishlist—sorry, make that cart.", "long_pause_offsets": [], "source_sha256": "3ab2686d22dc45c0d3a2e0f81b921d28b9b7258e124f38b5f03b36c11485e76c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'wishlist' but immediately corrects to 'cart'. The final destination is 'cart', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/false_start/false_start_04_destination_collection/13251792a342a8e9/speech.wav", "wer": 0.29411764705882354, "wav_sha256": "c3cc34db9a12c95a667d502cf2f6c9618547b958392b78f7f1845d2bb2acc03a", "seconds": 8.0, "spoken_utterances": [ "Move all um... watches from my amazon wish list to uh... wishlist—hmmm... sorry, make that cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'wish list' but immediately self-corrects to 'cart' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value ('cart') should be applied, and the original incorrect value ('wish list') should be discarded. The final intent matches the original instruction with the corrected target destination.", "heard_transcript": "Move all um watches from my Amazon wish list to uh wish list um sorry, make that cart." }, "candidate_id": "wav_375c59c75f2a9bd3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a revision from 'wishlist' to 'cart' after a pause. The correction is clear and the intonation shift is natural. However, the initial request is slightly awkward with the phrase 'to uh wishlist', and the speaker's voice sounds a bit more hesitant and less fluid than in the top two. The delivery is still credible but less smooth." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all watches from my amazon shopping list—wait, I meant wish list to cart.", "long_pause_offsets": [], "source_sha256": "3ffa4fafbe4d5c47f15d9db21d6d106620c18d088bc8962110e458bbf84c356a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'amazon shopping list' but immediately corrects to 'wish list' within the same utterance. The final target list is correctly specified as 'wish list', and all other elements (action, item, destination) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/false_start/false_start_03_source_collection/1ad4ec5f3c700165/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "cb6d013c6b3b13e79037c7dbb1ecc9857ccd1310068bd9c1b6ad799a6fd445b4", "seconds": 11.84, "spoken_utterances": [ "Move all uh, let's see... watches from my amazon shopping list—uh... wait, I meant wish list uh, let's see... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shopping list' but immediately self-corrects to 'wish list' after a pause and hesitation. According to the 'false_start' type criterion, the final corrected value ('wish list') overrides the initial mistaken value ('shopping list'), and the rest of the instruction remains consistent. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Move all uh let's see watches from my Amazon shopping list uh uh wait I meant wish list uh let's see to cart." }, "candidate_id": "wav_829558437637921f", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has multiple hesitations ('uh', 'let's see') and a false start ('wait I meant wish list'). While the self-correction is present, the delivery is less fluid and the pauses feel slightly longer and more disjointed. The speaker seems to be struggling to formulate the request, and the overall prosody is less natural compared to the others. The correction is less integrated and the flow is more fragmented." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Move watches from my amazon wish list to cart. Oh, I forgot to mention, move all watches.", "long_pause_offsets": [ 47 ], "source_sha256": "29fc221f0ce3c0c711cc8c7157ad6ac184c0139a714f82d2d02f6c07813ac02c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the 'all' modifier to the initial request, which was missing in the first utterance. The app (Amazon), source (wish list), destination (cart), and item type (watches) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/addition/addition_01/8a8acaf958d47776/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "61542bc82dc5a553fba13eb3fefe767368de3a0ba098cb8795bca26ef43473db", "seconds": 10.52, "spoken_utterances": [ "Move watches from um... my amazon wish list to cart. ", "Oh, I forgot to mention, move ... all watches." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was 'Move watches from my Amazon wish list to cart.' The speaker then added 'move all watches' after a pause, which clarifies the scope of the watches to be moved. The final intent combines the original action with the added constraint, resulting in 'Move all watches from my Amazon wish list to cart.' This matches the original instruction, which already specified 'all watches'. The addition does not introduce any conflicting or incorrect information.", "heard_transcript": "Move watches from um my Amazon wish list to cart. Oh, I forgot to mention, move all watches." }, "candidate_id": "wav_5adbb084c7b83aef", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Move all from my amazon wish list to cart. I left out a detail: only move watches.", "long_pause_offsets": [ 43 ], "source_sha256": "9ce288d9cb7bdaca235d0084f30419bd9eae72414de00b1221f1a1af0d144010", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that limits the action to only watches, which is consistent with the original instruction and does not contradict it." }, "audio": { "status": "pending", "spoken_utterances": [ "Move all from ... my amazon wish list um... to cart. ", "I left out a detail: only move uh... watches." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all watches from my amazon gift registry to cart. I misspoke about the source: move all watches from my amazon wish list to cart, not gift registry.", "long_pause_offsets": [ 55 ], "source_sha256": "1cba491add5de1c26bd921773990dd2e5447b2b6f12e0f91114467c699a05d78", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the source from 'gift registry' to 'wish list', aligning with the original instruction. The action, item, and destination remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/revision/revision_03_source_collection/5e9ffb289d35f918/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "e97ae6fb68552c51ecfbc5efa1348a5b411706c1140ebb82817587053bab097b", "seconds": 14.32, "spoken_utterances": [ "Move all watches from ... my amazon gift registry uh... to cart. ", "I misspoke about the source: move all watches from my amazon ... wish list to cart, not gift registry." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'gift registry' but immediately corrects themselves with 'I misspoke about the source' and clearly states the intended action as 'move all watches from my Amazon wish list to cart'. This is a revision where the final, corrected instruction matches the original reference. The correction is clear and unambiguous, so the task should be executed as revised.", "heard_transcript": "Move all watches from my Amazon gift registry uh to cart. I misspoke about the source, move all watches from my Amazon wish list to cart, not gift registry." }, "candidate_id": "wav_04af30934b368148", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request, then self-corrects with a natural-sounding 'I misspoke about the source' followed by a smooth, grammatically correct revision. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit. The pacing and intonation are very natural, with a believable pause before the correction." } ] }, "is_default": true }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Move all shoes from my amazon wish list to cart. Wait, move all watches from my amazon wish list to cart, not shoes.", "long_pause_offsets": [ 49 ], "source_sha256": "b35af779624dfd8baf484077f70274cb639cdcb5f11831bc5dd6b4a1d7057124", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial category from shoes to watches, preserving the app (Amazon), source (wish list), destination (cart), and action (move all)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/revision/revision_01_product_category/5f3bd4c4183d4091/speech.wav", "wer": 0.2692307692307692, "wav_sha256": "c632a9ff452f68d54f67a37994e032e71d8c9044972c3cdf022a25127d5af78b", "seconds": 15.48, "spoken_utterances": [ "Move all shoes from um... my amazon wish list ah... to cart. ", "Wait, move all um... watches from my amazon wish list to cart, not shoes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shoes' but immediately corrects to 'watches' after a pause and the word 'Wait'. This is a clear revision where the final intent is to move watches, matching the original instruction. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Move all shoes from um my Amazon wish list uh to cart. Wait, move all um watches from my Amazon wish list to cart, not shoes." }, "candidate_id": "wav_5956d847e3536f35", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'wait' before correcting to the right item. The correction is fluent and the intonation shift is natural. However, the initial 'shoes' is stated more confidently than the final 'watches', making the correction feel slightly less spontaneous than in audio_03. The flow is still very realistic, but the self-correction is less embedded in the initial thought process." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Move all watches from my ebay wish list to cart. Let me correct the app: move all watches from my amazon wish list to cart, not ebay.", "long_pause_offsets": [ 49 ], "source_sha256": "f0d894c2dd9ab976db90a96ce45f4f8185a7d86e2f9fe929bd6107cf56bbe97e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from ebay to amazon, preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/revision/revision_02_app/86f5ac24ce9aeae9/speech.wav", "wer": 0.27586206896551724, "wav_sha256": "22bc1c1840ec368b8950843d9a8c071c1b016e635c25f702708ab4b392f63270", "seconds": 13.48, "spoken_utterances": [ "Move all watches from hmmm... my ebay wish list um... to cart. ", "Let me correct the app: move all watches from my ... amazon wish list to cart, not ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially makes a mistake by saying 'eBay' but immediately self-corrects with 'Let me correct that' and provides the correct target as 'Amazon'. The revision is clear and final, and the corrected instruction matches the original reference. The control intent is to execute the corrected task.", "heard_transcript": "Move all watches from my eBay wish list to cart. Let me correct that. Move all watches from my Amazon wish list to cart, not eBay." }, "candidate_id": "wav_16ddfca31e1329ee", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then corrects 'eBay' to 'Amazon'. However, the correction is delivered with a slightly robotic, rehearsed cadence and lacks the natural hesitation or intonation shift seen in the other two. The 'let me correct that' is a very explicit self-correction phrase that sounds less spontaneous. The overall delivery is less fluid and more like a scripted edit than a natural thought process." } ] }, "is_default": false }, { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all watches from my amazon wish list to wishlist. Make that move all watches from my amazon wish list to cart, not wishlist.", "long_pause_offsets": [ 55 ], "source_sha256": "a7a82ad09a7afbb7ca2265de6d88f8c36e1d107adcc4ed9ee0d54efebd238eb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination from wishlist to cart, preserving the app (amazon), item type (watches), and source (wish list)." }, "audio": { "status": "pending", "spoken_utterances": [ "Move all watches from ... my amazon wish list to wishlist. ", "Make that move all watches from my amazon wish list to ... cart, not wishlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "de00bf7949c53c1a1d0eb55e67c75fdf9405b4eae1bf69a3ca6c470710c06efb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/cancellation/cancel_008/dd847dc2ffa21fa0/speech.wav", "wer": 0.08, "wav_sha256": "bcaaeecc09c5a665c28f247db12be994789fd3b649264a7811806781f42b3908", "seconds": 13.08, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart.", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly revokes it with 'Hang on, I want to cancel that task.' This constitutes a final cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart. Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "d02e3ad52d9ed45446825fd3bda909f9303e59acdd05ae896cec6056bdf655a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/ambiguous_halt/halt_023/e92d80531911bdfc/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "bbc884424a95fdcc12482437c5dc6ba20fad6b7ee488c040011e388f5f102702", "seconds": 12.6, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, I need a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial instruction to move watches from the Amazon wish list to the cart is not canceled or revised, and the speaker does not provide a new instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1821d3b69d7bbf622d99ade23b2dcf80ec90c492f1c352d604d7f114fbf56cb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_2/audio_natural/filler/original/345c162fb50a3e90/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "744ef41c7bc153abee4b372434d389d3bc6de1be7a75976dc128bdf9b5265bc7", "seconds": 8.44, "spoken_utterances": [ "Move all watches from um, let me think... my amazon wish list uh, let's see... to cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler phrases ('um', 'let me think', 'uh', 'let's see') but does not change the core request. The final audible intent is to move all watches from the Amazon wish list to the cart, matching the original instruction. The filler does not alter the goal, so the task should be executed.", "heard_transcript": "Move all watches from um, let me think, my Amazon wish list, uh, let's see, to cart." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_product_category", "proposed_instruction": "Move all chairs—I meant sofas from my amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "28f7f757c56b1a0d3d730eb683e96d166d4305677463c4c484df04f28f2a8b43", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'chairs' to 'sofas' within the same utterance, preserving the app (Amazon), action (move), source (cart), destination (wish list), and scope (all)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/false_start/false_start_01_product_category/94e8ba26b95768b9/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "88a9e8b1ca80fbf23b915f48967e6470cfb4ae81430e32137164c6829aea928e", "seconds": 8.16, "spoken_utterances": [ "Move all chairs—ah... I meant sofas um... from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'chairs' but immediately self-corrects to 'sofas' with 'I meant sofas'. This is a false start where the final corrected value replaces the initial mistaken one. The rest of the instruction matches the original, so the final intent is clear and executable.", "heard_transcript": "Move all chairs out. I meant sofas, um, from my Amazon cart to wish list." }, "candidate_id": "wav_e68ecd07af6db05f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('chairs'), immediately self-corrects with 'I meant sofas' in a natural, conversational tone, and then delivers the full, correct request smoothly. The correction is integrated into the flow without awkwardness, and the prosody sounds spontaneous and unscripted." } ] }, "is_default": true }, { "variant_id": "false_start_04_destination_collection", "proposed_instruction": "Move all sofas from my amazon cart to saved items—or rather, wish list.", "long_pause_offsets": [], "source_sha256": "1ce80d45e162b783153a1134737786fd9d1c621beaf15677490fab1e5c9faf5e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved items' but immediately corrects to 'wish list'. The final intended destination is 'wish list', matching the original instruction. The action, item type, and source remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/false_start/false_start_04_destination_collection/85ab1206131ca0c1/speech.wav", "wer": 0.3125, "wav_sha256": "efe70847095aa84a4b9519296fb067640b31fce8dea7b08cc04f2db02a2b3cc8", "seconds": 9.52, "spoken_utterances": [ "Move all sofas ah... from my amazon cart to saved items—uh... or rather, wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved items' but immediately self-corrects to 'wish list' after a pause and 'or rather'. According to the false_start type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction with the correct target ('wish list' instead of 'saved items').", "heard_transcript": "Move all sofas uh from my Amazon cart to saved items uh or rather wish list." }, "candidate_id": "wav_6e12be5271e3ecc2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with 'move all sofas' and hesitates with 'uh' before correcting 'from my Amazon cart to saved items' to 'wish list'. The self-correction is natural, and the pause and filler are placed in a way that suggests genuine recall. The delivery is slightly less fluid than audio_04 but still very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Move all sofas from my ebay—sorry, make that amazon cart to wish list.", "long_pause_offsets": [], "source_sha256": "700bd1ec65d8e4e34926be6cac3edc644e044d48741e5d6f14f4d5cbc795ec2b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon cart' within the same utterance, preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/false_start/false_start_02_app/4f61e4820b38515c/speech.wav", "wer": 0.2, "wav_sha256": "d4a1ef54f0a2136b1b53b060db292db619e7a3d30d08884da923ee2f02a2b3f4", "seconds": 10.56, "spoken_utterances": [ "Move all sofas from my ebay—um... sorry, make that ... amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon cart'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction with the correct target ('Amazon cart') and action ('move all sofas to wish list').", "heard_transcript": "Move all sofas from my eBay um, sorry, make that Amazon cart to wish list." }, "candidate_id": "wav_39f7be4eec75336a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker makes a false start ('eBay inbox'), then corrects to 'Amazon cart to wish list'. However, the correction 'Make that' sounds slightly more rehearsed or less spontaneous than in audio_04 or audio_01. The prosody is a bit flat, and the transition feels less organic, reducing its realism." } ] }, "is_default": false }, { "variant_id": "false_start_03_source_collection", "proposed_instruction": "Move all sofas from my amazon inbox—no, that should be cart to wish list.", "long_pause_offsets": [], "source_sha256": "a5d0eaf2e8662b5de9109936084ecf436ff0c30cec452220f94be8d3411089b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'inbox' to 'cart' within the same utterance, preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/false_start/false_start_03_source_collection/833ca92c816b707e/speech.wav", "wer": 0.25, "wav_sha256": "a600b7c701b3d437e6aa21d71b527d53e185243da2384fe2693613cd52c0043a", "seconds": 10.28, "spoken_utterances": [ "Move all sofas ... from my amazon inbox—um... no, that should be cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately self-corrects to 'cart to wish list'. This is a false start where the final corrected value overrides the initial mistake. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Move all sofas from my Amazon inbox. Um, no, that should be cart to wish list." }, "candidate_id": "wav_64e089152f6166a9", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker says 'move all sofas' and then hesitates with 'sh from my Amazon inbox'. The correction 'no, that should be cart to wish list' is delivered with a flat, almost robotic intonation. The pause and filler are unnatural, and the correction lacks the conversational flow of the others, making it sound less realistic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Move all sofas to wish list. Oh, and one more thing: move them from my Amazon cart.", "long_pause_offsets": [ 29 ], "source_sha256": "8062e2841858964da7cd64e9f9fdfab508799f4abc69683d6d2a65bf98b6a151", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing source ('from my Amazon cart') to the initial instruction, completing the full context. The action, target, and object remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/addition/addition_02/7d86f303c945f4e5/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "c738ed546d5510bcb1638a669f72a1097d54497da5b3480c9280793c0fc19a42", "seconds": 12.84, "spoken_utterances": [ "Move all um, let me think... sofas to wish list. ", "Oh, and one more thing: move them ... from my Amazon cart." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Move all um, let me think, sofas to wish list' is a filler-interrupted version of the original task. The speaker then adds the missing constraint 'move them from my Amazon cart' after a pause, which completes the original instruction. The addition clarifies the source of the sofas, matching the original intent. The final audible goal is fully aligned with the original instruction, so the addition type criterion is satisfied.", "heard_transcript": "Move all um, let me think, sofas to wish list. Oh, and one more thing, move them from my Amazon cart." }, "candidate_id": "wav_e9154ae626e25fad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a plausible self-initiated pause ('let me think'), followed by a clear initial request. After a brief pause, they add a clarifying detail ('Oh, and one more thing') with a conversational tone, smoothly integrating the additional instruction ('move them from my Amazon cart'). The prosody is varied, the pacing is human-like, and the addition feels spontaneous rather than forced." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Move sofas from my amazon cart to wish list. Oh, right, move all sofas from my Amazon cart to wish list.", "long_pause_offsets": [ 45 ], "source_sha256": "e34632c91bdfe07ea8cbf9b527d971bdd9bb81d916af2343bf5549c291daed38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing quantifier 'all' and corrects the capitalization of 'Amazon'; the action, object, and source destination remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/addition/addition_01/6ef638771bc8eed9/speech.wav", "wer": 0.34782608695652173, "wav_sha256": "c64150a66f582f7767b8b20a064265cec5a6bf07ddf58efcc6d0b169d1dd7d87", "seconds": 12.32, "spoken_utterances": [ "Move sofas uh... from my amazon cart hmmm... to wish list. ", "Oh, right, move all sofas from my Amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Move sofas' without the quantifier 'all', then corrects themselves with 'Oh right, move all sofas'. This is a clear revision where the final instruction includes the missing constraint 'all'. The addition of 'all' refines the initial request to match the original instruction exactly. The final intent is to execute the complete task as stated.", "heard_transcript": "Move sofas uh from my Amazon cart to wish list. Oh right, move all sofas from my Amazon cart to wish list." }, "candidate_id": "wav_58007f8cc7169902", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slight hesitation ('uh') and a minor self-correction ('from my Amazon cart'), then pauses before completing the request. The second part ('Oh right...') sounds like a belated realization, but the delivery is slightly rushed and the correction feels less organic than in audio_01. The overall flow is less natural, with a slightly robotic cadence and less convincing hesitation." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_destination_collection", "proposed_instruction": "Move all sofas from my amazon cart to shopping list. Change 'shopping list' to 'wish list'.", "long_pause_offsets": [ 53 ], "source_sha256": "ecc686430fba1d1fb1f36fb40e44664e457bb9251f253e2c67c48724382130f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the destination from 'shopping list' to 'wish list', preserving the action, item, and source." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/revision/revision_04_destination_collection/a6411487987031dd/speech.wav", "wer": 0.125, "wav_sha256": "357643a0bd82d7fe49dfc1cf5dd0ac1842c9d4720a4139d33832246ed2c9d828", "seconds": 18.92, "spoken_utterances": [ "Move all uh, let's see... sofas uh, let me think... from my amazon cart hmmm... to shopping list. ", "Change 'shopping list' to 'wish list'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'to shopping list' but then corrects it by saying 'Change shopping list to wish list.' This is a revision where the final intent is to move the sofas to the wish list, matching the original instruction. The correction is clear and final, so the task should be executed as revised.", "heard_transcript": "Move all uh let's see sofas uh let me think from my Amazon cart to shopping list. Change shopping list to wish list." }, "candidate_id": "wav_eef4189165cb18ed", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and complex revision flow. It begins with multiple fillers ('uh', 'let's see', 'let me think') and a drawn-out pause, suggesting genuine hesitation. The speaker then makes a clear initial request, followed by a pause and a believable correction of 'shopping list' to 'wish list'. The prosody is varied, with natural intonation shifts and a conversational rhythm, including a self-correcting 'change shopping list to wish list' that sounds like a spontaneous thought process rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Move all sofas from my ebay cart to wish list. Let me correct the app: use 'amazon', not 'ebay'.", "long_pause_offsets": [ 47 ], "source_sha256": "2015e1dad6637a4cf1e454724c8d985617078df54a9e0ec32bc5ce7b15c82638", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from 'ebay' to 'amazon', while retaining the action, item type, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/revision/revision_02_app/01e5b2a719d287f4/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "13ab73eccf3bb8942c0eaaa1ab840388defcebfd2f7924af3255bbfe4a759ace", "seconds": 13.92, "spoken_utterances": [ "Move all sofas um... from my ebay cart ah... to wish list. ", "Let me correct the app: use '... amazon', not 'ebay'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately corrects it to 'Amazon' after a pause, which matches the revision type criterion. The final audible intent is to move all sofas from the Amazon cart to the wish list, aligning with the original instruction.", "heard_transcript": "Move all sofas um from my eBay cart uh to wish list. Let me correct the app, use Amazon not eBay." }, "candidate_id": "wav_e61f18d102e778d9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible revision with a clear false start ('from my eBay cart') and a self-correction ('let me correct the app use Amazon not eBay'). The speaker uses a filler ('um') and a natural-sounding correction phrase. The prosody is slightly more robotic than audio_04, and the correction is less integrated into the flow, but the overall delivery is still realistic and reflects a genuine moment of correction." } ] }, "is_default": false }, { "variant_id": "revision_01_product_category", "proposed_instruction": "Move all chairs from my amazon cart to wish list. Wait, change 'chairs' to 'sofas'.", "long_pause_offsets": [ 50 ], "source_sha256": "4fd57a411d83f6d4481d35044f7332795ed0f62d4e48233d4d46eba73a4f4895", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the furniture category from 'chairs' to 'sofas', preserving the app, source, destination, and scope of 'all'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/revision/revision_01_product_category/5a64d7906d635784/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "205803830fbd4f0a0450439f6c913905f5f4e5a8879d6679b97dee0976dbec21", "seconds": 11.68, "spoken_utterances": [ "Move all let me think... chairs from my amazon cart to wish list. ", "Wait, change 'chairs' to '... sofas'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'chairs' but immediately corrects to 'sofas' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Move all let me think chairs from my Amazon cart to wish list. Wait, change chairs to sofas." }, "candidate_id": "wav_125d2b7b7f652634", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision where the speaker changes 'chairs' to 'sofas' after an initial request. The delivery is relatively fluent with a single filler ('wait'), but the revision is abrupt and lacks the natural hesitation or prosodic variation seen in the top two. The correction feels more like a simple edit than a spontaneous thought process, and the overall rhythm is slightly flat, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_03_source_collection", "proposed_instruction": "Move all sofas from my amazon wishlist to wish list. Make that 'cart', not 'wishlist'.", "long_pause_offsets": [ 53 ], "source_sha256": "77417da4c80b1d011dbe63f8a5a56e0a683f7e9c0f8c2e905174c7bdbc0f26da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'wishlist' to 'cart', preserving the action, item, and destination." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/revision/revision_03_source_collection/5f8681ff41bedf91/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "86594859895a3da4795edead4526aa9f382d461426f8fc3d59440a3391eb78bd", "seconds": 14.44, "spoken_utterances": [ "Move all sofas ... from my amazon wishlist um... to wish list. ", "Make that 'um, let's see... cart', not 'wishlist'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'wish list' but then corrects it to 'cart' after a pause and self-correction. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final intent is to move all sofas from the wish list to the cart, which matches the original instruction's action and target, with only the source and destination roles reversed. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Move all sofas from my Amazon wish list, um, to wish list. Make that, um, let's see, cart, not wish list." }, "candidate_id": "wav_482ed724cbcf2cd6", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a revision where the speaker corrects 'wishlist' to 'cart'. However, the phrasing 'make that um let's see cart not wishlist' is awkward and unnatural. The filler 'um' and 'let's see' are placed in a way that disrupts the flow, and the correction sounds like a forced insertion rather than a spontaneous thought. The prosody is monotonous, and the overall delivery feels less like a real person speaking and more like a scripted or rehearsed correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "4d4ce8afb9b99ae0c5ad81aaa492ae90f80721247638e76e8a8436dd3d830230", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/cancellation/cancel_037/0f2defce77972544/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "13ba680e86e2b7d2f318c9cf54f5cf732b68e3e1e41bcf72377b0efffeb964d4", "seconds": 10.8, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list.", "Wait, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but immediately corrects themselves with 'Wait, stop there' and explicitly states 'Cancel the whole task.' This constitutes a clear cancellation of the original instruction. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list. Wait, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "e16d4960cee5fee81601d6f605093d4236b5e739e848b71ff451a981a53ddc1e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/ambiguous_halt/halt_010/dba5f39bd81de734/speech.wav", "wer": 0.26666666666666666, "wav_sha256": "0e2d3eb52d3e59f76fe1d501c1d9912ce5a9e28733689f5a2c57de7cbfef011b", "seconds": 9.2, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original instruction and then audibly requests to pause with 'Hold on a second.' This constitutes a final, clear request to halt, matching the 'ambiguous_halt' type criterion. No cancellation or execution is requested.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "9bf2c8a_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "131a656060e4c52159c2ba0afd5cf39d674dda7a1f8d813b16e63cfed9632e90", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/9bf2c8a_3/audio_natural/filler/original/8fda3afedef963c2/speech.wav", "wer": 0.36363636363636365, "wav_sha256": "0ba9005e15a4e643c969bc90c6cfb1da156f088c29610dd5185f0bd2c2821309", "seconds": 5.48, "spoken_utterances": [ "Move all ... sofas uh... from my amazon cart to wish list." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'uh' but does not change the core task. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Move all sofas uh from my Amazon cart to wish list." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all artists of all jazz-genre—no, I meant classical-genre songs in any of my playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "9ac5e38c400d64dd1efe03423fa2e24e205c0ec2a1dea285462ea129f25194ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'jazz-genre' to 'classical-genre' within the same utterance. The rest of the instruction, including the app (Spotify), scope ('all artists'), and context ('in any of my playlists'), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/false_start/false_start_01_genre/252939907f9103b9/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "05488d130a0921791d35e2af859bcdde8a4f42187d8c7b8f2979176b53c793e9", "seconds": 10.84, "spoken_utterances": [ "Follow all artists of all jazz-genre—ah... no, I meant um... classical-genre songs in any of my playlists on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz genre' but immediately self-corrects to 'classical genre songs'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference instruction in both action ('Follow all artists') and target ('classical-genre songs in any of my playlists on Spotify').", "heard_transcript": "Follow all artists of all jazz genre. Ah, no, I meant um, classical genre songs in any of my playlists on Spotify." }, "candidate_id": "wav_ff091fd1c3796ad4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start and self-correction. The initial 'jazz genre' is clearly a slip, followed by a believable 'no, I meant um' with a filler that sounds like spontaneous hesitation. The correction to 'classical genre' is delivered with a slight shift in intonation and pacing, and the rest of the sentence flows smoothly. The prosody and rhythm are consistent with someone thinking aloud and correcting themselves in real time." } ] }, "is_default": true }, { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Follow all artists of all classical-genre songs in any of my albums—sorry, that should be playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "19e6aa79ec008b73487b181df86df93d17494e569b6756526ec59afd7acbcb3b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'albums' to 'playlists' within the same utterance, preserving the app, genre, and scope of artists. The final target is unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/false_start/false_start_02_source_collection/764ca1023ffb835a/speech.wav", "wer": 0.0, "wav_sha256": "4b7f3e15785eb650d0894fb56a31e360efa38f7ff6470a122f8e6ac496591fc1", "seconds": 11.36, "spoken_utterances": [ "Follow all artists of all classical-genre songs in ... any of my uh, let me think... albums—uh... sorry, that should be playlists on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction with the correction applied.", "heard_transcript": "Follow all artists of all classical-genre songs in any of my uh, let me think, albums uh, sorry, that should be playlists on Spotify." }, "candidate_id": "wav_f1382c20a6ddf223", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'albums' followed by a self-correction to 'playlists'. The 'uh' and 'let me think' are natural fillers, and the correction is delivered with a slight pause and a tone of realization. However, the correction is slightly less fluid than in audio_03, with a more abrupt shift and a slightly longer pause before the final word. The overall delivery is still very realistic but slightly less seamless than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow all artists of all classical-genre songs in any of my playlists on Apple Music—I should say Spotify.", "long_pause_offsets": [], "source_sha256": "17369891ebd00bd481f271026106d97179e29b0ac9f9807f12c1dd3d4de53b4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects Apple Music to Spotify within the same utterance, preserving the action, genre, scope, and playlist context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/false_start/false_start_03_app/6facc2ff5cd09774/speech.wav", "wer": 0.0, "wav_sha256": "8adc84c2d02b3ef89e9cf206b097cc241f15d6b6dc7e32ed12b08a7e20d5c870", "seconds": 10.04, "spoken_utterances": [ "Follow all artists of all classical-genre songs in any of my playlists on ... Apple Music—um... I should say ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the reference instruction exactly.", "heard_transcript": "Follow all artists of all classical-genre songs in any of my playlists on Apple Music. Um, I should say Spotify." }, "candidate_id": "wav_99ece2684b057b49", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a revision, correcting 'Apple Music' to 'Spotify'. The 'um' filler is natural, but the correction feels slightly more mechanical and less emotionally integrated than the other two. The shift from 'Apple Music' to 'Spotify' is abrupt, and the intonation doesn't fully convey the same sense of spontaneous realization. The pacing is also slightly more hesitant, with a longer pause after 'Apple Music' and a less fluid transition to the correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Follow all artists of all classical-genre songs in on Spotify. I left out a detail: this should apply only to my playlists.", "long_pause_offsets": [ 63 ], "source_sha256": "a4cf7bc8193cca50da0607cac6750041d7efa4c1a305e248818872ebac1aaeb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only to my playlists', which narrows the scope of the initial instruction without contradicting it. The app, genre, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/addition/addition_02/952fe4ab44a77a5a/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "07cae4b374c11d8f55e2dc37959f1c431b84d11783f9c04725f787b88f0cca23", "seconds": 16.88, "spoken_utterances": [ "Follow all artists of all uh... classical-genre songs in on um... Spotify. ", "I left out a detail: this should apply only to um, let's see... my playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a broad instruction to follow all classical artists across any playlists. After a pause and self-correction, they narrow the scope to 'my playlists' using 'This should apply only to...'. Since the type is 'addition', the final goal combines the original instruction with the added constraint. The final audible intent is a narrower version of the original, which matches the expected behavior for an addition. No other issues are present.", "heard_transcript": "Follow all artists of all uh classical genre songs in on um Spotify. I left out a detail. This should apply only to um let's see, my playlists." }, "candidate_id": "wav_745ffc675101b3b5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Follow all artists in any of my playlists on Spotify. Oh, I forgot to mention, only follow artists from classical-genre songs.", "long_pause_offsets": [ 54 ], "source_sha256": "7994ff7f849eb3acff394b95af733b7fb6c41485d795eff335b04fb979b3b7ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a genre restriction that narrows the initial broad request; the app, scope of playlists, and action remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all artists in uh... any of my playlists on Spotify. ", "Oh, I forgot to mention, only follow artists from ... classical-genre songs." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Follow all artists of all classical-genre songs in any of my playlists on Apple Music. Let me correct the app: follow the songs on Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "eea727e4b0b824862dd83db2ce8655ab25496422c4f60abbd97c6f101ae8734d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the genre, scope, and action. The final goal matches the original instruction with the correct app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/revision/revision_03_app/efba3c545ecaa6d3/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "e2266e0bca1ad40d3bbbbca6d9de903917f9287cee218ee457f1c774922262f7", "seconds": 15.92, "spoken_utterances": [ "Follow all artists of all uh... classical-genre songs in ... any of my playlists on um... Apple Music. ", "Let me correct the app: follow the songs on Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with only a minor paraphrase ('songs' vs 'artists of songs') that does not change the core task. The correction is clear and unambiguous.", "heard_transcript": "Follow all artists of all uh classical genre songs in any of my playlists on um Apple Music. Let me correct the app, follow the songs on Spotify not Apple Music." }, "candidate_id": "wav_72bd52d66a94b6db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant request, uses a believable 'let me think' filler, and then smoothly self-corrects the platform from Apple Music to Spotify. The prosody is consistent, the pauses are appropriately placed, and the correction is integrated into the same utterance without awkwardness. The speaker sounds like a real person formulating and refining a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Follow all artists of all classical-genre songs in any of my albums on Spotify. Make that all classical-genre songs in any of my playlists, not albums.", "long_pause_offsets": [ 80 ], "source_sha256": "aa904fe657b4bceed62bd2b27e28d2de864e7c9a8afa617b5d6f2682f7a341c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the target from 'artists of classical-genre songs' to 'classical-genre songs' and from 'albums' to 'playlists'. The final goal is no longer about following artists but about following songs, and the source context is playlists instead of albums." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/revision/revision_02_source_collection/3673197a7beebef2/speech.wav", "wer": 0.05405405405405406, "wav_sha256": "77e8b18ee53899fe8f8b7b181a50dd792dd9c1c374b6433e6e829e77ffa374f6", "seconds": 17.44, "spoken_utterances": [ "Follow all artists of all classical-genre songs in uh, let me think... any of my albums on uh... Spotify. ", "Make that all classical-genre songs in any of my um, let me think... playlists, not albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately corrects to 'playlists, not albums' after a pause and hesitation. This constitutes a revision where the final audible intent overrides the initial incorrect value. The final goal matches the original instruction with the corrected target ('playlists' instead of 'albums').", "heard_transcript": "Follow all artists of all classical-genre songs in a, let me think, any of my albums on Spotify. Make that all classical-genre songs in any of my, um, let me think, playlists, not albums." }, "candidate_id": "wav_82cb68aafc5cfd50", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision with multiple fillers ('let me think') and a correction from 'albums' to 'playlists'. While the speaker's delivery is natural and the correction is plausible, the flow is slightly less smooth than audio_03. The repetition of 'classical genre songs' and the backtracking to correct 'albums' to 'playlists' feels a bit more disjointed, and the second 'let me think' is less integrated than the first. The prosody remains natural, but the overall structure is more fragmented." } ] }, "is_default": false }, { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all artists of all rock songs in any of my playlists on Spotify. Wait, the genre should be classical-genre, not rock.", "long_pause_offsets": [ 72 ], "source_sha256": "9913327f8017176c6e39d92f5bfe7a739e4fa6ed2f38263137ed0f32cfab77e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the genre from 'rock' to 'classical-genre' while preserving the app, scope ('all artists'), and context ('in any of my playlists')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/revision/revision_01_genre/e66994148886474e/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "c0a60b668123e31250838f1271d12d9c6c09e22548f5ff10cbb847e9735aa32c", "seconds": 12.68, "spoken_utterances": [ "Follow all artists of all rock songs in um... any of my playlists on Spotify. ", "Wait, the genre should be ... classical-genre, not rock." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock songs' but immediately corrects themselves to 'classical genre, not rock' after a pause. This is a revision where the final intent is the corrected version. The original instruction was 'Follow all artists of all classical-genre songs in any of my playlists on Spotify.' The final audible intent matches this corrected instruction exactly, so the task is executed as revised.", "heard_transcript": "Follow all artists of all rock songs in um, any of my playlists on Spotify. Wait, the genre should be classical genre, not rock." }, "candidate_id": "wav_52380d40997be9ad", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker initially says 'rock songs' and then corrects to 'classical genre'. However, the correction is abrupt and lacks the natural flow of the other clips. The pause before 'wait' is long and the correction feels more like a separate thought than a self-repair within the same utterance. The prosody is less continuous, and the delivery sounds slightly more robotic or rehearsed compared to the others, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "80bd387e1b16a216a8db02575e5eacd6a1d873d935af757adceae0433d747171", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/cancellation/cancel_039/afcd8d7e81b3e92c/speech.wav", "wer": 0.08, "wav_sha256": "3aa7e120810d2059bbaebba4275c581ba6ebf02e45f83c11c83e95edd08c70a4", "seconds": 15.76, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify.", "Wait, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the task to follow artists of classical songs in playlists on Spotify. After a brief hesitation, they explicitly say 'Wait, don't go ahead with that. Cancel it.' This constitutes a clear cancellation of the entire task. The final control intent is cancellation, which matches the expected behavior for this type.", "heard_transcript": "Follow all artists of all classical genre songs in hmm any of my playlists on uh Spotify. Wait, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "432c351d0eaa67b3faa012f1e18475a286feb6cb5125995a632f9e979069a478", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/ambiguous_halt/halt_016/f574d0194ccd4ed6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3c667e57a128d7ab07d4154a9d2c6e4b0e5c96f886f7f1cdcb19e84e9c21a11b", "seconds": 14.96, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify.", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause request without cancellation or resumption. The initial instruction is not revoked or corrected, but the speaker explicitly halts further action.", "heard_transcript": "Follow all artists of all classical genre songs in any of my playlists on uh Spotify. Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e2fbe70145d03d745e38061acd69f23cd17b8e5f2eb216e3575e993cdfd55bc7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_1/audio_natural/filler/original/066fef5a5fbc1d78/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "9c88ee9211757627c5f454eb9d53e075b2b596d2d15025a2346506ed59d9eed1", "seconds": 10.96, "spoken_utterances": [ "Follow all artists of all ... classical-genre songs in hmmm... any of my playlists on uh... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('hmm', 'uh') but does not change the core task. The final intent remains to follow all artists of all classical-genre songs in any of the user's playlists on Spotify, matching the original instruction. The fillers do not alter the goal, so the task should be executed.", "heard_transcript": "Follow all artists of all classical genre songs in hmm any of my playlists on uh Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Follow all artists of all reggae-genre songs in any of my albums—make that playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "e0af0ccdd8a864e119d4d98afe2f319ae98b2d17ab903cfa8293d515403b94a2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately corrects it to 'playlists' within the same utterance. The final target is 'playlists', which matches the original instruction. All other elements—genre, action, scope, and app—remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/false_start/false_start_02_source_collection/5e96b2409c4004e6/speech.wav", "wer": 0.1, "wav_sha256": "abc2f3b7b5838002a448c57ffcd3dcaf118b1b1cea7cdd8620a85d0650447359", "seconds": 10.84, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in uh... any of my albums—hmmm... make that ... playlists on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with the phrase 'Make that playlists'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only the target object corrected from 'albums' to 'playlists'. All other conditions, including 'reggae-genre songs', 'all artists', 'any of my', and 'on Spotify', are preserved. Therefore, the task is executed as intended.", "heard_transcript": "Follow all artists of all reggae-genre songs in uh any of my albums. Hmm. Make that playlists on Spotify." }, "candidate_id": "wav_889c39346633ce5f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'albums', hesitates with 'hmm', then smoothly revises to 'playlists' with a clear, conversational 'make that' transition. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pacing and intonation suggest genuine thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all artists of all rock—I meant reggae-genre songs in any of my playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "19ef952d0ef0ad73c0e8dffc7406adfc991dda44526a1a59cf28c9a4fc289ffe", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'rock' to 'reggae-genre' within the same utterance, preserving the rest of the instruction including the app, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/false_start/false_start_01_genre/eaaea3d863b70800/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "0697a8902c07e47d15d69c89975000f65c453d5aaaff67dce880d8e93804f58e", "seconds": 11.08, "spoken_utterances": [ "Follow all artists of all rock—um... I meant reggae-genre songs in any of my playlists on um, let's see... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock' but immediately self-corrects to 'reggae'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('rock') is discarded. The final instruction matches the original reference instruction in action, target, and scope.", "heard_transcript": "Follow all artists of all rock um, I meant reggae genre songs in any of my playlists on um, let's see, Spotify." }, "candidate_id": "wav_d63643aaf7480ae5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'rock' followed by a self-correction to 'reggae'. The speaker uses 'I meant' to signal the correction, which is a natural phrase. However, the correction is slightly less fluid than audio_03, with a longer pause and a more abrupt shift. The 'let's see' filler is plausible but feels a bit more hesitant than the smoother revision in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow all artists of all reggae-genre songs in any of my playlists on Apple Music—or rather, Spotify.", "long_pause_offsets": [], "source_sha256": "0b0566785a7080754ee653b780c7105b9faa3007f4e936bd2cefe47646bb910b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify, which matches the original instruction. The genre, scope ('all artists', 'all reggae-genre songs'), and context ('in any of my playlists') are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/false_start/false_start_03_app/daeaf52ebda328fb/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "73d810e3b731b9bce731378b0f1fd0155ee5184e7d5517a35c27bcb30bc454bd", "seconds": 8.36, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in any of my playlists on Apple Music—um... or rather, uh... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction with the correct platform.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Apple Music. Um, or rather, uh, Spotify." }, "candidate_id": "wav_aaeb57cd3601e3b5", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'Apple Music' corrected to 'Spotify'. While the correction is present, the delivery is less natural. The speaker says 'uh rather uh' with a stutter-like repetition that sounds slightly unnatural. The correction lacks the smoothness of the other two, and the overall prosody feels flatter and less conversational, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Follow all artists of all reggae-genre songs on Spotify. I left out a detail: this should only apply to songs in my playlists.", "long_pause_offsets": [ 57 ], "source_sha256": "8bf4ef2e79c9cf82b4431c9d9a22426afb7ce89286f9e3e9e9161186e09af64a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint that the songs must be in the user's playlists, which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/addition/addition_02/b90d4fe68317d67a/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "ef83a4bcc98c3faa538185571e5d7c5578a36bee1b7d5d82184d76c35bf6fe92", "seconds": 14.8, "spoken_utterances": [ "Follow all artists of all ... reggae-genre songs on uh... Spotify. ", "I left out a detail: this should only apply to songs um... in my playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to follow all artists of all reggae-genre songs on Spotify. After a pause, they add a clarifying constraint: 'This should only apply to songs um in my playlists.' Since the addition is made after a longer pause and modifies the scope of the original request, the final intent is the combination of the original instruction with the added constraint. The final goal matches the reference instruction with the addition of a specific scope ('in my playlists'), which is a valid addition. Therefore, the task is executed with the combined intent.", "heard_transcript": "Follow all artists of all reggae-genre songs on uh Spotify. I left out a detail. This should only apply to songs um in my playlists." }, "candidate_id": "wav_f8313b4d112478b5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, fluent request, then naturally pauses and adds a crucial detail ('this should only apply to songs...') with a believable hesitation ('um') and a smooth, additive flow. The correction feels spontaneous and contextually grounded, fitting the 'addition' augmentation type well. The prosody is consistent, and the speaker's voice remains stable throughout, suggesting a single, continuous thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Follow all artists in any of my playlists on Spotify. Oh, I forgot to mention, only follow artists from reggae-genre songs.", "long_pause_offsets": [ 54 ], "source_sha256": "77a777b4587690ea944b63804e6289a3156c42adff606cd5d16a9a9360d10d35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the genre restriction 'reggae-genre songs' to the initial broad request, narrowing the scope without contradicting it. The app, playlist scope, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/addition/addition_01/5cf1c92b8e26b58b/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "02ec857330b836dbffce9ce6eab6563185e2b068314af016ee507f2082d88e56", "seconds": 13.44, "spoken_utterances": [ "Follow all artists in uh... any of my playlists on ... Spotify. ", "Oh, I forgot to mention, only follow artists from ... reggae-genre songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives the instruction to follow all artists in any playlist on Spotify. After a pause and an interjection, they add a constraint: 'only follow artists from reggae genre songs.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Follow all artists in uh any of my playlists on Spotify. Oh, I forgot to mention, only follow artists from reggae genre songs." }, "candidate_id": "wav_1ff1fce22132c359", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail, the initial request is less fluent and contains a slight stumble ('in uh any of my playlists'). The addition ('only follow artists from reggae genre songs') is delivered with a noticeable shift in intonation and a brief pause before the correction, which sounds more like a revision than a natural addition. The overall delivery feels slightly more rehearsed or segmented, with less organic flow compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all artists of all rock songs in any of my playlists on Spotify. I misspoke about the genre; use reggae-genre instead of rock.", "long_pause_offsets": [ 72 ], "source_sha256": "8719a872a59d6d12560d6e9a29ad815071f115247817fd53c98751c75b2e9854", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from 'rock' to 'reggae-genre', while retaining the app, scope ('all artists'), and context ('in any of my playlists')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/revision/revision_01_genre/9fcbd5dd9a474444/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "d5196a1b8f7d9d93b2940b876b25cd826ccc776dd1caf6a01f61308e8b572954", "seconds": 16.76, "spoken_utterances": [ "Follow all artists of all uh... rock songs in any of my playlists on uh, let me think... Spotify. ", "I misspoke about the genre; use ... reggae-genre instead of rock." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rock songs' but immediately corrects themselves with 'I misspoke about the genre. Use reggae genre instead of rock.' This constitutes a revision where the final intent is to follow artists of reggae songs, matching the original instruction. The correction is clear and applied to the earlier instruction, with no other conflicting conditions.", "heard_transcript": "Follow all artists of all uh rock songs in any of my playlists on uh let me think. Spotify. I misspoke about the genre. Use reggae genre instead of rock." }, "candidate_id": "wav_43e586437682c2d5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible revision flow. The speaker begins with a clear request, inserts a realistic 'let me think' pause, then self-corrects the genre from 'rock' to 'reggae' with a believable 'I misspoke about the genre' phrase. The correction is smoothly integrated into the same utterance, maintaining a natural rhythm and prosody. The speaker's hesitation and self-correction feel spontaneous and not rehearsed." } ] }, "is_default": true }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Follow all artists of all reggae-genre songs in any of my albums on Spotify. Change the source to playlists, not albums.", "long_pause_offsets": [ 77 ], "source_sha256": "72e8940976b9904f8dda4bc0cbdd0dcb50d5601ff8cf8c160f1dffbbe926ff9f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the source from 'albums' to 'playlists', preserving the genre, action, and scope of 'all artists'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/revision/revision_02_source_collection/795560b44c8e4ee0/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "4e78e69cd4e9f7dd498b1a8bb6fd5f760971ffcc815b78c5f499923967da6b08", "seconds": 15.16, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in um... any of my albums on uh, let me think... Spotify. ", "Change the source to playlists, not albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately corrects to 'playlists' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with only the source changed from 'albums' to 'playlists'. All other constraints (reggae-genre, artists, Spotify) are preserved.", "heard_transcript": "Follow all artists of all reggae-genre songs in um, any of my albums on uh, let me think, Spotify. Change the source to playlists, not albums." }, "candidate_id": "wav_b38ce5682b44e4f1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a 'let me think' pause and a clear correction from 'albums' to 'playlists'. The speaker's delivery is natural, with a slight hesitation and a smooth transition to the correction. The prosody and pacing are consistent with a person formulating a request and then adjusting it, though the initial request is slightly less fluent than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Follow all artists of all reggae-genre songs in any of my playlists on Apple Music. Use Spotify, not Apple Music.", "long_pause_offsets": [ 84 ], "source_sha256": "0e56aee3a63d06a27ad966454cd3e4df846c26d05de171f129712b380b9a4d9a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the genre, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/revision/revision_03_app/4fd562b9d6280382/speech.wav", "wer": 0.0, "wav_sha256": "c764fc4104ac540a34564cc45f69e0ad4409ccf2d43fbd8bea62b26dc72969f3", "seconds": 13.76, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Apple Music. ", "Use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the incorrect platform (Apple Music) but immediately corrects it to Spotify. According to the revision type criterion, the final instruction after the correction is the one that should be executed. The correction is clear and unambiguous, and the rest of the task (following all artists of all reggae-genre songs in any playlist) remains consistent. Therefore, the final intent matches the original reference instruction.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Apple Music. Use Spotify, not Apple Music." }, "candidate_id": "wav_4eb6bff5e10a0ab9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a revision, the flow feels less natural. The speaker states the incorrect platform ('Apple Music'), then corrects it to 'Spotify'. However, the correction is delivered as a separate, abrupt sentence rather than being integrated into the original request. The prosody is flatter and more robotic, and the revision lacks the smooth, conversational integration seen in the other clips, making it sound less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "fbb4d99727d9a6ad1ce394d12df160bcac1248c5ccc97ff74bab569e5aa0af3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/cancellation/cancel_012/6c4dede73afabcbb/speech.wav", "wer": 0.0, "wav_sha256": "3131b577f8a720b32a3a9abfa2bf0b901293fd38c8153dd5850d36ffc9095e34", "seconds": 13.04, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify.", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then after a pause and the phrase 'Wait, actually, I've changed my mind,' explicitly says 'Don't do that.' This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute, and the cancellation applies to the entire task as originally stated.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify. Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "9b9f40bbf7eb04423068fd36cd231eb99eae6220fd42899d26fd8d595fe66949", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/ambiguous_halt/halt_011/96383dbaefd9a8cd/speech.wav", "wer": 0.0, "wav_sha256": "f0464c6466c27948dfcde66291bfd5cc02fae18ccff35806a80b5ab81fc236a0", "seconds": 11.04, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify.", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on a second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task but asks for a temporary halt. The initial instruction is fully audible and complete, and the final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify. Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "eafd29ea67b1c644f7b34391419ed6675928bf854f4efefa22e97f534b32ba66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_2/audio_natural/filler/original/2e70c3c6849b50a9/speech.wav", "wer": 0.0, "wav_sha256": "9184d9ec1374be12f78ae93286d62cebbd6a8b120f7a566c5703f48df32384ed", "seconds": 7.76, "spoken_utterances": [ "Follow all artists of all reggae-genre songs in ... any of my playlists on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not change the goal. Since no such hesitation is present, the task is executed as stated.", "heard_transcript": "Follow all artists of all reggae-genre songs in any of my playlists on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_source_collection", "proposed_instruction": "Follow all artists of all indie-genre songs in any of my albums—no, make that playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "9f71b987c873e3a63065f989ce6d63e6a72b6e608b19b82a14de1ad87fd9579a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'albums' to 'playlists' within the same utterance. The final target is 'playlists', which matches the original instruction. All other conditions—genre, app, and scope—remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/false_start/false_start_02_source_collection/8d6cd57b5082367c/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "ae3d39d19d36217c372e6e905b9cae2a6ae82c210afade2de3f7212e83501ef4", "seconds": 10.28, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my ... albums—uh... no, make that playlists on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'albums' but immediately self-corrects to 'playlists' with 'no make that playlists'. This is a false start where the final corrected value is 'playlists'. The rest of the instruction matches the original. The final intent is clear and executable.", "heard_transcript": "Follow all artists of all indie genre songs in um any of my albums uh no make that playlists on Spotify." }, "candidate_id": "wav_d904a3fcfeafeff9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'um' and 'uh' before self-correcting 'albums' to 'playlists' with a smooth, plausible intonation shift. The correction is immediate and integrated into the flow, sounding like a genuine moment of recall. The prosody is consistent, and the delivery feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow all artists of all indie-genre songs in any of my playlists on Apple Music—sorry, I should have said Spotify.", "long_pause_offsets": [], "source_sha256": "a76560decd2275ebb5c9758c06e1f25971255338ea2709e553e6320f97eb3d5b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the music app from Apple Music to Spotify, while retaining all other conditions including the genre, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/false_start/false_start_03_app/0a2a6f974fb16f7e/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "bfd6cc7a0c6ea5b36805a418b05748f6208a8b627dd87f9d4228bcef182b69d4", "seconds": 10.8, "spoken_utterances": [ "Follow all artists of all indie-genre songs in any of my playlists on Apple Music—ah... sorry, I should have said ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value must be applied, and the original incorrect value is discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Follow all artists of all indie-genre songs in any of my playlists on Apple Music. Sorry, I should have said Spotify." }, "candidate_id": "wav_7d7c12b2c1e65d7f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a longer, more complex initial request with a natural pause and filler ('uh') before the correction. However, the correction 'sorry, I should've said Spotify' is delivered with a slightly more formal, rehearsed tone and a longer, less organic pause than in audio_01. While still realistic, the delivery feels slightly less spontaneous and more like a scripted self-correction." } ] }, "is_default": false }, { "variant_id": "false_start_01_genre", "proposed_instruction": "Follow all artists of all rock—I meant indie-genre songs in any of my playlists on Spotify.", "long_pause_offsets": [], "source_sha256": "0a4cdb82a118e0fa5f5327ca81ca77e335ce5fa04b9d805ec7de09432541b958", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'rock' to 'indie-genre' within the same utterance, preserving the app, scope ('all artists'), and context ('in any of my playlists')." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all artists of all rock—um... I meant indie-genre songs in any of my playlists on uh... Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Follow all artists in any of my playlists on Spotify. Oh, I forgot to mention, only follow artists from indie-genre songs.", "long_pause_offsets": [ 54 ], "source_sha256": "1e8ac3b71fd3ad642bcc68d51e7b389b6e87a39a7082fc47943bbaf26d4cec4c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing genre restriction, narrowing the initial broad request to only include artists from indie-genre songs, while preserving the app, scope, and action." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all artists in uh... any of my playlists on ... Spotify. ", "Oh, I forgot to mention, only follow artists from indie-genre songs." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Follow all artists of all indie-genre songs on Spotify. One more detail: only follow artists from songs in any of my playlists.", "long_pause_offsets": [ 56 ], "source_sha256": "316616a27910029c38a46fa96364d60c95b0174e80ba95276893be7217738d8e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a constraint that the indie-genre songs must be from the user's playlists, which refines the initial broad request. The core action, genre, and app remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all artists of all let's see... indie-genre songs on ... Spotify. ", "One more detail: only follow artists from songs in ... any of my playlists." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Follow all artists of all indie-genre songs in any of my playlists on Apple Music. Let me correct the app: follow all indie-genre songs in any of my playlists on Apple Music, not Spotify.", "long_pause_offsets": [ 83 ], "source_sha256": "927dda8977f3a8ba62d31c7b67c6102fffce260d4ecdc8528dd99e9724d6c6fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the target from 'artists' to 'songs' and corrects the app from Spotify to Apple Music. The original instruction required following artists, but the final goal is to follow songs." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/revision/revision_03_app/f909811add2b05a0/speech.wav", "wer": 0.05128205128205128, "wav_sha256": "db5dda4489b9a700d9488fd499eefa7930ade38172e82e499034597d7ab08308", "seconds": 18.2, "spoken_utterances": [ "Follow all artists of all um... indie-genre songs in uh... any of my playlists on um... Apple Music. ", "Let me correct the app: follow all indie-genre songs in any of my playlists on Apple Music, not Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Follow all artists of all indie-genre songs in any of my playlists on Spotify' but then immediately corrects the app from 'Spotify' to 'Apple Music'. The revision is clear and final, and the rest of the instruction remains consistent. The final intent matches the original instruction with the corrected app, satisfying the revision type criterion.", "heard_transcript": "Follow all artists of all um indie genre songs in uh any of my playlists on um Apple Music. Let me correct the app. Follow all indie genre songs in any of my playlists on Apple Music, not Spotify." }, "candidate_id": "wav_cbfca69f42434621", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable self-correction. The initial request is slightly hesitant with filler words ('um', 'uh'), which is realistic. The correction ('not Spotify') is smoothly integrated with a clear change in intonation and pacing, indicating a genuine moment of realization. The prosody and rhythm are consistent with spontaneous speech, and the repair is linguistically and pragmatically plausible." } ] }, "is_default": false }, { "variant_id": "revision_02_source_collection", "proposed_instruction": "Follow all artists of all indie-genre songs in any of my albums on Spotify. Make that all indie-genre songs in any of my albums, not playlists.", "long_pause_offsets": [ 76 ], "source_sha256": "d331bd9ae15825e968925115ac0f756a81473d5983a97b8fd2b355ce35efa13a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup revision changes the target from 'artists' to 'songs' and removes the 'all' scope over artists, while preserving the genre and source (albums). This alters the core action and object of the instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/revision/revision_02_source_collection/ad5ad27c68dbb132/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "981a6c3a045feaa91ee42cf4bd2aba8918fc1beb8b713aeb67ec29812859aaf3", "seconds": 16.68, "spoken_utterances": [ "Follow all artists of all ... indie-genre songs in uh, let me think... any of my albums on Spotify. ", "Make that all indie-genre songs in any of my albums, not uh... playlists." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Follow all artists of all indie-genre songs in any of my playlists on Spotify', then after a pause and self-correction, says 'Make that all indie-genre songs in any of my albums, not uh playlists'. This is a revision where the target changes from 'playlists' to 'albums'. The final intent is clear and matches the revised instruction. The original instruction is overridden by the correction, and the final goal is executable.", "heard_transcript": "Follow all artists of all indie-genre songs in a let me think any of my albums on Spotify. Make that all indie-genre songs in any of my albums, not uh playlists." }, "candidate_id": "wav_f0b8418d38330b12", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a realistic self-correction, but the repair is less fluid. The speaker says 'let me think' as a pause, which is natural, but the correction ('not uh playlists') includes a filler ('uh') that disrupts the clarity of the intended change. The phrase 'make that all indie genre songs in any of my albums' is slightly awkwardly phrased, and the final 'not uh playlists' sounds like a rushed, less confident correction. While still realistic, the delivery is less polished and the repair feels slightly more mechanical compared to audio_01." } ] }, "is_default": false }, { "variant_id": "revision_01_genre", "proposed_instruction": "Follow all artists of all rock songs in any of my playlists on Spotify. Wait, the genre should be indie-genre, not rock.", "long_pause_offsets": [ 72 ], "source_sha256": "409c02492256373fd272626f83335681f8d0c944ecc4cf9b20fb037efda15d67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from 'rock' to 'indie-genre', while retaining the app, scope ('all artists'), and context ('in any of my playlists')." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow all artists of all uh... rock songs in ah... any of my playlists on um... Spotify. ", "Wait, the genre should be indie-genre, not rock." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_032", "hide_variant_original": true, "source_sha256": "f81b3b35d440637ab27a0bf272f2641dc6c07fcc48d9b081d8753b6f4e6b6324", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/cancellation/cancel_032/e2962d0aeae4ee42/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "f3628cc9de57a95cd522a951e31f0a3a5344ca4820eae32c202e70252249d436", "seconds": 12.72, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify.", "Hold on, I don't need this anymore. Please cancel the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to issue a request but immediately self-cancels with the explicit phrase 'Hold on, I don't need this anymore. Please cancel the request.' This constitutes a clear and final cancellation of the task. The initial part of the audio is a false start, but the final control intent is unambiguous.", "heard_transcript": "Follow all artists of all indie-genre songs in um any of my playlists on let's see. Hold on, I don't need this anymore. Please cancel the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "604b6723e0843020634b15cc25c4d9d2794f4216923366d892bd35ba06396cfb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/ambiguous_halt/halt_006/18f2f83b5381590e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "3fe24a5180a8473d8f1ed9fdcdc152c8fdf36eaf562d44de329fffe8bf891fba", "seconds": 10.72, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a moment', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is incomplete and contains hesitation ('um'), but the final interjection is a pause command, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Follow all artists of all indie genre songs in um any of my playlists on let's see just a moment" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "b7a9ee9_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a2eeff774663533b3d94281c449044330bf29dcfa973ee92175671ae04c9de27", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/b7a9ee9_3/audio_natural/filler/original/8171b8e4177a5c5d/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "5c32a9a047976c59245c1de3a69619fd3e4da4f411cb48369bfab1cd3f8b617d", "seconds": 7.48, "spoken_utterances": [ "Follow all artists of all indie-genre songs in um... any of my playlists on let's see... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let's see') but does not change the core request. The final intent is to follow all artists of all indie-genre songs in any of the user's playlists on Spotify, which matches the original instruction. The hesitation and fillers are ignored as per the 'filler' type criterion.", "heard_transcript": "Follow all artists of all indie-genre songs in um any of my playlists on let's see" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I should say Venmo anyone I have sent or received money from this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "5562916415081a91ae13903221c683f143dafff5a9626e4d0ac56361fa1fbd6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, the condition about money history, and the unfriend directive." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/false_start/false_start_01_app/51145399e847ba40/speech.wav", "wer": 0.36, "wav_sha256": "7023f87177fddb3884b145413c771ad9bf18c79fe43613b7d5e28ed85c0a5f48", "seconds": 13.6, "spoken_utterances": [ "Befriend on CashApp—uh... I should say um... Venmo anyone I have sent or received money from let's see... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'I should say, um, Venmo'. This is a false start where the initial incorrect value is corrected. The final instruction matches the original reference, so the task should be executed as corrected.", "heard_transcript": "Befriend on Cash App. I should say, um, Venmo anyone I have sent or received money from, let's see, this month and unfriend everyone else." }, "candidate_id": "wav_430f051af49dbd50", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits a very natural false start with 'Befriend on Cash App' followed by a self-correction to 'Venmo'. The speaker uses a realistic filler 'um' and a brief pause before continuing. The correction is immediate and the flow remains smooth, with a plausible hesitation ('let's see') before completing the request. The prosody and pacing are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money to—sorry, I meant sent or received money from this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "2801c77e2c0f5c3268abef4649ae57f623d76fa2427d2cc017fe5a56a0fc5fa3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'to' to 'from' within the same utterance, which is a false start. The final intended preposition is 'from', matching the original instruction. All other conditions, including the app (Venmo), the time frame (this month), and the action (befriend and unfriend), remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/false_start/false_start_02_transaction_relation/ed05a66889f3eae9/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "f00a55900b44f048f4c80050a599044bbfe8729c13a9f4fa86b1c17524605314", "seconds": 17.56, "spoken_utterances": [ "Befriend on Venmo anyone I have sent or received money to—hmmm... sorry, I meant uh... sent or received money from ... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent or received money to', then immediately self-corrects to 'sent or received money from'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistake should be discarded. The final instruction matches the original reference in both action and scope, with only a minor grammatical rephrasing ('from' vs 'to') that does not change the semantic intent. The correction is immediate and clear, satisfying the false_start type requirement.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money to. Sorry, I meant sent or received money from this month and unfriend everyone else." }, "candidate_id": "wav_5bc2bbc54b98a5a8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'to' corrected to 'from', which is a believable word swap. The speaker uses a natural 'sorry I meant' repair. The pacing is slightly hesitant but the correction is integrated smoothly. The overall delivery is credible, though the initial error is less common than 'to/from' confusion, making it slightly less typical than audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from last month—no, that should be this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "c4e9d90e51a34868238c993168285d31bc99eeb0b3948d69b143b27598f05947", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month' within the same utterance, preserving the rest of the instruction including the action, app, condition, and the unfriend directive." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/false_start/false_start_03_time_reference/84fa1b60520b6f9e/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "d15f466a1db4ba49f91a32235640ecf9f4e78e4946fcea74572bef1f030f8b1b", "seconds": 13.68, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have sent or received money from last month—um... no, that should be ah... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'no that should be uh this month'. This is a clear false start where the final value is corrected. The rest of the instruction remains unchanged and matches the original. The final intent is to befriend contacts from 'this month' and unfriend others, which aligns with the original instruction after the correction.", "heard_transcript": "befriend on uh Venmo anyone I had sent or received money from last month um no that should be uh this month and unfriend everyone else" }, "candidate_id": "wav_e687a74c918c5190", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip features a revision where the speaker initially says 'last month' and then corrects to 'this month'. The correction is clear and the speaker uses 'no, that should be' which is a plausible way to self-correct. However, the initial phrase 'I had sent' is grammatically awkward compared to 'I have sent', and the correction feels slightly rushed, reducing the naturalness slightly." } ] }, "is_default": false }, { "variant_id": "false_start_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from this month and unfriend everyone—scratch that, make it unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "33b40d11cc426adcd9784fbfe20504de6276abdab25cb0ad5fc14c849612a8ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'unfriend everyone' but immediately corrects it to 'unfriend everyone else'. The correction replaces the broader 'everyone' with the more specific 'everyone else', which aligns with the original instruction's intent to retain only those contacts from whom money was sent or received." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/false_start/false_start_04_other_contacts_action/3bfa799373533250/speech.wav", "wer": 0.08, "wav_sha256": "590f7f87867a54f6e38996297970090940565e1726806b460aa9a29ebe661d17", "seconds": 15.68, "spoken_utterances": [ "Befriend on ... Venmo anyone I have sent or received money from this month and unfriend everyone—uh... scratch that, make it um... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the full task but then self-corrects with 'scratch that' and revises the final action to 'unfriend everyone else'. This is a false start where the final audible intent is the corrected version, which is the opposite of the original instruction. However, the type-specific criterion for 'false_start' is to execute the immediate self-correction. The final instruction is clear and actionable, so the task is to unfriend everyone else.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from this month and unfriend everyone uh, scratch that, make it um, unfriend everyone else." }, "candidate_id": "wav_14216fd015739b5a", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a revision where the speaker says 'unfriend everyone else' and then corrects to 'make it um unfriend everyone else'. The correction is awkward and redundant, as the phrase is repeated almost verbatim. The 'uh' filler is placed unnaturally, and the overall flow is less smooth. The speaker seems to be rephrasing without a clear reason, making the revision sound less spontaneous and more like a mechanical edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have this month and unfriend everyone else. One more thing: only befriend people I've sent or received money from.", "long_pause_offsets": [ 71 ], "source_sha256": "c10b2b09885ae63b650044d537e5cb1d7c2246b9be731b106e93ea57a93c9942", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction that only people with whom money was sent or received should be befriended. The initial instruction is broad ('anyone I have this month'), but the addition clarifies the condition. The action (befriend/unfriend) and scope (this month) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/addition/addition_02/17b324ecc0df98fe/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "34d871c6c286604318e297f9fb1ca77abdf28dc28f33bfe83626a87fd5974cdc", "seconds": 17.08, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have this month and ... unfriend everyone else. ", "One more thing: only befriend people I've let me think... sent or received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'anyone I have this month' which is incomplete, but then adds the clarifying condition 'only befriend people I've sent or received money from' after a pause. This addition correctly specifies the intended constraint, combining with the original instruction to form the complete and correct goal. The final intent matches the original instruction exactly.", "heard_transcript": "Befriend on uh Venmo anyone I have this month and unfriend everyone else. One more thing, only befriend people I've let me think sent or received money from." }, "candidate_id": "wav_784526c1d243963d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from and unfriend everyone else. Oh, I forgot to mention, only do this for transactions this month.", "long_pause_offsets": [ 88 ], "source_sha256": "088caa5bdc37db7f215f57a8129c170a869e260d1b28125bb20b162361a8135c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this month') to the initial broad instruction, narrowing the scope without contradicting it. The core action and target remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on uh... Venmo anyone I have sent or received money from and unfriend everyone else. ", "Oh, I forgot to mention, only do this for transactions um... this month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have sent or received money from this month and unfriend everyone else. Wait, the app should be Venmo, not PayPal.", "long_pause_offsets": [ 100 ], "source_sha256": "199ae49208ffca74caf758b9a75715659e8f9a7c3f8eae6b3e24c77c5cb37d1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the action, condition, and scope of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/revision/revision_01_app/87ca115e080c25af/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "acdb4e50142d818e8c85a519893eb4db68ff79da0d4de73dfa740acc4d60ee8a", "seconds": 16.96, "spoken_utterances": [ "Befriend on uh... PayPal anyone I have ... sent or received money from this month and unfriend everyone else. ", "Wait, the app should be ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions PayPal but immediately corrects themselves to Venmo after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to perform the action on Venmo, which matches the original reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Befriend on uh PayPal anyone I have sent or received money from this month and unfriend everyone else. Wait, the app should be Venmo, not PayPal." }, "candidate_id": "wav_abfcc8e95ed2e20a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker starts with a clear request, then pauses with a believable 'wait' before self-correcting the app name from 'PayPal' to 'Venmo'. The correction is integrated smoothly into the ongoing utterance with a natural tone of realization, and the prosody remains consistent throughout." } ] }, "is_default": true }, { "variant_id": "revision_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from this month and unfriend no one. I misspoke about the action: it should be 'unfriend everyone else', not 'unfriend no one'.", "long_pause_offsets": [ 92 ], "source_sha256": "8d21a5253798fd615302473b4ce90dca0bb3f916cc4e80b4b39ced9748697e68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the action from 'unfriend no one' to 'unfriend everyone else', which matches the original instruction. The first utterance is fully revised by the second." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/revision/revision_04_other_contacts_action/ecdb3316959e4990/speech.wav", "wer": 0.13513513513513514, "wav_sha256": "600563129fb8447f46ccf4ae7cac18874d2764413efda578a6acfac289eca922", "seconds": 19.32, "spoken_utterances": [ "Befriend on Venmo anyone I have uh... sent or received money from ... this month and um, let me think... unfriend no one. ", "I misspoke about the action: it should be 'unfriend everyone else', not 'unfriend no one'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unfriend no one' but immediately corrects themselves, stating 'I misspoke about the action. It should be unfriend everyone else, not unfriend no one.' This is a clear revision where the final intent is to unfriend everyone else, which aligns with the original instruction's 'unfriend everyone else' clause. The revision is spoken after a pause and self-correction, satisfying the revision type criterion.", "heard_transcript": "Befriend on Venmo anyone I have uh sent or received money from this month and um let me think, unfriend no one. I misspoke about the action. It should be unfriend everyone else, not unfriend no one." }, "candidate_id": "wav_5e8452e77ebefe91", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers a clear initial request and then pauses with 'let me think' before correcting a key action from 'unfriend no one' to 'unfriend everyone else'. The self-correction is plausible and the intonation reflects a genuine change of mind. The delivery is fluent and the repair feels spontaneous, though the correction is slightly more abrupt than in audio_03." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent or received money from last month and unfriend everyone else. Change 'this month' to 'last month' instead.", "long_pause_offsets": [ 99 ], "source_sha256": "0d6e06334c5d19c6f11b04b542441a9e12facae260cdc81571a76c9d8dfba4c1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'this month' to 'last month', which is the only difference from the original instruction. The action (befriend/reject), app (Venmo), and condition (sent/received money) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/revision/revision_03_time_reference/c1f4dc6f814f9bc7/speech.wav", "wer": 0.03125, "wav_sha256": "6e7c2724383d34d545d9af0440b0187d851a177ec28e270fa9034cc30e2a7694", "seconds": 19.48, "spoken_utterances": [ "Befriend on Venmo anyone I have sent or received money from um, let's see... last month and uh... unfriend everyone else. ", "Change 'um, let me think... this month' to 'last month' instead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this month' but then explicitly revises it to 'last month' after a pause and self-correction ('Change... this month to last month instead'). According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction with the corrected time frame.", "heard_transcript": "Befriend on Venmo anyone I have sent or received money from um let's see last month and unfriend everyone else. Change um let me think this month to last month instead." }, "candidate_id": "wav_714307e7c4d9fda3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request, then hesitates with 'let's see' and 'um' before revising the time frame from 'this month' to 'last month'. The revision is natural, but the pauses are longer and the delivery feels slightly more hesitant and less fluid than the top two. The correction is clear but the overall pacing is slower." } ] }, "is_default": false }, { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent or received messages from this month and unfriend everyone else. Make that 'sent or received money from', not 'sent or received messages from'.", "long_pause_offsets": [ 102 ], "source_sha256": "bf27ea81291048b7eafe37d64bfb0698bf78285d5e1e6eed0b5c0bcf1053882a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'sent or received messages from' to 'sent or received money from', preserving the app, action, and scope of the instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/revision/revision_02_transaction_relation/5779a78bc9669e86/speech.wav", "wer": 0.05405405405405406, "wav_sha256": "d9942770045776eddceea0dfd548ab933362b37cd84ac28f4f3802a38ed444bf", "seconds": 20.88, "spoken_utterances": [ "Befriend on um... Venmo anyone I have uh, let's see... sent or received messages from this month and um, let's see... unfriend everyone else. ", "Make that 'sent or received money from', not 'sent or received messages from'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent or received messages from' but immediately corrects it to 'sent or received money from' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "befriend on um Venmo anyone I have uh let's see sent or received messages from this month and um let's see unfriend everyone else. Make that sent or received money from, not sent or received messages from." }, "candidate_id": "wav_0c162c6b0296f79d", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio contains multiple self-corrections and hesitations. The speaker initially says 'sent or received messages' before correcting to 'money'. The correction is clear, but the initial error is more significant and the phrasing 'make that' sounds slightly less natural than 'change' or 'it should be'. The delivery is less fluid, with more fragmented pauses and a slightly less confident tone." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "1e2d2e3c163cf3cc4cfc62f241509ad2a43bdf51bb47905fb339c7216cc9d775", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/cancellation/cancel_037/b84eb75983e737fe/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "4298c69a36b9f5aae272b3571a8454d60cd95039b2019aaf9ffb261503aac78c", "seconds": 17.04, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else.", "Hang on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial request and then clearly and unambiguously cancels the entire task with 'Cancel the whole task.' This meets the cancellation type criterion, as the final audible intent is a complete revocation of the original instruction.", "heard_transcript": "Befriend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else. Hang on, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "f2d40145b64fd575378b4aeb02a061c0e6d521f1deefeea57e44a6fce2f735c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/ambiguous_halt/halt_022/e00932b55938b33d/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "eaada0c86528e6884177653bbf05f93ad4c3aa5ba41844c9480ae4b34856c001", "seconds": 16.08, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else.", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or a new instruction. The initial request is not executed, but the intent to pause is audible and final.", "heard_transcript": "befriend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f8d716f1d38b308596bdbf3e6763d4ae4feaf73eafd433ba6f7773fa9e21b036", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_1/audio_natural/filler/original/29463bc346ed6791/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "d638f0dd5e199db06b0bbf30a62f33f5b9ca1249d5442c414d66f4614d1f193b", "seconds": 12.04, "spoken_utterances": [ "Befriend on uh, let's see... Venmo anyone I have um... sent or received money from this month and ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('uh', 'let's see', 'um') but completes the full original instruction without any correction or change of intent. The final interjection is the completion of the task request, not a pause or cancellation. The hesitation and fillers are ignored as per the 'filler' type criterion, and the final audible task matches the original instruction.", "heard_transcript": "be friend on uh let's see Venmo anyone I have um sent or received money from this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have received money from—no, I meant sent money to this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "55edea40df4150e92824d71c99c28bb0eb654a7daece89c710f193953b739a49", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'received money from' to 'sent money to' within the same utterance. The final action and condition match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/false_start/false_start_02_transaction_relation/501928e09e6f9194/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "a1d046efa9bb70b01a7d63b89ef4ea374dc5a90d987b623ab1cd753bd0da7890", "seconds": 10.2, "spoken_utterances": [ "Befriend uh... on Venmo anyone I have received money from—hmmm... no, I meant sent money to uh... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received money from' but immediately self-corrects to 'sent money to'. According to the 'false_start' type criterion, the final corrected intent should be executed. The correction is immediate and unambiguous, and the rest of the instruction remains unchanged. Therefore, the final intent matches the original instruction.", "heard_transcript": "Befriend uh on Venmo anyone I had received money from hmm no I meant sent money to uh this month and unfriend everyone else." }, "candidate_id": "wav_d9aa09731bb83a33", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'befriend uh on Venmo anyone I had received money from' and immediately self-corrects to 'no I meant sent money to' with a believable hesitation ('hmm') and a smooth, conversational tone. The correction is integrated into the same utterance without awkward pauses or robotic phrasing, and the prosody reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last month—sorry, that should be this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "7d3e65ecb89eeca21d8986b1de0f3349450d36dbcbc1d33440a82ef316bf273c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month' within the same utterance, preserving the action, app, and the unfriend directive. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/false_start/false_start_03_time_reference/7e03269bb3cbb3de/speech.wav", "wer": 0.12, "wav_sha256": "76c49074bb83d607f341d3ae6fa4dfd2e91194b1d0088dc6f7706f64609010fe", "seconds": 9.88, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to last month—uh... sorry, that should be ah... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'Sorry, that should be uh this month'. According to the 'false_start' type criterion, the final corrected value ('this month') should be applied, and the original incorrect value ('last month') should be discarded. The rest of the instruction remains consistent. The final intent matches the original instruction with the corrected time frame.", "heard_transcript": "befriend, let's see, on Venmo, anyone I have sent money to last month. Sorry, that should be uh this month, and unfriend everyone else." }, "candidate_id": "wav_fafb218b4bc3239d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'be friend let's see on Venmo anyone I have set money to last month' followed by a self-correction to 'this month'. The correction is natural, but the initial phrase 'be friend let's see' sounds slightly more rehearsed or less spontaneous than audio_02. The prosody is slightly flatter, and the transition into the correction is less fluid, making it feel a bit more like a scripted correction than a genuine thought process." } ] }, "is_default": false }, { "variant_id": "false_start_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and remove all other contacts—scratch that, make it unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "3d054cad65b813786f114958cebe1a797b7426293a04b239483757e16164309e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'remove all other contacts' but immediately corrects it to 'unfriend everyone else'. Since 'unfriend' is the correct action for Venmo contacts, the final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/false_start/false_start_04_other_contacts_action/2deb201126244534/speech.wav", "wer": 0.0, "wav_sha256": "88e3ed66a11f81956629984b6239bab965315f40727c68d701bc7a8db9123f39", "seconds": 12.6, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to this month and ... remove all other contacts—um... scratch that, make it ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the full task, then immediately self-corrects with 'scratch that' and provides the final instruction to 'unfriend everyone else'. According to the 'false_start' type criterion, the final instruction after the self-correction must be executed. The final audible intent is to unfriend everyone else, which is a valid and complete command. The original instruction is overridden by the correction, and no other constraints or conditions are added or modified. Therefore, the final intent matches the expected control intent for a false_start.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and remove all other contacts. Um, scratch that, make it unfriend everyone else." }, "candidate_id": "wav_a2427837895b7d15", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start but with a more complex and less natural repair. The speaker says 'befriend on Venmo anyone I have sent money to this month and remove all other contacts' then says 'scratch that make it unfriend everyone else'. The 'scratch that' is a common phrase, but the delivery feels abrupt and the second part is spoken with a different, more declarative tone, suggesting a cut-and-paste or non-spontaneous edit. The prosody lacks the natural hesitation and flow of the other two, making it sound less realistic." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I should say Venmo anyone I have sent money to this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "22253c1803beb29ff606a05af8f9e9dd5c962787c43e62c5f22334d2b1fe6250", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, the condition (anyone sent money to this month), and the unfriend-all-else directive." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on CashApp—uh... I should say ... Venmo anyone I have sent money to let's see... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and. One more detail: unfriend everyone else.", "long_pause_offsets": [ 62 ], "source_sha256": "65b9c8d670ecb8063230ca1546bdd357085beb7d8c37e30a07ad1d993fa55f80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'unfriend everyone else' condition, completing the original instruction. The initial utterance already specified the befriending action and its condition, and the addition supplies the complementary unfriending action without altering the first part." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/addition/addition_02/5fd093d8acf8eca2/speech.wav", "wer": 0.12, "wav_sha256": "78463b39d6aee15e274ddf596f5fea0f84487da95904d14b1685fe3d6c536bda", "seconds": 14.08, "spoken_utterances": [ "Befriend um, let's see... on Venmo anyone I have sent money to uh, let me think... this month and. ", "One more detail: ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task and then adds the missing constraint 'unfriend everyone else' after a pause and self-interruption. This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is clear and complete, matching the reference instruction exactly.", "heard_transcript": "Befriend, um, let's see, on Venmo, anyone I have sent money to, uh, let me think, this month, and one more detail, unfriend everyone else." }, "candidate_id": "wav_701055d82f2ed8c8", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have sent money to and unfriend everyone else. Oh, I forgot to mention, only do this for people I've sent money to this month.", "long_pause_offsets": [ 74 ], "source_sha256": "cd26e0afeb5f9307ad5aedeba7d472e24d4e65d266adb0cade6e12cb8d6b9ce1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this month') to the initial broad instruction, narrowing the scope without contradicting it. The core action (befriend those sent money to, unfriend others) remains consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend ... on Venmo anyone I have sent money to and unfriend everyone else. ", "Oh, I forgot to mention, only do this for people I've sent money to um... this month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have sent money to last month and unfriend everyone else. Make that this month, not last month.", "long_pause_offsets": [ 85 ], "source_sha256": "05458ded87c854bfdc1f4048d30bc05ccde4dfa3b45cd6cc0ef3fb799bbba1c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this month', preserving the action and scope of befriending and unfriending." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/revision/revision_03_time_reference/26ec4dfaaa93c859/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "70b53ec8626b55539e4832d0ce6bd1334fe7fb4e390065e6fe016d62f9db24dc", "seconds": 13.68, "spoken_utterances": [ "Befriend uh... on Venmo anyone I have sent money to let me think... last month and unfriend everyone else. ", "Make that ... this month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'this month' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Befriend uh on Venmo anyone I have sent money to let me think last month and unfriend everyone else. Make that this month, not last month." }, "candidate_id": "wav_191893e05252a462", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('uh') and a believable self-correction of the time frame from 'last month' to 'this month'. The correction is smoothly integrated with a natural-sounding 'make that' phrase, and the overall prosody is consistent and conversational. The pacing and intonation suggest spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month and unfriend everyone else. Sorry, I meant anyone I have sent money to, not received money from.", "long_pause_offsets": [ 91 ], "source_sha256": "6aaf4fc5530b1e90c27170088e12575d706cb0ba427994ba907faab8abed9898", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial condition from 'received money from' to 'sent money to', preserving the app, action, and scope of 'this month'. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/revision/revision_02_transaction_relation/1f68a0592cfed328/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "8071a7af03b44b61226115de5a240c4e7b4fc1d0248356a86949dedba1aa6315", "seconds": 17.88, "spoken_utterances": [ "Befriend um, let me think... on Venmo anyone I have received money from um, let's see... this month and unfriend everyone else. ", "Sorry, I meant anyone I have um... sent money to, not received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'received money from' but immediately corrects themselves with 'Sorry, I meant anyone I have sent money to, not received money from.' This is a clear revision where the final intent matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as revised.", "heard_transcript": "Befriend, um, let me think, on Venmo, anyone I have received money from, um, let's see, this month and unfriend everyone else. Sorry, I meant anyone I have, um, sent money to, not received money from." }, "candidate_id": "wav_cbe618ee929eb379", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a natural hesitation ('um') and a self-initiated pause ('let me think'), which is followed by a clear request. The correction from 'received money from' to 'sent money to' is a plausible self-repair within the same utterance, and the speaker's tone remains consistent. However, the correction is slightly less fluid than in audio_03, with a brief pause before the correction, making it feel a bit more like a mid-sentence edit than a seamless thought process." } ] }, "is_default": false }, { "variant_id": "revision_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and unfollow everyone else. Change unfriend everyone else to unfollow everyone else.", "long_pause_offsets": [ 85 ], "source_sha256": "df260731293ea33584d6fa8f75c8420a08081dbf4e2778772ad2a47710186c51", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'unfriend' to 'unfollow', which is the only change. The condition 'anyone I have sent money to this month' and the app 'Venmo' remain unchanged. The final goal is consistent with the original instruction except for the action verb, which is now correctly specified as 'unfollow'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/revision/revision_04_other_contacts_action/bdc0a875a32543a5/speech.wav", "wer": 0.2, "wav_sha256": "1e65f01c783dce1d5ca96eed4e44caa8d14d5ad894edb0ccc5489ece8157f7b1", "seconds": 15.4, "spoken_utterances": [ "Befriend on Venmo anyone I have sent money to um, let's see... this month and unfollow everyone else. ", "Change uh, let me think... unfriend everyone else to unfollow everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unfriend everyone else' but then corrects it to 'unfollow everyone else' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original instruction's goal, with only the verb 'unfriend' replaced by 'unfollow' as the corrected action. All other conditions (Venmo, 'anyone I have sent money to this month') remain consistent. Therefore, the final intent is correctly executed.", "heard_transcript": "befriend on Venmo anyone I have sent money to um let's see this month and unfollow everyone else. Change uh let me think. unfriend everyone else to unfollow everyone else." }, "candidate_id": "wav_3dac3b5f2af0c0a0", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but then introduces a revision ('change uh let me think') that feels more like a meta-instruction to the system than a natural thought process. The correction from 'unfriend everyone else' to 'unfollow everyone else' is delivered with a flat, almost robotic intonation, and the 'let me think' is placed awkwardly after the initial command, breaking the flow. The overall delivery sounds less spontaneous and more like a system-generated prompt being modified." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have sent money to this month and unfriend everyone else. Wait, the action should be on Venmo, not PayPal.", "long_pause_offsets": [ 86 ], "source_sha256": "8ec8a4eaab0f0cfdd24cd566a5335ce10b5466eb5d2b81b8f8e86452057c48e9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the action (befriend/unfriend), the condition (anyone sent money to this month), and the scope (everyone else)." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on PayPal anyone I have sent money to um... this month and unfriend everyone else. ", "Wait, the action should be on uh... Venmo, not PayPal." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_037", "hide_variant_original": true, "source_sha256": "4e2dde7ef0ced5758dccd42a396e8c95f27584770a370063f6d0ebf78a5f89bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/cancellation/cancel_037/66df4b2774f01d14/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "5c740fd4508b2bc2b60662c5dc2dd1de971d9a350cf7c91bc69274c577691e05", "seconds": 15.52, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else.", "Hold on, stop there. Cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the initial instruction and then clearly states 'Cancel the whole task' as a final interjection, which constitutes a direct cancellation of the entire task. The cancellation is unambiguous and not followed by any resumption or modification of the original goal.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. Hold on, stop there. Cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_035", "hide_variant_original": true, "source_sha256": "119064280d27122ea1776de3f006ea06a098c47878d288f7005445bfd4ceb899", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/ambiguous_halt/halt_035/e357967abafa9321/speech.wav", "wer": 0.125, "wav_sha256": "d669579a2117b6d8465d2bc23028ac24577ffd8048e958efabb9acafc1a819f7", "seconds": 14.8, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else.", "Hold on before you do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on, before you do that,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or provide a new instruction, but rather halts execution temporarily. The initial instruction is not revoked, and no new goal is stated. Therefore, the final control intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. Hold on, before you do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b8704b517eb7e010364f63832ed1a65e5905feda395fb557f0de65af3d34e59f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_2/audio_natural/filler/original/c595b0f27364d459/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "ee9512726d09f6dfdbb6a559b149bc69e9a7562fe09e568b4df6383cbfc9ef0e", "seconds": 10.6, "spoken_utterances": [ "Befriend let's see... on Venmo anyone I have sent money to hmmm... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Befriend let's see' which is a hesitation or filler, but the core instruction remains unchanged. The final intent is to befriend those sent money to this month and unfriend others. The filler does not alter the goal, so the task should be executed as per the original instruction.", "heard_transcript": "Befriend let's see on Venmo anyone I have sent money to this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have received money from last month—no, that should be this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "cc232526b9f358d390edbf71f5b7a1a1d4b8bf82439985ff922c1a88d7c3b3d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this month' within the same utterance, preserving the action, app, and the unfriend-all-else condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/false_start/false_start_03_time_reference/ec25e9a617f3bb92/speech.wav", "wer": 0.12, "wav_sha256": "0a4d8f4c2a1f0e2ebe1c34b82f74133aaf5faf7fe0e907af3ba3a5c7b009c88f", "seconds": 15.48, "spoken_utterances": [ "Befriend on let me think... Venmo anyone I have received money from last month—hmmm... no, that should be ... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this month' with 'No, that should be this month'. According to the 'false_start' type criterion, the final corrected value ('this month') overrides the initial mistaken value ('last month'). The rest of the instruction, including 'befriend on Venmo anyone I have received money from' and 'unfriend everyone else', remains unchanged and is preserved. The final intent matches the original instruction with the corrected time frame.", "heard_transcript": "Befriend on Venmo anyone I have received money from last month. No, that should be this month, and unfriend everyone else." }, "candidate_id": "wav_54dc053714c7a20d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and realistic speech patterns. The speaker uses a genuine 'let me think' filler, followed by a spontaneous self-correction from 'last month' to 'this month' with a believable hesitation and intonation shift. The pauses are appropriately placed, and the overall flow mimics a real person formulating a complex request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Befriend on CashApp—I meant Venmo anyone I have received money from this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "03c46d98d5a87297ecdf3c5cf7552d79a44f9ddf7c2e8de4b2874ae65dd08c25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'CashApp' to 'Venmo' within the same utterance, preserving the action, target condition, and the unfriend-all-else directive." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/false_start/false_start_01_app/850443af53840079/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "10e411ce3add78cdc01b19c8136869db3876ece8c9ac4afbe359e4a7d04e5b35", "seconds": 8.48, "spoken_utterances": [ "Befriend on CashApp—uh... I meant ... Venmo anyone I have received money from this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' with 'I meant'. This is a false start where the first value is replaced by the second. The final instruction matches the original reference exactly, so the task should be executed as corrected.", "heard_transcript": "Befriend on Cash App. I meant, Venmo anyone I have received money from this month and unfriend everyone else." }, "candidate_id": "wav_964110ec02be4e03", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'I meant' followed by a corrected version of the request. The correction is natural, and the speaker integrates the change smoothly. However, the initial error ('cash app') and the correction are less integrated than in audio_03, and the overall delivery is slightly more robotic, lacking the organic hesitation seen in the top-ranked clip." } ] }, "is_default": false }, { "variant_id": "false_start_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month and remove all other contacts—scratch that, make it unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "959140ccdcec785f3fa8591b926341b6730849ef103b9cce4d401a190e864ee1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'remove all other contacts' but immediately corrects it to 'unfriend everyone else'. Since 'unfriend' is the correct action for Venmo contacts, the final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/false_start/false_start_04_other_contacts_action/f915a564ec08dce3/speech.wav", "wer": 0.125, "wav_sha256": "b14e9bc052981bef7d92bb45134fda9ee90b83314793e51ec01f2a2c8087d608", "seconds": 13.16, "spoken_utterances": [ "Befriend on Venmo anyone I have received money from this month and ... remove all other contacts—um... scratch that, make it ... unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states a task to befriend contacts and remove others, then immediately self-corrects with 'Scratch that, make it unfriend everyone else.' This constitutes a false_start where the final intent is the corrected version. The final instruction is clear and executable, matching the expected behavior for a false_start type.", "heard_transcript": "Befriend on Venmo anyone I have received money from this month and remove all other contacts. Scratch that, make it unfriend everyone else." }, "candidate_id": "wav_dcd57946075686cf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'Scratch that' and a correction, but the initial request is incomplete and awkwardly phrased ('Befriend on then mo'). The correction is abrupt, and the speaker does not fully rephrase the task. The delivery sounds less spontaneous and more like a rehearsed or fragmented instruction, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "false_start_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to—or rather, received money from this month and unfriend everyone else.", "long_pause_offsets": [], "source_sha256": "d74aa579df9a6abf406860a8da21a45b256b4be3733e4816890739e5f3b927f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent money to' but immediately corrects to 'received money from'. The final goal matches the original instruction exactly, including the app, action, condition, and scope." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on ... Venmo anyone I have ... sent money to—um... or rather, received money from this month and unfriend everyone else." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month and. I left out a detail: unfriend everyone else after befriending those you've received money from.", "long_pause_offsets": [ 68 ], "source_sha256": "7e55af03458e05f05fa88198ddada273b8ff82cea1693fd24cfedb9298f2ed32", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing unfriend instruction, which completes the original goal. The final action is to befriend the specified group and then unfriend all others, matching the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/addition/addition_02/6f91a2f2b5bb23ec/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "4a867687f07a307f8ac345283748b5944eee8ebbed5a74096d56ae75b620fa13", "seconds": 14.12, "spoken_utterances": [ "Befriend on uh... Venmo anyone I have received money from ... this month and. ", "I left out a detail: ... unfriend everyone else after befriending those you've received money from." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the initial instruction, then adds a clarifying detail ('Unfriend everyone else after befriending those you've received money from') after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final audible intent is complete and matches the reference instruction.", "heard_transcript": "Befriend on uh Venmo anyone I have received money from this month and I left out a detail. Unfriend everyone else after befriending those you've received money from." }, "candidate_id": "wav_a932fda4521336a5", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Befriend on Venmo anyone I have received money from and unfriend everyone else. Oh, I forgot to mention, only do this for people I've received money from this month.", "long_pause_offsets": [ 80 ], "source_sha256": "37f497576e9c102773b5423ab5a7d195d69cad62ea78b2ffc929cb5bf1e9f179", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction ('this month') to the initial request, narrowing the scope of the befriending action without contradicting the original intent. The core action and target remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Befriend on uh... Venmo anyone I have received money from and unfriend everyone else. ", "Oh, I forgot to mention, only do this for people I've received money from ... this month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_transaction_relation", "proposed_instruction": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. I misspoke about the transaction relation: it should be 'received money from', not 'sent money to'.", "long_pause_offsets": [ 85 ], "source_sha256": "0fc41415c4e157d0592c53211c06d2083ff75ca5b01dd03702067c2a94f3c2c4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the transaction relation from 'sent money to' to 'received money from', while retaining the app, time frame, and unfriend-all action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/revision/revision_02_transaction_relation/d32c27c6ee59e49c/speech.wav", "wer": 0.06060606060606061, "wav_sha256": "41968adb7661901dc88fbe65dcdd6cd3d425ac1919b2105f5882639754cf18e8", "seconds": 17.8, "spoken_utterances": [ "Befriend on ah... Venmo anyone I have sent money to ... this month and unfriend everyone else. ", "I misspoke about the transaction relation: it should be 'um... received money from', not 'sent money to'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the incorrect condition 'sent money to' but immediately self-corrects after a pause, revising it to 'received money from'. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Befriend on Venmo anyone I have sent money to this month and unfriend everyone else. I misspoke about the transaction relation. It should be um received money from, not sent money to." }, "candidate_id": "wav_ac0d4a2b36a00182", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request, then naturally pauses and self-corrects the transaction direction ('sent money to' to 'received money from') with a plausible 'I misspoke' opener. The correction is smoothly integrated into the same utterance, maintaining a natural conversational flow and realistic hesitation." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Befriend on PayPal anyone I have received money from this month and unfriend everyone else. Let me correct the app: befriend on PayPal, not Venmo.", "long_pause_offsets": [ 92 ], "source_sha256": "72d0a335ca93296f4e28b0baad3725fc713a9588cce68bf17d52d4023054d7bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the app from Venmo to PayPal, changing the platform where the action is to be performed." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/revision/revision_01_app/7570e49e4ed56cf8/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "7b30f11cf45116bd8de0b21669d0e7d48ab65ad6db480bd10ab541aea6c1e565", "seconds": 14.04, "spoken_utterances": [ "Befriend on hmmm... PayPal anyone I have received money from um... this month and unfriend everyone else. ", "Let me correct the app: befriend on ... PayPal, not Venmo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task for Venmo, then immediately corrects it to PayPal. The revision is clear and unambiguous, and the final instruction matches the original intent but with the correct app. The correction is applied to the entire task, and no other conditions are changed. Therefore, the final intent is to execute the task on PayPal.", "heard_transcript": "Befriend on Venmo anyone I have received money from this month and unfriend everyone else. Let me correct the app, befriend on PayPal, not Venmo." }, "candidate_id": "wav_e2cc074d09ee0571", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a clear revision to the app name ('PayPal' instead of 'Venmo') after the initial request. The correction is introduced with 'Let me correct the app,' which is a pragmatic and believable way to amend a detail. The prosody is consistent, and the pause before the correction feels like a genuine moment of recall." } ] }, "is_default": false }, { "variant_id": "revision_04_other_contacts_action", "proposed_instruction": "Befriend on Venmo anyone I have received money from this month and unfollow everyone else. That should be 'unfriend everyone else', not 'unfollow everyone else'.", "long_pause_offsets": [ 91 ], "source_sha256": "342823b30f7a15f9d3c343d764eccf707629c3e07d21e58ee636d990136d61de", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'unfollow' to 'unfriend', which is the same as the original instruction. The recipient scope and condition remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/revision/revision_04_other_contacts_action/00a7d54a50de60be/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "ac29a36034624135b6cfb7d9d27af22ddb2cd45b3f42996ccfd4be541aca5e20", "seconds": 14.8, "spoken_utterances": [ "Befriend on ... Venmo anyone I have received money from um... this month and unfollow everyone else. ", "That should be 'unfriend everyone else', not 'unfollow everyone else'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unfollow everyone else' but immediately corrects themselves with 'That should be unfriend everyone else, not unfollow everyone else.' This is a clear revision where the final intent is to unfriend, not unfollow. The correction is immediate and unambiguous, satisfying the revision type criterion. The rest of the instruction matches the original.", "heard_transcript": "Befriend on Venmo anyone I have received money from um this month and unfollow everyone else. That should be unfriend everyone else, not unfollow everyone else." }, "candidate_id": "wav_69866300939a8d1b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker initially says 'unfollow everyone else' and then corrects it to 'unfriend everyone else.' The correction is introduced with 'That should be,' which is a common way to revise a term. However, the correction is slightly abrupt and the prosody lacks the natural flow of the top two, making it feel a bit more like a mechanical edit than a spontaneous thought." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Befriend on Venmo anyone I have received money from last month and unfriend everyone else. Change 'this month' to 'last month' instead.", "long_pause_offsets": [ 91 ], "source_sha256": "2a46068284be80af06fd23a11e0ef9e0ccc190702a132eaca51b5e791841afcc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'this month' to 'last month', aligning the spoken instruction with the original intent but shifting the time frame. The action (befriend/reject) and target (people from whom money was received) remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/revision/revision_03_time_reference/6fd8003c0baabf34/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "39dae19e6aec6d24d120e7a8c6682464f159d0c00281c8fff809dfeb0c602d97", "seconds": 14.48, "spoken_utterances": [ "Befriend on ah... Venmo anyone I have received money from ... last month and unfriend everyone else. ", "Change 'uh... this month' to 'last month' instead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this month' but immediately corrects it to 'last month' in a revision. The final audible intent is to apply the action to 'last month', which matches the reference instruction after the revision. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Befriend on uh Venmo anyone I've received money from last month and unfriend everyone else. Change uh this month to last month instead." }, "candidate_id": "wav_a216fab2e3adacb5", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker makes a revision from 'this month' to 'last month' but does so in a separate sentence rather than integrating it into the original request. The phrasing 'Change this month to last month instead' sounds more like a post-hoc instruction than a natural self-correction within a single thought process, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_025", "hide_variant_original": true, "source_sha256": "181e9b7587e5c9e12777d158cb3dafe2250a5b208aff8caf7ae100fa947821c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/cancellation/cancel_025/be3b33abd9b320d0/speech.wav", "wer": 0.16, "wav_sha256": "8e1bd13b298fe469b0e43e374797b3158d39b3b645c09eeca901df304eaddd84", "seconds": 14.8, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else.", "Hang on, forget that request. Don't do it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original task, then clearly revokes it with 'Hang on, forget that request. Don't do it.' This constitutes a final, unambiguous cancellation of the entire task. The control intent matches the 'cancellation' type criterion.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else. Hang on, forget that request. Don't do it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "8c47bdfd38e2b6999b588211ccaf34e7531817518ee500e8880bbc0b7174a11a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/ambiguous_halt/halt_038/03e5432d867e15a2/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "264464defb21e853f4508322fda03b7701d59ad160cd8696d61396be7753eeb2", "seconds": 13.92, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the full original instruction and then makes a final interjection requesting to pause. The phrase 'Could you pause there for a second?' is a clear, audible request to hold the current operation, with no indication of cancellation or resumption. This matches the 'ambiguous_halt' type criterion, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c77c005_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7997c1c839172c662973b73a7bbd739a5cefb4bb96d7db9e2f7f7a571b20458c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c77c005_3/audio_natural/filler/original/126fc4eea33ef8ae/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "9cfa7eeb527a6cc230f9cce3fe88660b876d3ad73be501d0c462ed47e1cdb58a", "seconds": 10.12, "spoken_utterances": [ "Befriend on hmmm... Venmo anyone I have received money from uh... this month and unfriend everyone else." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core instruction. The final intent matches the original instruction exactly, with no correction or modification. The 'filler' type allows for such hesitations without altering the task goal.", "heard_transcript": "Befriend on Venmo anyone I have received money from uh this month and unfriend everyone else." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_collection_name", "proposed_instruction": "Play the least listened to song on Spotify from the The Velvet Echo—scratch that, make it Echo Chamber Chronicles album.", "long_pause_offsets": [], "source_sha256": "a9529b45a9f2f6fa79aed7944ba3fe0dae6e8d235585a03e38a26ac11233b8ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'The Velvet Echo' but immediately corrects it to 'Echo Chamber Chronicles' using 'scratch that, make it'. The final target album is correctly specified, and all other conditions (action, app, and song selection criteria) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/false_start/false_start_03_collection_name/d8b7bc10bd9dadf0/speech.wav", "wer": 0.0, "wav_sha256": "394290555adec032afc436ec9d0ae9e631dce4f272466ad61678536467f83abb", "seconds": 13.08, "spoken_utterances": [ "Play the least listened to song on ... Spotify from the uh... The Velvet Echo—um... scratch that, make it Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Velvet Echo' but immediately self-corrects with 'scratch that' and specifies 'Echo Chamber Chronicles album'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "Play the least listened to song on Spotify from the uh the Velvet Echo um scratch that, make it Echo Chamber Chronicles album." }, "candidate_id": "wav_1abfd8d9648854ad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, natural-sounding false start with a spontaneous-sounding 'scratch that' correction. The initial error ('The Velvet Echo') is plausible, and the correction is immediate and integrated into the flow. The prosody is consistent, and the self-correction feels like a genuine moment of recall, making it the most realistic false start." } ] }, "is_default": true }, { "variant_id": "false_start_01_song_selection", "proposed_instruction": "Play the most played—no, I mean least listened to song on Spotify from the Echo Chamber Chronicles album.", "long_pause_offsets": [], "source_sha256": "2e6a9f6da7327f0b0be273864655290778197dab53f4da20a53b685d463a2ae6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately corrects to 'least listened to'. The app, album, and action remain consistent. The correction is clear and final." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/false_start/false_start_01_song_selection/e57544c59c095638/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "da7e0fcf790177ca69436437d8f918c84e652a9c84b0aa1d0ffb9c19df158b38", "seconds": 13.32, "spoken_utterances": [ "Play the most played—ah... no, I mean least listened to song on hmmm... Spotify from the uh... Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'least listened to'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including the platform (Spotify) and album (Echo Chamber Chronicles), remains unchanged and is preserved. Therefore, the final audible intent matches the original instruction after the correction.", "heard_transcript": "Play the most played uh, no, I mean least listened to song on Spotify from the uh Echo Chamber Chronicles album." }, "candidate_id": "wav_358ed1b18e2ddcb5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker makes a false start ('most played') and corrects to 'least listened to' with a natural-sounding 'no, I mean' repair. The correction is immediate and the intonation reflects a genuine self-correction. The pacing and pauses are realistic, though the initial error is less plausible than audio_03's, making it slightly less realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play the least listened to song on Apple Music—I should say Spotify from the Echo Chamber Chronicles album.", "long_pause_offsets": [], "source_sha256": "129a213bde1debb92675213076fa061b9c801c460c1d3695651675ac67eb8d0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, the album, and the 'least listened to' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/false_start/false_start_02_app/dab27037a8970756/speech.wav", "wer": 0.19047619047619047, "wav_sha256": "907c2e268fdf33a4f2d18462c18b4b4c856791ab3446b02622b4983a4639aa4b", "seconds": 12.64, "spoken_utterances": [ "Play the ... least listened to song on hmmm... Apple Music—ah... I should say Spotify from the Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction matches the reference exactly.", "heard_transcript": "Play the least listened to song on um, Apple Music. I should say Spotify from the Echo Chamber Chronicles album." }, "candidate_id": "wav_d16af4af6639d1cb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with a false start ('Apple Music') and corrects to 'Spotify'. The correction is clear, but the initial error is less plausible than the others, and the 'I should say' repair feels slightly more rehearsed. The prosody is natural, but the flow is less smooth than the top two." } ] }, "is_default": false }, { "variant_id": "false_start_04_collection_kind", "proposed_instruction": "Play the least listened to song on Spotify from the Echo Chamber Chronicles playlist—I should have said album.", "long_pause_offsets": [], "source_sha256": "ae785830e408cc6dc8e8fc0f8dc5624b6c3776ba060ec6e9225d904754f50aba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'playlist' to 'album' within the same utterance, preserving the app, action, and song selection criteria." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/false_start/false_start_04_collection_kind/36b567a4e60b2cf8/speech.wav", "wer": 0.0, "wav_sha256": "cc8cd8b64499757a7adb95bdb8b2e7bae422ec4e0aff649941d96bf6f39e459e", "seconds": 10.36, "spoken_utterances": [ "Play the least listened to song on Spotify from the Echo Chamber Chronicles uh... playlist—um... I should have said album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects to 'album'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('playlist') is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "Play the least listened to song on Spotify from the Echo Chamber Chronicles uh playlist um I should've said album." }, "candidate_id": "wav_eb6d7d4c5b2057bc", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker makes a false start ('playlist') and corrects to 'album'. However, the correction is delivered with a noticeable pause and a hesitant 'I should've said album' that sounds less spontaneous and more like a rehearsed or read-aloud correction. The prosody is slightly off, and the overall delivery feels less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Play the song on Spotify from the Echo Chamber Chronicles album. Oh, I forgot to mention, play the least listened to song.", "long_pause_offsets": [ 65 ], "source_sha256": "3a4c255c81fcff071b02b462fc045f85e90445c023365945105462ec7e938a34", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least listened to' to the initial request, refining the selection from any song to the specific one with the lowest listen count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/addition/addition_01/8e1c626d58d22d44/speech.wav", "wer": 0.04, "wav_sha256": "652dbbdf7d99dd9dbc3d355cd473569b67286c03034e8b74956aba40aef7034d", "seconds": 13.44, "spoken_utterances": [ "Play the song on hmmm... Spotify from the um... Echo Chamber Chronicles album. ", "Oh, I forgot to mention, play the uh... least listened to song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests playing 'the song' from the specified album on Spotify, then adds the constraint 'least listened to song' after a pause. This matches the addition type criterion: the final goal combines the original instruction with the added information. The final intent is clear and executable.", "heard_transcript": "Play the song on um Spotify from the um Echo Chamber Chronicles album. Oh, I forgot to mention, play the uh least listened to song." }, "candidate_id": "wav_8f943520c21fbd18", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Play the least listened to song on from the Echo Chamber Chronicles album. Oh, I forgot to say, play it on Spotify.", "long_pause_offsets": [ 75 ], "source_sha256": "61547d64ec79bacd9df4bbf334a7042f36ae160701a62aaaafadc04b0f653cf5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app requirement; the initial utterance already specifies the album and the 'least listened to' criterion. The final goal combines both parts without contradiction." }, "audio": { "status": "pending", "spoken_utterances": [ "Play the ... least listened to song on from the um... Echo Chamber Chronicles album. ", "Oh, I forgot to say, play it on let me think... Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "Play the least listened to song on Apple Music from the Echo Chamber Chronicles album. Let me correct the app: play the song on Spotify, not Apple Music.", "long_pause_offsets": [ 87 ], "source_sha256": "a27b5fc655ba5979db5a7db50f2355b9992052b62b5729992202807ee98a0fae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the album and the 'least listened to song' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/revision/revision_02_app/ba7c48b4941bc577/speech.wav", "wer": 0.027777777777777776, "wav_sha256": "15e8f4070229850d0f7a573e0df5d9af5ecf3f4958073b03435c2511e41ae9aa", "seconds": 19.08, "spoken_utterances": [ "Play the um, let me think... least listened to song on Apple Music from the um... Echo Chamber Chronicles album. ", "Let me correct the app: play the song on let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the task on Apple Music but immediately corrects it to Spotify after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, with the correct app (Spotify) and album ('Echo Chamber Chronicles').", "heard_transcript": "Play the um, let me think, least listened to song on Apple Music from the um, Echo Chamber Chronicles album. Let me correct the app, play the song on, let me think, Spotify not Apple Music." }, "candidate_id": "wav_9517a11cd794125f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural 'let me think' filler, followed by a believable correction of the platform from Apple Music to Spotify. The self-repair is smooth, with a pause and a rephrased, corrected version that flows logically. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a scripted edit." } ] }, "is_default": true }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "Play the most played song on Spotify from the Echo Chamber Chronicles album. Wait, play the least listened to song, not the most played.", "long_pause_offsets": [ 77 ], "source_sha256": "98e77e7c38e417fc26a4622d32202d84d87ee704464337b498a0b33ee2d74136", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial instruction by replacing 'most played' with 'least listened to', preserving the app, album, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/revision/revision_01_song_selection/57222d32483fb3c9/speech.wav", "wer": 0.23333333333333334, "wav_sha256": "84f9c98bbd03258acfccbb4d17461da190940a363aa66ec894eda9628bb12694", "seconds": 16.16, "spoken_utterances": [ "Play the most played song on hmmm... Spotify from the uh... Echo Chamber Chronicles album. ", "Wait, play the um, let me think... least listened to song, not the most played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately corrects to 'least listened to' after a pause and self-interruption. This constitutes a revision where the final audible intent is the corrected version. The correction is clear and unambiguous, and the rest of the instruction (Spotify, Echo Chamber Chronicles album) remains unchanged. Therefore, the final intent matches the original reference instruction.", "heard_transcript": "Play the most played song on Spotify from the Echo Chamber Chronicles album. Wait, play the um, let me think, least listened to song, not the most played." }, "candidate_id": "wav_1f4ad6f2f011cb84", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip starts with a clear request but includes a self-correction from 'most played' to 'least listened to'. The 'let me think' pause is natural, and the correction is integrated into the same utterance. However, the correction is slightly less fluid than in audio_02, with a brief hesitation and a slightly more abrupt shift in phrasing, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_03_collection_name", "proposed_instruction": "Play the least listened to song on Spotify from the The Silent Waves album. Use the Echo Chamber Chronicles album instead of The Silent Waves.", "long_pause_offsets": [ 76 ], "source_sha256": "edb5792d5eb3cd92f705867dbca3114963f3455fa7c1390e561275dd169fe4d6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the album name from 'The Silent Waves' to 'Echo Chamber Chronicles', while preserving the action, app, and condition of playing the least listened to song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/revision/revision_03_collection_name/dbc5309e9382db38/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "fd0e57242acc560234bb170f23c98468a199468d43cfa762eed2928aa1144572", "seconds": 15.28, "spoken_utterances": [ "Play the um... least listened to song on Spotify from the uh... The Silent Waves album. ", "Use the ... Echo Chamber Chronicles album instead of The Silent Waves." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the least listened to song from the 'Silent Waves' album, then corrects this by stating 'Use the Echo Chamber Chronicles album instead of the Silent Waves.' This constitutes a revision where the final target album is clearly specified as 'Echo Chamber Chronicles.' The action (play least listened to song) and platform (Spotify) remain consistent. The revision is clear and final, satisfying the revision type criterion.", "heard_transcript": "Play the um least listened to song on Spotify from the uh the Silent Waves album. Use the Echo Chamber Chronicles album instead of the Silent Waves." }, "candidate_id": "wav_c328a5d7f79385cc", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a request for the 'least listened to song on Spotify from the Silent Waves album', then adds a revision to use 'Echo Chamber Chronicles' instead. The addition is awkwardly phrased as 'use the Echo Chamber Chronicles album instead of the Silent Waves', which sounds more like a post-hoc instruction than a natural thought process. The delivery lacks the smoothness of a self-correcting thought and feels more like a disjointed addition, reducing its realism." } ] }, "is_default": false }, { "variant_id": "revision_04_collection_kind", "proposed_instruction": "Play the least listened to song on Spotify from the Echo Chamber Chronicles playlist. That should be the album Echo Chamber Chronicles, not a playlist.", "long_pause_offsets": [ 86 ], "source_sha256": "309264e940066b2cac040313d69fa75ecfbbdaa52d3401791ea1f715bac6da39", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'playlist' to 'album', preserving the app, action, and song selection criteria." }, "audio": { "status": "pending", "spoken_utterances": [ "Play the uh... least listened to song on ... Spotify from the Echo Chamber Chronicles playlist. ", "That should be the um... album Echo Chamber Chronicles, not a playlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_020", "hide_variant_original": true, "source_sha256": "80804cf19dc2812a38f8ac09f4bba4845e4ccbf25cc997022117eafdec3a8ad3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/cancellation/cancel_020/bb1687a73898ac21/speech.wav", "wer": 0.25, "wav_sha256": "bc0149fc7323f4cfe632457c9b32ee4ad1d86d3d39303b01d275478a0e0eedba", "seconds": 13.08, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album.", "Hang on, never mind, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hang on, never mind, cancel that request,' which constitutes a direct and unambiguous cancellation of the previously stated task. The final control intent is to cancel, matching the expected behavior for the 'cancellation' type. No execution of the original task is implied.", "heard_transcript": "Play the um least listened to song on uh Spotify from the um Echo Chamber Chronicles album. Hang on, never mind, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "4866f4f47f8cba78e9f2e6c40e6bab18776afb112780dfb97281c1e73b46041b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/ambiguous_halt/halt_037/0f6a5d1c666b4283/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "c050cc9ebc0728c9fef87b58bdd93c0a30531168b1bd1aff765537b13fb20f58", "seconds": 11.72, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album.", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a clear final interjection 'Pause for a moment' after the completion of the initial request. This constitutes an audible request to pause, with no indication of cancellation or permission to resume. The initial request to play a song is not revoked, but the final control command is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Play the least listened to song on uh Spotify from the um Echo Chamber Chronicles album. Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1db01108f1c3ec79329e1822a6fd3e07459703dd3b36558ecf38046892ee16ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_1/audio_natural/filler/original/4fb7e632479bf551/speech.wav", "wer": 0.35294117647058826, "wav_sha256": "8b7b7d21c38188fe5c0cec9c147499a621d808536f7f8ef72d48b6b6ec6791f3", "seconds": 8.28, "spoken_utterances": [ "Play the hmmm... least listened to song on uh... Spotify from the um... Echo Chamber Chronicles album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') and pauses but does not change the core request. The final intent matches the original instruction, with only hesitation markers present. The filler type allows ignoring such hesitations, so the task is executed as intended.", "heard_transcript": "Play the um least listened to song on uh Spotify from the um Echo Chamber Chronicles album." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_selection", "proposed_instruction": "Play the most played—sorry, I meant most listened to song on Spotify from my Woodstock Reimagined: Festival Vibes playlist.", "long_pause_offsets": [], "source_sha256": "94a10c6b2064c4b8b7c803d9a0c83cf4feae3deacf746301f693f3539b477a21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'most played' to 'most listened to' within the same utterance, preserving the app, playlist, and action. The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/false_start/false_start_01_song_selection/28ad8a10201fd4f7/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "167ce66c2f5908790e573991750d891d382aa307bd63e5dc7d98de36f99904f6", "seconds": 13.28, "spoken_utterances": [ "Play the most played—hmmm... sorry, I meant ah... most listened to song on hmmm... Spotify from my Woodstock Reimagined: Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'most listened to'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction, including the platform (Spotify) and playlist name, remains consistent and clear. The correction is immediate and unambiguous, so the final intent matches the original instruction.", "heard_transcript": "Play the most played [inaudible] sorry, I meant uh most listened to song on [inaudible] Spotify from my Woodstock Reimagined: Festival Vibes playlist." }, "candidate_id": "wav_6b85307ab11abe27", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a very natural false start with 'most played' corrected to 'most listened to song' and 'Spotify' corrected to 'Spotify from my Woodstock Reimagined Festival Vibes playlist'. The self-correction is immediate and fluid, with a believable hesitation and intonation shift. The speaker's prosody is consistent and the pacing reflects genuine thought processing." } ] }, "is_default": true }, { "variant_id": "false_start_04_collection_kind", "proposed_instruction": "Play the most listened to song on Spotify from my Woodstock Reimagined: Festival Vibes album—no, I mean playlist.", "long_pause_offsets": [], "source_sha256": "ec22392f84002a1eade63d4e393ab6d572423e52a6aed0e6f9307c62021df0da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'album' but immediately corrects to 'playlist'. The final target is the playlist, preserving the app, action, and ranking criteria." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/false_start/false_start_04_collection_kind/e55ee474e7183719/speech.wav", "wer": 0.0, "wav_sha256": "ff039134f2deb37452494396681beee5d79e751905b740259f61108ef62b4d8c", "seconds": 10.8, "spoken_utterances": [ "Play the most listened to song on Spotify ... from my Woodstock Reimagined: Festival Vibes album—uh... no, I mean playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'album' but immediately self-corrects to 'playlist'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('album') is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "Play the most listened to song on Spotify from my Woodstock Reimagined Festival Vibes album. Uh, no, I mean playlist." }, "candidate_id": "wav_49fd1a160a3e5f7d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision: the speaker initially says 'album' and then corrects to 'playlist'. The correction is delivered with a natural 'uh, no, I mean' structure, and the intonation reflects a spontaneous change of mind. The prosody is smooth and the correction is integrated into the flow without sounding forced." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play the most listened to song on Apple Music—I should say Spotify from my Woodstock Reimagined: Festival Vibes playlist.", "long_pause_offsets": [], "source_sha256": "55a7cffc9f17c313927e7242ee1fb09f75e00e008019777e9226e07890646772", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the app, playlist, and action. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/false_start/false_start_02_app/1baf114072dd28c5/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "8355e5fb7f325767b7a17e48baacf8a6e3b0d8ef5fcfe45a1b34f4786d45891d", "seconds": 11.6, "spoken_utterances": [ "Play the ... most listened to song on uh... Apple Music—ah... I should say Spotify from my Woodstock Reimagined: Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the playlist name, is preserved. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Play the most listened to song on uh Apple Music uh. I should say Spotify, for my Woodstock Reimagined: Festival Vibes playlist." }, "candidate_id": "wav_c6178e99e13ee242", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip includes a self-correction from 'Apple Music' to 'Spotify'. The correction is made with a 'I should say' phrase, which is a common and realistic way to self-correct. However, the initial part of the sentence is slightly rushed, and the correction feels a bit more rehearsed than the others, making it less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_03_collection_name", "proposed_instruction": "Play the most listened to song on Spotify from my Summer Vibes—scratch that, make it Woodstock Reimagined: Festival Vibes playlist.", "long_pause_offsets": [], "source_sha256": "138c33366c50822491327f3bce3a3e8db648b3c13e2065da53033585508ecf3a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Summer Vibes' but immediately corrects it to 'Woodstock Reimagined: Festival Vibes'. The final playlist name matches the original instruction, and all other conditions (action, app, and song selection criteria) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/false_start/false_start_03_collection_name/0e1bf830cce36b5f/speech.wav", "wer": 0.22727272727272727, "wav_sha256": "4afca27d5d28f4ac63b3b19f769cd437d453af3f02f2a149952e8045c05cb527", "seconds": 13.88, "spoken_utterances": [ "Play the ... most listened to song on Spotify from my Summer Vibes—uh... scratch that, make it ... Woodstock Reimagined: Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Summer Vibes' but immediately self-corrects with 'Scratch that, make it Woodstock Reimagined: Festival Vibes playlist.' This is a clear false start where the final target playlist is explicitly corrected. The final intent matches the original instruction, so the task should be executed with the corrected playlist.", "heard_transcript": "Play the most listened to song on Spotify from my Summer Vibes uh. Scratch that, make it Woodstock Reimagined: Festival Vibes playlist." }, "candidate_id": "wav_d1584800f509a62c", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a revision where the speaker first says 'Summer Vibes' and then corrects to 'Woodstock Reimagined Festival Vibes playlist'. While the correction is present, the initial phrase is delivered flatly and the correction is abrupt, lacking the natural hesitation or intonation shift seen in the others. The overall delivery sounds more like a read-aloud with a minor edit than a spontaneous correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Play the most listened to song on from my Woodstock Reimagined: Festival Vibes playlist. I should also mention, play it on Spotify.", "long_pause_offsets": [ 89 ], "source_sha256": "334a2439887ff5ac32e84bb986a8d7a4d070eca948baf58ab20e61a37ea68633", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify', which was implied in the original instruction but not explicitly stated. The core action, playlist, and ranking criterion remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/addition/addition_02/bf24f218e37b0540/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "1fcb1a1450ba50b84e07b238859a3ccdeb36cde3b54c9fb5359bc86be1ee7bb4", "seconds": 15.4, "spoken_utterances": [ "Play the uh... most listened to song on ... from my Woodstock Reimagined: Festival Vibes playlist. ", "I should also mention, play it on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is partially garbled ('Play the uh most listened to song on from my...'), but the final addition 'I should also mention, play it on Spotify' clarifies the platform. The addition specifies the context for the song playback, combining with the original request to form a complete and coherent task. The final intent matches the reference instruction with the addition of the platform, which is a valid refinement under the 'addition' type.", "heard_transcript": "Play the uh most listened to song on from my Woodstock Reimagined: Festival Vibes playlist. I should also mention, play it on Spotify." }, "candidate_id": "wav_918dd892785a9aee", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Play the song on Spotify from my Woodstock Reimagined: Festival Vibes playlist. Oh, I forgot to say, play the most listened to song.", "long_pause_offsets": [ 80 ], "source_sha256": "9b27faba25d4a7211a17503032624dc9f50dcd4992c3ea6c8555fb5d6a312123", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'most listened to' to the initial request, refining the scope of the song to be played. The playlist and app remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Play the song on uh... Spotify from my Woodstock Reimagined: Festival Vibes playlist. ", "Oh, I forgot to say, play the ... most listened to song." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_song_selection", "proposed_instruction": "Play the most played song on Spotify from my Woodstock Reimagined: Festival Vibes playlist. Wait, that should be the most listened to song, not the most played.", "long_pause_offsets": [ 92 ], "source_sha256": "ec3b4488f319a3f66e9ca8121020aefddc014d5b6a8d1b03eced42add6fd38cf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the phrase 'most played' to 'most listened to', which is semantically equivalent in this context. The app, playlist, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/revision/revision_01_song_selection/9933c7206863a222/speech.wav", "wer": 0.1, "wav_sha256": "f51ccf7676eace0974a10f7757e4deafe722a87e28697eef3454515138a9e8f8", "seconds": 17.12, "spoken_utterances": [ "Play the let's see... most played song on Spotify uh... from my Woodstock Reimagined: Festival Vibes playlist. ", "Wait, that should be the ... most listened to song, not the most played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played song' but immediately corrects themselves to 'most listened to song' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to play the 'most listened to song', which matches the original instruction. All other conditions, including the app (Spotify) and playlist, remain unchanged.", "heard_transcript": "Play the uh let's see most played song on Spotify uh from my Woodstock Reimagined: Festival Vibes playlist. Wait, that should be the most listened to song, not the most played." }, "candidate_id": "wav_03a3c95347891491", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and spontaneous revision. The speaker begins with a slightly hesitant 'uh' and 'let's see,' then makes a clear mistake ('most played') which is immediately self-corrected with 'Wait, that should be the most listened to song.' The correction is fluent and integrated into the same utterance, with a natural pause and intonation shift. The prosody is consistent, and the flow feels like a genuine moment of recall and correction." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Play the most listened to song on Apple Music from my Woodstock Reimagined: Festival Vibes playlist. Sorry, I meant Spotify, not Apple Music.", "long_pause_offsets": [ 101 ], "source_sha256": "9339b3dd4ce123b4ec57dd8bc71a4b645bc4ad669f0b76829e316fe734c2e35f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the playlist and the action of playing the most listened to song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/revision/revision_02_app/cdc114d647888cb5/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "edd6e8c8afa67e40053c1f0611a4ef347e2e4ebb0e0a492fd36ba457872be9b5", "seconds": 14.4, "spoken_utterances": [ "Play the um... most listened to song on ah... Apple Music from my Woodstock Reimagined: Festival Vibes playlist. ", "Sorry, I meant uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested the task on Apple Music but immediately corrected it to Spotify. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final intent is to execute the task on Spotify, which matches the original reference instruction. All other constraints (song, playlist) remain unchanged.", "heard_transcript": "Play the um most listened to song on uh Apple Music from my Woodstock Reimagined: Festival Vibes playlist. Sorry, I meant uh Spotify, not Apple Music." }, "candidate_id": "wav_c06e18efbbbf0ca8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a revision, but the correction is less integrated. The speaker says 'Sorry, I meant uh Spotify not Apple Music,' which is a clear correction but delivered as a separate sentence rather than a smooth self-repair within the original request. The 'uh' filler before 'Spotify' sounds slightly awkward and less spontaneous than the 'Wait' in audio_01. The overall delivery is still realistic but less fluid in its revision flow." } ] }, "is_default": false }, { "variant_id": "revision_03_collection_name", "proposed_instruction": "Play the most listened to song on Spotify from my Summer Vibes 2023 playlist. Change the playlist to \"Woodstock Reimagined: Festival Vibes\" instead of \"Summer Vibes 2023\".", "long_pause_offsets": [ 78 ], "source_sha256": "c259012148f529c309b62dc7af7dc171b450fcce9fac31a98b613f9b1cd4353b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the playlist name from 'Summer Vibes two thousand twenty-three' to 'Woodstock Reimagined: Festival Vibes', preserving the action and the 'most listened to song' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/revision/revision_03_collection_name/1727859d7db34e66/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "60a8317afbb0f9998a360435ebc68bcec65311b60294615c12eae89f1737fab2", "seconds": 18.6, "spoken_utterances": [ "Play the um, let me think... most listened to song on Spotify from my Summer Vibes two thousand twenty-three playlist. ", "Change the playlist to \"uh... Woodstock Reimagined: Festival Vibes\" instead of \"Summer Vibes two thousand twenty-three\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the playlist as 'Summer Vibes 2023' but immediately corrects it to 'Woodstock Reimagined: Festival Vibes' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to play the most listened to song from the corrected playlist, which matches the original reference instruction.", "heard_transcript": "Play the um, let me think, most listened to song on Spotify from my Summer Vibes 2023 playlist. Change the playlist to uh, Woodstock Reimagined: Festival Vibes instead of Summer Vibes 2023." }, "candidate_id": "wav_2c5c55d83713841a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip features an addition rather than a revision. The speaker starts with 'let me think,' then makes a request, and later adds a new instruction about changing the playlist. The second part is delivered as a separate, slightly abrupt statement. The flow feels less like a single spontaneous thought process and more like two distinct requests stitched together, reducing its realism compared to the other two clips that show more natural self-correction within one utterance." } ] }, "is_default": false }, { "variant_id": "revision_04_collection_kind", "proposed_instruction": "Play the most listened to song on Spotify from my Woodstock Reimagined: Festival Vibes album. Make that from my Woodstock Reimagined: Festival Vibes playlist, not album.", "long_pause_offsets": [ 94 ], "source_sha256": "52509434f18d5f8ab479997e023e53be11c90634a1164d559a4ecd5d6035a211", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'album' to 'playlist', preserving the app, action, and ranking criterion." }, "audio": { "status": "pending", "spoken_utterances": [ "Play the uh... most listened to song on Spotify ... from my Woodstock Reimagined: Festival Vibes album. ", "Make that from my Woodstock Reimagined: Festival Vibes um... playlist, not album." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "6147b2d0288de8b36992d526cd2402617612c995bf5f928298e226f6320f4c53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/cancellation/cancel_038/bfbaef0b133a05f2/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "45bfd4be35bccac4ac8604211c4fb8d368c13f5f8ca4c57c1da560b75cbb2443", "seconds": 12.52, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist.", "Hang on, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially issues a request to play a song from a playlist, then immediately follows with 'Hang on, please stop and cancel that request,' which is a clear and unambiguous cancellation of the prior task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist. Hang on, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_023", "hide_variant_original": true, "source_sha256": "705fb2d8927888440ee5c39b97ce8c19bed0109d4bc1e2dcac2f9da658368d63", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/ambiguous_halt/halt_023/1b7b621af27aaed6/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "27e4cce5c7598879715142c9fa048fba28780302a8ba6bab4ab69b1ec2665a61", "seconds": 11.52, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist.", "Hold on, I need a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The final interjection 'Hold on, I need a second' is a clear request to pause or wait, with no indication of cancellation or resumption. The initial request is fully audible and complete, and the final control intent is a pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist. Hold on, I need a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "44523e640e617f18394f25016dcbf2c4e9d5048e2295cfa71c25cae1673be7d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_2/audio_natural/filler/original/ce6b094e0af9b6b5/speech.wav", "wer": 0.125, "wav_sha256": "aa2be070c12f4f02c4bae665013b81d60d0615670b50d358dc16978ea7dee7ea", "seconds": 7.6, "spoken_utterances": [ "Play the ... most listened to song on Spotify um... from my Woodstock Reimagined, Festival Vibes playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word ('um') but does not change the core task. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Play the most listened to song on Spotify um from my Woodstock Reimagined: Festival Vibes playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_collection_name", "proposed_instruction": "Play the most listened to song on Spotify from the The Velvet Underground—no, I meant Velvet Underground album.", "long_pause_offsets": [], "source_sha256": "a02d8294a564bd9e96e633c82cab67f78f0e8d3c6e3556ec5c7f088ae27af1bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misstates the album name as 'The Velvet Underground' but immediately corrects it to 'Velvet Underground'. The final intended album name matches the original instruction, and all other requirements (action, app, and song selection criteria) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/false_start/false_start_03_collection_name/1d3a0c57deb99047/speech.wav", "wer": 0.3181818181818182, "wav_sha256": "18bff83e8e622e3c39ba65b9e4462c6d5f1b9c11ac6ae1887d485fa3cfc61c33", "seconds": 13.08, "spoken_utterances": [ "Play the ah... most listened to song on uh... Spotify from the The Velvet Underground—um... no, I meant Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces 'Velvet Underground' as 'Develvet Underground' but immediately self-corrects with 'No, I meant Velvet Underground album.' This is a classic false_start where the final, corrected version of the target is clearly stated. The control intent is to execute the corrected request, which matches the original instruction.", "heard_transcript": "Play the uh most listened to song on uh Spotify from the Develvet Underground um. No, I meant Velvet Underground album." }, "candidate_id": "wav_f9b6691f75c0f270", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker initially says 'Velvet Undergroundm' with a clear mispronunciation, then immediately self-corrects with 'No, I meant Velvet Underground album' in a way that sounds like a genuine, spontaneous recovery. The prosody is consistent, the pause before the correction is natural, and the intonation reflects a real-time realization of the error." } ] }, "is_default": true }, { "variant_id": "false_start_01_song_selection", "proposed_instruction": "Play the most played—I mean, most listened to song on Spotify from the Velvet Underground album.", "long_pause_offsets": [], "source_sha256": "59a209e37e4d9254a7d4323aa69f8e3a2fdb908c6a381640b3c6068c5ca03ce2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'most played' to 'most listened to' within the same utterance, preserving the app, album, and action. The final intended goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/false_start/false_start_01_song_selection/57cc128317bbfd82/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "db9fe604e72f139103920f5362428a3717c4eec02b4b94dc267d44de0419d69e", "seconds": 10.76, "spoken_utterances": [ "Play the most played—uh... I mean, most listened to song on Spotify from the hmmm... Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'most listened to'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only a minor grammatical correction ('played' to 'listened to') that does not change the meaning. The target (Velvet Underground album) and platform (Spotify) are preserved. Therefore, the task is executed as intended.", "heard_transcript": "Play the most played uh, I mean most listened to song on Spotify from the Velvet Underground album." }, "candidate_id": "wav_e91cd63d63dccbff", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'most played' corrected to 'most listened to'. The correction is natural and the speaker's intonation suggests a moment of self-awareness. However, the initial error ('played') is less likely than a mispronunciation, and the correction feels slightly more rehearsed than in audio_02. The overall delivery is still very realistic." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Play the most listened to song on Apple Music—sorry, make that Spotify from the Velvet Underground album.", "long_pause_offsets": [], "source_sha256": "7b0ab311b6cd4ac92bac50ea6efa31fe802997e2ec43fe2ded7e5f5d9b7401c9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance; the app, action, and album remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/false_start/false_start_02_app/778fc9e767881684/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "0a4301570c664bd31800fdb4f474e702e79b5d75fddf775b16308f25ff0593d8", "seconds": 10.24, "spoken_utterances": [ "Play the um... most listened to song on Apple Music—hmmm... sorry, make that Spotify from the um... Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "Play the um most listened to song on Apple Music. Hmm. Sorry, make that Spotify from the um Velvet Underground album." }, "candidate_id": "wav_c38ffef5df464a28", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a correction from Apple Music to Spotify, which is a change of platform rather than a self-repair of a word or value. While the 'sorry, make that Spotify' is a plausible follow-up, the initial error is not a mispronunciation or misstatement but a different service, making the correction less typical of a spontaneous false start. The delivery is natural, but the nature of the correction is less aligned with the 'false_start' augmentation type." } ] }, "is_default": false }, { "variant_id": "false_start_04_collection_kind", "proposed_instruction": "Play the most listened to song on Spotify from the Velvet Underground playlist—or rather, album.", "long_pause_offsets": [], "source_sha256": "4fea31fd672e2fe109db23fe3d0140f4a5cbdba43026e32a5dccb3bec937520c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately corrects to 'album' within the same utterance. The final intended target is the album, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/false_start/false_start_04_collection_kind/74c1c8828376b7af/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "316bf8a338a91f7afecfe674116a0cf13bc7efce8248cc349659805af6d0b739", "seconds": 11.16, "spoken_utterances": [ "Play the most listened to song on ... Spotify from the ... Velvet Underground playlist—um... or rather, album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'playlist' but immediately self-corrects to 'album' after a brief hesitation ('um rather'). According to the 'false_start' type criterion, the final corrected value ('album') should be applied, and the original incorrect value ('playlist') should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Play the most listened to song on Spotify from the Velvet Underground playlist um rather album." }, "candidate_id": "wav_636f5c5813a4dcf7", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio is the least realistic because it lacks any self-correction or repair. It ends with a hesitant 'um' and a pause, suggesting the speaker is searching for words, but there is no actual false start or correction. The delivery is flat and lacks the natural flow of a speaker recovering from an error, making it sound more like a hesitant read-aloud than a spontaneous utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Play the most listened to song on from the Velvet Underground album. I should also mention, do this on Spotify.", "long_pause_offsets": [ 69 ], "source_sha256": "62628513d25e1328709ca0069eecffc220c266570d1ebd52be64876018f84803", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the platform 'Spotify', which was implied but not explicitly stated in the original instruction. The core action, target album, and ranking criterion remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/addition/addition_02/d786761b3db0c735/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "cab87056e75c80f45e2bd14b888162b46a62fb9d8d946e71fd8acc89fab5d24f", "seconds": 11.28, "spoken_utterances": [ "Play the ah... most listened to song on from the uh... Velvet Underground album. ", "I should also mention, do this on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is slightly hesitant with fillers ('uh') but clearly requests playing the most listened to song from the Velvet Underground album. The addition 'do this on Spotify' is a clear specification of the platform, which is a valid addition to the original instruction. The final intent combines both the original request and the added constraint, matching the expected behavior for an 'addition' type task.", "heard_transcript": "Play the uh most listened to song on from the uh Velvet Underground album. I should also mention do this on Spotify." }, "candidate_id": "wav_06048f1bd4ce7cef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural prosody and a believable addition. The initial request is slightly hesitant with 'uh' fillers, which is realistic. The addition 'I should also mention do this on Spotify' is smoothly integrated with a natural pause and intonation shift, sounding like a spontaneous afterthought rather than a rehearsed or robotic insertion." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Play the song on Spotify from the Velvet Underground album. Oh, I forgot to say, play the most listened to song.", "long_pause_offsets": [ 60 ], "source_sha256": "6848f1ad74670b7922e2d6370c7815d9c9e060b79b43a27c3cf361c48d308eb7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most listened to' to the initial request, refining the scope of the song to be played. The app, album, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/addition/addition_01/37cea2a0a75414b4/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "81730ef153fdc9d0b46a10b4435d41be4371221a13165ebf571d346f1d391d28", "seconds": 13.52, "spoken_utterances": [ "Play the song on um... Spotify from the Velvet Underground album. ", "Oh, I forgot to say, play the hmmm... most listened to song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to play a song from the Velvet Underground album on Spotify. The speaker then adds the constraint 'play the most listened to song', which narrows the scope of the original request. The final intent combines the original request with the added detail, matching the reference instruction. The addition is clear and directly modifies the initial task without introducing conflicting or contradictory information.", "heard_transcript": "Play the song on um Spotify from the Velvet Underground album. Oh, I forgot to say, play the most listened to song." }, "candidate_id": "wav_6a5ee9a1b05d1456", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but includes a noticeable pause and filler ('um') that feels slightly unnatural in pacing. The second part, 'Oh, I forgot to say play the most listened to song,' is a revision that sounds more like a scripted correction than a spontaneous thought. The intonation and flow are less fluid compared to audio_01, making it feel less like a natural, in-the-moment utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_collection_kind", "proposed_instruction": "Play the most listened to song on Spotify from the Velvet Underground playlist. Let me correct: it should be the Velvet Underground album, not playlist.", "long_pause_offsets": [ 80 ], "source_sha256": "59ddeab867f14550ae77d6e58a1b8d318ee99627c5ff53e69b64aa9dd784ba23", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the source from 'playlist' to 'album', preserving the artist, app, and action. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/revision/revision_04_collection_kind/0d7e28cea339635c/speech.wav", "wer": 0.0, "wav_sha256": "8f5ef3d4b20dea3888f437ad931501a7e149cd9e84f7232865556be82e83393f", "seconds": 16.8, "spoken_utterances": [ "Play the uh... most listened to song on ... Spotify from the let's see... Velvet Underground playlist. ", "Let me correct: it should be the Velvet Underground album, not playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially refers to the 'Velvet Underground playlist' but immediately corrects themselves to 'the Velvet Underground album, not playlist'. This constitutes a revision where the final, corrected target is the album. The rest of the instruction, including the action (play), the metric (most listened to), and the platform (Spotify), remains unchanged. The revision is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Play the uh most listened to song on Spotify from the let's see, Velvet Underground playlist. Let me correct, it should be the Velvet Underground album, not playlist." }, "candidate_id": "wav_ba332f356f60914b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a plausible initial request, then hesitates with 'let's see' before correcting 'playlist' to 'album' with a clear, self-aware, and smoothly integrated self-correction. The prosody is consistent, and the correction feels like a genuine moment of recollection rather than a rehearsed edit." } ] }, "is_default": true }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "Play the most played song on Spotify from the Velvet Underground album. Wait, that should be the most listened to song, not the most played.", "long_pause_offsets": [ 72 ], "source_sha256": "6fdaa7e9b7379a050c6d99d4916769dc1619ec1e87eca7ea406c1fa2d13b5b25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the phrase 'most played' to 'most listened to', which is the same as the original instruction. The app, album, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/revision/revision_01_song_selection/f4fe79b27c2e3a41/speech.wav", "wer": 0.0, "wav_sha256": "1a0115114f343816912adc6da5f79b2472f527c47d066524b0cd92d7f3ed260e", "seconds": 14.28, "spoken_utterances": [ "Play the uh... most played song on Spotify from the Velvet Underground album. ", "Wait, that should be the ... most listened to song, not the most played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'most listened to' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to play the most listened to song, which matches the original instruction. The correction is clear and unambiguous, and the rest of the constraints (Spotify, Velvet Underground album) remain unchanged.", "heard_transcript": "Play the uh most played song on Spotify from the Velvet Underground album. Wait, that should be the most listened to song, not the most played." }, "candidate_id": "wav_6172119c811643db", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision where the speaker first says 'most played' and then corrects it to 'most listened to'. The pause and 'wait' are natural, and the correction is immediate and grammatically sound. The prosody is slightly less fluid than audio_03, with a slightly more robotic intonation, but the revision flow is still very realistic." } ] }, "is_default": false }, { "variant_id": "revision_03_collection_name", "proposed_instruction": "Play the most listened to song on Spotify from the The Beatles album. Change the collection to Velvet Underground, not The Beatles.", "long_pause_offsets": [ 70 ], "source_sha256": "a3fa8e39e340bce3b614089c440b5db745249416f78af254380caf4f17bc7e48", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the album from The Beatles to Velvet Underground, while retaining the action and the metric (most listened to song)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/revision/revision_03_collection_name/079b78b220895c49/speech.wav", "wer": 0.20833333333333334, "wav_sha256": "12def669431132b77eed573a950d4fb034d2240331e21f65cfda1ad6b990579f", "seconds": 14.04, "spoken_utterances": [ "Play the um... most listened to song on Spotify from the ... The Beatles album. ", "Change the collection to uh... Velvet Underground, not The Beatles." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests the most listened to song from The Beatles album, but then revises the target collection to the Velvet Underground. According to the revision type criterion, the correction spoken after the initial instruction overrides the earlier value. The final audible intent is to play the most listened to song from the Velvet Underground album, which matches the original reference instruction. All other constraints are preserved.", "heard_transcript": "Play the um most listened to song on Spotify from the the Beatles album. Change the collection to uh Velvet Underground, not the Beatles." }, "candidate_id": "wav_f915242e41173ef9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision from 'The Beatles' to 'Velvet Underground'. The correction is clear, but the initial request is delivered with a flat, slightly robotic prosody and lacks the natural hesitation or filler that would make it sound spontaneous. The revision itself is abrupt and lacks the subtle intonation shift of a genuine change of mind." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Play the most listened to song on Apple Music from the Velvet Underground album. Use Spotify instead of Apple Music.", "long_pause_offsets": [ 81 ], "source_sha256": "6cab7a1a30ebd198bab44211f6f8fede74eac1d28cb72f8c03d694c6d85639df", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the rest of the instruction including the artist, album, and the 'most listened to' criterion." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/revision/revision_02_app/7342cbdc0001d58b/speech.wav", "wer": 0.0, "wav_sha256": "8f94f11a6ede6cba17ac7bdcfd4e2bcd236a1495697b63c22ecbd768804aec0f", "seconds": 15.08, "spoken_utterances": [ "Play the most listened to song on uh, let's see... Apple Music from the um, let me think... Velvet Underground album. ", "Use Spotify instead of Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects to 'Spotify' after a pause, which is the revision type. The final intent is to play the most listened to song on Spotify from the Velvet Underground album, matching the original instruction. The correction is clear and final, so the task should be executed as revised.", "heard_transcript": "Play the most listened to song on uh let's see, Apple Music from the um let me think, Velvet Underground album. Use Spotify instead of Apple Music." }, "candidate_id": "wav_1253ea0886fcafbd", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has multiple issues. The speaker starts with 'Apple Music' and then corrects to 'Spotify', but the initial mention of 'Apple Music' is delivered with a flat, unnatural tone. The phrase 'let me think' is awkwardly placed and sounds like a forced filler. The correction is abrupt, and the overall delivery lacks the natural rhythm and prosody of a spontaneous speaker." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "ad948726c13fba957cba4ce55c4a955211111da4d8d69fffbac1efae22cd295e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/cancellation/cancel_008/6078ba593fc204bf/speech.wav", "wer": 0.04, "wav_sha256": "b7436a7700d3465877bec63e3021c0a574328e3bebdb1186766dd67db9c02928", "seconds": 12.96, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album.", "Hold on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with a request to play a song from the Velvet Underground album on Spotify, followed by a clear and final statement, 'Hold on, I want to cancel that task,' which constitutes a cancellation of the initial request. The final intent is to cancel, not to execute the original task. The cancellation is explicit and leaves no ambiguity about the speaker's intent.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album. Hold on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "37e668a22ffd2364c842e2e3acba268c2f8cd887c26eb294dc5a000f8599d190", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/ambiguous_halt/halt_034/e6e26c15057512bc/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "67f03cad47dab5bda6503867d6d9cdc1e8d899ad37c8ffc81415f9235c2db002", "seconds": 12.6, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the task but then explicitly says 'Wait, I was just thinking,' which is a clear request to pause and reconsider. This matches the 'ambiguous_halt' type, where the final interjection is a pause request without cancellation or resumption. The initial task is not executed, but the intent to pause is audible and final.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album. Wait, I was just thinking." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "c901732_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "5567648ff51b7508c6feb99614da9c97bf5083fc1e2e3b5316e1b5384fef771d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/c901732_3/audio_natural/filler/original/34d6354a63da80f4/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "83a79d25c2e3325edaf90f24b5b115308696e2f3899b5c1cfb3a7d8420e3df97", "seconds": 8.08, "spoken_utterances": [ "Play the uh, let's see... most listened to song on Spotify from the uh... Velvet Underground album." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('uh', 'let's see') and pauses but does not change the core request. The final intent is to play the most listened to song on Spotify from the Velvet Underground album, which matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Play the uh let's see most listened to song on Spotify from the uh Velvet Underground album." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_billing_period", "proposed_instruction": "Make me an amazon account with yearly—no, that should be monthly prime membership.", "long_pause_offsets": [], "source_sha256": "d00fb345af0d08306c1fb593705fe582cc69ad5b2c768ebeedccc543f5195595", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly' but immediately corrects it to 'monthly' within the same utterance. The final intended membership type is 'monthly', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/false_start/false_start_02_billing_period/ad85fe3c9c0dadfb/speech.wav", "wer": 0.23529411764705882, "wav_sha256": "d0f9e15b67e559c8fe041cb407b0da7dd643d2dd8bb980898e10c17cd841abce", "seconds": 8.48, "spoken_utterances": [ "Make me an amazon account with uh... yearly—um... no, that should be hmmm... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly' but immediately self-corrects to 'monthly' with 'no, that should be monthly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as requested.", "heard_transcript": "Make me an Amazon account with a yearly um, no, that should be monthly Prime membership." }, "candidate_id": "wav_35f9922ec3db2141", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with 'uh', then self-corrects 'yearly' to 'monthly' with a natural-sounding 'no, that should be' and a brief pause. The correction is integrated smoothly into the flow, and the final phrase is delivered with consistent prosody and a believable change-of-mind intonation. The filler and repair are well-placed and sound spontaneous." } ] }, "is_default": true }, { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Make me an amazon account with monthly basic membership—scratch that, make it prime membership.", "long_pause_offsets": [], "source_sha256": "07334e15bda025c98e552e65b733249b92f4ae8f9c47389f9cfba4fa8fabcd30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately corrects it to 'prime membership'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/false_start/false_start_03_subscription_type/140569467bfd9151/speech.wav", "wer": 0.0, "wav_sha256": "bf68285269c49ded1df401253792d105743251edd992471482048952f4beaaa4", "seconds": 9.2, "spoken_utterances": [ "Make me an amazon account with monthly basic membership—um... scratch that, make it ... prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly basic membership' but immediately self-corrects with 'scratch that, make it Prime membership.' According to the 'false_start' type criterion, the final corrected instruction must be executed, and the initial mistaken value is discarded. The final intent matches the original instruction's goal of creating an Amazon account with Prime membership, with only the term 'basic' replaced by 'Prime' as a direct correction.", "heard_transcript": "Make me an Amazon account with monthly basic membership. Um, scratch that, make it Prime membership." }, "candidate_id": "wav_665c540d22b1a618", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a complete request, then uses 'um' as a hesitation before a revision. The 'scratch that' is a pragmatic and natural way to cancel and restart, and the correction to 'Prime membership' is delivered with a clear shift in intonation. The flow is coherent, though the revision is slightly more abrupt than in audio_03, making it feel a bit less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Make me an ebay—I should say amazon account with monthly prime membership.", "long_pause_offsets": [], "source_sha256": "93f3af967fa398818b0410aa4eab81cf5dc17a62a3f2c45a20137297b426e20f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the account creation action and the monthly prime membership requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/false_start/false_start_01_app/bf85ac87300409e0/speech.wav", "wer": 0.0, "wav_sha256": "1def98bd2e06856c25e5d6b6377125a7232b0f7506ceb2f127ac46d4aefac284", "seconds": 8.68, "spoken_utterances": [ "Make me an ebay—um... I should say ah... amazon account with um... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'I should say uh Amazon'. This is a clear false start where the final intent is the corrected version. The rest of the instruction, including 'monthly Prime membership', remains unchanged and is preserved. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Make me an eBay um, I should say uh Amazon account with um monthly Prime membership." }, "candidate_id": "wav_f4af2c3539a40502", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with 'eBay' and immediately self-corrects to 'Amazon' with 'I should say'. While this is a plausible false start, the correction is delivered with a slightly unnatural, clipped intonation and the 'um' after 'eBay' sounds more like a filler than a genuine hesitation. The overall delivery feels less spontaneous and more rehearsed compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Make me an amazon account with prime membership. Oh, right, make it a monthly prime membership.", "long_pause_offsets": [ 49 ], "source_sha256": "3e5e827ab34110b171daa96c9d0c7b3b4bd8dc9029f3878660e36e03a67977c6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the specific term 'monthly' to clarify the Prime membership type, which was unspecified in the initial utterance. The core action and target remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/addition/addition_01/f65ad621b4f7db5c/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "6f714879ca6b5af3fac7a9c7c3c5b04c30fb9d2294773f8c308f4df3ff15d0a7", "seconds": 10.24, "spoken_utterances": [ "Make me an ah... amazon account with prime membership. ", "Oh, right, make it a ... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to create an Amazon account with Prime membership. The speaker then added 'monthly' as a specific detail to the Prime membership type. The final intent combines the original request with the added constraint, resulting in a complete and coherent task that matches the reference instruction. The addition is clear and does not contradict the original request.", "heard_transcript": "Make me an Amazon account with Prime membership. Oh right, make it a monthly Prime membership." }, "candidate_id": "wav_828bd5db7550a0c6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers the initial request fluently and with natural prosody. After a brief pause, they self-correct with 'Oh right' and revise the membership type to 'monthly prime membership'. The revision is pragmatic and plausible, with a believable change-of-mind intonation and clear reference back to the earlier request. The flow feels spontaneous and conversational." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Make me an amazon account with monthly. One more thing: include the prime membership.", "long_pause_offsets": [ 40 ], "source_sha256": "54ce8521dd2bb2b72911274f129b0f266c2e2dabba29f20dffaa3182a7fee514", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'Prime membership' requirement, combining it with the initial 'monthly' specification to form a coherent final goal matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/addition/addition_02/ea39467329e2b402/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "7d018cc84fd759b309b2b2353589cbff734f3c265ccac70dc566e3b71df023f1", "seconds": 11.64, "spoken_utterances": [ "Make me an uh... amazon account with uh, let's see... monthly. ", "One more thing: include the um... prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make me an uh Amazon account with uh let's see monthly' and then adds 'One more thing, include the um Prime membership.' This is a clear addition of the 'Prime membership' constraint to the initial request. The final intent combines both the account creation and the monthly Prime membership, matching the original instruction. The hesitation and fillers ('uh', 'let's see', 'um') do not change the final goal, and the addition is explicitly stated after a pause.", "heard_transcript": "Make me an uh Amazon account with uh let's see monthly. One more thing, include the um Prime membership." }, "candidate_id": "wav_fc9ef6689f60f4c8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a natural hesitation ('uh') and a pause ('let's see'), suggesting genuine thought. They state 'monthly' and then add 'one more thing' before requesting the Prime membership. However, the addition is awkwardly phrased and delivered with a flat, slightly robotic tone. The transition feels less organic than in audio_02, and the overall delivery lacks the natural conversational rhythm of a spontaneous request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Make me an amazon account with monthly basic membership. That should be prime membership, not basic membership.", "long_pause_offsets": [ 57 ], "source_sha256": "7c49a5e2ce3cfeaf942f3f830bf66542991f758ae00d8ebb72745defc2e0a2c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the membership type from 'basic' to 'prime', while retaining the account creation and monthly duration. The original instruction's intent is preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/revision/revision_03_subscription_type/824cdfe7b67b01de/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "28f8573f3f86f567fe46de0483416608726825b918ce91a8108f420a5fcb7a4a", "seconds": 11.24, "spoken_utterances": [ "Make me an uh... amazon account with monthly basic membership. ", "That should be um... prime membership, not basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but immediately corrects it to 'Prime membership' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to create an Amazon account with monthly Prime membership, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "Make me an ah Amazon account with monthly basic membership. That should be um Prime membership, not basic membership." }, "candidate_id": "wav_267c5f678500cd1b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear request, hesitates with 'uh' and 'um', then self-corrects the membership type from 'basic' to 'Prime' in a natural, spontaneous manner. The correction is integrated smoothly into the same utterance, with a believable shift in intonation and pacing. The flow feels like a genuine moment of recollection, and the prosody remains consistent throughout." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "Make me an ebay account with monthly prime membership. Wait, make that an eBay account, not Amazon.", "long_pause_offsets": [ 55 ], "source_sha256": "748df6ae3b67e942610103786c7d7eeb783dd071c7492b10cbdddb7bd937bd4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup corrects the app from Amazon to eBay, but the original instruction specified Amazon. The final goal is for eBay, which contradicts the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/revision/revision_01_app/5ba3fd3aa0f57029/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "36a9bc17c17f8d17676c186aec0388500e45ae2237ee228425c9bc49f96880b0", "seconds": 13.2, "spoken_utterances": [ "Make me an ebay account with uh, let's see... monthly prime membership. ", "Wait, make that an um... eBay account, not uh... Amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon account' but immediately corrects it to 'eBay account' after a pause and self-correction. The final intent is to create an eBay account with monthly Prime membership, which matches the original instruction's structure but with the correct target platform. The revision is clear and final, so the task should be executed as revised.", "heard_transcript": "Make me an eBay account with uh let's see monthly Prime membership. Wait, make that an um eBay account, not a Amazon." }, "candidate_id": "wav_9d33b15afe206658", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, hesitates with 'let's see', then corrects 'Amazon' to 'eBay' mid-sentence. The self-repair is plausible and integrated into the flow, though the correction is slightly more abrupt than in audio_03. The prosody and pacing are natural, with a believable moment of realization, but the correction feels a bit more like a simple swap than a deep recollection." } ] }, "is_default": false }, { "variant_id": "revision_02_billing_period", "proposed_instruction": "Make me an amazon account with yearly prime membership. Make that a yearly billing period instead of monthly.", "long_pause_offsets": [ 56 ], "source_sha256": "263c12f5de1cb38d9073f60a68be59ad80a4998c9ae1e19147c10b44b5cd5895", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the billing period should be yearly, which is consistent with the initial request for 'yearly prime membership'. The original instruction requested 'monthly prime membership', but the final spoken instruction specifies 'yearly', and the followup confirms this change." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/revision/revision_02_billing_period/49a5663b059cdfd8/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "fce8f67471573827d7b9591fdc46745dd59f658deb7c7b9efca273a0372ef5e9", "seconds": 12.28, "spoken_utterances": [ "Make me an amazon account with hmmm... yearly prime membership. ", "Make that a let's see... yearly billing period instead of ... monthly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests a monthly Prime membership, then revises the billing period to yearly after a pause. The revision is clear and final, and the rest of the instruction (creating an Amazon account) remains unchanged. The final intent matches the expected revision outcome.", "heard_transcript": "Make me an Amazon account with monthly Prime membership. Make that a let's see, yearly billing period instead of monthly." }, "candidate_id": "wav_3de7edd8f6a25545", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, hesitates with 'hmm', then attempts a revision. However, the revision is not a self-correction of the initial request but a new, separate instruction ('make that a let's see yearly billing period instead of monthly'). This feels less like a spontaneous correction and more like a disjointed follow-up, breaking the natural flow of a single thought process. The prosody is consistent, but the pragmatic structure is less realistic as a single, coherent utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_004", "hide_variant_original": true, "source_sha256": "020a4cdb13ee4bf3acef433259c7416a54e28cb92d8c1ea1b58035687b7403d1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/cancellation/cancel_004/95880e3a9c39a70a/speech.wav", "wer": 0.2, "wav_sha256": "b63cb3b8ffc49526f3eefd4e47cc62007ae35ac22c88e1b9e6b677f9bf123e0d", "seconds": 9.04, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership.", "Wait, cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues a request to create an Amazon account with monthly Prime membership, then immediately follows with 'Wait, cancel that request.' This constitutes a clear and final cancellation of the initial task. The cancellation is unambiguous and applies to the entire task as requested.", "heard_transcript": "Make me an Amazon account with monthly Prime membership. Wait, cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "aac045e55352edc799321403fafe675eb1ea8b086ecec4a4f59dd0a5c7f04286", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/ambiguous_halt/halt_018/2b686123d145d9d2/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "1a6b2b9c66f9857b144a5e74e5d3394a64effcd031094e2201fd852084ef24ba", "seconds": 8.76, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, then ends with a clear request to pause. The final interjection 'Could you pause for a second?' constitutes an unambiguous halt command. The type is 'ambiguous_halt', which requires only a pause request, not execution or cancellation. The final intent is to pause, which matches the type's expected control intent.", "heard_transcript": "Make me an Amazon account with monthly Prime membership. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "96bebf8f156c171bda68a052f21ec6172c028d62ec563e87fcaeab32f68b57ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_1/audio_natural/filler/original/aaf15e943f747e15/speech.wav", "wer": 0.2727272727272727, "wav_sha256": "dd6bcf6bb265a8b7af5b47b610c9cf1516a01e780293ee65ade8ebbc0ea83651", "seconds": 5.12, "spoken_utterances": [ "Make me an ah... amazon account with hmmm... monthly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Amazon account with monthly Prime membership' with a slight hesitation and correction ('monthly Prime membership' instead of 'monthly Prime membership'), but the final, corrected utterance matches the original instruction. The hesitation and self-correction are typical filler behavior and do not alter the final goal. The task is executed as intended.", "heard_transcript": "Make me an Amazon account with monthly Prime membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Make me a Apple Music—I should say spotify account with yearly premium membership.", "long_pause_offsets": [], "source_sha256": "e927da6c0de48bd305559e3865b019484c8c34c44bb20a6c9716c69d55f2aff6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the action, app, and membership type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/false_start/false_start_01_app/1b9147c71c8f9c87/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "57271b68891d6b6dd7676be931ee69b5ef21e67887c3e9456cd47bff90a136a0", "seconds": 7.76, "spoken_utterances": [ "Make me a Apple Music—uh... I should say spotify account with ... yearly premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The rest of the instruction, including 'yearly premium membership', remains unchanged and is preserved. Therefore, the final audible intent matches the original instruction, and the task should be executed.", "heard_transcript": "Make me an Apple Music uh, I should say Spotify account with yearly premium membership." }, "candidate_id": "wav_508097a933604589", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a natural-sounding self-correction from 'Apple Music' to 'Spotify account' with a believable 'I should say' phrase. The prosody is consistent, the correction is smoothly integrated into the sentence, and the overall delivery sounds like a spontaneous, real-time formulation of a request without robotic pauses or unnatural phrasing." } ] }, "is_default": true }, { "variant_id": "false_start_02_billing_period", "proposed_instruction": "Make me a spotify account with monthly—no, that should be yearly premium membership.", "long_pause_offsets": [], "source_sha256": "1a68c315e304f67576c2d1d56d67df150e4f3b6c462ad214562dd85fcb46bab5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly' but immediately corrects it to 'yearly'. The final intended membership type is 'yearly', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/false_start/false_start_02_billing_period/a54b9d50b38bd9b3/speech.wav", "wer": 0.0625, "wav_sha256": "fc2556a464700a3df1f466811d1c1bb0cb5b6b1b4653a647fc80acb37bdfea91", "seconds": 9.36, "spoken_utterances": [ "Make me a uh... spotify account with ... monthly—um... no, that should be yearly premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly' but immediately self-corrects to 'yearly' with 'no, that should be yearly premium membership.' This is a clear false start where the final intent is the corrected version. The original instruction matches the final audible intent, so the task should be executed as requested.", "heard_transcript": "Make me a uh Spotify account with monthly um no, that should be yearly premium membership." }, "candidate_id": "wav_e51ae641387c1d3c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision where the speaker initially says 'monthly' and then corrects to 'yearly premium membership'. The correction is clear and the intonation reflects a change of mind. However, the initial phrase 'monthly um' has a slightly longer and less natural filler pause compared to audio_03, and the correction is a bit more abrupt, making it slightly less fluid than the self-correction in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Make me a spotify account with yearly free trial—scratch that, make it premium membership.", "long_pause_offsets": [], "source_sha256": "0686a46a0ec08abaebfbd5f51b529d011d7ffade4ef6b8f9c025ad76e5ed06b7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests 'yearly free trial' but immediately corrects it to 'premium membership'. The final intended goal is to create a Spotify account with premium membership, which matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/false_start/false_start_03_subscription_type/af2c0eced9cfde08/speech.wav", "wer": 0.15, "wav_sha256": "4c13c7d35a3633165e99f2f5b1bb694681de91a3584b465e962b2e32056b7e69", "seconds": 12.64, "spoken_utterances": [ "Make me a uh... spotify account with yearly um, let's see... free trial—uh... scratch that, make it premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'free trial' but immediately self-corrects with 'scratch that' and specifies 'premium membership'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction, so the task is executed as intended.", "heard_transcript": "Make me a um Spotify account with yearly um let's see free trial uh scratch that make it premium membership." }, "candidate_id": "wav_93f73e871a546d64", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'free trial' followed by a 'scratch that' correction. While the self-correction is present, the phrase 'scratch that' is a more abrupt and less conversational way to retract a statement compared to 'I should say' or 'no, that should be'. Additionally, the initial hesitation and filler words ('um') are more drawn out and less integrated into the flow, making the delivery sound slightly less spontaneous and more like a rehearsed or awkward attempt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Make me a spotify account with premium membership. Oh, I forgot to mention, make it a yearly premium membership.", "long_pause_offsets": [ 51 ], "source_sha256": "72813c494d281bfaf461e0dff6fa32363549abe50f18801b7a3164d1a9708ba0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'yearly' duration to the premium membership, which was omitted in the initial utterance. The final goal combines the initial request with the added constraint, resulting in the same intent as the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Make me a ... spotify account with premium membership. ", "Oh, I forgot to mention, make it a um... yearly premium membership." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Make me a spotify account with yearly. I left out a detail: it should be a premium membership.", "long_pause_offsets": [ 39 ], "source_sha256": "a28be02649aa5383abd6c715abbd85e696aaa0e745d1090ce1ebf2d96cef0ef2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'premium' detail to the initial 'yearly' request, combining both to form the complete intended goal." }, "audio": { "status": "pending", "spoken_utterances": [ "Make me a um... spotify account with uh, let's see... yearly. ", "I left out a detail: it should be a um... premium membership." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Make me a apple music account with yearly premium membership. Let me correct the app: use Spotify, not Apple Music.", "long_pause_offsets": [ 62 ], "source_sha256": "c10bff37a15e6504d9534d5e30b43bdf933f9d089c35b7127777da90b0ea47bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining the account creation and yearly premium membership requirements." }, "audio": { "status": "pending", "spoken_utterances": [ "Make me a let's see... apple music account with uh... yearly premium membership. ", "Let me correct the app: use um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_billing_period", "proposed_instruction": "Make me a spotify account with monthly premium membership. Make that yearly premium membership, not monthly.", "long_pause_offsets": [ 59 ], "source_sha256": "7a96455e4c2ec71efc5d4692869293f093c420f68a1dff3039cf7b78b5936fa4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the membership duration from monthly to yearly, aligning with the original instruction. The app and account creation action remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Make me a uh... spotify account with um... monthly premium membership. ", "Make that ah... yearly premium membership, not monthly." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Make me a spotify account with yearly free tier. Change the subscription type to premium membership instead of free tier.", "long_pause_offsets": [ 49 ], "source_sha256": "3b9a63d2d12658438f0c83c13eaed8098776d2e5610695f0458cac50187d1564", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the subscription type from 'free tier' to 'premium membership', preserving the app, account creation action, and yearly duration." }, "audio": { "status": "pending", "spoken_utterances": [ "Make me a let me think... spotify account with uh... yearly free tier. ", "Change the subscription type to um... premium membership instead of free tier." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "cbed9b68b427bdc58fa8bf8341130d7c8b709585fd45524fe09656ce6fc6fefa", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/cancellation/cancel_001/573e889c748e845f/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "ac801395574937df5e67d5a3c2be152ab7461c618b5049aaac5d908cf7311a5d", "seconds": 10.12, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership.", "Oh, wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'cancel that' after initially stating the task. This constitutes a direct cancellation of the entire task as requested in the reference. The final control intent is a cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Make me a Spotify account with a yearly premium membership. Oh wait, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_028", "hide_variant_original": true, "source_sha256": "df5e94ad0a78348c896c46b1e050917a00b5827c0da1fc0b33b31db40214b633", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/ambiguous_halt/halt_028/67fe27c2daf86acf/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "dd1433c3e65d2329e74a3fbe4f731b28e4fdef8fa8e91c1a9bd3de23f2e1cdcf", "seconds": 10.76, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership.", "Wait, let me get my thoughts together." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me get my thoughts together,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial request is not revoked, and the speaker does not resume or correct it. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "Make me a Spotify account with a yearly premium membership. Wait, let me get my thoughts together." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "84d613d692b1b5c873dbd157e56a7b83ae44ad93d11f0f926b1fec23b804e53e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_2/audio_natural/filler/original/a51efbb454e2fd27/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "aafbf613c9dae99605b6160fef44177536eb1f93d8b5772de4f6885246b01b67", "seconds": 6.32, "spoken_utterances": [ "Make me a uh... Spotify account with ah... yearly premium membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Make me a', which is a hesitant start, but immediately proceeds to state the full, correct instruction without any correction or change in intent. The hesitation and filler word 'a' do not alter the final goal, which matches the original instruction. Therefore, the task is executed as intended.", "heard_transcript": "Make me a Spotify account with a yearly premium membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_subscription_type", "proposed_instruction": "Make me an amazon account with yearly basic membership—scratch that, make it prime membership.", "long_pause_offsets": [], "source_sha256": "bacce68ffa0608dbeb4d7b8a7816729d9f8a40d153f4c714744bd73aedd993c5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly basic membership' but immediately corrects it to 'prime membership'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/false_start/false_start_03_subscription_type/9bd0fd7de1a32505/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d56e68d239d17ed951a20286eb310404ae7dad771b8b5a96260153440d8d8a7e", "seconds": 10.8, "spoken_utterances": [ "Make me an amazon account with let me think... yearly basic membership—um... scratch that, make it prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'yearly basic membership' but immediately self-corrects with 'scratch that' and specifies 'Prime membership'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent is clear and executable.", "heard_transcript": "Make me an Amazon account with, let me think, yearly basic membership. Um, scratch that. Make it Prime membership." }, "candidate_id": "wav_cf968bd4a1850811", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and layered self-correction. It begins with a clear false start ('Make me an Amazon account with'), followed by a believable 'let me think' filler, then a specific but incorrect detail ('yearly basic membership'). The speaker then self-corrects with 'scratch that' and a smooth, natural revision to 'make it Prime membership'. The prosody includes appropriate hesitation, a change in intonation during the correction, and a plausible final phrasing. The flow mimics a real person reconsidering a detail mid-sentence." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Make me an ebay—I should say amazon account with yearly prime membership.", "long_pause_offsets": [], "source_sha256": "0ec3b9cf11571286cdb65b3cf57c891ebbcd2b695ac35d6ddec97493b72d2634", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the action, account type, and membership condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/false_start/false_start_01_app/fbe72434dabda9d4/speech.wav", "wer": 0.1875, "wav_sha256": "04f8bf4ecb8e2b3adfd956b6bf9f664a589eca85c03251920c9a880a6db2de35", "seconds": 10.0, "spoken_utterances": [ "Make me an uh... ebay—um... I should say amazon account with uh... yearly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Make me an uh, eBay um, I should say Amazon account with uh, yearly Prime membership." }, "candidate_id": "wav_817b7b1585f5a138", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a false start ('Make me an uh eBay um...') and a self-correction to 'Amazon account'. The correction is plausible, but the filler words ('uh', 'um') are less natural and more stilted than in audio_03. The second part of the request ('with uh yearly Prime membership') is delivered with a flat, hesitant tone and lacks the smooth revision flow seen in audio_03. The overall delivery feels slightly more robotic and less like spontaneous speech." } ] }, "is_default": false }, { "variant_id": "false_start_02_billing_period", "proposed_instruction": "Make me an amazon account with monthly—no, that should be yearly prime membership.", "long_pause_offsets": [], "source_sha256": "43ae887607c27c60fed7495faa1d5f8b28624175b2a84f4ac83d8a861dbcfc40", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly' but immediately corrects to 'yearly'. The final intended membership type is 'yearly Prime', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/false_start/false_start_02_billing_period/771b0f541ee8fcff/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "14cd3efca53e17cf779e1e3911c840e04858b1b27813bcd29601206c56c51f9e", "seconds": 12.12, "spoken_utterances": [ "Make me an amazon account with ... monthly—hmmm... no, that should be ... yearly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly' but immediately self-corrects to 'yearly' with 'no, that should be yearly'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Make me an Amazon account with monthly, hmm, no, that should be yearly Prime membership." }, "candidate_id": "wav_ab62fdc65f129431", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start ('Make me an Amazon account with monthly...') and a self-correction ('no, that should be yearly Prime membership'). While the correction is clear, the transition is abrupt and lacks the natural hesitation or filler words that would make it sound more spontaneous. The speaker's tone is flat and the pacing is too even, making the correction feel more like a read-aloud correction than a genuine moment of reconsideration. The prosody is less varied compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Make me an amazon account with yearly. I left out a detail: include prime membership.", "long_pause_offsets": [ 39 ], "source_sha256": "5b746600bf35c515d665a5a2ebb4af88794a5899fd6e4e358b00dd2b18356855", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'prime membership' requirement, completing the initial request for a yearly Amazon account. The final goal combines both the account creation and the membership type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/addition/addition_02/4390f97d80f02592/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "9e9d4da7b4daacef624eb0850a2fffec0bb7fdd818b802dcadf6d1ae5dd95861", "seconds": 12.52, "spoken_utterances": [ "Make me an um, let me think... amazon account with ... yearly. ", "I left out a detail: include prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make me an um, let me think, Amazon account with yearly' and then adds 'include Prime membership' after a pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added detail. The final audible intent is to create an Amazon account with yearly Prime membership, which matches the original instruction. The hesitation and self-correction do not change the final goal.", "heard_transcript": "Make me an um, let me think, Amazon account with yearly. I left out a detail, include Prime membership." }, "candidate_id": "wav_94bad1a94d15d74c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a plausible self-pause ('let me think'), followed by a clear initial request. The speaker then remembers a missing detail and adds it smoothly with a self-corrective tone ('I left out a detail'), which is a realistic and common pattern in spontaneous speech. The prosody is consistent, and the addition flows naturally without awkwardness or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Make me an amazon account with prime membership. Oh, and one more thing: make it a yearly prime membership.", "long_pause_offsets": [ 49 ], "source_sha256": "17d85af24c48fa87c2694e888145703d56446f0081dbd3b98cc78d3c40d0a832", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the 'yearly' duration to the initial 'Prime membership' request, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/addition/addition_01/94e24c493f46748e/speech.wav", "wer": 0.1, "wav_sha256": "76e1a230a5daaa773b396c83e34e697b78366d3a8d68ebb4a57123a639e7779a", "seconds": 12.36, "spoken_utterances": [ "Make me an um... amazon account with prime membership. ", "Oh, and one more thing: make it a ... yearly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction was to make an Amazon account with Prime membership. The speaker then added 'one more thing' and specified 'yearly Prime membership'. This is a clear addition that narrows the scope of the original request. The final intent is to create an Amazon account with a yearly Prime membership, which matches the original instruction with the added specificity. Therefore, the addition is correctly applied and the task is complete.", "heard_transcript": "Make me an Amazon account with Prime membership. Oh, and one more thing, make it a yearly Prime membership." }, "candidate_id": "wav_de067ba30e72a676", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a detail ('Oh, and one more thing'), the delivery feels slightly less spontaneous. The initial request is fluent, but the addition is delivered with a more abrupt transition and a slightly rehearsed cadence. The prosody is less varied, and the 'one more thing' transition lacks the natural hesitation or cognitive processing seen in audio_01, making it sound more like a rehearsed or read instruction than a genuine moment of recall." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_billing_period", "proposed_instruction": "Make me an amazon account with monthly prime membership. Make that yearly prime membership, not monthly.", "long_pause_offsets": [ 57 ], "source_sha256": "b436cecd0a43c131af82a925a6b912fe675b078e08396ce5b034e96e1a8a6a69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the membership type from monthly to yearly, preserving the account creation and Amazon app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/revision/revision_02_billing_period/76b14fb71ddfda69/speech.wav", "wer": 0.15, "wav_sha256": "4f17b1f0fa90c4ccfbc961ea13608a7a004e73bbeb2158601af743a6ddd6b59b", "seconds": 13.36, "spoken_utterances": [ "Make me an let's see... amazon account with um... monthly prime membership. ", "Make that hmmm... yearly prime membership, not monthly." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'monthly Prime membership' but immediately corrects it to 'yearly Prime membership, not monthly'. This is a clear revision where the final intent is to request a yearly membership. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Make me an let's see Amazon account with monthly Prime membership. Make that yearly Prime membership, not monthly." }, "candidate_id": "wav_c3baae213be67cd0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The initial hesitation ('Let's see') is followed by a clear, self-corrected revision of the membership duration. The correction ('Make that... yearly prime membership not monthly') is smoothly integrated into the utterance with a natural pitch drop and emphasis on 'yearly', showing a believable internal thought process. The prosody is consistent and the pacing feels like a person formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_03_subscription_type", "proposed_instruction": "Make me an amazon account with yearly basic membership. Use prime membership instead of basic membership.", "long_pause_offsets": [ 56 ], "source_sha256": "113ee2b9811c71f5adaa0d459b85c56291b36754456008d9cdb7476c8b7918b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly replaces 'basic membership' with 'prime membership', aligning the final goal with the original instruction. The account creation and yearly duration remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/revision/revision_03_subscription_type/9a977298f97f98fc/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "327b922f7375ccc304a9d0b60eaf37b3ac218ac7464f25d46304cddf2226ab19", "seconds": 12.44, "spoken_utterances": [ "Make me an ah... amazon account with yearly basic membership. ", "Use um, let me think... prime membership instead of basic membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'basic membership' but then corrects it to 'Prime membership' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to create an Amazon account with yearly Prime membership, which matches the original instruction's intent and is a direct revision of the membership type.", "heard_transcript": "Make me an Amazon account with yearly basic membership. Use um, let me think, Prime membership instead of basic membership." }, "candidate_id": "wav_ea9bca7ef063c79e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker shows a clear revision with a self-correction ('use... Prime membership instead of basic membership'). The 'let me think' filler is placed naturally before the correction. However, the initial request is slightly less fluent, with a minor stumble ('Make me an ah...'). The correction is delivered with a slight pitch rise, which is plausible, but the overall delivery feels a bit more hesitant than audio_03, making it slightly less realistic." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Make me an ebay account with yearly prime membership. Change amazon to ebay.", "long_pause_offsets": [ 54 ], "source_sha256": "2cd1024904915222a4baf65fd5e7716594b3a6b1bc4b5c22e1b7186ddf896506", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly changes the app from Amazon to eBay, which contradicts the original instruction to create an Amazon account." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/revision/revision_01_app/c562458ee78ac18e/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "abbb9eeee266d707a2e663cace1696b33d16bddd5e48e805a3d5754b8ed86c2a", "seconds": 12.6, "spoken_utterances": [ "Make me an ah... ebay account with yearly prime membership. ", "Change ... amazon to ... ebay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Make me an uh, Amazon account with yearly prime membership' but then immediately corrects 'Change Amazon to eBay.' This is a revision where the final intent is to create an eBay account, not an Amazon one. The correction is clear and applied to the original instruction, so the task is executed as revised.", "heard_transcript": "Make me an uh, eBay account with yearly prime membership. Change Amazon to eBay." }, "candidate_id": "wav_516d4fc83932bfa5", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker's delivery is the least realistic. The initial request is cut off mid-word ('Make me an uh...'), and the revision ('change Amazon to eBay') is stated as a separate, abrupt command rather than being integrated into the original request. The prosody is flat and the transition between the two parts is jarring, sounding more like a post-hoc instruction than a spontaneous correction. The lack of a natural flow or self-correction within a single utterance reduces its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_015", "hide_variant_original": true, "source_sha256": "7573baf7f47fa312dd5812b0184aec2464c9e24be507c48f168fb9ed48d4a8b2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/cancellation/cancel_015/f6897404283869ff/speech.wav", "wer": 0.0, "wav_sha256": "fb98ce16271d8dd88d5e1b24ae9480896ceb08d38f6afbf12effd2e4741f532c", "seconds": 11.28, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership.", "Wait, actually, cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to create an Amazon account with a yearly Prime membership, but immediately corrects themselves by saying 'Wait, actually, cancel the whole thing.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction.", "heard_transcript": "Make me an Amazon account with yearly Prime membership. Wait, actually, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_038", "hide_variant_original": true, "source_sha256": "c653b23a42c9965931bda70a4327a827c61deba6dd0bcc4a01f5a495d9bc83b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/ambiguous_halt/halt_038/34516a7966b65997/speech.wav", "wer": 0.0, "wav_sha256": "38bc1932aacc54b4045784475b950c8cd6d826d8c5c066a854d24948164fc59a", "seconds": 10.88, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership.", "Could you pause there for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Could you pause there for a second?', which is a clear request to hold or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request. The original instruction is not canceled or revised, but the speaker explicitly asks to pause, so the control intent is to pause.", "heard_transcript": "Make me an Amazon account with yearly Prime membership. Could you pause there for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cb4ee7a_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7a5653ebcf66151db4928aa746241504404acaccbd1155d524c4f46f21dd5705", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cb4ee7a_3/audio_natural/filler/original/e778bb53e5a56c41/speech.wav", "wer": 0.0, "wav_sha256": "7dab79974a0a05970f9eb1d18943f17b762eab303d174868d6cfda35d5bf79dc", "seconds": 6.4, "spoken_utterances": [ "Make me an ... amazon account with ... yearly prime membership." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'Make me an', which is a hesitant start, but immediately proceeds to state the full, correct instruction without any correction or change in intent. The hesitation is a filler and does not alter the final goal. The final audible intent matches the original instruction exactly.", "heard_transcript": "Make me an Amazon account with yearly Prime membership." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_artist_selection", "proposed_instruction": "Who is the least followed—no, that should be most followed artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "cb81def7a8436650d25427e9885ea1bc373a7bfe3235949cf118fbd3462d3a8f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least followed' but immediately corrects it to 'most followed'. The rest of the query, including the app (Spotify) and the scope ('artist I follow'), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/false_start/false_start_01_artist_selection/0b1e6190ea64bcf4/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "daf95b134fadf0c06db3b7d83b7e99804b7b988a0778aee04345ad9abe87491a", "seconds": 9.76, "spoken_utterances": [ "Who is the um... least followed—uh... no, that should be most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least followed' but immediately self-corrects to 'most followed' with 'no, that should be most followed'. This is a clear false start where the final intended value is 'most followed'. The rest of the instruction matches the original, so the final intent is executable and correct.", "heard_transcript": "Who is the um least followed uh no, that should be most followed artist I follow on Spotify?" }, "candidate_id": "wav_be3e8a97b1a47f46", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible self-correction. The initial 'least followed' is clearly a slip, and the immediate, smooth correction to 'most followed' with a self-aware 'no, that should be' is highly realistic. The filler 'um' and 'uh' are well-placed, and the overall prosody reflects genuine hesitation and recollection rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_artist_reference", "proposed_instruction": "Who is the most followed artist I don't follow—I should say artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "61e3b494ce8ce6d0e636cc4fd1594c947d5c240f17c8fe28d5d0e41fcca8cc8e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately corrects it to 'artist I follow' within the same utterance. The final intended target is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/false_start/false_start_02_artist_reference/0acfe861afce61f5/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "f9857d4a626a262b81828ebf3554e875c5c6509a6259f70147ba5b83faddeb45", "seconds": 9.56, "spoken_utterances": [ "Who is the most followed let me think... artist I don't follow—uh... I should say artist I follow on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately self-corrects to 'artist I follow on Spotify'. This is a false start where the final utterance clearly corrects the initial mistake. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Who is the most followed, let me think, artist I don't follow, uh, I should say artist I follow on Spotify?" }, "candidate_id": "wav_e2ac499fcc64a51f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker shows a clear false start with 'most followed', then corrects to 'artist I follow on Spotify'. The 'let me think' and 'uh' fillers are natural, and the correction is immediate and fluent. However, the correction is less self-aware than audio_03's; it's more of a simple rephrasing than a recognition of a mistake, making it slightly less realistic in terms of spontaneous self-correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music—scratch that, make it Spotify?", "long_pause_offsets": [], "source_sha256": "6f8a9ab87d7d63f7f365aec26091db4f1ab5ec053a731e0a1d223e879385e7ae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The final app is Spotify, matching the original instruction. The action and scope (most followed artist among those I follow) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/false_start/false_start_03_app/dcef542b69cc9661/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "c0c1860202b6e344c93ccbd651a71725a39e00f16a8e36f7d79ddcfce7a28091", "seconds": 8.52, "spoken_utterances": [ "Who is the most followed artist I follow on Apple Music—uh... scratch that, make it hmmm... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects with 'scratch that, make it Spotify'. This is a clear false start where the final audible intent is to execute the original instruction but with the platform changed from Apple Music to Spotify. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Who is the most followed artist I follow on Apple Music uh, scratch that, make it Spotify?" }, "candidate_id": "wav_2cfb3f906c8caa25", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with a clear request but then says 'scratch that' and changes the platform. While this is a valid revision, the delivery feels slightly abrupt and less natural. The 'uh' after 'Apple Music' is short and doesn't convey the same depth of hesitation as the others. The correction is pragmatic but lacks the organic flow of the other two, making it the least realistic in terms of natural speech patterns." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Who is the most followed on Spotify? Oh, I forgot to mention, I mean the artist I follow.", "long_pause_offsets": [ 37 ], "source_sha256": "10ac0732ac67790538f37ad4f875f2436374b8fa747ffed08d840ac6a5d5e8b0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'artist I follow' to the initial broad question, narrowing the scope to only artists the user follows. The core action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/addition/addition_01/c2404edefcdcdd68/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "ac79264c01a13b13683f9eaaa1ff0780384c6316b6ae5155245b98061818574e", "seconds": 12.4, "spoken_utterances": [ "Who is the uh... most followed on Spotify? ", "Oh, I forgot to mention, I mean the let me think... artist I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request 'Who is the most followed on Spotify?' is incomplete. The speaker then adds the missing constraint 'artist I follow' after a pause and self-correction, completing the original instruction. The addition clarifies the scope of 'most followed' to be limited to the user's own followed artists, which is the intended final goal. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the uh most followed on Spotify? Oh, I forgot to mention, I mean the, let me think, artist I follow." }, "candidate_id": "wav_033a926ee528f501", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music? Let me correct the app: the question is about Spotify, not Apple Music.", "long_pause_offsets": [ 57 ], "source_sha256": "df06d67adaa44ef8fbf143fdc61b3afb0905dcbdb8d59a8ab1f7196627fb5e1f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (most followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/revision/revision_03_app/43c1566a9ae427c4/speech.wav", "wer": 0.0, "wav_sha256": "ee33e0cf8cdf24c2cc9bd2c32186f44ce5c714a86471c08d607d2fba83cd141c", "seconds": 12.24, "spoken_utterances": [ "Who is the most followed artist I follow on uh... Apple Music? ", "Let me correct the app: the question is about let's see... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects the app to Spotify. The revision is clear and unambiguous, and the final intent matches the original instruction. The type is 'revision', and the correction is applied as required.", "heard_transcript": "Who is the most followed artist I follow on uh Apple Music? Let me correct the app. The question is about, let's see, Spotify, not Apple Music." }, "candidate_id": "wav_0c136f1c702e44c7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects the platform from 'Apple Music' to 'Spotify' with a believable hesitation ('let me correct the app') and a brief, plausible thought process ('the question is about... let's see'). The correction is smoothly integrated into the utterance, and the prosody reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "revision_02_artist_reference", "proposed_instruction": "Who is the most followed artist I listen to on Spotify? Wait, the query is about the artist I follow, not the artist I listen to.", "long_pause_offsets": [ 56 ], "source_sha256": "e1ffb1b103b70fb60faad408134394e84114de5244268af02f67f260308b0e38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the action from 'listen to' to 'follow', which matches the original instruction. The rest of the query, including the app and the intent to find the most followed artist, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/revision/revision_02_artist_reference/af09892e1073a7eb/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "030301474ad326dd744cf35a0e176c9d4f15710eaa1079c3f6f344c238041bc5", "seconds": 12.08, "spoken_utterances": [ "Who is the hmmm... most followed artist I listen to on um... Spotify? ", "Wait, the query is about the artist I follow, not the artist I listen to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'most followed artist I listen to' but immediately corrects themselves to 'the artist I follow'. This is a revision where the speaker self-corrects the target of the query. The final intent is to find the most followed artist the user follows, which matches the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Who is the um most followed artist I listen to on um Spotify? Wait, the query is about the artist I follow, not the artist I listen to." }, "candidate_id": "wav_7b7071040ada8aa3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant phrasing ('who is the...') and then corrects 'listen to' to 'follow' after a brief pause and self-aware realization. The correction is natural, and the speaker's intonation reflects a genuine change of mind. However, the initial phrasing is less fluent than audio_03, and the correction is slightly more abrupt, making it less smooth overall." } ] }, "is_default": false }, { "variant_id": "revision_01_artist_selection", "proposed_instruction": "Who is the most popular artist I follow on Spotify? Make that the most followed artist, not the most popular.", "long_pause_offsets": [ 52 ], "source_sha256": "0d696a20e7c84bc71c9a0e1906712768e3c2db6011ffd1e239c5a3c7a49abce5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'most popular' to 'most followed', which is the exact term from the original instruction. The app and scope of 'artists I follow' are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/revision/revision_01_artist_selection/7e908d129b87ec53/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "3a4e8a2a787ea31d8feca52bc54401b5ea0140c58264a8fa27df5dc87aae22b1", "seconds": 11.64, "spoken_utterances": [ "Who is the ah... most popular artist I follow on Spotify? ", "Make that the ... most followed artist, not the most popular." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'most popular artist' but immediately corrects themselves to 'most followed artist' after a pause. This is a revision where the final intent is to find the most followed artist, which matches the original instruction. The correction is clear and unambiguous, and the final control intent is to execute the revised task.", "heard_transcript": "Who is the most popular artist I follow on Spotify? Make that the most followed artist, not the most popular." }, "candidate_id": "wav_c12ec6a9e74c0b39", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but then self-corrects 'popular' to 'followed' with a noticeable pause and a slightly unnatural, rehearsed-sounding delivery. The correction feels more like a mechanical edit than a spontaneous thought process, and the prosody lacks the natural flow of the other two. The overall delivery is less convincing as a real-time, unscripted utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "d135287deeccdeeef8e93ca3b1ee887ee3de8a61bdfff4357bb0916b7a62fa6e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/cancellation/cancel_012/642e81ded75b6248/speech.wav", "wer": 0.17391304347826086, "wav_sha256": "7684ef3ce5b1f08e904cbb38b74c2ca15852216d70c9527df68f0817eea5911d", "seconds": 11.04, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially starts the task but then clearly revokes it with 'Wait, actually, I've changed my mind. Don't do that.' This constitutes a direct cancellation of the original request. The final intent is to cancel, not to execute or pause.", "heard_transcript": "Who is the um, let me think, most followed artist I follow on Spotify? Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_003", "hide_variant_original": true, "source_sha256": "b2533de5616b97bfec6b4969efa50bd1d63824492179d874768b30efe9206ae8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/ambiguous_halt/halt_003/0996e10c5e039c9f/speech.wav", "wer": 0.25, "wav_sha256": "20472b8d6e6ccac2dedc75de0f23f119bde8738806aad09f50fd4c294d4340f6", "seconds": 8.64, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?", "Hang on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on,' which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed. This matches the 'ambiguous_halt' type, where the final interjection is a pause command.", "heard_transcript": "Who is the um, let me think. Most followed artist I follow on Spotify. Hang on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "8f37a2fbcdc4be4a5615c56bde3e2787b960037deab9aa7166402a83d15e0e9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_1/audio_natural/filler/original/8289fa2671261680/speech.wav", "wer": 0.2857142857142857, "wav_sha256": "cf11951443ac5249810f5bd97be7802c3d0bf757e0ac8935e855db606109b6d9", "seconds": 5.8, "spoken_utterances": [ "Who is the um, let me think... most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler phrases 'um' and 'let me think' but does not alter the core request. The final intent remains unchanged and matches the original instruction. The type is 'filler', which permits such hesitations and backchannels as long as the task goal is preserved.", "heard_transcript": "Who is the um, let me think, most followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_artist_selection", "proposed_instruction": "Who is the most followed—no, I mean least followed artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "cf616eee8c6a973eaeddc3abcd01f76721d0cc3e8678807fca65e63f12a9f380", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most followed' but immediately corrects to 'least followed' within the same utterance. The final intended comparison is 'least followed', and all other elements (scope: 'artist I follow', app: 'Spotify') match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/false_start/false_start_01_artist_selection/343a484b44bcbaed/speech.wav", "wer": 0.0, "wav_sha256": "ee628fc7529147a52044538d9811692e2440e1fcbee92c5aa30713cc04e9753c", "seconds": 10.04, "spoken_utterances": [ "Who is the most followed—um... no, I mean ... least followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most followed' but immediately self-corrects to 'least followed'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the most followed um, no, I mean, least followed artist I follow on Spotify?" }, "candidate_id": "wav_46dbad165adfd86f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural and plausible false start. The initial request 'most followed' is clearly spoken, followed by a genuine-sounding hesitation ('um') and a self-correction to 'least followed'. The intonation and pacing of the correction ('No, I mean...') are highly realistic, and the overall flow is continuous and unforced. The speaker's voice and prosody remain consistent throughout, with no signs of artificiality or disjointedness." } ] }, "is_default": true }, { "variant_id": "false_start_02_artist_reference", "proposed_instruction": "Who is the least followed artist I don't follow—sorry, I meant artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "2c9cc969b4f5315e9eb6c020a697efab7882bc8693c3a4c92fa0859e263c4d67", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately corrects to 'artist I follow' within the same utterance. The final intended target is the set of artists the user follows, and the action is to identify the least followed one." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/false_start/false_start_02_artist_reference/304ddd5563404131/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "046399e9886e16624ec2e96e37ef14c5bce550f5c160b2577839e9b953cd3b9b", "seconds": 10.12, "spoken_utterances": [ "Who is the um... least followed artist I don't follow—uh... sorry, I meant artist I follow on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately self-corrects to 'artist I follow'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the um least followed artist I don't follow um sorry, I meant artist I follow on Spotify?" }, "candidate_id": "wav_5911412371cd8502", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a self-correction ('I don't follow... sorry, I meant artist') that is also realistic. However, the initial phrase 'Who is the um...' is slightly more hesitant and drawn out than in audio_02, and the correction feels a bit more abrupt. The speaker also uses 'um' as a filler twice, which is natural, but the overall delivery is less fluid and the correction is less seamlessly integrated into the sentence compared to audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Who is the least followed artist I follow on Apple Music—scratch that, make it Spotify?", "long_pause_offsets": [], "source_sha256": "28cf27a8207f1c97c82c2d232cbefebe95acb35817dbe2a653d9076bb3d73758", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The action (finding the least followed artist) and the scope (artists I follow) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/false_start/false_start_03_app/83fc5ed3c2e08e1d/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "80641aec53606cc527f92c98b12bfa284759837211ec9b165266540ffde762b6", "seconds": 9.68, "spoken_utterances": [ "Who is the uh... least followed artist I follow on Apple Music—um... scratch that, make it Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects with 'scratch that, make it Spotify.' The final audible intent is to query Spotify, which matches the original instruction. The false_start type requires applying the correction, which is done here.", "heard_transcript": "Who is the uh least followed artist I follow on Apple Music? Um, scratch that, make it Spotify." }, "candidate_id": "wav_bf313e93ae2348c9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start with 'Who is the uh...' and then corrects 'Apple Music' to 'Spotify'. While the correction is present, the initial hesitation ('uh') and the correction are less smoothly connected. The phrase 'uh... least followed artist I follow on Apple Music' is delivered with a slightly robotic or read-aloud cadence, and the correction feels more like a mechanical switch than a spontaneous thought process. The prosody is flatter, and the overall delivery is less natural compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Who is the artist I follow on Spotify? Oh, I forgot to mention, I want the least followed artist.", "long_pause_offsets": [ 39 ], "source_sha256": "eecb0d2eec1c7e86810c2f9ee273c1032e70635c6175eba6125824f798f29617", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least followed' to the initial question about artists I follow, refining the scope to the specific artist with the minimum follower count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/addition/addition_01/0ef27f9ac602ae7e/speech.wav", "wer": 0.0, "wav_sha256": "688a3ab7d1594064f954b80fb04929a35d6ac327d20c9fe25c2fbf3ed0946230", "seconds": 11.76, "spoken_utterances": [ "Who is the artist I follow on ... Spotify? ", "Oh, I forgot to mention, I want the ... least followed artist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction 'Who is the artist I follow on Spotify?' is a broad request. The speaker then adds the constraint 'I want the least followed artist' after a pause, which narrows the scope. The final intent is to identify the least followed artist from the user's followed artists, matching the original reference instruction. The addition refines the initial request without contradicting it, satisfying the addition type criterion.", "heard_transcript": "Who is the artist I follow on Spotify? Oh, I forgot to mention, I want the least followed artist." }, "candidate_id": "wav_936d365d48c88e1c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, coherent request with natural prosody and a believable, spontaneous-sounding addition. The initial request is grammatically complete, followed by a pause and a self-correcting addition ('Oh, I forgot to mention...') that is smoothly integrated with appropriate intonation and pacing. The flow feels like a real-time thought process, making it highly realistic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Who is the least followed on Spotify? I didn't mention this, but it's the artist I follow.", "long_pause_offsets": [ 38 ], "source_sha256": "1f579dd8086264db904b3d813eb73f9d06d83e3c74fc68171b70379f5f79cfdf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the query refers to artists within the user's own followed list, which was implied but not explicitly stated in the initial utterance. The final goal combines the initial broad question with the added restriction, matching the original instruction's intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/addition/addition_02/043788c5f17e777c/speech.wav", "wer": 0.15, "wav_sha256": "93e5d6b44dfb36116c43aa8e54adb6e8c84620241d3867667f9acae194d966b8", "seconds": 12.36, "spoken_utterances": [ "Who is the uh... least followed on hmmm... Spotify? ", "I didn't mention this, but it's the um... artist I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial part of the audio is a slightly hesitant but clear version of the original instruction. The speaker then adds a clarifying phrase, 'artist I follow,' which is a direct addition to the original request. The final intent is to find the least followed artist from the subset of artists the user follows, which matches the addition type criterion. The addition clarifies the scope of 'least followed' without changing the core action or target.", "heard_transcript": "Who is the uh least followed on hmm Spotify? I didn't mention this, but it's the um artist I follow." }, "candidate_id": "wav_42e1851a36a5747e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip begins with a natural hesitation ('uh') and a plausible self-correction ('I didn't mention this, but it's the...'), but the delivery is less fluid. The speaker stumbles over the phrase 'least followed on hmm Spotify,' with an awkward pause and filler that disrupts the rhythm. The addition feels slightly forced and less integrated than in audio_02, reducing its realism despite a similar structure." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Who is the least followed artist I follow on Apple Music? Let me correct the app: the least followed artist I follow on Spotify, not Apple Music.", "long_pause_offsets": [ 58 ], "source_sha256": "efacde022b2c80d5e16e123d5e1b787e250786417ee9af1773f9c116a4b28ce1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (least followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/revision/revision_03_app/0629ad33dec72288/speech.wav", "wer": 0.02857142857142857, "wav_sha256": "56a870bebb1dff142707b074ccf26c82c7c9d06b51ff6d5060cda9d355411f08", "seconds": 18.96, "spoken_utterances": [ "Who is the uh, let's see... least followed artist I follow on um, let me think... Apple Music? ", "Let me correct the app: the least followed artist I follow on um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects themselves to Spotify, which is the final and intended target platform. The revision is clear and unambiguous, and the rest of the instruction remains consistent with the original. Therefore, the final intent matches the original instruction and the revision criterion is satisfied.", "heard_transcript": "Who is the uh, let's see, least followed artist I follow on um, let me think, Apple Music? Let me correct the app. The least followed artist I follow on um, Spotify, not Apple Music." }, "candidate_id": "wav_0df70748f7cd4cef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a natural, multi-stage thought process with plausible fillers ('uh', 'let me see', 'let me think') and a clear, self-corrected revision of the platform from 'Apple Music' to 'Spotify'. The correction is integrated smoothly into the ongoing utterance with a natural shift in intonation and pacing, suggesting genuine recall and adjustment rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "revision_02_artist_reference", "proposed_instruction": "Who is the least followed artist I don't follow on Spotify? Make that the artist I follow, not the artist I don't follow.", "long_pause_offsets": [ 60 ], "source_sha256": "f5f0388f012b627fd6ac32e1c59b95670321b82a7142abdbed71ded6528f2625", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial condition from 'I don't follow' to 'I follow', preserving the rest of the query including the app and the intent to find the least followed artist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/revision/revision_02_artist_reference/075ad2e0a502c2a4/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "1a8597690c3858b5df4eb4c537af450111ae639bf4680ac60b0956562bad2dfd", "seconds": 16.68, "spoken_utterances": [ "Who is the uh, let me think... least followed artist I don't follow on um, let's see... Spotify? ", "Make that the artist I follow, not the artist I don't follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the least followed artist that they do not follow, but then immediately corrects this to 'the artist I follow, not the artist I don't follow.' This constitutes a revision where the final intent is the corrected version. The original instruction matches the final corrected intent, so the task is executed as revised.", "heard_transcript": "Who is the uh, let me think, least followed artist that I don't follow on um, let's see, Spotify? Make that the artist I follow, not the artist I don't follow." }, "candidate_id": "wav_e8b70d86108378b8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also shows a natural hesitation and self-correction, changing 'artist that don't follow' to 'artist I follow'. The flow is conversational, with fillers and pauses that fit a spontaneous thought process. However, the correction is slightly less fluid than audio_01, with a brief pause and a more explicit 'Make that...' phrasing, which is still realistic but slightly less seamless." } ] }, "is_default": false }, { "variant_id": "revision_01_artist_selection", "proposed_instruction": "Who is the most followed artist I follow on Spotify? Wait, I meant the least followed artist, not the most followed.", "long_pause_offsets": [ 53 ], "source_sha256": "032d469ebe5d84ea89edd379a1d0fcd8e50aabb9083d902527f46746ed9fc373", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the comparison operator from 'most followed' to 'least followed', preserving the app, scope of followed artists, and the intent to identify a single artist by follower count." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/revision/revision_01_artist_selection/c551f6776fec700a/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "a33aa3a00145832bfd196a330d1c302b0a833db265ed621312ba52145060b4ad", "seconds": 15.08, "spoken_utterances": [ "Who is the um... most followed artist I follow on uh... Spotify? ", "Wait, I meant the um... least followed artist, not the most followed." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'most followed artist' but immediately corrects themselves to 'least followed artist' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction exactly, and the correction is clear and unambiguous.", "heard_transcript": "Who is the um most followed artist I follow on uh Spotify? Wait, I mean the um least followed artist, not the most followed." }, "candidate_id": "wav_8d624095f3143426", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with a clear request but then corrects themselves from 'most followed' to 'least followed'. While the correction is plausible, the phrasing 'Wait, I mean the um least followed artist' introduces a slightly awkward, less natural transition. The 'um' and 'wait' suggest hesitation, but the overall delivery feels a bit more like a rehearsed or less spontaneous correction compared to the other two, making it the least realistic in terms of natural, continuous thought flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "19cedf95620d1a38d5ba389523ddd44a0a9c25bff4064f9bb65475d560637b58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/cancellation/cancel_019/5ea554f839f051cb/speech.wav", "wer": 0.0, "wav_sha256": "6ae621e77e3ed1145bc7f3818bf4d8d85120a4c98e21329415c3c45d19b0b5e6", "seconds": 11.6, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?", "Hang on, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, never mind, don't do that,' which is a clear and unambiguous cancellation of the initial request. The final interjection explicitly revokes the task, satisfying the cancellation type criterion.", "heard_transcript": "Who is the least followed artist I follow on Spotify? Hang on, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "4b1693b5b6095ef0b48f85ffca0855fb19b31417787671ec9c055033d01727b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/ambiguous_halt/halt_002/09d4532b88f8089d/speech.wav", "wer": 0.0, "wav_sha256": "7fcda7be0f36b21d3e993cc4170a6583ab841d1bd70de6a59a232ca8d01e8cf2", "seconds": 9.88, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or modified, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "Who is the least followed artist I follow on Spotify? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a01889086f5f6207363dfc7bfb23a565c9e9308811c2654f9d8c1977ea54492a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_2/audio_natural/filler/original/074aba93571633e5/speech.wav", "wer": 0.0, "wav_sha256": "43f1586ff27b777f3ca27727f7cfdb0bd30ed718fc1bcdd2611ac1bcb7624e4d", "seconds": 6.6, "spoken_utterances": [ "Who is the ... least followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or pause that would indicate a pause or cancellation. The speaker delivers the full request clearly and without correction. As this is a 'filler' type, any initial hesitation or pause within the instruction is ignored, and the final task meaning matches the original instruction.", "heard_transcript": "Who is the least followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music—make that Spotify?", "long_pause_offsets": [], "source_sha256": "3e482f0ec7197ae22fc15bd4f1e06507b814071d953b1b91524489648cd2ce66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says Apple Music but immediately corrects to Spotify. The action (finding the most followed artist) and the scope (artists I follow) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/false_start/false_start_03_app/70b4768f7953a357/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "7ec4c13d7c5f8e05989f90b2657322b316aaac1d68d0a19e84cde6d7d969aeca", "seconds": 7.88, "spoken_utterances": [ "Who is the ... most followed artist I follow on Apple Music—um... make that Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Who is the most followed artist I follow on Apple Music? Um, make that Spotify." }, "candidate_id": "wav_b783420938b428db", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and fluent self-correction. The initial request is slightly hesitant ('who is the...'), followed by a clear, spontaneous correction ('make that Spotify') that flows directly from the prior statement. The prosody is consistent, and the correction feels like a genuine moment of realization rather than a rehearsed edit. The pause before the correction is brief and plausible, fitting a moment of recall." } ] }, "is_default": true }, { "variant_id": "false_start_01_artist_selection", "proposed_instruction": "Who is the least followed—no, I mean most followed artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "2d9781a80cd73afc4951d6feb79a11a08f468d3bf6137bc1e584ab05626eba6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'least followed' to 'most followed' within the same utterance. The rest of the query, including the app (Spotify) and the scope (artists I follow), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/false_start/false_start_01_artist_selection/dd060ee8c6eec36d/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "949315ac961c4d18cd21a1930baaf52ec636bdc443530a2c9710f599bf5a8666", "seconds": 8.36, "spoken_utterances": [ "Who is the uh... least followed—um... no, I mean most followed artist I follow on Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least followed' but immediately self-corrects to 'most followed' with 'no, I mean'. The final audible intent is the corrected version, which matches the original instruction. The false_start type requires applying the correction, which is done here.", "heard_transcript": "Who is the um least followed um no, I mean most followed artist I follow on Spotify?" }, "candidate_id": "wav_4fdf8306cce88d4e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker also shows a clear self-correction from 'least followed' to 'most followed', which is a strong indicator of realism. However, the correction is less fluid than audio_03. The initial hesitation ('um') and the pause before the correction are longer and more pronounced, making the transition feel slightly more forced. The overall delivery is credible but less smooth than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_artist_reference", "proposed_instruction": "Who is the most followed artist I don't follow—sorry, I meant artist I follow on Spotify?", "long_pause_offsets": [], "source_sha256": "c190f5aba568120010c2f9e538e8147b0ebda1a5ba82504cd26b0ef904c68806", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately corrects to 'artist I follow'. The final intent matches the original instruction, which asks for the most followed artist among those the user follows." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/false_start/false_start_02_artist_reference/6dd72d334dd7690b/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "fe7a0fb40fad06ee72fd44fc068e38fbdb8296d622bb7f985b25e6be98ea8dad", "seconds": 9.48, "spoken_utterances": [ "Who is the most followed artist I don't follow—um... sorry, I meant uh... artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'artist I don't follow' but immediately self-corrects with 'sorry, I meant the artist I follow on Spotify'. This is a false start where the final audible intent clearly corrects the initial mistake. The final instruction matches the original reference and should be executed.", "heard_transcript": "Who is the most followed artist I don't follow um sorry, I meant the artist I follow on Spotify?" }, "candidate_id": "wav_64c1faa4735065be", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a self-correction from 'I don't follow' to 'I follow', but the correction is less natural. The initial phrase 'I don't follow' is spoken with a flat, almost robotic intonation, and the correction 'sorry I meant the artist I follow' sounds like a rehearsed or read-aloud edit rather than a spontaneous thought. The prosody lacks the natural variation of a real-time correction, and the transition feels abrupt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Who is the most followed on Spotify? Oh, I forgot to say, I mean the artist I follow.", "long_pause_offsets": [ 37 ], "source_sha256": "52c5f3994bf3915823179b210e3f065f8c034da3fa38aaba21bd334452d57449", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'artist I follow' to the initial broad question, narrowing the scope to only artists the user follows. The core action and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/addition/addition_01/7d2ea8e5bad2d253/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "e534d9179ab2e49413d17d3bbc48c144b1e9e079a535d995e98968af5608cb4e", "seconds": 14.12, "spoken_utterances": [ "Who is the um... most followed on uh, let me think... Spotify? ", "Oh, I forgot to say, I mean the ... artist I follow." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the most followed artist on Spotify, then adds a clarifying phrase 'I mean the artist I follow' after a pause. This addition narrows the scope of the original request, and the final intent is a combined version of both. The addition is clear and directly modifies the initial request to specify the user's personal context, matching the expected behavior for an 'addition' type. The final goal is coherent and executable.", "heard_transcript": "Who is the um most followed on uh let me think Spotify? Oh, I forgot to say, I mean the artist I follow." }, "candidate_id": "wav_82b0e8303309459e", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Who is the most followed artist I follow on Apple Music? Let me correct the app: it should be Spotify, not Apple Music.", "long_pause_offsets": [ 57 ], "source_sha256": "8d8c1d626351067f15b5ca0e833ca742997616e1cd8ab022a5cec7431967a18f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, preserving the rest of the query including the scope (artists I follow) and the comparison (most followed)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/revision/revision_03_app/45e0625e6b4a86e0/speech.wav", "wer": 0.0, "wav_sha256": "5ad8ceaeb3899658d72ac27a775f4d5cc2b5fb64ece40dad80283872b912f92c", "seconds": 11.84, "spoken_utterances": [ "Who is the ... most followed artist I follow on let me think... Apple Music? ", "Let me correct the app: it should be Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Apple Music' but immediately corrects it to 'Spotify' after a pause, which aligns with the revision type criterion. The final audible intent is to perform the task on Spotify, matching the original instruction. The correction is clear and the final goal is consistent with the reference.", "heard_transcript": "Who is the most followed artist I follow on let me think, Apple Music? Let me correct the app, it should be Spotify, not Apple Music." }, "candidate_id": "wav_2b8ffe671f07fa0d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_01_artist_selection", "proposed_instruction": "Who is the least followed artist I follow on Spotify? Wait, the most followed artist I follow, not the least followed.", "long_pause_offsets": [ 54 ], "source_sha256": "0ea3ee425c36e5851d5701c621490debdb676f64ef9b4965478adab39e336b8c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the initial request from 'least followed' to 'most followed', preserving the app (Spotify), scope ('artists I follow'), and the action (identifying the artist)." }, "audio": { "status": "pending", "spoken_utterances": [ "Who is the ... least followed artist I follow on uh... Spotify? ", "Wait, the most followed artist I follow, not the least followed." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_artist_reference", "proposed_instruction": "Who is the most followed artist I don't follow on Spotify? Make that the artist I follow, not the artist I don't follow.", "long_pause_offsets": [ 59 ], "source_sha256": "c35999e6597b22e9c831db0928c2157b0f51d0c43893cc2950219950a616123c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial condition from 'I don't follow' to 'I follow', preserving the rest of the query including the app and the intent to find the most followed artist among the specified group." }, "audio": { "status": "pending", "spoken_utterances": [ "Who is the most followed artist I don't follow on uh, let's see... Spotify? ", "Make that the um, let's see... artist I follow, not the artist I don't follow." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "cd451c406d278775060e054438161f9f70bb8b132455983030bd5278105a7ec6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/cancellation/cancel_023/33f4672d2aa6938a/speech.wav", "wer": 0.0, "wav_sha256": "d1d65541beaa99da6d22b88baebdeca657198ef95f2dc901315b94200619d7d5", "seconds": 12.08, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?", "Oh, wait, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Oh wait, never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly instructs not to execute the prior request, satisfying the cancellation type criterion.", "heard_transcript": "Who is the most followed artist I follow on Spotify? Oh wait, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "6159b7f8af247a5fbf74392dcb46fa9b0e4404fe248362bd2ce164393e4ef56c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/ambiguous_halt/halt_006/f23c31d396dbad19/speech.wav", "wer": 0.0, "wav_sha256": "88c3eade080e95b3712894f0a78152e3f288c7b34425319ed7366cfc3d954677", "seconds": 9.84, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a moment,' which is a clear request to wait or pause. The initial instruction is not canceled or revised, and the final interjection is a temporary halt, not a cancellation. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Who is the most followed artist I follow on Spotify? Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cef9191_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9e69d5cb832bb4b3b3b62eaf38ab8534113c4c60276e141a25d3ef28efe20f30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cef9191_3/audio_natural/filler/original/75f6140c430ff3bc/speech.wav", "wer": 0.0, "wav_sha256": "59f903cce77eedd3ee4d4bee49bfa87f2db652696a110fe4df92d8a86eca06c7", "seconds": 6.56, "spoken_utterances": [ "Who is the ... most followed artist I follow on ... Spotify?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction with no hesitation, backchannel, or pause that changes the goal. The speaker delivers the full request clearly and without correction. As per the 'filler' type criterion, any hesitation or pause within the instruction is ignored, and the final task meaning must match the original instruction. The final intent is fully aligned with the original instruction.", "heard_transcript": "Who is the most followed artist I follow on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_note_description", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Daily Tasks—no, I mean Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "c14fa01b45cedd68564f24dacf94d22fede2182384908f1333d417f938c2de68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'Daily Tasks' to 'Bucket List Simple Note' within the same utterance, preserving the action, note description, and completion status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/false_start/false_start_02_note_description/64a6c064b83729af/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "8b488d9fe3a749778fce09aef41cc9a1a2e2390d4a93478cd5c5b947de1f0d29", "seconds": 13.64, "spoken_utterances": [ "Mark um, let's see... \"Learning to cook a signature dish from scratch\" in my Daily Tasks—um... no, I mean ... Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misidentifies the location as 'daily tasks' but immediately self-corrects to 'bucket list simple note'. According to the 'false_start' type criterion, the final corrected target should be applied. The action ('mark as done') and the task description ('learning to cook a signature dish from scratch') are preserved. The correction is immediate and unambiguous, satisfying the type-specific requirement.", "heard_transcript": "Mark, um, let's see, learning to cook a signature dish from scratch in my daily tasks, um, no, I mean, bucket list simple note as done." }, "candidate_id": "wav_d30d668982ebebd9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction flow. The speaker begins with a filler ('um'), hesitates with 'let's see', then makes a clear false start ('daily tasks') before a smooth, self-correcting revision ('no, I mean bucket list'). The prosody is consistent, and the correction is integrated into the ongoing thought process without awkwardness. The pacing and intonation suggest genuine spontaneous formulation." } ] }, "is_default": true }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Google Keep—scratch that, make it Simple Note as done.", "long_pause_offsets": [], "source_sha256": "890111ee44ac7c76365e351cd796f51d35885f811e78211e770a26786f4e8ff9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Google Keep but immediately corrects to Simple Note. The note title and action (mark as done) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/false_start/false_start_03_note_app/b4dda34f2c4e082d/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "9080dbc21064fed7815b79845e28b392ff18912bcec743e9b39d974b9951ac89", "seconds": 10.24, "spoken_utterances": [ "Mark \"Learning to cook a signature dish from scratch\" in my uh, let me think... Bucket List Google Keep—uh... scratch that, make it Simple Note as um... done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Bucket List Google Keep' but immediately self-corrects with 'scratch that' and specifies 'Simple Note'. According to the 'false_start' type criterion, the final corrected target should be applied. The action (marking as done) and the item remain unchanged. The correction is clear and immediate, so the final intent is to execute the task on 'Simple Note'.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh let me think, Bucket List, Google Keep, uh scratch that, make it Simple Note as um done." }, "candidate_id": "wav_75670ffcf53f3250", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start ('Google Keep') followed by a self-correction ('scratch that, make it Simple Note'). The repair is natural, and the speaker's hesitation ('uh, let me think') is placed plausibly. However, the correction is slightly more abrupt than in audio_03, and the overall flow feels a bit more rehearsed or less fluid in the transition from the false start to the correction." } ] }, "is_default": false }, { "variant_id": "false_start_01_item_text", "proposed_instruction": "Mark \"Learning to play the piano\"—sorry, I meant \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "3ea309e5536e9ed979c3fcfbdbce5a362baa7cb262b92ac4e00e3795780ebf4a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the task description from 'Learning to play the piano' to 'Learning to cook a signature dish from scratch'. The app ('Bucket List Simple Note'), action ('mark as done'), and context are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/false_start/false_start_01_item_text/3fc61f91df9faf82/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "42744f0456059245c260877075b423ec89aa70da9fd1a1e3db2f02868359bc14", "seconds": 13.24, "spoken_utterances": [ "Mark \"Learning to play the piano\"—uh... sorry, I meant \"Learning to cook a signature dish from scratch\" in my um, let me think... Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'learning to play the piano' but immediately self-corrects to 'learning to cook a signature dish from scratch'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the target note and action, remains unchanged and clear. Therefore, the final audible intent matches the original instruction after correction.", "heard_transcript": "Mark learning to play the piano, uh, sorry, I meant learning to cook a signature dish from scratch in my um, let me think, bucket list simple note as done." }, "candidate_id": "wav_ba45186b58cffa77", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a false start ('piano'), then corrects to 'learning to cook'. However, the correction is delivered as a separate, slightly disjointed sentence rather than a smooth integration into the original thought. The 'uh, sorry I meant' phrase is a bit unnatural in this context, and the overall flow feels less like a spontaneous thought process and more like a rehearsed correction." } ] }, "is_default": false }, { "variant_id": "false_start_04_completion_state", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note as pending—I should have said done.", "long_pause_offsets": [], "source_sha256": "be1af0f944c27fa10975ac0c11aab6b4bfea8a2559195eda9df376e60ebec1da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker explicitly corrects 'pending' to 'done' within the same utterance, indicating that 'done' is the intended state. The action, target note, and task description remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/false_start/false_start_04_completion_state/627ffb474a211fe4/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "502c7035f88b5fc1af791173046919254861ce34df8089becd96d2aa623bd198", "seconds": 9.56, "spoken_utterances": [ "Mark \"Learning to cook a signature dish from scratch\" in my ... Bucket List Simple Note as pending—uh... I should have said done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pending' but immediately self-corrects with 'I should have said done.' This is a false start where the final audible intent is to mark the task as done. The correction is immediate and unambiguous, so the final control intent is to execute the corrected version of the task.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note as pending. I should have said done." }, "candidate_id": "wav_1ebe40476807c08f", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has a false start ('pinning') followed by a self-correction ('I should have said done'). However, the correction is delivered as a separate, declarative sentence rather than being integrated into the original request. The prosody and pacing of the correction sound more like a post-hoc realization than a natural part of the thought process, making it the least realistic in terms of spontaneous speech flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note. One more detail: mark it as done.", "long_pause_offsets": [ 85 ], "source_sha256": "3bfd81f40127f49c67281ccab8c96c9436a865cd0358475a8e40284f3cb03df0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the 'mark it as done' instruction, which is consistent with the original goal of marking the note as done. The initial utterance already specifies the note and its content, and the addition completes the intended action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/addition/addition_02/3cdc69f7ad8e36ac/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "fa96e7a45b87d05e16fd4a3b2ddf6d097f6d7875873daa8ba0144a12c9046d9a", "seconds": 13.88, "spoken_utterances": [ "Mark ... \"Learning to cook a signature dish from scratch\" in my um... Bucket List uh... Simple Note. ", "One more detail: mark it as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to mark a specific task in a note as done. The speaker then adds the detail 'One more detail, mark it as done,' which reinforces and clarifies the original intent. Since the addition is consistent with the original instruction and does not introduce conflicting or contradictory information, the final intent is to execute the task as originally stated. The addition serves to emphasize the completion status, not to alter the target or action.", "heard_transcript": "Mark learning to cook a signature dish from scratch in my um bucket list uh simple note. One more detail, mark it as done." }, "candidate_id": "wav_22508d41319b3404", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request to mark a task as done. After a brief pause and filler ('um'), they add a contextual detail ('in my bucket list simple note') in a way that sounds like a spontaneous memory. The prosody is consistent, and the addition flows naturally without awkwardness or robotic phrasing. The speaker's hesitation and correction feel authentic and integrated into the flow." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" as done. Oh, I forgot to mention, mark it in my Bucket List Simple Note.", "long_pause_offsets": [ 63 ], "source_sha256": "81f97460ed8b49ea80a75cc85a5c5244f0df42c7ab8dd365f2e4f1552cdab643", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the required context of the note app, preserving the action and the note content from the initial instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/addition/addition_01/40d779da9f98ae22/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "d9b75b69bd3e5c4367acd565c950ac7b57a37a25680bdefc5f191a51900b4b87", "seconds": 15.24, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" as um... done. ", "Oh, I forgot to mention, mark it in my uh... Bucket List Simple Note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Mark, hmm, learning to cook a signature dish from scratch as um, done', which is a slightly hesitant but complete instruction. Then, after a pause, they add 'Oh, I forgot to mention, mark it in my uh, bucket list simple note', which provides the specific location for the task. Since this is an addition that specifies where to mark the task, the final intent combines the original action with the new context. The final instruction matches the original reference, so the addition is correctly interpreted and executed.", "heard_transcript": "Mark, hmm, learning to cook a signature dish from scratch as um, done. Oh, I forgot to mention, mark it in my uh, bucket list simple note." }, "candidate_id": "wav_44eccb6bc2845ffd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar request but uses a more hesitant 'hmm' and a slightly unnatural pause before 'as um done'. The addition of 'Oh I forgot to mention' is a plausible revision, but the phrase 'mark it in my bucket list simple note' sounds slightly forced and less fluent than audio_01. The prosody is less smooth, and the overall delivery feels a bit more rehearsed or less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_item_text", "proposed_instruction": "Mark \"Learning to cook a signature dish from home\" in my Bucket List Simple Note as done. Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\".", "long_pause_offsets": [ 90 ], "source_sha256": "4df65ee2b5f6dbbd65415246f7c927cbe6cc5b584e3fb2982dd0a85a7781b856", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the item description from 'from home' to 'from scratch', preserving the app, action, and status. The correction is clear and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/revision/revision_01_item_text/63af8cfb45bec2cc/speech.wav", "wer": 0.0, "wav_sha256": "cceba2ff2ff0bfe815062fe1b49a293383f8bd9ac811d2ff04ead3c8be549d7e", "seconds": 18.16, "spoken_utterances": [ "Mark ... \"Learning to cook a signature dish from home\" in my ... Bucket List Simple Note as done. ", "Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the item as 'from home' but immediately corrects it to 'from scratch' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Mark \"Learning to cook a signature dish from home\" in my Bucket List Simple Note as done. Wait, the item should be \"Learning to cook a signature dish from scratch\", not \"Learning to cook a signature dish from home\"." }, "candidate_id": "wav_0ce658818ccb3858", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a clear initial request, then pauses with 'Wait' before smoothly correcting the specific phrase 'from home' to 'from scratch'. The correction is integrated into the same utterance with a natural change of mind intonation, and the flow is continuous and believable without awkwardness or robotic delivery." } ] }, "is_default": true }, { "variant_id": "revision_03_note_app", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Evernote as done. Let me correct the app: use \"Simple Note\", not Evernote.", "long_pause_offsets": [ 90 ], "source_sha256": "5a4e75669d5ad14c7ff26dd7195e69782365a78794b670c2c3d68a94b794319a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Evernote to Simple Note, preserving the note title, action, and list context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/revision/revision_03_note_app/82a93b06bc7f0ada/speech.wav", "wer": 0.22580645161290322, "wav_sha256": "c89c6a594d3f55585e7e726dbecac426a9b8beb7adf53d5c12d5cb98f90c0d05", "seconds": 16.76, "spoken_utterances": [ "Mark uh... \"Learning to cook a signature dish from scratch\" in my hmmm... Bucket List Evernote as uh, let's see... done. ", "Let me correct the app: use \"Simple Note\", not Evernote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately corrects it to 'Simple Note' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to mark the task as done in Simple Note, which matches the original instruction's intent except for the app name. The correction is clear and final, so the task should be executed in Simple Note.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Evernote as done. Let me correct the app, use Simple Note not Evernote." }, "candidate_id": "wav_a850a44c28b76aa2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a realistic revision with a believable self-correction. The speaker starts with a clear request, then hesitates with 'hmm' and 'let's see' before correcting the app name from 'Evernote' to 'SimpleNote'. The fillers are placed naturally, and the correction is delivered as a spontaneous realization rather than a forced edit, making it sound like a real-time thought process." } ] }, "is_default": false }, { "variant_id": "revision_04_completion_state", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Bucket List Simple Note as todo. Change the completion state to \"done\" instead of \"todo\".", "long_pause_offsets": [ 93 ], "source_sha256": "4d821a5bb8b79ba2623ed190520cbc0e104b59259c6a8ea015b0a5a9166d7cc4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the completion state from 'todo' to 'done', preserving the note title and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/revision/revision_04_completion_state/85f083c9d5638bcb/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "9aa7fb41446e2999d66d13544ff1e3b407d761236332263fac71ddbeae818003", "seconds": 16.96, "spoken_utterances": [ "Mark \"Learning to cook a signature dish from scratch\" in my um... Bucket List ... Simple Note as todo. ", "Change the completion state to \"uh... done\" instead of \"todo\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as to do' but immediately corrects it to 'Change the completion state to uh done instead of todo.' This is a revision where the final instruction clearly overrides the initial one. The final intent is to mark the task as 'done', which matches the expected control intent for a revision. All other constraints (task name, list name) remain consistent.", "heard_transcript": "Mark learning to cook a signature dish from scratch in my um bucket list simple note as to do. Change the completion state to uh done instead of todo." }, "candidate_id": "wav_d52726d1eb26f0f2", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but with less natural flow. The speaker starts with a clear request, then pauses with 'um' before correcting the completion state from 'to do' to 'done'. While the correction is plausible, the phrasing 'change the completion state to uh done' sounds slightly awkward and less fluid than the top two, with a less organic integration of the correction." } ] }, "is_default": false }, { "variant_id": "revision_02_note_description", "proposed_instruction": "Mark \"Learning to cook a signature dish from scratch\" in my Daily Goals Simple Note as done. Make that \"Bucket List\", not \"Daily Goals\".", "long_pause_offsets": [ 93 ], "source_sha256": "8dcda64c4e9aa3d20f8981b30e9cae2b8850e217397a0bfb98b23b5f11de3cb4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the note name from 'Daily Goals' to 'Bucket List', preserving the action, item, and completion status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/revision/revision_02_note_description/adf5f58b801e1b68/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "901099286c8706a000cf7b8d5da0ed1ba0eebb0f227d15b32ef9a77a5cc6a3ce", "seconds": 13.56, "spoken_utterances": [ "Mark um... \"Learning to cook a signature dish from scratch\" in my ... Daily Goals Simple Note as uh... done. ", "Make that \"Bucket List\", not \"Daily Goals\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the note location as 'daily goals' but immediately corrects it to 'bucket list' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to mark the task as done in the Bucket List Simple Note, which matches the original instruction. All other constraints are preserved.", "heard_transcript": "Mark um, learning to cook a signature dish from scratch in my daily goals simple note as a done. Make that bucket list, not daily goals." }, "candidate_id": "wav_985553d95c3d8e67", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip is the least realistic. It begins with a clear request, then abruptly corrects 'daily goals' to 'bucket list' without any natural hesitation or filler. The correction is delivered as a flat, immediate statement rather than a spontaneous self-repair, making it sound more like a read-aloud correction than a natural revision. The lack of hesitation or flow reduces its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_024", "hide_variant_original": true, "source_sha256": "fff6cac785542edebe8927b1fa341a4c7d0b80ea1fbacdaa4f1a7c8c80f8d696", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/cancellation/cancel_024/b4678578ec3a325b/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "cf5f7447e4b30f6298805f642ceae97c1a59720482c524bb9bdc94cee27e3873", "seconds": 16.72, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done.", "Wait, oh, never mind. Cancel the whole thing." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original task but immediately corrects themselves with 'Wait, oh never mind, cancel the whole thing,' which is a clear and final cancellation of the task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done. Wait, oh never mind, cancel the whole thing." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "8a3c1921fb3648dff8889f71af49249a9911534fda636b7132e9c86418213b7a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/ambiguous_halt/halt_006/b18990a40cbdcc83/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "54f1c8cfb753471ec4b86382d9646e3fa302a23fd2cecff417f04d27b78cc8fd", "seconds": 14.92, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Just a moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final interjection indicates a temporary halt rather than cancellation or execution. The initial instruction is not revoked, and the speaker does not resume or correct it, so the final intent is to pause.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done. Just a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "50234f82310b915614798d96be76412f39368fd59592cb97345b30a452bed248", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_1/audio_natural/filler/original/6d834e580b4f3476/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "c51898a97323ec1df73ece730cb7e63a0cb4389458f8c4fa18d25ecb7e9893ae", "seconds": 11.88, "spoken_utterances": [ "Mark hmmm... \"Learning to cook a signature dish from scratch\" in my ah... Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a brief hesitation ('uh') but does not alter the core instruction. The final intent matches the original request exactly, with no correction or change in goal. The hesitation is a filler and does not constitute a change in control intent.", "heard_transcript": "Mark \"Learning to cook a signature dish from scratch\" in my uh Bucket List Simple Note as done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_item_text", "proposed_instruction": "Mark \"Going to the beach\"—I meant \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done.", "long_pause_offsets": [], "source_sha256": "6d85864e1a82050840793e48f6709357b0aee7226fece9b4ad6b6e5f7be393e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the note title from \"Going to the beach\" to \"Taking a solo backpacking trip\" within the same utterance. The app, action, and status ('not done') remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/false_start/false_start_01_item_text/f421e4653da53f89/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "1d76a1fe1ed3dcfd75cbe8fd1f143c723af8d3749c0d2e89a97c1272f47622a6", "seconds": 12.72, "spoken_utterances": [ "Mark um, let me think... \"Going to the beach\"—uh... I meant \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Going to the beach' as a false start, then immediately self-corrects with 'I meant taking a solo backpacking trip'. The correction is clear and immediate, and the rest of the instruction matches the original. The final intent is to mark the correct task as not done, so the false_start type is satisfied.", "heard_transcript": "Mark, um, let me think. Going to the beach. Ah, I meant taking a solo backpacking trip in my bucket list simple note as not done." }, "candidate_id": "wav_926354e956d7e542", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and plausible false start. The speaker begins with a clear hesitation ('um, let me think'), then makes a distinct, unrelated initial attempt ('going to the beach') before self-correcting with 'I meant...' to deliver the correct request. The correction is smooth, the intonation reflects genuine recollection, and the pacing feels spontaneous rather than rehearsed. The filler words and pauses are well-placed and contribute to the realism of a person formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_04_completion_state", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as done—no, I mean not done.", "long_pause_offsets": [], "source_sha256": "1cb23347006447888fda5d6cc4c70fbd060310569c251d9e365859c767849c9c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'as done' but immediately corrects with 'no, I mean not done', so the final intended action is to mark the item as not done." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/false_start/false_start_04_completion_state/74149e416f7f1985/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "cf5e65f6f5613536eb0adaa38e490441068cb95b84e0c0eb4fe25347737ba4d3", "seconds": 10.8, "spoken_utterances": [ "Mark \"Taking a solo backpacking trip\" in uh, let me think... my Bucket List Simple Note as done—um... no, I mean ... not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as done' but immediately self-corrects to 'not done'. According to the false_start type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final audible intent matches the original instruction exactly.", "heard_transcript": "Mark taking a solo backpacking trip in uh let me think. My bucket list simple note as done um. No, I mean not done." }, "candidate_id": "wav_1b887b1e9f8bd2a7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with a clear self-correction ('no, I mean not done'). The initial hesitation ('uh, let me think') and the correction are natural. However, the correction is less integrated than in audio_04; the speaker seems to abruptly switch from 'done' to 'not done' without a smooth transition, and the final 'not done' is delivered with a slightly flat intonation, making the correction feel a bit more mechanical than the more fluid correction in audio_04." } ] }, "is_default": false }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Evernote—scratch that, make it Simple Note as not done.", "long_pause_offsets": [], "source_sha256": "e71860f04a0337eff55077843ca93d8c4371a0d44702b67634bab5629ac9a369", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Evernote but immediately corrects to Simple Note. The action, item, and status remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/false_start/false_start_03_note_app/a75f4e6adbc8e9ea/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "f300874dcaa5f7dad17a53dd91171e3125ef860dae3ab277bca076c9b72fa33c", "seconds": 14.36, "spoken_utterances": [ "Mark uh... \"Taking a solo backpacking trip\" in my Bucket List Evernote—um... scratch that, make it hmmm... Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately self-corrects with 'scratch that make it Simple Note'. According to the 'false_start' type criterion, the final corrected target app 'Simple Note' should be used, and the original instruction with the corrected target is fully executable. The action and item remain consistent.", "heard_transcript": "Mark uh taking the solo backpacking trip in my bucket list Evernote um scratch that make it Simple Note as not done." }, "candidate_id": "wav_7398f8504d0ec9da", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip features a false start with a self-correction ('scratch that, make it...'). The initial hesitation ('uh') and the correction are present, but the correction is less plausible. The speaker first says 'Evernote' and then immediately corrects to 'Simple Note'. The correction feels abrupt and less like a genuine recollection and more like a forced edit. The 'scratch that' phrase is also less natural in this context compared to 'I meant...' in audio_04." } ] }, "is_default": false }, { "variant_id": "false_start_02_note_description", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Travel Goals—sorry, I should have said Bucket List Simple Note as not done.", "long_pause_offsets": [], "source_sha256": "86defb81d945525e59bc9f6e949b5f525ab1ef163ce68c0c9f66b404d649cedd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the app name from 'Travel Goals' to 'Bucket List Simple Note' within the same utterance. The action, item, and status remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/false_start/false_start_02_note_description/8c2b3de8efe3d4cd/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "66a11f6e88d22513177d10d69055fe89fa1978ccd49cdbab3ea26c8713fcbb42", "seconds": 12.44, "spoken_utterances": [ "Mark \"Taking a solo backpacking trip\" in my Travel Goals—um... sorry, I should have said hmmm... Bucket List Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'travel goals' but immediately self-corrects to 'bucket list simple note'. According to the 'false_start' type criterion, the final corrected target should be applied. The action (marking as not done) and the item (Taking a solo backpacking trip) remain consistent. The correction is clear and immediate, so the final intent matches the original instruction.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my travel goals, um, sorry, I should have said, bucket list simple note as not done." }, "candidate_id": "wav_f3cd377e7e5c0a06", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start with a self-correction ('sorry, I should have said...'). The initial hesitation ('um') and the correction are present, but the correction is less integrated. The speaker says 'travel goals' and then corrects to 'bucket list'. The correction feels more like a simple word substitution rather than a genuine recollection. The intonation is relatively flat, and the overall delivery sounds less spontaneous and more rehearsed compared to the other clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note. Oh, I forgot to add, mark it as not done.", "long_pause_offsets": [ 69 ], "source_sha256": "69961315a95abb9b21c2b3c264864d2a4318e980d4b1bb354438c623153309ee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'not done' status after the initial marking action, completing the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/addition/addition_01/2280a904a8619603/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "bd747e527702c5e53b00b1151d02c18e42a58c8ec1967acc3bac157cef3d4fed", "seconds": 16.48, "spoken_utterances": [ "Mark hmmm... \"Taking a solo backpacking trip\" in um... my Bucket List Simple Note. ", "Oh, I forgot to add, mark it as let's see... not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to mark the item as 'not done' but then adds the explicit phrase 'mark it as, let's see, not done' after a pause, which clarifies and completes the original instruction. The final audible intent matches the original instruction exactly, with no contradiction or ambiguity.", "heard_transcript": "Mark, hmm, taking a solo backpacking trip in um, my bucket list simple note. Oh, I forgot to add, mark it as, let's see, not done." }, "candidate_id": "wav_7fede9813e526ac8", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('um', 'oh I forgot to add') and a believable self-correction ('mark it as let's see not done') that flows smoothly within the utterance. The prosody is conversational, with appropriate pauses and intonation shifts that suggest spontaneous thought. The addition of the missing detail feels like a genuine memory, not a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in as not done. I should also mention, do this in my Bucket List Simple Note.", "long_pause_offsets": [ 54 ], "source_sha256": "820ad8359f04e81d952e726b804e54975633455c9feec70db79c78d4c66ea477", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The first utterance partially specifies the task but omits the note name. The followup correctly supplies the missing context, completing the full instruction with the correct app and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/addition/addition_02/0d1e238eef04b26d/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "0ba062b495af7d5bb97d1b9a668fc47389cf7744c3a89bdd9204eb2e57915055", "seconds": 13.76, "spoken_utterances": [ "Mark ah... \"Taking a solo backpacking trip\" in as not done. ", "I should also mention, do this in my ... Bucket List Simple Note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete, missing the target app 'Bucket List Simple Note'. The speaker then adds this information after a pause, completing the full task. The addition correctly specifies the app, and the final intent matches the original instruction. The type is 'addition', and the final combined goal is complete and correct.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in as not done. I should also mention, do this in my Bucket List Simple Note." }, "candidate_id": "wav_e8624f0e28189b71", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker includes a plausible addition ('I should also mention do this in my bucket list simple note'), the delivery is less fluid. The phrase 'do this in my bucket list simple note' is awkwardly phrased and delivered with a slightly robotic cadence. The self-correction is less integrated into the flow, and the overall prosody lacks the natural hesitation and recovery seen in audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_note_description", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Travel Goals Simple Note as not done. Let me correct the note description: use \"Bucket List\", not \"Travel Goals\".", "long_pause_offsets": [ 82 ], "source_sha256": "d1f3509c99ec11faf0c75d96e4f34244ad0d3da4404d9b5b521b345dd6b67196", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the note name from 'Travel Goals' to 'Bucket List', preserving the action, item, and status. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/revision/revision_02_note_description/9fba94e12798c81a/speech.wav", "wer": 0.12903225806451613, "wav_sha256": "02089912c392f400e12dfb471d81639a7e5bf609f27a6407211a169c51f2aeae", "seconds": 19.6, "spoken_utterances": [ "Mark uh, let's see... \"Taking a solo backpacking trip\" in uh... my Travel Goals Simple Note as not done. ", "Let me correct the note description: use \"... Bucket List\", not \"Travel Goals\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'travel goals' but immediately corrects it to 'bucket list' after a pause, which is the revision type. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Mark uh let's see taking a solo backpacking trip in uh my travel goals simple note as not done. Let me correct the note description, use bucket list not travel goals." }, "candidate_id": "wav_1e66aa9b3d77bfad", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous prosody with believable hesitation and self-correction. The speaker uses 'uh' and 'let me see' as genuine fillers, pauses naturally, and smoothly revises 'travel goals' to 'bucket list' with a plausible justification ('let me correct the note description'). The intonation and rhythm mimic a real person thinking aloud and adjusting their request, making it the most realistic." } ] }, "is_default": true }, { "variant_id": "revision_04_completion_state", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as done. Wait, mark it as \"not done\", not \"done\".", "long_pause_offsets": [ 77 ], "source_sha256": "d278a881b42bf0a28ee5a826bc4422b77840570cf210017dfe6f2017f6fcf597", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the status from 'done' to 'not done', preserving the action, item, and note context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/revision/revision_04_completion_state/0b8406c12195ab57/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "1d36a9c80ab4313436cb9281dcda15d096074fab2a8e0c08cf2cbbb8f01746f9", "seconds": 14.72, "spoken_utterances": [ "Mark let me think... \"Taking a solo backpacking trip\" in my Bucket List Simple Note as done. ", "Wait, mark it as \"hmmm... not done\", not \"done\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as done' but immediately corrects themselves with 'Wait, mark it as hmm, not done, not done.' This constitutes a revision where the final audible intent is to mark the item as 'not done'. The correction is clear and immediate, satisfying the revision type criterion. The final control intent is to execute the revised task.", "heard_transcript": "Mark, let me think. Taking a solo backpacking trip in my bucket list simple note as done. Wait, mark it as hmm, not done, not done." }, "candidate_id": "wav_1c8066ae1e264ca9", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision with a false start ('wait, mark it as...') and a self-correction ('not done, not done'). The speaker's hesitation and repetition sound spontaneous. However, the delivery is slightly more robotic than audio_03, with less natural phrasing in the correction and a slightly abrupt transition between the initial and revised commands." } ] }, "is_default": false }, { "variant_id": "revision_01_item_text", "proposed_instruction": "Mark \"Hiking in the Alps\" in my Bucket List Simple Note as not done. Change the item to \"Taking a solo backpacking trip\" instead of \"Hiking in the Alps\".", "long_pause_offsets": [ 69 ], "source_sha256": "a625a74907fd30f0dfc73d484b2cdf935094aafc1311c312ea96a676f3a949e3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "unchecked", "reason": "현재 문장과 일치하는 semantic check 결과가 없습니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/revision/revision_01_item_text/bb497bec5ef9c7f8/speech.wav", "wer": 0.12121212121212122, "wav_sha256": "d31f00ae6d2e5ed504d5b3efb1b784f8f253d74cd2c2c76d680c01ecdce506f3", "seconds": 20.16, "spoken_utterances": [ "Mark uh, let's see... \"Hiking in the Alps\" in ... my Bucket List Simple Note as not done. ", "Change the item to uh... \"Taking a solo backpacking trip\" instead of \"Hiking in the Alps\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'hiking in the Alps' but immediately corrects to 'taking a solo backpacking trip' after a pause, which is the final and intended task. The revision is clear and the final goal matches the original instruction. The correction is applied to the earlier instruction, preserving the 'not done' status and the app context.", "heard_transcript": "Mark uh, let's see, hiking in the Alps in my bucket list simple note as not done. Change the item to uh, taking a solo backpacking trip instead of hiking in the Alps." }, "candidate_id": "wav_a84aea0d0470fe44", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker uses fillers ('uh', 'let's see') and pauses, but the revision is less natural. The phrase 'change the item to...' sounds more like a mechanical instruction than a spontaneous thought. The correction is delivered with a flat, declarative tone, lacking the intonation shift of a genuine change-of-mind, making it less realistic than the others." } ] }, "is_default": false }, { "variant_id": "revision_03_note_app", "proposed_instruction": "Mark \"Taking a solo backpacking trip\" in my Bucket List Evernote as not done. Use \"Simple Note\" instead of \"Evernote\" for the note app.", "long_pause_offsets": [ 78 ], "source_sha256": "040033195fb4a1319810e2cacd9d5af3ea8e93275c6ea31f2750ffdbeff98a36", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the note app from Evernote to Simple Note, preserving the note title, action, and status. The original instruction's app was Evernote, but the correction specifies Simple Note as the intended app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/revision/revision_03_note_app/6098f22d4c0c17f6/speech.wav", "wer": 0.08, "wav_sha256": "71c97b5d08f893e1d1bba44fa4e4ece1c895c72de46dbfea86a5fdcf17d65e42", "seconds": 14.56, "spoken_utterances": [ "Mark \"Taking a solo backpacking trip\" in ... my Bucket List Evernote as not done. ", "Use \"um... Simple Note\" instead of \"Evernote\" for the note app." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately corrects to 'Simple Note' after a pause and filler word 'um'. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to use 'Simple Note', which matches the original instruction's target app. All other conditions are preserved.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Evernote as not done. Use um, Simple Note instead of Evernote for the note app." }, "candidate_id": "wav_00122782e8de42d3", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has the least natural delivery. The speaker starts with a clear, fluent sentence, then adds a correction ('use um...') that sounds like a post-hoc instruction rather than a spontaneous realization. The 'um' is awkwardly placed and the correction lacks the prosodic cues of a real-time thought process. The overall flow feels more like a scripted or rehearsed task than a natural utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "d842df6653a527e511d21050b4b102b42b6c2524bbc7d56f204f03954a51046a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/cancellation/cancel_003/7f47edfed656f2cf/speech.wav", "wer": 0.0, "wav_sha256": "71dd74b1c2e4d093da888a18177460900b74a83c885f64744a412070c23fe7a7", "seconds": 12.84, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done.", "Hang on, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio begins with the full original instruction, followed by a clear and final self-correction: 'Hang on, please cancel that.' This constitutes a direct cancellation of the entire task. The final audible intent is to cancel, not to execute the original request. The cancellation is unambiguous and complete, satisfying the 'cancellation' type criterion.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done. Hang on, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "921619842b7c2ac3eefdfb345c251b1caaf8590d21d32255d2f3fce750d913b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/ambiguous_halt/halt_021/f6387170c44213e5/speech.wav", "wer": 0.0, "wav_sha256": "dc75d715a969760e6c71af425032da57a96a7af563223810c89f36c14f51622d", "seconds": 12.72, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The initial task is not revoked, and the speaker does not provide a new goal, so the correct control intent is 'pause'.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "36555ac41d1b7b80164c287a816090d996d131af8f09cfab60baf9e9a752bb2a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_2/audio_natural/filler/original/da42fda639b01ebb/speech.wav", "wer": 0.0, "wav_sha256": "5d7b69f8e361c3f7c44e3609e89e40344ef323a8a5fd34a33c32ef7532e9a42e", "seconds": 8.52, "spoken_utterances": [ "Mark ... \"Taking a solo backpacking trip\" in ... my Bucket List Simple Note as not done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the 'filler' type allows for minor hesitations or pauses that do not alter the goal. Since the final utterance is a complete and correct execution of the task, the verdict is pass.", "heard_transcript": "Mark \"Taking a solo backpacking trip\" in my Bucket List Simple Note as not done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_item_text", "proposed_instruction": "Mark \"Hiking Mount Everest\"—I meant \"Witnessing a total solar eclipse\" in my Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "9fb6f2786dfb49f1ea968d8c6fce64a0dbf4d7f61db101e009537d0085e2fe66", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the item name from \"Hiking Mount Everest\" to \"Witnessing a total solar eclipse\" within the same utterance. The app, action, and status ('as done') remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/false_start/false_start_01_item_text/c2e785b824ed0c36/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "be5bc0da0ad6b0fd8eddbc1cb1d2b1e8b473dba693b45e352d672c35dfcd72f4", "seconds": 11.52, "spoken_utterances": [ "Mark um... \"Hiking Mount Everest\"—hmmm... I meant \"Witnessing a total solar eclipse\" in uh... my Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hiking Mount Everest' but immediately self-corrects to 'witnessing a total solar eclipse'. According to the 'false_start' type criterion, the final intent is the corrected version, and the initial mistake is discarded. The final instruction matches the original reference exactly, so the task is executed as intended.", "heard_transcript": "Mark um, hiking Mount Everest. Hmm. I meant witnessing a total solar eclipse in uh my Bucket List Simple Note as done." }, "candidate_id": "wav_bee2b5bde89b90bc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. It begins with a false start ('hiking Mount Everest'), followed by a believable self-correction ('I meant...') with a realistic 'hmm' hesitation. The correction is delivered with a natural shift in intonation and pacing, and the final request is clear and complete. The prosody and rhythm are consistent with someone thinking aloud and adjusting their request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_note_app", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Google Keep—scratch that, make it Simple Note as done.", "long_pause_offsets": [], "source_sha256": "23ed989ddadc97c1caf1b6606168a5a318154305dbe26bec57fb6a09287e3078", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Google Keep but immediately corrects to Simple Note. The action, item, and status remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/false_start/false_start_03_note_app/e1ece46a4f126700/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ae78d757fc7b91429a16c57e4fdccb485baeeae2c2fdb567e6a1cf1c67ce9a23", "seconds": 9.68, "spoken_utterances": [ "Mark uh... \"Witnessing a total solar eclipse\" in my Bucket List Google Keep—um... scratch that, make it Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misnames the app as 'Google Keep' but immediately self-corrects to 'Simple Note' after a pause and the phrase 'Scratch that'. According to the 'false_start' type criterion, the final corrected instruction must be executed, and the initial incorrect value is discarded. The final goal matches the original instruction's intent, with only the app name corrected from 'Google Keep' to 'Simple Note'.", "heard_transcript": "Mark uh witnessing a total solar eclipse in my bucket list Google Keep um. Scratch that, make it Simple Note as done." }, "candidate_id": "wav_742331759c1eaffa", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start ('Google Keep') followed by a self-correction ('Simple Note'). The speaker's intonation and pacing are natural, and the correction is immediate and smooth. The only minor issue is that the initial error ('Google Keep') is not a plausible alternative to 'Simple Note' in the same context, making the correction slightly less pragmatic, but the delivery itself is realistic." } ] }, "is_default": false }, { "variant_id": "false_start_04_completion_state", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Simple Note as pending—no, that should be done.", "long_pause_offsets": [], "source_sha256": "b4310120930c2dc1136737e64d07310d8cee257e5be51798c3d4426ac80d8b7e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pending' but immediately corrects it to 'done'. The final intended action is to mark the note as done, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/false_start/false_start_04_completion_state/e9c1d9969370cbef/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "041b79216570a063a261e64164821a598505b29209819b9410c5a4fa02d3ec94", "seconds": 15.48, "spoken_utterances": [ "Mark uh... \"Witnessing a total solar eclipse\" in my Bucket List Simple Note as let's see... pending—uh... no, that should be done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pending' but immediately self-corrects to 'done'. According to the 'false_start' type criterion, the final corrected value should be applied. The final intent is to mark the task as 'done', which matches the original instruction. The correction is clear and immediate, with no other conflicting conditions.", "heard_transcript": "Mark uh witnessing a total solar eclipse in my bucket list simple note as let's see pending uh no, that should be done." }, "candidate_id": "wav_2c1513bbabb9af14", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start and a self-correction, but the correction is less natural. The speaker says 'let's see pending' before correcting to 'done', which sounds like a forced or rehearsed phrase rather than a spontaneous thought. The prosody is slightly off, and the correction feels more like a mechanical fix than a genuine change of mind." } ] }, "is_default": false }, { "variant_id": "false_start_02_note_description", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Travel Goals—sorry, I should have said Bucket List Simple Note as done.", "long_pause_offsets": [], "source_sha256": "784cf3b0631139f3291a9b0d9eeb6b452439942c3e2b7b8057f638504cf42a74", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially misnames the note as 'Travel Goals' but immediately corrects it to 'Bucket List Simple Note'. The action, item, and status remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/false_start/false_start_02_note_description/c2b3eca1c7c723d8/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "ca91cc668793f8356c46a2d157b0b61cbda01728cf416a302b5240e5f0b94bf5", "seconds": 11.68, "spoken_utterances": [ "Mark uh... \"Witnessing a total solar eclipse\" in my Travel Goals—um... sorry, I should have said Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the note as 'travel goals' but immediately self-corrects to 'bucket list simple note'. According to the 'false_start' type criterion, the final corrected target must be applied. The action ('mark as done') and the item ('Witnessing a total solar eclipse') are preserved, and the correction is clear and immediate. The final intent matches the original instruction.", "heard_transcript": "Mark a witnessing a total solar eclipse in my travel goals. Um, sorry, I should've said bucket list simple note as done." }, "candidate_id": "wav_f719b0eb03af6617", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start ('travel goals') and a correction ('bucket list Simple Note'), but the correction is awkward and unnatural. The speaker says 'sorry I should have said' which is a very formal and rehearsed way to self-correct, breaking the flow. The prosody is stilted, and the overall delivery sounds less spontaneous and more like a scripted example." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Simple Note. Oh, I forgot to mention, mark it as done.", "long_pause_offsets": [ 71 ], "source_sha256": "d68a3490f9188391473dc89e4e09b274554f9cea1c6f46a3607ed9fb0184d8e7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'mark it as done' instruction, which completes the original task. The note title and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/addition/addition_01/5eddfa1a41e3e581/speech.wav", "wer": 0.16, "wav_sha256": "a783726ae0f6bd49a65b6353462d0fce5a5a050db4336b30ed5c2edf000329db", "seconds": 13.88, "spoken_utterances": [ "Mark \"Witnessing a total solar eclipse\" in ... my Bucket List Simple Note. ", "Oh, I forgot to mention, mark it um, let me think... as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to mark an item in a note. The speaker then adds 'as done' after a brief hesitation, which is a clear addition to the original task. The final intent is to mark the specified item as done, which matches the expected behavior for an 'addition' type. The correction is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "Mark witnessing a total solar eclipse in my bucket list simple note. Oh, I forgot to mention, mark it, um, let me think, as done." }, "candidate_id": "wav_68634ef5dffb4703", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then pauses with a believable 'oh I forgot to mention' and 'let me think' before adding the correction 'mark it um as done'. The prosody is consistent, the fillers are placed naturally, and the self-correction flows smoothly within the utterance, making it sound like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in as done. One more thing: do this in my Bucket List Simple Note.", "long_pause_offsets": [ 52 ], "source_sha256": "f63ebf5f55ff3d88fd1e85c95f16ed6ad06218542ede0ba82306dfe3e7870906", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The first utterance partially specifies the action and item but omits the target note. The followup adds the missing context, completing the full instruction. The final goal matches the original." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/addition/addition_02/e40047ec37697a89/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "93fcb1b893f5b9e6f435e72760de5fb72050334ac832285554835f0f560321b9", "seconds": 15.08, "spoken_utterances": [ "Mark um, let's see... \"Witnessing a total solar eclipse\" in as done. ", "One more thing: do this in uh, let me think... my Bucket List Simple Note." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is partially obscured by hesitation ('Mark um, let's see...') but the core action and item are audible. The addition 'do this in my bucket list simple note' clarifies the target note, which was implied but not fully specified in the original instruction. The final intent combines the original task with the added context, matching the reference instruction exactly. The hesitation and self-correction do not alter the final goal.", "heard_transcript": "Mark um, let's see, witnessing a total solar eclipse in as done. One more thing, do this in uh, let me think, my bucket list simple note." }, "candidate_id": "wav_a9cc4a8107df61e3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also includes a self-correction, the initial request is less fluent and contains a more abrupt 'let me see' followed by a fragmented delivery. The second part of the utterance, 'one more thing, do this in uh let me think my bucket list simple note', sounds more like a disjointed afterthought with a less natural flow and a slightly awkward phrasing, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_item_text", "proposed_instruction": "Mark \"Hiking Mount Everest\" in my Bucket List Simple Note as done. Wait, the item should be \"Witnessing a total solar eclipse\", not \"Hiking Mount Everest\".", "long_pause_offsets": [ 67 ], "source_sha256": "954738b984e7be680f0812aa22f903494b7efe9497e712812f0b0c959e100937", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly identifies the intended item to mark as done, replacing the initially mentioned 'Hiking Mount Everest' with 'Witnessing a total solar eclipse'. The app ('Bucket List Simple Note') and action ('mark as done') remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/revision/revision_01_item_text/70ce748b63748f16/speech.wav", "wer": 0.0, "wav_sha256": "34bf9c5e8fcf6fe152dee0fdd4043e021288b7bb692165c5fdcd9fef3cd3d380", "seconds": 16.2, "spoken_utterances": [ "Mark \"Hiking Mount Everest\" in um... my Bucket List Simple Note as done. ", "Wait, the item should be \"uh... Witnessing a total solar eclipse\", not \"Hiking Mount Everest\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the item as 'hiking Mount Everest' but immediately corrects it to 'witnessing a total solar eclipse' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Mark hiking Mount Everest in um my Bucket List Simple Note as done. Wait, the item should be uh witnessing a total solar eclipse, not hiking Mount Everest." }, "candidate_id": "wav_1ead70089dd9c80e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural hesitation ('um') and a believable self-correction. The revision ('the item should be...') is smoothly integrated into the same utterance, with a pause and filler ('uh') before the correction, matching a spontaneous thought process. The prosody is consistent, and the correction flows logically without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_02_note_description", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Travel Goals Simple Note as done. Make that the note is \"Bucket List\", not \"Travel Goals\".", "long_pause_offsets": [ 80 ], "source_sha256": "593f3835db00c19c615a2b4940143608c27080ba20180c1615acc6222e0d3b9a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the note name from 'Travel Goals' to 'Bucket List', preserving the action, item, and completion status." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/revision/revision_02_note_description/e3325fd62be11a8c/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "67101d5320d69c621ba9ff3a596244a6f7221070f868739869c7c4f6b0659cb4", "seconds": 17.36, "spoken_utterances": [ "Mark uh, let's see... \"Witnessing a total solar eclipse\" in uh... my Travel Goals Simple Note as done. ", "Make that the note is \"um... Bucket List\", not \"Travel Goals\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my travel goals simple note' but immediately corrects it to 'bucket list' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly.", "heard_transcript": "Mark uh let's see witnessing a total solar eclipse in uh my travel goals simple note as done. Make that the note is um bucket list not travel goals." }, "candidate_id": "wav_dad2cf61ae6a125b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a hesitant 'uh' and 'let's see', then makes a clear request. The revision is introduced with a natural pause and filler ('make that the note is um'), followed by a believable correction. The flow is coherent, though the correction is slightly more abrupt than in audio_01, and the 'um' filler is less integrated into the sentence." } ] }, "is_default": false }, { "variant_id": "revision_03_note_app", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Evernote as done. Let me correct the app: use \"Simple Note\", not Evernote.", "long_pause_offsets": [ 76 ], "source_sha256": "78e0752ea6176410a6c990ceb2c353242e37fcd175f75cf72554c7938ee310a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Evernote to Simple Note, preserving the note title and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/revision/revision_03_note_app/0a99c2dcc9919a5b/speech.wav", "wer": 0.16, "wav_sha256": "1d6918a923c43ff9355ff0e1efab1628c04db0ca53b388cc6145e710ba7ff361", "seconds": 15.6, "spoken_utterances": [ "Mark \"Witnessing a total solar eclipse\" in uh... my Bucket List Evernote as done. ", "Let me correct the app: use \"um... Simple Note\", not Evernote." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Evernote' but immediately corrects to 'Simple Note' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to mark the task as done in Simple Note, which matches the original instruction except for the app name. The correction is clear and final, so the task should be executed in the correct app.", "heard_transcript": "Mark witnessing a total solar eclipse in uh my bucket list Evernote is done. Let me correct the app, use um Simple Note, not Evernote." }, "candidate_id": "wav_0844e538625d93f6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then introduces a correction with 'let me correct the app'. The phrasing is slightly less natural as 'correct the app' sounds more like a meta-instruction than a spontaneous realization. The filler 'um' is used, but the overall flow feels more like a task correction than a natural thought process, and the correction is less smoothly embedded." } ] }, "is_default": false }, { "variant_id": "revision_04_completion_state", "proposed_instruction": "Mark \"Witnessing a total solar eclipse\" in my Bucket List Simple Note as todo. Change the completion state to \"done\" instead of \"todo\".", "long_pause_offsets": [ 79 ], "source_sha256": "9d03f4f72f498e4603f5f6ae95285d4d96e0be778a0446ef3ded82ea5824376b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the completion state from 'todo' to 'done', preserving the note title and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/revision/revision_04_completion_state/7b0944e9cf0e4bde/speech.wav", "wer": 0.18518518518518517, "wav_sha256": "0780b6192a7e927beafcbc66b5f7c3cf06084b6d7e62280d32f494185abe41a2", "seconds": 18.84, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in ... my Bucket List Simple Note as todo. ", "Change the completion state to \"uh, let's see... done\" instead of \"todo\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'as to do' but immediately self-corrects to 'done instead of todo'. This is a clear revision where the final intended value is 'done'. The original instruction is to mark the item as 'done', which matches the final audible intent. The revision is applied correctly, and the task is fully executable.", "heard_transcript": "Mark um witnessing a total solar eclipse in my bucket list simple note as to do. Change the completion state to uh let's see done instead of todo." }, "candidate_id": "wav_7e567476b6c9eb19", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request but then introduces a revision with 'change the completion state to uh let's see done'. The phrase 'let's see' is used as a filler but is awkwardly placed mid-correction. The correction is less fluent and sounds more like a mechanical update than a natural thought process. The prosody is slightly off, and the flow feels disjointed compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_029", "hide_variant_original": true, "source_sha256": "c4e5af0eea02ab89abf32c6e3ffb5d515f6fe7e91697662cebf204e871fcabc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/cancellation/cancel_029/0dfe1800804c820a/speech.wav", "wer": 0.04, "wav_sha256": "e351f98e7250ace8dbdf88b7927b93669dffb7a787b39cd763a6f6fb528cfabf", "seconds": 15.44, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done.", "Hang on, I'm withdrawing that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original task but then clearly states 'Hang on, I'm withdrawing that request,' which constitutes a direct cancellation of the entire task. The final audible intent is to cancel, not to execute or pause. The cancellation is unambiguous and applies to the whole task.", "heard_transcript": "Mark um witnessing a total solar eclipse in uh let me think my bucket list simple note as done. Hang on, I'm withdrawing that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_017", "hide_variant_original": true, "source_sha256": "d100f9d1347074a3eba07112545583f5dfe593faa49ce4d6c8fa655e25ed2b8f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/ambiguous_halt/halt_017/651603dbdd92a286/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "9c1aed03b7d4c098b966bd4309c5b032f37d114a03eaeda232a4c864163a7ad4", "seconds": 14.36, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done.", "One moment, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a clear request to pause with 'One moment please' following a hesitation and self-interruption ('let me think'). This constitutes an unambiguous halt request. The original instruction is not canceled or revised, and the speaker does not complete the task. The final interjection is a pause directive, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Mark um, witnessing a total solar eclipse in uh, let me think, my bucket list simple note as done. One moment please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "cf6abd2_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "2702e1d27a2c3278a1efeba3182daca7513ae35b251ee4ecf32f63ea59155728", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/cf6abd2_3/audio_natural/filler/original/e6e0255326aff7c4/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "fb03f58878bed37e413027d8de4aa4c65be32b734dcb47cddadaa0cd10c9fc6e", "seconds": 10.92, "spoken_utterances": [ "Mark um... \"Witnessing a total solar eclipse\" in uh, let me think... my Bucket List Simple Note as done." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let me think') and pauses but does not change the core task. The final audible intent is to mark the specified task as done in the specified note. The hesitation and self-interruption do not alter the goal, so the filler type criterion is satisfied.", "heard_transcript": "Mark um, witnessing a total solar eclipse in uh, let me think, my bucket list simple note as done." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my friends—no, I mean roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "8701a82923531144f4703f2395a746d8982729c5f1da2e345dbaa29133fa3285", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'friends' to 'roommates' within the same utterance, preserving the app (Venmo), action (approve), scope (all), and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/false_start/false_start_02_requester_group/650465dc45bd3fe2/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "c7ddb996c7561b8df2619dd77cef43f7638021b990ae28a973dbaf9c21a60e5d", "seconds": 13.2, "spoken_utterances": [ "Approve all venmo payment requests from my hmmm... friends—um... no, I mean ... roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'friends' but immediately self-corrects to 'roommates' with 'no, I mean'. According to the 'false_start' type criterion, the final corrected value ('roommates') should be applied, and the original incorrect value ('friends') should be discarded. The rest of the instruction matches the original. Therefore, the final intent is clear and executable.", "heard_transcript": "Approve all Venmo payment requests from my um friends um no, I mean roommates from this calendar month." }, "candidate_id": "wav_eae585138e7ca5a2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible self-correction. The speaker begins with 'my friends' and then clearly self-corrects to 'roommates' with a believable hesitation and intonation shift. The pause and filler 'um' before the correction sound spontaneous, and the overall flow is continuous and human-like. The correction is integrated smoothly into the sentence without awkwardness." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last calendar month—wait, make that this calendar month.", "long_pause_offsets": [], "source_sha256": "f4ecb08e7d6404e173081e95bf65039109a44ddd1b98ce1de611eb62332576d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'this calendar month' within the same utterance. The final intended time reference is 'this calendar month', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/false_start/false_start_03_time_reference/e93b5e2ecf680f78/speech.wav", "wer": 0.05, "wav_sha256": "c4ad999d8f8adc5be78e01def27719e6355960618875e949495c933118e3f364", "seconds": 11.52, "spoken_utterances": [ "Approve all venmo payment requests from um... my roommates from last calendar month—uh... wait, make that ... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately self-corrects to 'this calendar month' after a pause and 'wait make that'. According to the false_start type criterion, the final corrected value must be applied. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Approve all Venmo payment requests from um my roommates from last calendar month uh wait make that this calendar month." }, "candidate_id": "wav_5c4043e2cfb62c10", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start with 'last calendar month' corrected to 'this calendar month'. The speaker uses a natural 'wait' to signal the correction, and the intonation shift is credible. However, the initial phrase 'approve all Venmo payment requests from' is slightly rushed, and the correction feels a bit more abrupt than in audio_03, making it slightly less fluid overall." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Approve all PayPal—sorry, I meant venmo payment requests from my roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "d2faeb74db61e54ede66501d9a1196a92e159b2ecc9fa94e320eb264e56d05be", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the recipient group, time frame, and action. The final intended app is Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/false_start/false_start_01_app/a4420d5cce8b0a60/speech.wav", "wer": 0.2777777777777778, "wav_sha256": "59bb73fab591cd3238706a403caeedf319ae34c2be723e7f7da58094f55b06c4", "seconds": 7.88, "spoken_utterances": [ "Approve all PayPal—uh... sorry, I meant um... venmo payment requests from my roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pay' but immediately self-corrects to 'Venmo'. The final intent is clear and matches the original instruction with the correction applied. The 'false_start' type requires applying the immediate self-correction, which has been done here.", "heard_transcript": "Approve all pay uh um sorry, I meant um Venmo payment requests from my roommates from this calendar month." }, "candidate_id": "wav_58b74f8f54c47645", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with 'approve all pay' and immediately stumbles, saying 'uh um'. The correction 'sorry, I meant um Venmo payment requests' is awkward and sounds rehearsed or unnatural. The phrase 'I meant um' is an unusual way to self-correct, and the overall flow is disrupted by the stutter and unnatural phrasing, making it sound less spontaneous than the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Approve all venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month.", "long_pause_offsets": [ 54 ], "source_sha256": "973aa8196080eda7df24eb7a4ce01f67543e0cdfa0306aeadb97c8271b837d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial request; the action, app, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/addition/addition_01/9a3bd5ef2b9bc8df/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "521e5f44f879bdef3c131fba703196eea1eed0a312ab1841a84cc6f55633f8d5", "seconds": 12.92, "spoken_utterances": [ "Approve all ah... venmo payment requests from um... my roommates. ", "Oh, I forgot to mention, only approve requests from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests approval of all Venmo payment requests from their roommates. After a pause and an interjection, they add the constraint 'only approve requests from this calendar month.' This addition narrows the scope of the original request, and the final intent is to approve only those requests that fall within the current calendar month. The addition is clear and modifies the original instruction without introducing conflicting or ambiguous conditions. Therefore, the final intent matches the expected behavior for an 'addition' type task.", "heard_transcript": "Approve all uh Venmo payment requests from um my roommates. Oh, I forgot to mention, only approve requests from this calendar month." }, "candidate_id": "wav_430db484a7516350", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding instruction ('Approve all Venmo payment requests from...') and then adds a restriction ('Oh, I forgot to mention...') with a believable hesitation ('um') and a smooth, self-contained follow-up. The prosody is consistent, and the addition feels like a spontaneous afterthought rather than a rehearsed or awkward insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Approve all venmo payment requests from from this calendar month. One more detail: only approve requests from my roommates.", "long_pause_offsets": [ 66 ], "source_sha256": "95a75229e1c63353aa1c00d1d9a14abd003c7ebeee8514220a4f6ddbf0d30452", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the scope to only requests from roommates, which is consistent with the original instruction's intent to approve such requests. The original instruction already specified 'from this calendar month' and 'all venmo payment requests', so the addition refines the recipient condition without contradicting or removing any prior requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/addition/addition_02/1cfc89c8036ff2d3/speech.wav", "wer": 0.05, "wav_sha256": "b48bba7a01053bd83d1b2f9ab17be731c21a12fa1f92f74ecad0c5e4e60d37a3", "seconds": 12.28, "spoken_utterances": [ "Approve all ... venmo payment requests from from this calendar month. ", "One more detail: only approve requests from um... my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Approve all Venmo payment requests from from this calendar month.' The speaker then adds a clarifying detail: 'only approve requests from um my roommates.' This addition narrows the scope of the original instruction by specifying the source of the requests. The final intent is a combination of the original instruction and the added constraint, which matches the reference instruction. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Approve all Venmo payment requests from from this calendar month. One more detail, only approve requests from um my roommates." }, "candidate_id": "wav_661f774fa28b14a7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers the initial request with a slight stumble ('from from this calendar month'), indicating a minor self-correction or hesitation. The addition of the roommate restriction comes after a pause, but the phrasing ('One more detail') and the 'um' filler sound slightly more rehearsed or less fluid than in audio_01. The overall delivery is less natural due to the initial repetition and slightly less organic flow in the follow-up." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last month. Make that this calendar month, not last month.", "long_pause_offsets": [ 70 ], "source_sha256": "79c3fc67bcc745a71b39426144e8af1b7caf5da2d2a1c5f17649ffd226ce4b24", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last month' to 'this calendar month', preserving the app, action, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/revision/revision_03_time_reference/a62acd36f0e3cbe3/speech.wav", "wer": 0.125, "wav_sha256": "d89e6bba4c108f35a0a8c617cd32c81064a1508c6871f7a64d1f7754b75d242b", "seconds": 14.12, "spoken_utterances": [ "Approve all let me think... venmo payment requests from uh... my roommates from last month. ", "Make that ah... this calendar month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but then corrects it to 'this calendar month, not last month' after a pause. This is a revision where the final intent is the corrected version. The final goal matches the original instruction, so the revision is correctly applied.", "heard_transcript": "Approve all, let me think, Venmo payment requests from uh my roommates from last month. Make that uh this calendar month, not last month." }, "candidate_id": "wav_8071a21f1cf14e58", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), then delivers a clear initial request. The revision ('Make that... this calendar month, not last month') is a smooth, self-corrected continuation with a believable change-of-mind intonation and clear reference back to the original request. The prosody is consistent, and the correction flows naturally within the same utterance." } ] }, "is_default": true }, { "variant_id": "revision_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family from this calendar month. Sorry, I meant my roommates, not family.", "long_pause_offsets": [ 76 ], "source_sha256": "4ae98d02ac2ab0af3308486f406697a5d82b91e636f14c7e88fc2a033e0c1a5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'roommates', preserving the app, action, and time scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/revision/revision_02_requester_group/d18a214307bba999/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "ef9d020431400397d88ee5081bef3f2ed131cc4dfcf687897261b13ad2419776", "seconds": 13.0, "spoken_utterances": [ "Approve all um... venmo payment requests from let's see... my family from this calendar month. ", "Sorry, I meant my ... roommates, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately corrects themselves to 'my roommates, not family' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to approve requests from roommates, which matches the original instruction. All other conditions (Venmo, payment requests, this calendar month) remain unchanged.", "heard_transcript": "Approve all um Venmo payment requests from let's see, my family from this calendar month. Sorry, I meant my roommates, not family." }, "candidate_id": "wav_3f5ea68d6b6782e8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a filler ('um') and a brief pause, then formulates the request. The revision ('Sorry, I meant my roommates, not family') is a complete sentence that clearly corrects the earlier statement. However, the correction is delivered as a separate, slightly abrupt utterance rather than a seamless continuation, making it feel more like a post-hoc correction than a spontaneous self-repair within the same thought flow." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Approve all PayPal payment requests from my roommates from this calendar month. Wait, use Venmo instead of PayPal for the payment requests.", "long_pause_offsets": [ 80 ], "source_sha256": "ab7995084f7778725fc3f22891aab230188a1419437a44d82665e60c0100b0a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the recipient group (roommates) and time frame (this calendar month)." }, "audio": { "status": "pending", "spoken_utterances": [ "Approve all uh, let me think... PayPal payment requests from my roommates um, let me think... from this calendar month. ", "Wait, use uh... Venmo instead of PayPal for the payment requests." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_011", "hide_variant_original": true, "source_sha256": "dbb22e8962efa7355a34cbaeb828e0c8e9e17e0c8ee5155ae661762f4c16599f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/cancellation/cancel_011/31fabe3a166f4794/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "c01de2ca818583d1c6f784eb78a416f2fac39784ebdb47f8b64a616fa01d7c31", "seconds": 13.12, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month.", "Oh, wait, I've changed my mind. Cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states a complete task to approve Venmo requests, then immediately follows with 'Oh wait, I've changed my mind. Cancel that.' This constitutes a clear and final cancellation of the original task. The cancellation is not ambiguous and is not followed by any resumption or new instruction. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Oh wait, I've changed my mind. Cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_040", "hide_variant_original": true, "source_sha256": "39d7f7d93cdcef825b96b376c0f38a7bdc3a02aefd553a19b258d69dfa9708eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/ambiguous_halt/halt_040/8fa22d602f3758f9/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "b4c58aae35ef79bda21209d34cb723bed938ce70400d2699f0cdbb0b581da84f", "seconds": 11.36, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month.", "Hang on just a minute." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, just a minute,' which is a clear request to pause or wait. There is no cancellation of the task, and the speaker does not resume or modify the instruction. This matches the 'ambiguous_halt' type, where the final interjection is a pause request.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hang on, just a minute." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a74caea109a60677728002ffff93a94324ba2c05c54f3349604fec4f973085ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_1/audio_natural/filler/original/4ac76bffbbd40d96/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "20102f3ae4a3cd0e90e51570ffe9e8438ff732594e549154671df7d4df2545f5", "seconds": 7.96, "spoken_utterances": [ "Approve all Venmo payment requests from hmmm... my roommates ... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a brief hesitation ('um') but the final, complete instruction matches the original reference exactly. The hesitation is a filler and does not alter the task goal. The instruction is fully audible and unambiguous, so the final intent is to execute the approval task as stated.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family—sorry, I meant roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "b8e9c9e5fdb923e129bea27dfca9cba625588555e24557da16dde040aaf8d7db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'roommates' within the same utterance, preserving the app (Venmo), action (approve), scope (all payment requests), and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/false_start/false_start_02_requester_group/56c67bf7411eef08/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "57998f00a157245e8ecd4946e56ef9e94fd0cb6ee6388cb4fc279b8af23d39ce", "seconds": 12.2, "spoken_utterances": [ "Approve all um... venmo payment requests from ... my family—hmmm... sorry, I meant roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'roommates' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value ('roommates') should be applied, and the original incorrect value ('family') should be discarded. The rest of the instruction, including 'Venmo payment requests' and 'this calendar month', remains unchanged and correct. Therefore, the final intent matches the original instruction, and the task should be executed.", "heard_transcript": "Approve all um Venmo payment requests from my family. Sorry, I meant roommates from this calendar month." }, "candidate_id": "wav_631e4c80d93f41a5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'approve all um Venmo payment requests from my family,' then self-corrects with 'sorry, I meant roommates from this calendar month.' The correction is smooth, with a believable shift in intonation and a clear reference back to the initial request. The pacing and prosody are consistent with spontaneous speech, and the repair feels integrated rather than forced." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last month—no, that should be this calendar month.", "long_pause_offsets": [], "source_sha256": "35f088abe89b1d1999001fb53f660c4d98275585e91ae878ac0e38a0325c889d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last month' to 'this calendar month' within the same utterance, preserving the app, action, recipient group, and scope of approval." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/false_start/false_start_03_time_reference/239157b7698a8708/speech.wav", "wer": 0.0, "wav_sha256": "ba4c50f97527f429db74576cfd27d9f2af0e81ae69a04393eae31574cc836030", "seconds": 12.44, "spoken_utterances": [ "Approve all venmo payment requests from my roommates ... from last month—um... no, that should be uh... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately self-corrects to 'this calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Approve all Venmo payment requests from my roommates from last month. Um, no, that should be uh this calendar month." }, "candidate_id": "wav_547eceacaa8bab9f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with a self-correction, but the repair is less fluid. The speaker says 'from last month' and then corrects to 'this calendar month.' While the correction is audible and plausible, the transition feels slightly abrupt, and the initial phrase 'from last month' is delivered with a more hesitant, less natural cadence. The overall delivery is credible but less polished than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Approve all PayPal—I should say venmo payment requests from my roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "b7ea7c76f02aef8468424142862edd5e173b232aa9d5f158c8717ff17ab3f5fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the recipient group, time frame, and action. The final intended app is Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/false_start/false_start_01_app/a47e4c8d78820fd3/speech.wav", "wer": 0.15789473684210525, "wav_sha256": "2fa969cee371f48d42dde6435d82f8d66f77e1d1018c2eedfa873a9787bd65b1", "seconds": 11.68, "spoken_utterances": [ "Approve all uh... PayPal—hmmm... I should say venmo payment requests from my roommates uh... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say Venmo'. This is a false start where the final, corrected value is 'Venmo'. The rest of the instruction, including the target ('roommates'), action ('approve'), and constraint ('this calendar month'), remains consistent with the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "Approve all uh PayPal mm I should say Venmo payment requests from my roommates uh from this calendar month." }, "candidate_id": "wav_1b683ca7cbe9947e", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start but the repair is less convincing. The speaker says 'PayPal' and then corrects to 'Venmo,' but the correction is delivered with a flat, hesitant tone and lacks the natural flow of a spontaneous self-correction. The prosody is less varied, and the transition feels more like a robotic edit than a genuine speech error. The overall delivery is the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Approve all venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month.", "long_pause_offsets": [ 54 ], "source_sha256": "973aa8196080eda7df24eb7a4ce01f67543e0cdfa0306aeadb97c8271b837d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial request; the action, app, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/addition/addition_01/5a29339d637d50bf/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "d38ee2782a7b27d9ea5919a7985c805676b36ba6f700c61ced999f22bd6ee98d", "seconds": 14.12, "spoken_utterances": [ "Approve all let me think... venmo payment requests from my roommates. ", "Oh, I forgot to mention, only approve requests from hmmm... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Approve all Venmo payment requests from my roommates' and then adds a constraint 'only approve requests from this calendar month' after a longer pause. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is a narrowed version of the original instruction, which is the expected behavior for an addition. The original instruction is not fully met without the addition, but the addition is clearly spoken and final, so the combined goal is complete and matches the reference.", "heard_transcript": "Approve all let me think Venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month." }, "candidate_id": "wav_ac792937889f785e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('Approve all Venmo payment requests from my roommates') and then adds a restriction ('only approve requests from this calendar month') after a believable hesitation ('Oh, I forgot to mention...'). The addition is smoothly integrated with a natural pause and filler ('um'), and the correction feels spontaneous rather than forced. The prosody is consistent, and the speaker's intonation reflects genuine recollection." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Approve all venmo payment requests from from this calendar month. Oh, right, these are only from my roommates.", "long_pause_offsets": [ 66 ], "source_sha256": "1bd1e8d8db0d9e8c583c8c6e2ef49d336096217f65035a7b17ca9e59945e7b80", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction that the requests must be from roommates, preserving the app, action, and time scope from the initial utterance." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/addition/addition_02/279c99d85336beee/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "aacf7504be82d482ac62329d8f602a92e8dc93c1a87a00ad07924e6bbc2b357c", "seconds": 12.72, "spoken_utterances": [ "Approve all ... venmo payment requests from ... from this calendar month. ", "Oh, right, these are only from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Approve all Venmo payment requests from from this calendar month' and then adds 'Oh right, these are only from my roommates.' This is an addition that narrows the scope of the original instruction. The final intent is to approve all Venmo payment requests from roommates from this calendar month, which matches the original instruction with the added constraint. Therefore, the task is executed as intended.", "heard_transcript": "Approve all Venmo payment requests from from this calendar month. Oh right, these are only from my roommates." }, "candidate_id": "wav_4d733c4ae02ec707", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then adds a correction ('Oh, right, these are only from my roommates') after a short pause. However, the initial request already includes 'from this calendar month', making the correction seem like a clarification rather than a natural addition. The flow feels slightly less organic because the correction appears to rephrase the original intent rather than add a new constraint. The prosody is natural, but the pragmatic sequence is less convincing than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last month. That should be from this calendar month, not last month.", "long_pause_offsets": [ 70 ], "source_sha256": "ff9da03f1317465fffb22f8287c4b15c9458092d9303d3f4c4becaf0d7685357", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'this calendar month', while preserving the action, app, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/revision/revision_03_time_reference/97e38a18f9d68183/speech.wav", "wer": 0.0, "wav_sha256": "56297150db0d03724b625f0c20dc54e50251012be7dc54991d7c9d312d13e59d", "seconds": 14.92, "spoken_utterances": [ "Approve all venmo payment requests from uh, let's see... my roommates um... from last month. ", "That should be from this calendar month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'from last month' but immediately corrects themselves to 'from this calendar month, not last month'. This is a clear revision where the final intent is the corrected version. The original instruction matches the final audible intent, so the task is executed as revised.", "heard_transcript": "Approve all Venmo payment requests from uh let's see my roommates um from last month. That should be from this calendar month, not last month." }, "candidate_id": "wav_963f5fa10ad05e4e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, continuous flow with a believable self-correction. The initial request is clear, followed by a plausible hesitation ('let's see'), a pause, and a smooth revision from 'last month' to 'this calendar month'. The prosody is consistent, and the correction is integrated naturally into the sentence without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family from this calendar month. Make that from my roommates, not family.", "long_pause_offsets": [ 76 ], "source_sha256": "a04d7fefe02af00ba2f32ccb6c86221a5356cecb3f6e91516fdfd978ae4dec03", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'roommates' while preserving the app (Venmo), action (approve), and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/revision/revision_02_requester_group/159c9f823b40d932/speech.wav", "wer": 0.08, "wav_sha256": "ce550805a4c87238550993276c7a77a954fba7b7447e01835f335ceef907d69c", "seconds": 15.0, "spoken_utterances": [ "Approve all uh... venmo payment requests from um, let me think... my family from this calendar month. ", "Make that from my uh... roommates, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately corrects to 'my roommates, not family' after a pause. This is a clear revision where the final intent is to approve requests from roommates, matching the original instruction. The correction is self-contained and does not introduce new constraints or remove existing ones.", "heard_transcript": "Approve all uh Venmo payment requests from um let me think my family from this calendar month. Make that from my uh roommates, not family." }, "candidate_id": "wav_e2e8e3fc27859217", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a realistic revision process with a clear false start ('my family') and a self-correction ('roommates'). The speaker uses a believable filler ('let me think') and a natural-sounding correction. The prosody is slightly less fluid than audio_01, with a minor stumble on 'uh' before 'roommates', but the overall delivery remains credible and spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Approve all PayPal payment requests from my roommates from this calendar month. Wait, the app should be Venmo, not PayPal.", "long_pause_offsets": [ 80 ], "source_sha256": "d69dd9a345b099c626cc3aacabcda5d89c17f3047f424849bc53d6f3db1e5699", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the recipient group (roommates), time frame (this calendar month), and action (approve)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/revision/revision_01_app/87d8ba1a25f83b6b/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "96438b909aff91fa42b92e5b382370ddde8b92f5e581ff53ce787ad94d9dfb71", "seconds": 15.56, "spoken_utterances": [ "Approve all hmmm... PayPal payment requests from uh... my roommates from this calendar month. ", "Wait, the app should be ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects themselves to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to approve Venmo requests, which matches the original reference instruction. All other conditions (from roommates, this calendar month) are preserved.", "heard_transcript": "Approve all um, PayPal payment requests from uh, my roommates from this calendar month. Wait, the app should be Venmo, not PayPal." }, "candidate_id": "wav_5a087c29d8c96aa0", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a revision, the delivery feels less natural. The speaker starts with 'PayPal' and corrects to 'Venmo', but the correction is abrupt and lacks the smooth integration seen in the others. The 'wait' is a bit too sharp, and the prosody shifts noticeably, making the correction sound more like a mechanical edit than a spontaneous realization. The overall flow is less continuous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "f1a9d3d575bec66d315bbced99ce046cbfacd468e30e11f96a2e9ed626bfd2d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/cancellation/cancel_018/773bae80039bc85c/speech.wav", "wer": 0.23076923076923078, "wav_sha256": "6a734f32bf54c8a8a09f53ba690ba7b15d042e27db648110c3b8d046151aeeb3", "seconds": 13.88, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month.", "Hold on, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to approve Venmo requests, but then clearly revokes it by saying 'Hold on, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hold on, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "6feb41706b92c78962fc94432bdfcedd39238da16c3935f909123d5ba0d649f0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/ambiguous_halt/halt_010/0ddb3aa9134ad847/speech.wav", "wer": 0.3333333333333333, "wav_sha256": "310543f113173c096cbd9f924a90ae9e9e6673e953dc124b14d02299c077bf75", "seconds": 11.12, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month.", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Hold on a second,' which is a clear request to pause or wait. There is no cancellation of the task, nor is there any indication that the task should be resumed. The final audible intent is to halt execution temporarily, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Approve all Venmo payment requests from my roommates from this calendar month. Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b7a83e650c661eda7759521e7eae794efc16827723bafbbedc5cd5129465dcab", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_2/audio_natural/filler/original/dc47f5f5df514aa1/speech.wav", "wer": 0.42857142857142855, "wav_sha256": "163e21f52de684897b607ac98f42bfa9b6626e563d6bf520f1dfbf4ef8f751a2", "seconds": 7.64, "spoken_utterances": [ "Approve all um... Venmo payment requests from my roommates uh... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or modifications. The filler type allows ignoring hesitations and backchannels, so the task is executed as stated.", "heard_transcript": "Approve all um Venmo payment requests from my roommates uh from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Approve all PayPal—I should say venmo payment requests from my roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "b7ea7c76f02aef8468424142862edd5e173b232aa9d5f158c8717ff17ab3f5fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the recipient group, time frame, and action. The final intended app is Venmo." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/false_start/false_start_01_app/c35fadff821d3748/speech.wav", "wer": 0.0, "wav_sha256": "a161250ff704b0bda23f1887f8dd89a966bd17b60a6c9e5a0358ff11e3509393", "seconds": 8.24, "spoken_utterances": [ "Approve all PayPal—uh... I should say ... venmo payment requests from ... my roommates from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say'. This is a false start where the final, corrected value is 'Venmo'. The rest of the instruction matches the original. The correction is immediate and unambiguous, so the final intent is to approve Venmo requests, not PayPal.", "heard_transcript": "Approve all PayPal uh, I should say Venmo payment requests from my roommates from this calendar month." }, "candidate_id": "wav_d254d3ee714f62c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural false start with 'PayPal' followed by a smooth self-correction to 'Venmo'. The intonation and pacing are consistent with spontaneous speech, and the correction flows seamlessly into the rest of the sentence without awkwardness or hesitation. The prosody is continuous and the delivery sounds like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last calendar month—scratch that, make it this calendar month.", "long_pause_offsets": [], "source_sha256": "f8bb769b91c73e6fc55386f675107e15b92750df28c68a68b8369bc33e81d98b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately corrects it to 'this calendar month' using 'scratch that, make it'. The final time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/false_start/false_start_03_time_reference/5beeb60dc088a6be/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "72905e0a418a70f6f5a0bd0aee2a7403c07b237c5a610cd912a947a32dfeda1b", "seconds": 12.76, "spoken_utterances": [ "Approve all uh, let's see... venmo payment requests from my roommates from last calendar month—um... scratch that, make it uh... this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar month' but immediately self-corrects with 'scratch that, make it uh this calendar month.' This is a clear false start where the final value replaces the initial one. The final intent matches the original instruction exactly, so the task should be executed as stated.", "heard_transcript": "Approve all uh let's see Venmo payment requests from my roommates from last calendar month. Um, scratch that, make it uh this calendar month." }, "candidate_id": "wav_8af9eb3dd18f0e5a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a revision rather than a false start. The speaker initially says 'last calendar month' and then corrects to 'this calendar month'. However, the correction is less fluid—the pause before 'scratch that' and the restart with 'make it uh' sound slightly more hesitant and less integrated than the correction in audio_01. The overall delivery is still realistic but slightly less natural in its flow." } ] }, "is_default": false }, { "variant_id": "false_start_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family—no, that should be roommates from this calendar month.", "long_pause_offsets": [], "source_sha256": "36a4ff78e59c1727d1f4c07df31e1377c377e2f1b3f2efef5f8a48954a4295d9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'roommates' within the same utterance, preserving the app (Venmo), action (approve), scope (all), and time frame (this calendar month)." }, "audio": { "status": "pending", "spoken_utterances": [ "Approve all venmo payment requests from my family—uh... no, that should be um... roommates from this calendar month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Approve all venmo payment requests from from this calendar month. One more detail: only approve requests from my roommates.", "long_pause_offsets": [ 66 ], "source_sha256": "95a75229e1c63353aa1c00d1d9a14abd003c7ebeee8514220a4f6ddbf0d30452", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction that narrows the scope to only requests from roommates, which is consistent with the original instruction's intent to approve such requests. The original instruction already specified 'from this calendar month' and 'all venmo payment requests', so the addition refines the recipient condition without contradicting or removing any prior requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/addition/addition_02/9bfb1710136caeef/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "dfe09f7de9629231b84dcae965b6d5c8f3311cdc5f0e11599c515300e0c42f54", "seconds": 12.6, "spoken_utterances": [ "Approve all ... venmo payment requests from ... from this calendar month. ", "One more detail: only approve requests from uh, let me think... my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to approve all Venmo payment requests from this calendar month. Then, after a pause and a self-correction ('let me think'), they add the constraint 'Only approve requests from my roommates.' This is a clear addition that narrows the scope of the original instruction. The final intent combines the original instruction with the added constraint, matching the reference instruction exactly. Therefore, the task passes.", "heard_transcript": "Approve all Venmo payment requests from from this calendar month. One more detail. Only approve requests from uh let me think. My roommates." }, "candidate_id": "wav_25c3099e72d7d500", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Approve all venmo payment requests from my roommates. Oh, I forgot to mention, only approve requests from this calendar month.", "long_pause_offsets": [ 54 ], "source_sha256": "973aa8196080eda7df24eb7a4ce01f67543e0cdfa0306aeadb97c8271b837d93", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial request; the action, app, and recipient group remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "Approve all venmo payment requests from ... my roommates. ", "Oh, I forgot to mention, only approve requests from ... this calendar month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Approve all PayPal payment requests from my roommates from this calendar month. Wait, use Venmo instead of PayPal for the payment requests.", "long_pause_offsets": [ 80 ], "source_sha256": "ab7995084f7778725fc3f22891aab230188a1419437a44d82665e60c0100b0a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the payment app from PayPal to Venmo, while retaining the recipient group (roommates) and time frame (this calendar month)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/revision/revision_01_app/f04ce5992db6a1f2/speech.wav", "wer": 0.04, "wav_sha256": "7178dd9917f43fc21c4e8db98ec8a1611f6b7aea88c3d345b5394e8d7cfcfd85", "seconds": 13.88, "spoken_utterances": [ "Approve all PayPal payment requests from let me think... my roommates from this calendar month. ", "Wait, use ... Venmo instead of PayPal for the payment requests." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction ('Wait, use Venmo instead of PayPal'). This is a revision where the final intent is to use Venmo. The rest of the instruction, including the target ('my roommates') and time frame ('this calendar month'), remains unchanged. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Approve all PayPal payment requests from let me think my roommates from this calendar month. Wait, use Venmo instead of PayPal for the payment requests." }, "candidate_id": "wav_c43b1bbbfccd3372", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') before stating the initial request. The correction ('wait, use Venmo instead of PayPal') is delivered with a clear shift in intonation and a self-repair that sounds spontaneous. The flow from initial request to correction is smooth and credible, with no unnatural pauses or robotic delivery. The speaker's voice remains consistent and the prosody reflects genuine thought process." } ] }, "is_default": true }, { "variant_id": "revision_02_requester_group", "proposed_instruction": "Approve all venmo payment requests from my family from this calendar month. Sorry, I meant my roommates, not family.", "long_pause_offsets": [ 76 ], "source_sha256": "4ae98d02ac2ab0af3308486f406697a5d82b91e636f14c7e88fc2a033e0c1a5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'roommates', preserving the app, action, and time scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/revision/revision_02_requester_group/4764ca84e36b2b30/speech.wav", "wer": 0.2, "wav_sha256": "c41873f6b8a2d27cee3afe097827032f548dbd134d75f9bac9b48b2b1cf6c72a", "seconds": 11.12, "spoken_utterances": [ "Approve all hmmm... venmo payment requests from my family ... from this calendar month. ", "Sorry, I meant my roommates, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects to 'roommates' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to approve requests from roommates, which matches the original instruction with the corrected target. All other constraints (Venmo, payment requests, this calendar month) remain unchanged.", "heard_transcript": "Approve all Venmo payment requests from my family from this calendar month. Sorry, I meant my roommates, not family." }, "candidate_id": "wav_46c8780de1c52d2e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a filler ('huh') and a slightly hesitant delivery, which is realistic. The correction ('sorry, I meant my roommates not family') is a believable self-repair with a natural tone of realization. However, the initial request is less fluent than audio_01, and the filler 'huh' is slightly awkward. The correction is clear but the overall pacing feels a bit more disjointed compared to the smoother flow in audio_01." } ] }, "is_default": false }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Approve all venmo payment requests from my roommates from last calendar month. Make that this calendar month, not last calendar month.", "long_pause_offsets": [ 79 ], "source_sha256": "4e450a4474abf9d84eb0680a43cfcfea97d1eba0988235c152d0b78b7307d082", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last calendar month' to 'this calendar month', preserving the action, app, and recipient scope." }, "audio": { "status": "pending", "spoken_utterances": [ "Approve all venmo payment requests from ... my roommates from last calendar month. ", "Make that ... this calendar month, not last calendar month." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_038", "hide_variant_original": true, "source_sha256": "407b6770174e37a3177bf27be16c4f96f0394caa2fee7227e2e8bd3185770f6b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/cancellation/cancel_038/2bd0f382858c2c4e/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "d14aa0bca1ad336036e39f8d970701068b9ed34e0bc2e261115ee6027de07d29", "seconds": 15.32, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month.", "Oh, wait, please stop and cancel that request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to execute the original instruction but then clearly revokes it with 'Oh wait, please stop and cancel that request.' This constitutes a final, unambiguous cancellation of the task. The original instruction is irrelevant after this cancellation.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month. Oh wait, please stop and cancel that request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_026", "hide_variant_original": true, "source_sha256": "3cc44651638e916bca3d59f70891d99323a68b3e9f977b40db61b78b1333fdfc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/ambiguous_halt/halt_026/7e9569d2f3abbe52/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "d5b7781a759996ca23e33479e3a7bc4dbf1f1b9da8b1ad98e3a719653b81a096", "seconds": 14.36, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month.", "Just a moment, let me think about that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original instruction but then repeatedly hesitates ('uh', 'let me think') and ends with 'Just a moment, let me think about that,' which is a clear request to pause. The final interjection indicates a temporary halt, not a cancellation or a change in the task goal. The instruction is not completed, but the speaker's intent is to pause, not to execute or cancel. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month. Just a moment, let me think about that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d18139b_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "fbb1d5e5aac5380aff925eab5efb03d79d0d005db1f2bbb9aee745e0ca216657", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d18139b_3/audio_natural/filler/original/b45a591f5f3e3524/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5e7c93de69355e2919d7041a2e4bc7463d2796c2e8f7178aa44328be663af7fa", "seconds": 10.28, "spoken_utterances": [ "Approve all ah... Venmo payment requests from uh, let's see... my roommates let me think... from this calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let me think') and pauses but does not change the core request. The final intent matches the original instruction, with only hesitation and backchanneling. The 'filler' type allows for such hesitations as long as the final goal remains unchanged.", "heard_transcript": "Approve all uh Venmo payment requests from uh let's see, my roommates, let me think, from this calendar month." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists—I should say songs and albums I have ever liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8759de467cb3ca2bcb9ab277eaffbc0c85b65411dd68822377ac1186fd9abeef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but immediately corrects to 'songs and albums'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/false_start/false_start_01_media_scope/7661328389d659f0/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "7939369532d83d1d7a9d241323c99cd4c68938c0f02771c9ea8575f81a2c590f", "seconds": 9.52, "spoken_utterances": [ "Follow artists of all the uh... tracks and playlists—um... I should say songs and albums I have ever liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlist them', which is a false start. They immediately self-correct to 'songs and albums I have ever liked on Spotify'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistake should be discarded. The final intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "Follow artists of all the uh tracks and playlist them. I should say, songs and albums I have ever liked on Spotify." }, "candidate_id": "wav_533d1405a50a4a59", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and fluent self-correction. The initial false start 'uh tracks and playlists' is immediately and smoothly repaired to 'songs and albums' with a believable 'I should say' transition. The prosody is consistent, the pace is conversational, and the correction feels spontaneous rather than rehearsed. The speaker also correctly specifies 'on Spotify' without further revision, maintaining task clarity." } ] }, "is_default": true }, { "variant_id": "false_start_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved—no, I meant liked on Spotify.", "long_pause_offsets": [], "source_sha256": "0663ddb00d150b1e43ec6c54028a8f69170dd139ded58805a7a8b9b9e6de42a6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended action is to follow artists of songs and albums the user has liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/false_start/false_start_02_interaction_type/671f569267896264/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "025fc74969f31c230d6fdb7063ca088a6505f4aecd9bd1af72e7203b14124843", "seconds": 9.08, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever saved—hmmm... no, I meant liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' with 'No, I meant liked on Spotify.' This is a false start where the final audible intent is the corrected version. The correction is clear and immediate, and the final instruction matches the original reference. Therefore, the task should be executed as per the corrected intent.", "heard_transcript": "Follow artists of all the songs and albums I have ever saved. No, I meant liked on Spotify." }, "candidate_id": "wav_9fe68e9ca9b19b18", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start ('saved') and a self-correction ('liked on Spotify'), but the repair is less fluid than audio_03. The speaker says 'no I meant liked on Spotify' which is a slightly more abrupt and less natural phrasing than 'I should say'. The prosody is otherwise consistent, and the correction is audible and plausible, but the transition feels a bit more mechanical compared to the smoother flow in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music—sorry, that should be Spotify.", "long_pause_offsets": [], "source_sha256": "08a7c2656af1a18feb4f23c61d2bb406671659b866c963b3213ea4ffea4ffd17", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects the music service from Apple Music to Spotify, preserving the action, scope, and target entity." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/false_start/false_start_03_app/fd26b080227635d0/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "5acf4881f0419ef4a4adf1c561c0251ceebe2e0e39c73928957ffd71a968770b", "seconds": 8.68, "spoken_utterances": [ "Follow artists of all the songs and albums I have ever liked on Apple Music—uh... sorry, that should be ah... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference exactly.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Apple Music. Uh, sorry, that should be uh Spotify." }, "candidate_id": "wav_af1cb422105bcc1f", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a false start ('Apple Music') and corrects it to 'Spotify', but the correction is less convincing. The phrase 'uh sorry that should be uh Spotify' includes two filler words and a hesitant 'uh' before the correction, making the delivery sound less confident and more like a rehearsed or awkward correction. The prosody is slightly off, and the overall flow is less natural compared to the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Follow artists of I have ever liked on Spotify. Oh, I forgot to mention, this should apply only to all the songs and albums.", "long_pause_offsets": [ 48 ], "source_sha256": "846cee9956c9197a779aad28d711499de5e6e79c355fcebea05387de65d3f6ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the action should apply only to 'all the songs and albums', which refines the initial ambiguous phrase 'artists of I have ever liked'. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/addition/addition_01/9c6b86f174b379d0/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "3108cd8f41699cb453edbe0b9b31d0540c4027381e5123601a76ac16cca96b06", "seconds": 14.04, "spoken_utterances": [ "Follow artists of um... I have ever liked on uh... Spotify. ", "Oh, I forgot to mention, this should apply only to um, let's see... all the songs and albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is incomplete ('Follow artists of um, I have ever liked on uh Spotify'), but the speaker adds the missing constraint 'all the songs and albums' after a pause. This addition completes the original instruction, making the final goal match the reference. The type is 'addition', and the final combined task matches the original instruction exactly.", "heard_transcript": "Follow artists of um, I have ever liked on uh Spotify. Oh, I forgot to mention, this should apply only to um, let's see, all the songs and albums." }, "candidate_id": "wav_75da393e2c53e3d3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous flow with a believable addition. The initial request is stated with slight hesitation ('um'), followed by a pause and a plausible, smoothly integrated restriction ('Oh, I forgot to mention...'). The prosody is consistent, and the correction feels spontaneous rather than forced. The pacing and intonation suggest genuine thought processing." } ] }, "is_default": true }, { "variant_id": "addition_03", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on. And make sure this is on Spotify.", "long_pause_offsets": [ 65 ], "source_sha256": "9a1f4bac5984b082d7f96f185bfeafc5d80d46612e5869d00c03e7faf6f12644", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'on Spotify', which was implied but not stated in the first utterance. The core action and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/addition/addition_03/b2da4f5e1a436845/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "5b3cbc33138dada3bbbc95a92e408c95ac8beb738eb8a9afee7417d5aa99c58c", "seconds": 11.92, "spoken_utterances": [ "Follow artists of um... all the songs and albums uh... I have ever liked on. ", "And make sure this is on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow artists of all the songs and albums I have ever liked on', which is incomplete. The speaker then adds 'and make sure this is on Spotify' after a pause, completing the task with the required platform. This matches the addition type criterion: the final goal combines the initial instruction with the added constraint. The final intent is clear and executable.", "heard_transcript": "Follow artists of um all the songs and albums uh I have ever liked on and make sure this is on Spotify." }, "candidate_id": "wav_e761ae215ec5f1e3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The delivery is mostly fluent but slightly less natural than audio_01. The speaker starts with a clear request, then adds a clarifying clause ('and make sure this is on Spotify') after a longer pause. While the addition is plausible, the second part sounds more like a rehearsed or read instruction than a spontaneous thought. The prosody is slightly flat, and the transition between clauses is less organic." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Follow artists of all the songs and albums on Spotify. I didn't mention this, but only follow artists of songs and albums I have ever liked.", "long_pause_offsets": [ 55 ], "source_sha256": "99af53a8d4796ca49a46d8f0ffb91343131252a38bac9cecb456f6c5dfd4b3d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'I have ever liked', which narrows the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/addition/addition_02/1e714a8196c7ad94/speech.wav", "wer": 0.1, "wav_sha256": "7b2727a4fef3d32dd0a5a4ba759eb548b0b6c58d847ccd9edf87e0de2bcf4b82", "seconds": 15.8, "spoken_utterances": [ "Follow artists of let's see... all the songs and albums on uh... Spotify. ", "I didn't mention this, but only follow artists of songs and albums um... I have ever liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow artists of all the songs and albums on Spotify.' This is followed by a longer pause and an addition: 'I didn't mention this, but only follow artists of songs and albums um I have ever liked.' This addition narrows the scope of the original instruction by specifying that only artists of songs and albums the user has ever liked should be followed. The final intent combines the original request with the added constraint, matching the reference instruction. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Follow artists of let's see, all the songs and albums on uh Spotify. I didn't mention this, but only follow artists of songs and albums um I have ever liked." }, "candidate_id": "wav_97ac70adaa5509a2", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a slightly awkward phrasing ('artists of let's see...') and then delivers the main request. The addition at the end is abrupt and sounds like a pasted-on afterthought rather than a natural recollection. The intonation is inconsistent, and the final clause lacks the natural flow of a spontaneous correction. The overall delivery feels less cohesive and more like a fragmented thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists I have ever liked on Spotify. Change the media scope to 'songs and albums' instead of 'tracks and playlists'.", "long_pause_offsets": [ 77 ], "source_sha256": "de075b75bc2ce0dddb63a34ceead7edb74302fae3a2ddd597f7e23ec392e9573", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the media scope from 'tracks and playlists' to 'songs and albums', aligning with the original instruction. The action and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/revision/revision_01_media_scope/5bb2d20f3667646a/speech.wav", "wer": 0.12121212121212122, "wav_sha256": "afde0a8540fbf0eb10f0e50c5bc3a1027c5547dcedb40d5a5d7e1993d7fd6bd0", "seconds": 17.0, "spoken_utterances": [ "Follow artists of ah... all the tracks and playlists uh, let me think... I have ever liked on Spotify. ", "Change the media scope to 'um... songs and albums' instead of 'tracks and playlists'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but then corrects to 'songs and albums' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent matches the original reference instruction, so the task passes.", "heard_transcript": "Follow artists of uh all the tracks and playlists uh let me think. I have ever liked on Spotify. Change the media scope to um songs and albums instead of tracks and playlists." }, "candidate_id": "wav_9706b1aa911bc8e3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses a realistic 'uh' and 'let me think' filler, followed by a clear, self-corrected revision ('change the media scope to...') that sounds like a spontaneous thought process. The prosody is consistent, and the correction is smoothly integrated into the ongoing utterance without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "revision_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "4318ed46b578d6e45566504ba8a7a1abe49a3bee04ce61cb297190d47a4bacbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the rest of the instruction including the scope of 'all the songs and albums I have ever liked'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/revision/revision_03_app/49cf4bc86e7f0d54/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "1668ae4bb8f2da5483ad91e9372e86f9b50749c13c937f8fd6514fcba593300b", "seconds": 12.8, "spoken_utterances": [ "Follow artists of um... all the songs and albums let me think... I have ever liked on uh... Apple Music. ", "Wait, the app should be Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects the target app to Spotify. Since this is a revision with a clear self-correction, the final intent is to execute the task on Spotify, matching the original instruction's goal. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Follow artists of um all the songs and albums let me think I have ever liked on uh Apple Music. Wait, the app should be Spotify not Apple Music." }, "candidate_id": "wav_e85bf91df74a3504", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable self-correction ('wait, the app should be Spotify not Apple Music'), but the initial phrasing ('follow artists of em all the sons and albums') contains a noticeable mispronunciation ('sons' instead of 'songs') and a slightly unnatural 'let me think' pause. The correction is clear and plausible, but the overall delivery is less fluent than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved on Spotify. I misspoke about the interaction: it should be 'liked', not 'saved'.", "long_pause_offsets": [ 73 ], "source_sha256": "e1fd81c9595b2ac0558a061a1d8f012718f1027846aef35bff0ecc3fca5fc0f4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the interaction type from 'saved' to 'liked', while retaining the rest of the instruction including the app, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/revision/revision_02_interaction_type/1cda2317c5fd78c7/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "9584bc56c5cfeacd37a5118545f8efd5788d6dd0d083cbb1433ed8df4c635240", "seconds": 16.52, "spoken_utterances": [ "Follow artists of all the songs and albums hmmm... I have ever saved on um... Spotify. ", "I misspoke about the interaction: it should be 'let's see... liked', not 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves to 'liked' after a pause, indicating a revision. The final intent is to follow artists of liked songs and albums, which matches the original instruction. The correction is clear and the final goal is executable.", "heard_transcript": "Follow artists of all the songs and albums I have ever saved on Spotify. I misspoke about the interaction. It should be liked, not saved." }, "candidate_id": "wav_1b8f7a84f7ba4ba8", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a flat, robotic delivery with unnatural pauses and a lack of natural speech rhythm. The speaker says 'I have ever saved' which is semantically odd, and the correction ('I misspoke about the interaction...') is delivered in a rehearsed, explanatory tone rather than a spontaneous self-repair. The prosody is monotonous, and the overall impression is less like a real-time thought process and more like a scripted correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_009", "hide_variant_original": true, "source_sha256": "e166825667fc670d3b0c125f4b26a4cb54e277fd2e400304735f8154b403c324", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/cancellation/cancel_009/9815681b3ce210d6/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "8af6acdd2f76ca3296ac665843534a33c3b3234d879ac6fbf68ff1cb7c5a2ea2", "seconds": 12.56, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify.", "Hang on, cancel that altogether." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but immediately follows with 'Hang on, cancel that altogether,' which is a clear and unambiguous cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify. Hang on, cancel that altogether." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_030", "hide_variant_original": true, "source_sha256": "560f71b252045b27376c661480bbf2466658d12ea1b247d2f15e87fca9dd8171", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/ambiguous_halt/halt_030/e70ca37c4b97c33c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "730a17513c69eeb929fe959156b7ab6e58048166849d9180673860e13d649443", "seconds": 11.52, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify.", "Hold on, actually..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, actually...', which is a clear request to pause and reconsider the task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or execution. The initial request is not completed, but the final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify. Hold on, actually..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4477e664b62b203b430e648c733322275d027a513b9a99ba971bc1c0bb9bcdff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_1/audio_natural/filler/original/2161fbf520479464/speech.wav", "wer": 0.125, "wav_sha256": "0f5c8ea7d92fa9c8862f424c69a23edfa32d984e3a8d586f6266c65662f9760b", "seconds": 8.28, "spoken_utterances": [ "Follow artists of uh... all the songs and albums ... I have ever liked on hmmm... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'uh' as a filler, then proceeds to deliver the full original instruction without any correction, cancellation, or pause. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "Follow artists of uh all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved—sorry, I meant liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8a6078a6c0167c65d7136f1b45233af212bf6f365e0f462344ee7b8074cdd287", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked'. The final intended action is to follow artists of songs and albums the user has liked, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/false_start/false_start_02_interaction_type/4fa6b6022e00d842/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "5b4ecdd1f2036351ed49e27da76f68b6841de5d81e51a0b5b5ab58eae1d3b3bc", "seconds": 10.88, "spoken_utterances": [ "Follow artists of ... all the songs and albums I have ever saved—hmmm... sorry, I meant ... liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked' after a brief pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('saved') is discarded. The final instruction matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "Follow artists of all the songs and albums I have ever saved. Hmm. Sorry, I meant liked on Spotify." }, "candidate_id": "wav_4d4aa5057661a8d4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'saved' to 'liked' with a plausible 'sorry I meant' phrase. The correction is smoothly integrated, and the prosody remains consistent throughout. The pause before the correction is brief and realistic, suggesting genuine hesitation rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "false_start_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists—I should say songs and albums I have ever liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8759de467cb3ca2bcb9ab277eaffbc0c85b65411dd68822377ac1186fd9abeef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but immediately corrects to 'songs and albums'. The final intended goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/false_start/false_start_01_media_scope/5e538ba927bd7995/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "b7a10de25eb11bb8ecd7168d4d21603da20491772a84228bdb2c00d863da6a58", "seconds": 10.8, "spoken_utterances": [ "Follow artists of all the tracks and playlists—um... I should say songs and albums hmmm... I have ever liked on um... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlist some', which is a false start. They immediately self-correct to 'songs and albums'. The final intent matches the original instruction, with only a minor grammatical variation ('playlist some' to 'albums') that does not change the meaning. The correction is immediate and clear, satisfying the false_start type criterion.", "heard_transcript": "Follow artists of all the tracks and playlist some. I should say songs and albums. Hmm. I have ever liked on um Spotify." }, "candidate_id": "wav_9a275548416b8982", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also self-corrects from 'tracks and playlists' to 'songs and albums', the correction is less fluid. The phrase 'I should say' is slightly less natural than 'sorry I meant', and the overall delivery feels more hesitant and less confident. The pause before the correction is longer and more awkward, and the intonation during the correction lacks the smoothness of audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music—no, that should be Spotify.", "long_pause_offsets": [], "source_sha256": "c15754550252915e863bd5fd5be8f9dc269f17eb1219add1bdcf8568a9e1b0a4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The action, scope (all liked songs and albums), and target (artists) remain consistent with the original instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Apple Music—uh... no, that should be Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Follow artists of all the songs and albums on Spotify. I didn't mention this, but only follow artists of songs and albums I have ever liked.", "long_pause_offsets": [ 55 ], "source_sha256": "99af53a8d4796ca49a46d8f0ffb91343131252a38bac9cecb456f6c5dfd4b3d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'I have ever liked', which narrows the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/addition/addition_02/405e8c8bc9c04db3/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "22efba7153651ff273886d1a993cd6b1f21c6a94daf86ce4f943b788c50b9d4c", "seconds": 12.76, "spoken_utterances": [ "Follow artists of ah... all the songs and albums on ... Spotify. ", "I didn't mention this, but only follow artists of songs and albums ... I have ever liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Follow artists of all the songs and albums on Spotify.' This is followed by a longer pause and an addition: 'I didn't mention this, but only follow artists of songs and albums I have ever liked.' This addition narrows the scope of the original request by specifying that only artists of songs and albums the user has ever liked should be followed. The final intent is a combination of the original instruction and the added constraint, which matches the reference instruction. Therefore, the task is executed as intended.", "heard_transcript": "Follow artists of uh all the songs and albums on Spotify. I didn't mention this, but only follow artists of songs and albums I have ever liked." }, "candidate_id": "wav_968c16b2a58f708b", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Follow artists of I have ever liked on Spotify. Oh, I forgot to mention, this should apply only to all the songs and albums.", "long_pause_offsets": [ 48 ], "source_sha256": "846cee9956c9197a779aad28d711499de5e6e79c355fcebea05387de65d3f6ba", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup clarifies that the action should apply only to 'all the songs and albums', which refines the initial ambiguous phrase 'artists of I have ever liked'. The final goal matches the original instruction exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow artists of um... I have ever liked on uh... Spotify. ", "Oh, I forgot to mention, this should apply only to um... all the songs and albums." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved on Spotify. Wait, I meant 'liked' rather than 'saved'.", "long_pause_offsets": [ 73 ], "source_sha256": "8632071240fa4ba1cd7330bc6c75e1932ce3961a8b09c8df7d6db7d7c107005a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'saved' to 'liked', preserving the app, scope (all songs and albums), and action (follow artists)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/revision/revision_02_interaction_type/df0940570eee8af1/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "37d9af1581194a3f0198620d5b01bc139d5aa61ad6e0315636453690bd4d0dc9", "seconds": 11.4, "spoken_utterances": [ "Follow artists of ... all the songs and albums uh... I have ever saved on Spotify. ", "Wait, I meant '... liked' rather than 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as intended.", "heard_transcript": "Follow artists of all the songs and albums uh I have ever saved on Spotify. Wait, I meant liked rather than saved." }, "candidate_id": "wav_5fb57da87060aeb6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural hesitation ('uh') and a self-repair ('Wait, I meant...') that sounds spontaneous and integrated into the flow. The correction is immediate and the intonation reflects a genuine moment of recollection. The pacing is conversational, and the overall delivery is fluent and continuous without robotic pauses or distortions." } ] }, "is_default": true }, { "variant_id": "revision_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists I have ever liked on Spotify. Change the media scope to 'songs and albums' instead of 'tracks and playlists'.", "long_pause_offsets": [ 77 ], "source_sha256": "de075b75bc2ce0dddb63a34ceead7edb74302fae3a2ddd597f7e23ec392e9573", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the media scope from 'tracks and playlists' to 'songs and albums', aligning with the original instruction. The action and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/revision/revision_01_media_scope/6f979657fa9ed35b/speech.wav", "wer": 0.11428571428571428, "wav_sha256": "0b71e87fe3d51da44349c50ae34211666b883b1c253be698b38fd9b5366b5873", "seconds": 18.48, "spoken_utterances": [ "Follow artists of all the tracks and playlists um... I have ever liked on um, let's see... Spotify. ", "Change the media scope to 'uh, let me think... songs and albums' instead of 'tracks and playlists'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but then corrects to 'songs and albums' after a pause and self-correction ('let me think'). According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original instruction, so the task is executed as intended.", "heard_transcript": "Follow artists of all the tracks and playlists um I have ever liked on um but see Spotify. Change the media scope to uh let me think songs and albums instead of tracks and playlists." }, "candidate_id": "wav_ee8209eb7e9749a6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker attempts a revision, the delivery is less natural due to multiple fillers ('um', 'uh'), a longer hesitation ('let me think'), and a more disjointed structure. The phrase 'Change the media scope to...' sounds like a separate thought rather than a smooth continuation of the original request, making the revision feel less integrated. The prosody is less fluid, and the speaker seems to be formulating the correction more slowly, reducing the sense of spontaneity." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music. Let me correct the app: follow artists on Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "45b4f01a2c0b278f180d9d2e94d6d1772d1c9def6fae2449a18b02d6db2a5ade", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining the original scope of artists from liked songs and albums." }, "audio": { "status": "pending", "spoken_utterances": [ "Follow artists of all the songs and albums ... I have ever liked on ... Apple Music. ", "Let me correct the app: follow artists on um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_017", "hide_variant_original": true, "source_sha256": "bf262404db2655c530c6d83a703afdda5a457b7078782b7d956bd8dc29b94d38", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/cancellation/cancel_017/7490e09f62877755/speech.wav", "wer": 0.0, "wav_sha256": "7bb4e8929221b040c4d1fffd1d1aac73476b27387d630603bfba117bb759a8fc", "seconds": 10.32, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify.", "Wait, I don't want that done anymore. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request, then clearly revokes it by stating 'Wait, I don't want that done anymore. Cancel it.' This constitutes a direct cancellation of the task. The final control intent is to cancel, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify. Wait, I don't want that done anymore. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "1c111b8f68666f8433ab2b4db6f49922acbede966415ca9dece2ceca0a56b7f1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/ambiguous_halt/halt_024/41dc5bf6c17b550a/speech.wav", "wer": 0.0, "wav_sha256": "6ee230b1bb179b92f9a2050e354fbdc2f16cfafa4865f3f73fb0885d2c17be93", "seconds": 9.4, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify.", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a moment to think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause directive. The original instruction is not canceled or revised, and the speaker does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify. Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e59887305786f4274ca9bde3d3898c173efda0744eaa633ff1e85870014e7ad3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_2/audio_natural/filler/original/abfbc219f8463c5e/speech.wav", "wer": 0.0, "wav_sha256": "c47f3c64b11951efcbca60f55a15486d832c4296e3e3d35ae8d861092aaa70da", "seconds": 5.68, "spoken_utterances": [ "Follow artists of ... all the songs and albums ... I have ever liked on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction exactly, and the filler type allows for minor hesitations or pauses that do not change the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved—I meant liked on Spotify.", "long_pause_offsets": [], "source_sha256": "c943377f239c95be0a4ca7cc2a5f82c61291cb2c65a7b7cc5ba675ebd35417bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects to 'liked' within the same utterance. The final intended action is to follow artists of liked content, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/false_start/false_start_02_interaction_type/8974bb6b5ac391a9/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "98e97ab4ac626df71364a59cbd8f818f462b6ff4e2c25f94149023b4b244526f", "seconds": 13.12, "spoken_utterances": [ "Follow artists of uh... all the songs and albums uh, let's see... I have ever saved—uh... I meant liked on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately self-corrects to 'liked'. According to the 'false_start' type criterion, the final corrected value ('liked') should be applied, and the original incorrect value ('saved') should be discarded. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "Follow artists of uh all the songs and albums uh let's see, I have ever saved uh, I meant liked on Spotify." }, "candidate_id": "wav_e584d1d80bbade13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker begins with a clear request, uses a plausible filler ('uh') to pause, then self-corrects 'saved' to 'liked' with a smooth, spontaneous-sounding repair. The intonation and rhythm are consistent with someone formulating a thought aloud, and the correction feels integrated into the sentence rather than tacked on. The pacing and prosody are very realistic." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music—make that Spotify.", "long_pause_offsets": [], "source_sha256": "ae40ad62e56a177fabe924eab004897c01dc5f4b0e481bab840e8cd86eb9a2a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The action, scope (all liked songs and albums), and target (artists) remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/false_start/false_start_03_app/b4ef92abf4c4d0d2/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "2bcb73f62d9dd76b51303913d4f715ad9f687d66500ffa39b4983cfe38812142", "seconds": 10.0, "spoken_utterances": [ "Follow artists of all the songs and albums let's see... I have ever liked on Apple Music—um... make that Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction with the correct platform.", "heard_transcript": "Follow artists of all the songs and albums let's see I have ever liked on Apple Music um make that Spotify." }, "candidate_id": "wav_3d7963cbae3a74b7", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip is realistic but slightly less so than audio_03. The speaker starts with a clear request, hesitates with 'let's see', then corrects 'Apple Music' to 'Spotify'. The correction is natural, but the initial phrase 'follow artists of all the songs and albums' is delivered in a slightly more robotic, read-aloud manner compared to audio_03. The pause and filler are plausible, but the overall delivery lacks the same level of conversational fluidity." } ] }, "is_default": false }, { "variant_id": "false_start_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists—I should have said songs and albums I have ever liked on Spotify.", "long_pause_offsets": [], "source_sha256": "8e0f9e9914c0a9dcb1b278031055cd2bb387b7873b71d3fd0530b34da8ff12da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker explicitly corrects 'tracks and playlists' to 'songs and albums I have ever liked' within the same utterance, indicating the intended scope is the original instruction. The final goal matches the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/false_start/false_start_01_media_scope/f61779a2368d8860/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "b9d599e7b80a9df54a6662ff01c50248116c3dfd7be757aac1692191188d3e41", "seconds": 13.6, "spoken_utterances": [ "Follow artists of all the um... tracks and playlists—hmmm... I should have said songs and albums I have ever liked on uh, let me think... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a false start, saying 'tracks and playlists' instead of 'songs and albums'. After a self-correction, they clearly state the intended goal: 'I should have said songs and albums I have ever liked on Spotify.' This matches the original instruction exactly. The hesitation and initial mistake are corrected, and the final intent is clear and executable.", "heard_transcript": "Follow artists of all the um tracks and playlists. Um, I should have said songs and albums I have ever liked on uh, let me think, Spotify." }, "candidate_id": "wav_18f912cbc78c4ddf", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the least realistic delivery. The speaker starts with a clear request but then uses multiple fillers ('um', 'mm', 'uh') and a long hesitation ('let me think') that disrupts the flow. The correction is also less smooth, with a noticeable break between the initial request and the revision. The overall prosody is more fragmented and less continuous, making it sound less like spontaneous speech and more like a rehearsed or disjointed attempt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Follow artists of all the songs and albums on Spotify. I didn't mention this, but it's only for the ones I have ever liked.", "long_pause_offsets": [ 55 ], "source_sha256": "1b7756f0719335ac203c257aeefb890e1c084d63125a96b9128ac2cdb4ca2e8d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'I have ever liked', which narrows the scope of the initial broad request to only those songs and albums that were liked. The action, app, and target remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/addition/addition_02/5be4112ecb5282c1/speech.wav", "wer": 0.0625, "wav_sha256": "6fe2cdc1ada3678ebe976308befd73fff8485065578915bbbf71b6e3c1b6fd4b", "seconds": 17.52, "spoken_utterances": [ "Follow artists of um, let's see... all the songs and albums on ah... Spotify. ", "I didn't mention this, but it's only for the ones uh, let me think... I have ever liked." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially gives the instruction to follow artists of all songs and albums on Spotify. After a pause and a self-correction, they add the constraint 'but it's only for the ones I have ever liked'. This is a clear addition that narrows the scope of the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Follow artists of um, let's see, all the songs and albums on uh, Spotify. I didn't mention this, but it's only for the ones uh, let me think, I have ever liked." }, "candidate_id": "wav_c1e3816d5729540e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and varied prosody, with believable pauses and fillers ('um', 'let me think') that align with spontaneous speech. The addition of 'I didn't mention this, but it's only for the ones...' is a clear, pragmatic follow-up that sounds like a remembered detail being added mid-thought. The intonation and rhythm feel organic, not rehearsed or robotic." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Follow artists of I have ever liked on Spotify. Oh, and one more thing: only follow artists of all the songs and albums.", "long_pause_offsets": [ 48 ], "source_sha256": "37759619e96e7fb03b5e5f8957f19c8fe05d9a937faee9b55482e29fcff8f88f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing phrase 'all the songs and albums' to clarify the scope of 'artists of I have ever liked', resulting in the complete and intended goal matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/addition/addition_01/00330766f4fbf8a3/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "ada13d517daffd7d72cef32fcfb19d0922ec1b32b5822ede28729574df7f6e56", "seconds": 15.4, "spoken_utterances": [ "Follow artists of let's see... I have ever liked on uh... Spotify. ", "Oh, and one more thing: only follow artists of all the songs and albums." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the original instruction, then adds a clarifying constraint ('only follow artists of all the songs and albums') after a pause and interjection. This matches the 'addition' type, where the final goal combines the initial request with the added information. The final intent is clear and executable, and the addition refines the scope without contradicting the original.", "heard_transcript": "Follow artists of let's see, I have ever liked on uh Spotify. Oh, and one more thing, only follow artists of all the songs and albums" }, "candidate_id": "wav_c807a12af702acc5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a natural flow with a clear addition ('Oh, and one more thing...') that sounds like a remembered restriction. The speaker uses fillers ('uh') and has a conversational rhythm. However, the initial phrasing 'Follow artists of let's see' is slightly less fluid than audio_01, and the addition is delivered with a slightly more abrupt transition, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on. Just to specify, do this on Spotify.", "long_pause_offsets": [ 65 ], "source_sha256": "683aaf2c28619e21c16a4c7fcba1cd5794642be7f5cd342e8bcffff7b2181371", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup specifies the platform as Spotify, which was implied but not explicitly stated in the initial utterance. The core action and scope remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/addition/addition_03/c525e16e05507947/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "0406a4452425d830c67038965a915129e6068ae7a1346df3b42df57d4837109a", "seconds": 13.4, "spoken_utterances": [ "Follow artists of ... all the songs and albums uh... I have ever liked on. ", "Just to specify, do this on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the reference. The speaker then adds 'just to specify, do this on Spotify' as a clarifying addition, which is consistent with the 'addition' type. The final intent is to execute the original task on Spotify, which aligns with the reference and the addition. No contradictions or errors are present in the final audible intent.", "heard_transcript": "Follow artists of all the songs and albums I have ever liked on, just to specify, do this on Spotify." }, "candidate_id": "wav_cfa34b9cd4d16b92", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker's delivery is the most robotic and flat. The prosody lacks natural variation, and the phrasing is very clipped and rehearsed. The addition ('just to specify, do this on Spotify') is delivered without the typical intonation or hesitation of a spontaneous correction. The overall impression is less like a person thinking aloud and more like a pre-written prompt being read." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "Follow artists of all the songs and albums I have ever liked on Apple Music. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 77 ], "source_sha256": "4318ed46b578d6e45566504ba8a7a1abe49a3bee04ce61cb297190d47a4bacbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Apple Music to Spotify, while retaining the rest of the instruction including the scope of 'all the songs and albums I have ever liked'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/revision/revision_03_app/9add0b9e221a07c3/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "efe9474c6310c17edbdd311f36fbe676dec56b6401faaed9061ee6d9103f5cc4", "seconds": 17.08, "spoken_utterances": [ "Follow artists of all the songs and albums let me think... I have ever liked on uh... Apple Music. ", "Wait, the app should be hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects themselves to 'Spotify' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to follow artists on Spotify, which matches the original instruction. The correction is clear and unambiguous, so the task should be executed as revised.", "heard_transcript": "Follow artists of all the songs and albums that me think I have ever liked on uh Apple Music. Wait, the app should be hmm Spotify, not Apple Music." }, "candidate_id": "wav_3aa2feefb4cb52ca", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous revision flow. The speaker begins with a clear request, hesitates with 'let me think', then corrects the platform from 'Apple Music' to 'Spotify' with a believable 'wait' and a self-repair that sounds like a genuine moment of recollection. The prosody is consistent, and the correction is integrated smoothly into the utterance without awkwardness." } ] }, "is_default": true }, { "variant_id": "revision_02_interaction_type", "proposed_instruction": "Follow artists of all the songs and albums I have ever saved on Spotify. I misspoke about the interaction: it should be 'liked', not 'saved'.", "long_pause_offsets": [ 73 ], "source_sha256": "e1fd81c9595b2ac0558a061a1d8f012718f1027846aef35bff0ecc3fca5fc0f4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the interaction type from 'saved' to 'liked', while retaining the rest of the instruction including the app, scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/revision/revision_02_interaction_type/a688bd41089bde7f/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "978a980e7bdd8f4d4331b233e7aa96ac856b956831427a77846dfc39bd545aeb", "seconds": 19.48, "spoken_utterances": [ "Follow artists of ... all the songs and albums um, let me think... I have ever saved on hmmm... Spotify. ", "I misspoke about the interaction: it should be 'liked', not 'saved'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'saved' but immediately corrects themselves with 'I misspoke about the interaction, it should be liked not saved.' This is a clear revision where the final intent is to follow artists of liked content, not saved content. The correction is spoken after a pause and directly addresses the error in the initial request, satisfying the revision type criterion.", "heard_transcript": "Follow artists of all the songs and albums um let me think. I have ever saved on whom. Spotify. I misspoke about the interaction, it should be liked not saved." }, "candidate_id": "wav_ac588d27a0c3af3a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a plausible false start and self-correction. The speaker initially says 'saved' but then corrects to 'liked' with a natural 'I misspoke' followed by a self-repair. The hesitation and correction are believable, though the 'let me think' pause feels slightly more rehearsed than in audio_03. The overall delivery is fluent and realistic." } ] }, "is_default": false }, { "variant_id": "revision_01_media_scope", "proposed_instruction": "Follow artists of all the tracks and playlists I have ever liked on Spotify. Change the media scope to 'songs and albums' instead of 'tracks and playlists'.", "long_pause_offsets": [ 77 ], "source_sha256": "de075b75bc2ce0dddb63a34ceead7edb74302fae3a2ddd597f7e23ec392e9573", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the media scope from 'tracks and playlists' to 'songs and albums', aligning with the original instruction. The action and recipient remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/revision/revision_01_media_scope/807d41fdea181974/speech.wav", "wer": 0.1, "wav_sha256": "37dd0b20a5e179d7cd1d4ab5b2082113a5d00302e5103b473a74c98b0d194fcc", "seconds": 18.6, "spoken_utterances": [ "Follow artists of hmmm... all the tracks and playlists uh... I have ever liked on Spotify. ", "Change the media scope to 'ah... songs and albums' instead of 'tracks and playlists'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'tracks and playlists' but immediately corrects to 'songs and albums' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to follow artists of songs and albums, which matches the original instruction's intent but with the correct media scope.", "heard_transcript": "Follow artists of all the tracks and playlists uh I have ever liked on Spotify. Change the media scope to uh songs and albums instead of tracks and playlists." }, "candidate_id": "wav_6a500b221c26ca76", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but feels less natural due to its structure. The speaker first makes a request, then pauses and explicitly states a change to the media scope. The phrasing 'change the media scope to...' sounds more like a meta-instruction than a spontaneous thought process. The delivery is clear but lacks the organic flow of the other two, making it sound more like a scripted or rehearsed correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "9b7df41f219a3471f9618ba995ce645edc28a6227e39259d6d2373a6da559ee5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/cancellation/cancel_028/2569f13bfd06f5cc/speech.wav", "wer": 0.0, "wav_sha256": "70b026f642bff60c719e9fdea166e9b867310f08f320a3db492bc9adae6ecbe0", "seconds": 13.4, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify.", "Oh, wait, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first states the original instruction, then immediately follows with 'Oh wait, disregard that instruction. Don't carry it out,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly revokes the request, satisfying the cancellation type criterion.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify. Oh wait, disregard that instruction. Don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_008", "hide_variant_original": true, "source_sha256": "64f8df398973341b0801dcdbf9ce193f227104f02fb15045bc13efcc8a1baf94", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/ambiguous_halt/halt_008/942762b7cc4bf8ac/speech.wav", "wer": 0.0, "wav_sha256": "ebe3c6c35b582899a7e88aca3251153b5245bc714515b896f9c1971df38cc69b", "seconds": 11.08, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify.", "Wait a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'Wait a second,' which constitutes a clear request to pause or hold. The initial instruction is not canceled, but the final control intent is to pause, matching the 'ambiguous_halt' type criterion. No execution occurs, and the task is not resumed.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify. Wait a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "d4e9306_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3eeb1f1012a0b330bffc37a078a54be3aebe5d70aa6839443f1fd3789d1afb33", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/d4e9306_3/audio_natural/filler/original/16bd6c31aba54c36/speech.wav", "wer": 0.0, "wav_sha256": "32af3d19a77e2046c4b50c72c3de6c6ffad3e0e30ef63573d6d603f9815734d5", "seconds": 7.76, "spoken_utterances": [ "Follow artists of um... all the songs and albums I have ever liked on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'um', a common filler, but continues with the full original instruction without any correction or change in intent. The final audible content matches the original instruction, with only a minor hesitation that does not alter the task goal. Therefore, the final intent is to execute the original task.", "heard_transcript": "Follow artists of um all the songs and albums I have ever liked on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of February—sorry, I meant January this year?", "long_pause_offsets": [], "source_sha256": "b0ada947be65758c7b3ee2c765edc7fa4c3900de3eea685490e85fea32f38a10", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'February' but immediately corrects to 'January' using 'sorry, I meant'. The final intended month is January, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/false_start/false_start_02_month_name/92619fd8c0689f67/speech.wav", "wer": 0.1, "wav_sha256": "e6d42569dd1b79a8c8df1e6a35edf18455f00b206eefaf9942a294b1fd59166b", "seconds": 9.48, "spoken_utterances": [ "How many ... venmo friends have I made uh... since the start of February—hmmm... sorry, I meant January this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'February' but immediately self-corrects to 'January this year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many Venmo friends have I made uh since the start of February? Hmm. Sorry, I meant January this year." }, "candidate_id": "wav_b8e68e9a32f1d722", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question, then self-corrects from 'February' to 'January this year' with a smooth, spontaneous-sounding repair. The intonation and pacing are consistent with a person recalling information, and the correction is integrated into the same utterance without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "false_start_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of January last year—no, that should be this year?", "long_pause_offsets": [], "source_sha256": "14d868ad3db0f5e069d2957cc54c713aedcc75adbf64b219517d1a4283d891a3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year' within the same utterance. The final intended time reference is 'this year', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/false_start/false_start_03_year_reference/886eda279bc3e174/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "bcc699a64f9d66c51495da74505de7f978b296912219ea1fe75946874529194a", "seconds": 9.76, "spoken_utterances": [ "How many uh... venmo friends have I made since the start of January last year—um... no, that should be ... this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many uh Venmo friends have I made since the start of January last year? No, that should be this year." }, "candidate_id": "wav_be819b12d76184f6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar question but makes a more complex and less fluent correction. They first say 'January last year' and then correct it to 'this year' with a noticeable pause and a hesitant 'no'. The correction feels slightly more forced and less integrated than in audio_01, and the overall delivery has a slightly more robotic or rehearsed quality." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "How many PayPal—I should say venmo friends have I made since the start of January this year?", "long_pause_offsets": [], "source_sha256": "e3d6632706a4c6bedef0895a03702d3f9e353cbc78442bb6798eb34c555fe344", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the query including the app, the metric (number of friends), and the time frame." }, "audio": { "status": "pending", "spoken_utterances": [ "How many um... PayPal—uh... I should say venmo friends have I made um... since the start of January this year?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How many friends have I made since the start of January this year? One more thing: I'm asking about Venmo friends.", "long_pause_offsets": [ 67 ], "source_sha256": "ec37d1eb0e90fb289eccba5d371f4981f67a6f69105a52f7f6bccc5fc409e169", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Venmo' to the initial question, which was otherwise identical in scope and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/addition/addition_02/6f2b1fecd33ca0dd/speech.wav", "wer": 0.08, "wav_sha256": "c2e73073bca9a05d18338303193d4fa7cfbf4a31bfc44ab404c27cfc67a65cf5", "seconds": 12.12, "spoken_utterances": [ "How many friends have I made uh, let me think... since the start of January this year? ", "One more thing: I'm asking about ... Venmo friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many friends have I made since the start of January this year?' which is a broad request. The speaker then adds 'I'm asking about Venmo friends' as a clarifying addition. The final intent combines the original time frame with the specific target 'Venmo friends'. The addition narrows the scope of the initial request, and the final goal matches the reference instruction. Therefore, the addition type is satisfied.", "heard_transcript": "How many friends have I made, uh, let me think, since the start of January this year. One more thing, I'm asking about Venmo friends." }, "candidate_id": "wav_722e5f5aa87a8a3c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant thought process ('uh, let me think') and a plausible self-correction from 'friends' to 'Venmo friends'. The addition of the time constraint ('since the start of January this year') is smoothly integrated with a conversational tone and appropriate prosody. The pacing and intonation suggest genuine spontaneous formulation rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How many venmo friends have I made? Oh, I forgot to mention, count only the Venmo friends made since the start of January this year.", "long_pause_offsets": [ 36 ], "source_sha256": "d3d9dcceed86c217d72db04735de8ccef35c2ca37a80096886133316295ebfee", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing time constraint that narrows the scope of the initial question. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/addition/addition_01/b5df921c00f222e3/speech.wav", "wer": 0.0, "wav_sha256": "b1129f49c9715b3afefd5cd2be6eb294f559bc86fb43f8315535b91721e5e602", "seconds": 12.32, "spoken_utterances": [ "How many ... venmo friends have I made? ", "Oh, I forgot to mention, count only the Venmo friends made ... since the start of January this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial question is incomplete, but the speaker adds the missing constraint 'since the start of January this year' after a pause. The addition clarifies the time frame, and the final goal matches the original instruction. The type is 'addition', and the final intent is to execute the combined task.", "heard_transcript": "How many Venmo friends have I made? Oh, I forgot to mention, count only the Venmo friends made since the start of January this year." }, "candidate_id": "wav_d76f478304bed352", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker clearly adds a time restriction, the initial request is delivered in a flat, somewhat robotic tone. The addition ('Oh, I forgot to mention...') is grammatically correct but sounds slightly abrupt and less naturally integrated than in audio_01. The prosody lacks the subtle hesitation and conversational flow of a truly spontaneous speaker, making it less realistic despite being semantically complete." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "How many PayPal friends have I made since the start of January this year? Wait, I meant Venmo, not PayPal.", "long_pause_offsets": [ 74 ], "source_sha256": "69d8944c3131fc3aad1d1acb4f3d17f3ea155270201a194f4c9b517297a82eb2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the rest of the query including the time frame and the action of counting friends." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/revision/revision_01_app/43ad48ff00556372/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "6100bc26004a22825b8a6bd80d35f171fc4261fb54c39c6cb6dbf82bbc971028", "seconds": 12.64, "spoken_utterances": [ "How many um... PayPal friends have I made ah... since the start of January this year? ", "Wait, I meant Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to query Venmo friends, which matches the original instruction. All other conditions, including the time frame 'since the start of January this year', are preserved.", "heard_transcript": "How many um, PayPal friends have I made uh, since the start of January this year? Wait, I meant Venmo, not PayPal." }, "candidate_id": "wav_74a0cc5367df3bfb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, plausible request, then self-corrects from 'PayPal' to 'Venmo' with a natural hesitation ('wait, I meant...') and a smooth, conversational tone. The correction is integrated into the flow without awkwardness, and the prosody reflects genuine recall and correction." } ] }, "is_default": true }, { "variant_id": "revision_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of January last year? That should be since the start of January this year, not last year.", "long_pause_offsets": [ 73 ], "source_sha256": "d847a961415b7fc82a47d103963f73e08703759db3b596d4eed17b58b9a6e16b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last year' to 'this year', preserving the app (Venmo), action (counting friends), and the starting point (start of January)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/revision/revision_03_year_reference/dbb576ca100a2f5e/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "dd3941fd85c5b745ea0a4bee53a6129c87c1db20f4cc75a8a4c7d0229d2baaa6", "seconds": 13.12, "spoken_utterances": [ "How many uh... venmo friends have I made um... since the start of January last year? ", "That should be since the start of January ... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'last year' but immediately self-corrects to 'this year' after a pause, which is the revision type. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, so the task should be executed with the corrected time frame.", "heard_transcript": "How many uh Venmo friends have I made um since the start of January last year? That should be since the start of January this year, not last year." }, "candidate_id": "wav_136c7b8b023d5a41", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a revision from 'last year' to 'this year' with a self-correcting phrase ('That should be...'). The delivery is natural, with a slight pause and intonation shift indicating reconsideration. However, the correction is less fluid than audio_02, with a slightly more abrupt transition and less conversational phrasing." } ] }, "is_default": false }, { "variant_id": "revision_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of February this year? Make that since the start of February this year, not January.", "long_pause_offsets": [ 74 ], "source_sha256": "07b22f890285b5d23b616c5681192c7e615abed157203e2b31a1e4082ae66b9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly confirms the revised time frame and rejects the original January reference, preserving the action, app, and scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_1/audio_natural/revision/revision_02_month_name/5de2e2b30d0ff362/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "eafc99f5fb46e8697f4cecf16893de2bf050ce4ee0cc5cfbfb42d9dccb5cfa8b", "seconds": 14.52, "spoken_utterances": [ "How many venmo friends have I made um... since the start of February this year? ", "Make that since the start of um, let's see... February this year, not hmmm... January." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about January, then corrects themselves to February after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent is clear and matches the corrected value, so the task should be executed as revised.", "heard_transcript": "How many Venmo friends have I made um since the start of February this year? Wait, that since the start of um let's see, February this year, not January." }, "candidate_id": "wav_00ba9926674c3226", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request, then hesitates ('wait, that since the start of...') and corrects from 'February' to 'January'. The correction is less smooth, with a noticeable stutter ('but see') and a more fragmented, less fluent delivery. The prosody sounds slightly rehearsed or unnatural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "f147ba657d10c57bc54d7c6d271daf18b546b2696af798e479b9394e66f14ce6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How many uh... Venmo friends have I made um... since the start of January this year?", "Hang on, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "dac78d9_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "e44e6d53e957ee570446e1ebd264e1921f1fe3f43490c24b626a06ff682fd759", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How many uh... Venmo friends have I made um... since the start of January this year?", "Give me a second..." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "dac78d9_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "cad212074b7204e2ef4a04bbb79e8da41a702b1f8b2afea315b03e8242a9cf4d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "pending", "spoken_utterances": [ "How many uh... Venmo friends have I made um... since the start of January this year?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "selection_status": "no_eligible_candidate" }, { "task_id": "dac78d9_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "How many PayPal—I should say venmo friends have I made since the start of October last year?", "long_pause_offsets": [], "source_sha256": "142d48e52cb4cb0d5bc54dcddddadf97877cfe4d6e7bd1a88819f0098d2fb1c8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the query including the app, the action, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/false_start/false_start_01_app/bd5093a0c798b001/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "c2d5bcf8217f6d93ed9de1647fa4e72f1375f532770709bf500516c3a2376e7f", "seconds": 11.48, "spoken_utterances": [ "How many PayPal—hmmm... I should say uh... venmo friends have I made um, let's see... since the start of October last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say uh Venmo friends'. This is a clear false start where the initial incorrect value is replaced by the correct one. The final intent matches the original instruction exactly, and the control intent is to execute the corrected task.", "heard_transcript": "How many PayPal, I should say uh Venmo friends have I made um let's see, since the start of October last year?" }, "candidate_id": "wav_abd8f09415de9abc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The false start 'PayPal' is immediately and smoothly corrected to 'Venmo' with a believable self-repair. The subsequent 'let's see' filler is a very plausible hesitation before continuing the request, and the overall prosody, pacing, and intonation sound like a genuine person formulating a question in real time." } ] }, "is_default": true }, { "variant_id": "false_start_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of October this year—sorry, make that last year?", "long_pause_offsets": [], "source_sha256": "5d9513d3471d90dcb94e96315c2cd56777d8acdf604ed34afce59af91d12a7dc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'this year' but immediately corrects it to 'last year' within the same utterance. The final intended time reference is 'last year', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/false_start/false_start_03_year_reference/fb089d4dc9e41fcd/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "21d402d4b9ea820509435c2210c31681882c6888ac6f707a32e2376ea812c89a", "seconds": 9.44, "spoken_utterances": [ "How many uh... venmo friends have I made since the start of October this year—um... sorry, make that last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this year' but immediately self-corrects to 'last year' after a pause and an apology. According to the 'false_start' type criterion, the final corrected value ('last year') overrides the initial mistaken value ('this year'), and the rest of the instruction remains unchanged. The final audible intent matches the original instruction.", "heard_transcript": "How many uh Venmo friends have I made since the start of October this year? Um, sorry, make that last year." }, "candidate_id": "wav_8c40dd4e24c56566", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start ('PayPal') and a self-repair ('Venmo'), but the correction is less fluid than in audio_03. The speaker says 'uh' before the initial word, which is a common filler, but the overall delivery is slightly more hesitant and less continuous. The pause before 'sorry' and the correction feel a bit more abrupt, making it less natural than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of September—no, I mean October last year?", "long_pause_offsets": [], "source_sha256": "cd7b8284d9fa2aeff8017e24b0108ecc3dae54b3cf95487a93449b5e61c98c89", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'September' but immediately corrects to 'October' within the same utterance. The final intended month is October, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/false_start/false_start_02_month_name/9fe0e7068ea88ab3/speech.wav", "wer": 0.15, "wav_sha256": "023b3d7394a1757842ac297c32e4cc1322af9245bfac90817609e8b6ee0284e9", "seconds": 7.56, "spoken_utterances": [ "How many hmmm... venmo friends have I made since the start of September—um... no, I mean ... October last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'September' but immediately self-corrects to 'October last year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many Venmo friends have I made since the start of September? No, I mean October last year." }, "candidate_id": "wav_d12d78bff5ebb4cb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker has a false start ('fanmo') and a self-repair ('Venmo'), but the correction is less smooth and the overall delivery is more disjointed. The initial word is mispronounced, and the correction is delivered with a slightly robotic or rehearsed quality. The prosody is flatter, and the transition from the false start to the correction lacks the natural flow of the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many venmo friends have I made? Oh, I forgot to mention, this is for the period since the start of October last year.", "long_pause_offsets": [ 36 ], "source_sha256": "73d248bb4f07b3c2cd39dc7b9128d99015558f7b6fc239104cf1987ed3475c20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint, which is the only missing requirement from the original instruction. The action and target remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "How many uh... venmo friends have I made? ", "Oh, I forgot to mention, this is for the period um, let me think... since the start of October last year." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "dac78d9_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of September last year? Make that since the start of September last year, not October.", "long_pause_offsets": [ 75 ], "source_sha256": "90c09785dcaef1bee2c512b44a66e8cc398997043b504defa8061ddeeb967cdc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the month from October to September, while retaining the app (Venmo), the action (counting friends), and the time frame (start of the month last year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/revision/revision_02_month_name/a638667565b21763/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "66bbb8f857a1c41f97604b07a24fcb110eba83c077abe25ac0ba6b7030fc7e6c", "seconds": 12.92, "spoken_utterances": [ "How many venmo friends have I made ... since the start of September last year? ", "Make that since the start of um... September last year, not ah... October." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about October, then corrects to September after a pause and self-repair. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final audible intent is to query September, which matches the revised goal. All other constraints remain unchanged.", "heard_transcript": "How many Venmo friends have I made since the start of September last year? Make that since the start of um uh September last year, not uh October." }, "candidate_id": "wav_3e3d87982ce49603", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question. After a brief pause and filler ('um'), they self-correct 'October' to 'September' with a plausible hesitation and a smooth, continuous flow. The correction is integrated into the same utterance without awkwardness, and the prosody reflects genuine recall and adjustment." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "How many PayPal friends have I made since the start of October last year? Wait, the app should be Venmo, not PayPal.", "long_pause_offsets": [ 74 ], "source_sha256": "b9ae5915359b1218cb28242bf42a0fb471db3380f9744711f7f86aa934067dda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the rest of the query including the time frame and the action of counting friends." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/revision/revision_01_app/f23704067dac7171/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "0c3a5b4784c84c6c59707ed1841ea1047a35d2b9050c86423c401fdd576da58a", "seconds": 11.32, "spoken_utterances": [ "How many uh... PayPal friends have I made since the start of October last year? ", "Wait, the app should be um... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about PayPal but immediately corrects the app to Venmo. According to the revision type criterion, the correction spoken after the pause must be applied to the earlier instruction. The final audible intent is to query Venmo friends since the start of October last year, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints are altered.", "heard_transcript": "How many uh PayPal friends have I made since the start of October last year? Wait, the app should be um Venmo, not PayPal." }, "candidate_id": "wav_1762e5ad4a1e40b2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant 'uh' and then makes a clear error ('PayPal' instead of 'Venmo'). After a short pause, they correct themselves to 'Venmo' with a believable 'wait' and a natural-sounding self-repair. The flow is mostly smooth, though the initial error is more jarring than a simple misstatement of a date." } ] }, "is_default": false }, { "variant_id": "revision_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of October this year? Change the reference to last year, not this year.", "long_pause_offsets": [ 73 ], "source_sha256": "c11c554e8aecafdd253be90ad9802c270d08c9167f1fb5ac0755b5b034d841c3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the reference year from 'this year' to 'last year', preserving the app, action, and time frame structure." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/revision/revision_03_year_reference/0586b1e3d5d2b3aa/speech.wav", "wer": 0.12, "wav_sha256": "9d9ec60693d0de5663eb5b06c1534cf1ab07b98d4e85ae136e2d1a26c62dd196", "seconds": 12.44, "spoken_utterances": [ "How many um... venmo friends have I made hmmm... since the start of October this year? ", "Change the reference to ... last year, not this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'this year' but immediately corrects to 'last year' after a pause. This is a revision where the later instruction overrides the earlier one. The final audible intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "How many um Venmo friends have I made um since the start of October this year? Change the reference to last year, not this year." }, "candidate_id": "wav_12ad68bf94a27f65", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a hesitant 'um' and then makes a clear error ('this year' instead of 'last year'). After a pause, they correct it with 'change the reference to last year'. However, the correction is delivered as a separate, declarative sentence rather than a natural self-repair within the original request, making it sound less spontaneous and more like a post-hoc instruction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "3ae5d8197bcee1f280b27768978e24208a009b554936e62d03bc992756d8fb16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/cancellation/cancel_027/4ab7b424e7deee8a/speech.wav", "wer": 0.0, "wav_sha256": "64573fe3bdd70190f7f17526ff8fc03f3fcac184e0422f53ad30ae5e7ec27576", "seconds": 11.36, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?", "Hang on, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request, then immediately follows with 'Hang on, please disregard my last request,' which is a clear and unambiguous cancellation of the prior task. The final audible intent is to cancel, satisfying the cancellation type criterion.", "heard_transcript": "How many Venmo friends have I made since the start of October last year? Hang on, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_011", "hide_variant_original": true, "source_sha256": "6647f5842ae3ecaa765e2a6d232f6bce6e779cedde564412e67e50f9b5a62e39", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/ambiguous_halt/halt_011/5d66b0a3592b5888/speech.wav", "wer": 0.0, "wav_sha256": "3414d69cddaa8217bbd656c276147a307dbce179a97c6af0509070888ef819b8", "seconds": 10.2, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?", "Hang on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on a second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled, and the speaker does not resume or modify it. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "How many Venmo friends have I made since the start of October last year? Hang on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ef1db8970b7906047548b633ced4b771863ec68f643e3220f711fcc9f554df42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_2/audio_natural/filler/original/0bbe7ecc60af8433/speech.wav", "wer": 0.0, "wav_sha256": "86b96e22a6507d5a6e718a87297a4518dbfe3af7ce2b95951ebcb46217fc3b7d", "seconds": 6.8, "spoken_utterances": [ "How many ... Venmo friends have I made ... since the start of October last year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full original instruction clearly and without modification. The final intent matches the original instruction exactly, and the 'filler' type allows for such minor hesitations or pauses that do not change the goal. Therefore, the task is executed as intended.", "heard_transcript": "How many Venmo friends have I made since the start of October last year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "How many PayPal—sorry, I meant venmo friends have I made since the start of March this year?", "long_pause_offsets": [], "source_sha256": "ab0e1b9c05cee1517dae909e63bac1a16e28ee5894afc233df767be6f52b60fd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the query including the app, the action, and the time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/false_start/false_start_01_app/f3b098483c1a0bf7/speech.wav", "wer": 0.0, "wav_sha256": "e5f2bcba82d4ecf7e4042b1b7608d57baf841c0b0f2760b40302944a8877b640", "seconds": 9.76, "spoken_utterances": [ "How many PayPal—uh... sorry, I meant ... venmo friends have I made ... since the start of March this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'sorry, I meant'. The final audible intent is to ask about Venmo friends, which matches the original instruction. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "How many PayPal uh, sorry, I meant Venmo friends have I made since the start of March this year?" }, "candidate_id": "wav_25949ea26ab2c57b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous flow. The speaker begins with a different service (PayPal), immediately self-corrects with 'sorry, I meant Venmo' in a conversational tone, and then continues with the full, correct request. The false start and correction are seamlessly integrated, and the prosody remains consistent and human-like throughout." } ] }, "is_default": true }, { "variant_id": "false_start_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of April—no, make that March this year?", "long_pause_offsets": [], "source_sha256": "07402f6a2cf51cd441c94d001c708f579c75019b86382df89ce7b1d3c012d869", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'April' but immediately corrects to 'March' using 'make that March'. The final intended month is March, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/false_start/false_start_02_month_name/af80ce1963d53a4c/speech.wav", "wer": 0.05, "wav_sha256": "8b7689e6b64542ac957caa57019779c84a30efd1b19212e33955c41d0730fc99", "seconds": 10.24, "spoken_utterances": [ "How many uh... venmo friends have I made ... since the start of April—um... no, make that March this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'April' but immediately self-corrects to 'March' with 'no, make that March this year'. This is a clear false start where the final value is the corrected one. The rest of the instruction matches the original, so the final intent is to execute the corrected request.", "heard_transcript": "How many uh Venmo friends have I made since the start of April um no, make that March this year?" }, "candidate_id": "wav_e1b138dd50fe50fe", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start and self-correction ('since the start of April... no, make that March this year'). The correction is natural, and the speaker's intonation reflects a genuine moment of recollection. However, the initial phrase 'Benmo' is mispronounced, and the correction is slightly more abrupt than in audio_03, making it feel a bit less fluid." } ] }, "is_default": false }, { "variant_id": "false_start_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of March last year—sorry, that should be this year?", "long_pause_offsets": [], "source_sha256": "137f5195127e472600a15cd9429281857daf47b22f986c66b81925dddfae3e5d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year' within the same utterance. The final intended time reference is 'this year', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/false_start/false_start_03_year_reference/c3d3e2ab44d66bf6/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "2394654836eb64adb2523b66b1059a1a563e215d004c3fe29700f08366134d0a", "seconds": 9.56, "spoken_utterances": [ "How many uh... venmo friends have I made since the start of March last year—hmmm... sorry, that should be this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year' after a pause and an interjection. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many uh Venmo friends have I made since the start of March last year? Hmm. Sorry, that should be this year." }, "candidate_id": "wav_e418435f35c1855a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start and correction ('since the start of March last year... sorry, that should be this year'), but the correction is delivered with a noticeable pause and a hesitant 'hmm' before the revision. The prosody feels slightly more rehearsed or less spontaneous compared to the others, and the correction is less integrated into the flow, making it sound less natural overall." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many venmo friends have I made? Oh, I forgot to mention, I want to know about the period since the start of March this year.", "long_pause_offsets": [ 36 ], "source_sha256": "0ba9b36ec2af38baacdb267f1f26d7e2a95cb5f1779d4216bfa6b1c7c51f38e6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint to the initial question, combining both parts into a single, complete query that matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/addition/addition_01/ec62921b16fbce14/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "25ea048b27e4836fe08eb783efbd49c9d4335a314ebe36e9b757942a21ff3065", "seconds": 15.04, "spoken_utterances": [ "How many ah... venmo friends have I made? ", "Oh, I forgot to mention, I want to know about the period uh, let's see... since the start of March this year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, asking only 'How many Venmo friends have I made?'. After a pause and self-correction, the speaker adds the missing time constraint 'since the start of March this year'. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, so the task should be executed.", "heard_transcript": "How many uh Venmo friends have I made? Oh, I forgot to mention, I want to know about the period uh let's see, since the start of March this year." }, "candidate_id": "wav_b489bdac48058377", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a slightly hesitant but natural-sounding 'uh' before stating the core question. The phrasing 'Oh, I forgot to mention' is a very common and realistic way to add a detail after an initial request. The subsequent addition of the time period is delivered with a believable 'let's see' pause and a smooth, conversational tone, maintaining a consistent rhythm and prosody throughout. The overall flow feels like a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many friends have I made since the start of March this year? I should also mention, I'm asking about Venmo friends.", "long_pause_offsets": [ 65 ], "source_sha256": "bede41e7e04ba63067292cc618814a34261a982c5c711e96445326f6ac3a2fea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Venmo' to the initial question, which was otherwise identical in scope and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/addition/addition_02/1cce2ddafdb851fc/speech.wav", "wer": 0.04, "wav_sha256": "534242f1c5f800f018065b7485c5cb71b20e6095d4ff7db2c815217efc3f84ff", "seconds": 13.52, "spoken_utterances": [ "How many friends have I made um... since the start of March this year? ", "I should also mention, I'm asking about let's see... Venmo friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'friends' in general, then adds the specific constraint 'Venmo friends' after a pause and self-correction. According to the 'addition' type criterion, the final goal is the combination of the original instruction with the added information. The final intent matches the reference instruction exactly, so the task passes.", "heard_transcript": "How many friends have I made um since the start of March this year? I should also mention, I'm asking about, let's see, Venmo friends." }, "candidate_id": "wav_9b55a7c9c0cabc04", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a natural hesitation ('um') and a slightly rambling, self-correcting structure ('I should also mention, I'm asking about...'). However, the addition of 'Venmo friends' is delivered as a separate, almost clipped sentence rather than a smooth continuation of the initial thought. The transition feels less organic than in audio_02, and the overall delivery is slightly more disjointed, making it less realistic in its flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_month_name", "proposed_instruction": "How many venmo friends have I made since the start of April this year? I misspoke about the month; it should be March, not April.", "long_pause_offsets": [ 71 ], "source_sha256": "733087b94fabf6629a00fdc12a66bb9d3c852ae38ded0840fb81e26a5d129ccf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the month from April to March, preserving the app (Venmo), the action (counting friends), and the time frame (since the start of the month this year)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/revision/revision_02_month_name/3459b46812a41592/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "bdc82806c71156c9e9001e0315cbc2e00e3810bbc5b5f4cc0aa6bf6cd5b257da", "seconds": 13.88, "spoken_utterances": [ "How many uh... venmo friends have I made ... since the start of April this year? ", "I misspoke about the month; it should be um... March, not April." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'April' but immediately corrects themselves to 'March' after a pause, which aligns with the revision type criterion. The final audible intent is to query Venmo friends since the start of March, matching the original instruction. The correction is clear and applied to the earlier instruction.", "heard_transcript": "How many uh Venmo friends have I made since the start of April this year? I misspoke about the month, it should be um March, not April." }, "candidate_id": "wav_d748ae9259f87790", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear initial request with a natural hesitation ('uh') and a plausible self-correction of the month. The revision ('it should be um March not April') is spoken with a believable change-of-mind intonation, and the flow from the initial request to the correction is smooth and continuous, without any jarring edits or unnatural pauses. The prosody and rhythm are consistent with spontaneous speech." } ] }, "is_default": true }, { "variant_id": "revision_03_year_reference", "proposed_instruction": "How many venmo friends have I made since the start of March last year? That should be 'this year', not 'last year'.", "long_pause_offsets": [ 71 ], "source_sha256": "b65c26d4d5d61aa858d652b37176afcd7665c6d2290fc60f38ada7ee24084c3d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects 'last year' to 'this year', preserving the app, action, and date range context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/revision/revision_03_year_reference/b676a078a923513b/speech.wav", "wer": 0.0, "wav_sha256": "dda2ec9379ee0d9939d2f5b28d9da80b43eba43eadea8d6ed647d34081d0d1f2", "seconds": 12.92, "spoken_utterances": [ "How many um... venmo friends have I made ... since the start of March last year? ", "That should be 'uh... this year', not 'last year'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier incorrect value. The final intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "How many um Venmo friends have I made since the start of March last year? That should be uh this year, not last year." }, "candidate_id": "wav_40da684de133af96", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision from 'last year' to 'this year'. The speaker hesitates with 'um' and 'uh' in a way that sounds natural, and the correction is delivered with a slight shift in intonation that suggests a genuine realization. However, the correction is slightly less fluid than in audio_02, with a brief pause and a slightly more abrupt delivery of the corrected phrase, making it feel a bit less seamless." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "How many PayPal friends have I made since the start of March this year? Change Venmo to PayPal.", "long_pause_offsets": [ 72 ], "source_sha256": "d1d07c9827bb7381eccc7e969918f4bd9b0087488b9fa78a9bd2160e62821030", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Venmo to PayPal, aligning the final goal with the original instruction's structure but updating the app name. The time frame and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/revision/revision_01_app/0b565a8ff42dd19e/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "c57a017ceae4893d43238d58c68677f4088dd9abcd5aa284a081ba81e0e53398", "seconds": 13.04, "spoken_utterances": [ "How many um, let's see... PayPal friends have I made since the start of March this year? ", "Change uh... Venmo to PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Venmo friends but immediately corrects the target from Venmo to PayPal. The revision is clear and unambiguous, and the rest of the instruction (timeframe: since the start of March this year) remains unchanged. The final intent matches the revised instruction, satisfying the revision type criterion.", "heard_transcript": "How many um, let's see, PayPal friends have I made since the start of March this year? Change uh Venmo to PayPal." }, "candidate_id": "wav_78306bed774df7fb", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker begins with a filler ('um') and a self-initiated pause ('let's see'), which is realistic, but the correction is not a simple revision of a previously stated value. Instead, the speaker changes the entire platform from 'Venmo' to 'PayPal' mid-sentence, which is a more significant and less common type of correction. The delivery of the correction ('change uh Venmo to PayPal') sounds more like a post-hoc instruction than a spontaneous self-correction, and the prosody is slightly less natural, with a flatter intonation and a less organic flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_010", "hide_variant_original": true, "source_sha256": "4b1dc3258357742a82f413f9c86e73e766b3d153866dd9235e7ab55c9baf602b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/cancellation/cancel_010/c313c10d0a0d3524/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "846e45f46bf09779dca9096e9a0c81dfe797f3ceb69430c988365b3fb534a7e9", "seconds": 12.48, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?", "Oh, wait, cancel the request I just made." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a cancellation directive ('cancel the request I just made') spoken immediately after the initial request. This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final intent is to cancel, not to execute the original request.", "heard_transcript": "How many Venmo friends have I made since the start of March this year? Oh wait, cancel the request I just made." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_010", "hide_variant_original": true, "source_sha256": "94c7c0d6221ae029f5cd8dd2489d9cb486fb51a1ac173ee6906b3d24598ac877", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/ambiguous_halt/halt_010/6784e29036479009/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "0ac257b74e49849319f1b38d6ccb2e3aa6863936eec0d371d5b094ae4649f46b", "seconds": 11.48, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?", "Hold on a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on a second,' which is a clear request to pause or wait. This matches the expected behavior for an 'ambiguous_halt' type, where the final interjection is a pause command and not a cancellation or execution. The initial request is not revoked, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is correctly identified as 'pause'.", "heard_transcript": "How many Venmo friends have I made since the start of March this year? Hold on a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "dac78d9_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4c6334c1909ce3c142f3211e5600425870553dc79d69ab2c215bd8da12527288", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/dac78d9_3/audio_natural/filler/original/aefbb4e6c00e2f4f/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "3aa3a64d6e63fb84e415fba24c482d479d22a7f59050cf507495b41d55011926", "seconds": 7.88, "spoken_utterances": [ "How many ... Venmo friends have I made hmmm... since the start of March this year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker delivers the full original instruction without hesitation, backchanneling, or self-correction. The audio contains no pauses or fillers that alter the task's goal. The final intent matches the original instruction exactly, satisfying the 'filler' type criterion where hesitations are ignored and the remaining task meaning must match the original.", "heard_transcript": "How many Venmo friends have I made since the start of March this year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family—sorry, I meant roommates.", "long_pause_offsets": [], "source_sha256": "0b1cbdde7cd5f686b149fd4e54cd9442af0dae703a580d2a07ab27450b9b835e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'roommates' within the same utterance, preserving the app, time scope, and action. The final intended target is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/false_start/false_start_04_counterparty_group/0f69821f7ce58e61/speech.wav", "wer": 0.05, "wav_sha256": "3793333547910bc8ec61d312b92628227c5893fccfd1ef1acf44c4ca70b2c851", "seconds": 7.8, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing year to and from my family—um... sorry, I meant ... roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'roommates' with 'sorry, I meant roommates.' This is a false start, and the final audible intent is the corrected version. The correction is immediate and unambiguous, so the task should be executed with 'roommates' as the target.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing year to and from my family um, sorry, I meant roommates." }, "candidate_id": "wav_92792fe12b214a07", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip features a clear false start ('family') followed by a natural self-correction ('sorry, I meant roommates') with a believable tone of mild embarrassment. The initial hesitation and correction are smoothly integrated into the flow, and the speaker's voice remains consistent and conversational throughout. The pacing and intonation suggest spontaneous speech rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the last year—no, that should be ongoing year to and from my roommates.", "long_pause_offsets": [], "source_sha256": "c9045750eeb8d449127c77d1e5fc58064a4427d83d35eb01a4fd0c12bcea6165", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last year' to 'ongoing year' within the same utterance, preserving the app (Venmo), scope (all transactions), direction (to and from), and recipient group (roommates)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/false_start/false_start_02_time_reference/2c1b3a9935038a38/speech.wav", "wer": 0.0, "wav_sha256": "b1edd8825c153cb0aa3dfd82d7299f4d19360196091163059f7bc34ece66622e", "seconds": 10.52, "spoken_utterances": [ "Like all the uh, let me think... venmo transactions of the last year—um... no, that should be ongoing year to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'ongoing year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, so the task should be executed as corrected.", "heard_transcript": "Like all the uh, let me think. Venmo transactions of the last year. Um, no, that should be ongoing year to and from my roommates." }, "candidate_id": "wav_11cbdc8fd3d102be", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip includes a false start ('like all the uh...') and a self-correction ('no, that should be ongoing year...'). The correction is immediate and the speaker continues without restarting, which is realistic. However, the initial hesitation is slightly longer and less fluid than in audio_04, and the correction feels a bit more abrupt, making it slightly less natural in flow." } ] }, "is_default": false }, { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to—or rather, to and from my roommates.", "long_pause_offsets": [], "source_sha256": "61a45fe163b4417fcae4d92b3c4ec8a564c3cb7b66c36ad151d0c3088c483103", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially restricts the scope to 'only to' but immediately corrects to 'to and from' using 'or rather'. The final intended scope includes both sending and receiving transactions with roommates, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/false_start/false_start_03_transaction_direction/2eb9feed84604518/speech.wav", "wer": 0.05, "wav_sha256": "7db9a77d81094928fa69d8d667e2e53477f5ac1dcec3b8ff94dd6e1114a945cb", "seconds": 9.52, "spoken_utterances": [ "Like all the venmo transactions of the ongoing year uh... only to—um... or rather, to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to' but immediately self-corrects with 'or rather to and from my roommates'. This is a false_start type, where the final intent is the corrected version. The correction is clear and unambiguous, and the final goal matches the original instruction. The hesitation and filler words ('uh', 'um') do not change the final task intent.", "heard_transcript": "Like all the Venmo transactions of the ongoing year uh only to um or rather to and from my roommates." }, "candidate_id": "wav_84be8d87047ed561", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a false start ('only to...') and a self-correction ('or rather to and from...'). While the correction is grammatically sound, the phrasing 'or rather' sounds slightly more formal and less spontaneous than typical conversational self-correction. The delivery is somewhat hesitant and the transition between the false start and correction is less seamless than in the top two." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Like all the PayPal—I should have said venmo transactions of the ongoing year to and from my roommates.", "long_pause_offsets": [], "source_sha256": "167ce532952ec8cb5ce0703835e1cc7c667e58e3f11a0a67cb07522372f8ba20", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to venmo within the same utterance, preserving the app, scope, and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/false_start/false_start_01_app/34fc4993252b6189/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "8d6d9ed625275b3dca034b5e0486042b67d4e98b52ab58a84996381c93ce6a35", "seconds": 13.08, "spoken_utterances": [ "Like all the PayPal—um... I should have said hmmm... venmo transactions of the um... ongoing year to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should've said Venmo'. This is a false_start type, where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to like all Venmo transactions of the ongoing year to and from the roommates.", "heard_transcript": "Like all the PayPal um, I should've said Venmo transactions of the um ongoing year to and from my roommates." }, "candidate_id": "wav_bc9164ee72561ccd", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has multiple fillers ('um', 'mm') and a self-correction ('I should have said...'), but the correction is not integrated into the original sentence. Instead, it's a separate clause that feels tacked on, breaking the natural flow. The speaker seems to be rephrasing the entire request rather than correcting a single word, which makes the delivery feel less like a spontaneous thought process and more like a rehearsed or disjointed edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Like all the venmo transactions of the ongoing year. Oh, and one more thing: only include transactions to and from my roommates.", "long_pause_offsets": [ 53 ], "source_sha256": "39b8b0ed0baaf0b74b8bf15fe3f0c836d1e7dc36e0be7e90a59ddf00ac4e26cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a restriction to the initial broad request, narrowing the scope to only transactions involving roommates. The action, app, and time frame remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/addition/addition_03/dd9469c1667a7422/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "275f1fac746fb4804aab757abbb711921a562b5c96478a8131ce5804bacf70a1", "seconds": 14.6, "spoken_utterances": [ "Like all the uh, let's see... venmo transactions of the ... ongoing year. ", "Oh, and one more thing: only include transactions um... to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Like all the Venmo transactions of the ongoing year'. After a pause and an interjection ('Oh, and one more thing'), the speaker adds a constraint: 'only include transactions to and from my roommates'. This is a clear addition that narrows the scope of the original request. The final intent combines both the original and the added constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Like all the uh, let's see. Venmo transactions of the ongoing year. Oh, and one more thing, only include transactions um, to and from my roommates." }, "candidate_id": "wav_99ce8b540bc3e702", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses a realistic 'uh' filler, a plausible 'let's see' hesitation, and a smooth addition of a new constraint ('only include transactions...') after a brief pause. The prosody is consistent, and the addition feels like a genuine afterthought rather than a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Like all the venmo transactions of the to and from my roommates. I didn't mention this, but it should be for the ongoing year.", "long_pause_offsets": [ 65 ], "source_sha256": "0e4907a50e3389d505fa45db920ef14be36805be25a2142f77f069e5672fce5e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'for the ongoing year' to the initial request, which already specified the app (Venmo), action (like), and participants (to and from my roommates). The final goal combines both utterances without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/addition/addition_02/64566a074771b178/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "49d294646108415c41d4efae6a9fa00e76b83aa88826812aa9f37ad17a7b7c7a", "seconds": 13.32, "spoken_utterances": [ "Like all the venmo transactions of the um, let me think... to and from my roommates. ", "I didn't mention this, but it should be for the um... ongoing year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates and says 'let me think' but does not change the core request. They then add the constraint 'ongoing year' after a pause, which is an addition to the original instruction. The final intent combines the original request with the added time constraint, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "Like all the Venmo transactions of the um, let me think, to and from my roommates. I didn't mention this, but it should be for the um, ongoing year." }, "candidate_id": "wav_71a99dc8d402d816", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable self-correction with 'let me think' and a pause, followed by a plausible addition of a time frame ('ongoing year'). The prosody is natural, and the repair is integrated into the utterance. However, the addition is slightly less smoothly integrated than in audio_03, and the speaker's voice seems to shift slightly in tone after the pause." } ] }, "is_default": false }, { "variant_id": "addition_01", "proposed_instruction": "Like venmo transactions of the ongoing year to and from my roommates. Oh, I forgot to mention, like all the Venmo transactions.", "long_pause_offsets": [ 70 ], "source_sha256": "56cf0dc3f411f3e9a6af1db754f810d100ab2a3d9c7bc0c0bf07dbd41932dab1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly adds the missing phrase 'all the Venmo transactions', which clarifies the scope of the initial request. The final goal matches the original instruction in action, app, time frame, and parties involved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/addition/addition_01/88d17bec92c00b1f/speech.wav", "wer": 0.12, "wav_sha256": "d6726ae94e1ef2a21a5e56e4bb1c6226f4609a87e036e3023901aaf416185b34", "seconds": 14.4, "spoken_utterances": [ "Like uh... venmo transactions of the ongoing year um... to and from my roommates. ", "Oh, I forgot to mention, like uh... all the Venmo transactions." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction specifies 'Venmo transactions of the ongoing year to and from my roommates'. After a pause and an interjection, the speaker adds 'all the Venmo transactions'. This is a broader, more general request that combines with the original. The final intent is to retrieve all Venmo transactions, which satisfies the addition type criterion by expanding the scope of the original request.", "heard_transcript": "Like uh Venmo transactions of the ongoing year um to and from my roommates. Oh, I forgot to mention, like um all the Venmo transactions." }, "candidate_id": "wav_670f9d9d71bb87de", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a more disjointed feel. The speaker starts with a clear request, then after a pause, adds a correction that seems to repeat the initial idea rather than adding a new constraint. The transition between the two parts feels less natural, and the second part sounds more like a rephrasing than a genuine addition. The prosody also has a slight flatness compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to my roommates. Wait, it should be to and from, not only to.", "long_pause_offsets": [ 74 ], "source_sha256": "d49d4b9a3ee52276dc3fea33bcdaeafb25ddc10bb1bf50241b71738b180aeebd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the direction from 'only to' to 'to and from', which matches the original instruction. The app, time scope, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/revision/revision_03_transaction_direction/1b5121e7a26b512e/speech.wav", "wer": 0.0, "wav_sha256": "40fac7461572612a76a4fb407dc05b6eeda0bfe1a2d4084af9e27745c18ce81a", "seconds": 12.4, "spoken_utterances": [ "Like all the venmo transactions of the ongoing year um... only to my roommates. ", "Wait, it should be let's see... to and from, not only to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to my roommates' but then corrects themselves with 'to and from, not only to' after a pause. This is a revision where the final intent is the corrected version. The original instruction matches the final intent, so the task is executed as revised.", "heard_transcript": "Like all the Venmo transactions of the ongoing year, um, only to my roommates. Wait, it should be, let's see, to and from, not only to." }, "candidate_id": "wav_2749405c73e6a624", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a natural, multi-stage revision. The initial request is clear, followed by a believable self-correction ('wait it should be...') with a slight pause and intonation shift. The correction is smoothly integrated into the same utterance, maintaining a consistent rhythm and prosody. The flow feels like a genuine moment of recollection and adjustment." } ] }, "is_default": true }, { "variant_id": "revision_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the last year to and from my roommates. Make that the ongoing year, not last year.", "long_pause_offsets": [ 75 ], "source_sha256": "981a439687b639097868ec211f29bd593b05d160f2d61f9565537cbd97a83d45", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last year' to 'ongoing year', preserving the app, action, direction, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/revision/revision_02_time_reference/fcab0a3945df9d9b/speech.wav", "wer": 0.04, "wav_sha256": "e06b371cf9b381423db81ce7584ba8f7c7523ba7d4dd3929293cc2c49ce65e41", "seconds": 14.12, "spoken_utterances": [ "Like all the um... venmo transactions of the let's see... last year to and from my roommates. ", "Make that the ... ongoing year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'ongoing year' with 'Make that the ongoing year, not last year.' This is a clear revision where the final intent is to use 'ongoing year' instead of 'last year'. The rest of the instruction remains unchanged and matches the original. The revision type criterion is satisfied.", "heard_transcript": "Like all the um Venmo transactions of the let's see, last year to and from my roommates. Make that the ongoing year, not last year." }, "candidate_id": "wav_bfb0802f3c20bab8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker makes a clear revision from 'last year' to 'ongoing year' with a natural-sounding self-correction. The prosody is consistent, and the pause before the correction feels like a moment of thought. The delivery is fluent and the revision is integrated smoothly, though slightly less conversational than audio_03 due to the lack of a filler word like 'wait'." } ] }, "is_default": false }, { "variant_id": "revision_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family. That should be roommates, not family.", "long_pause_offsets": [ 75 ], "source_sha256": "f4e0a0c90c04b254e7b353a1fc4a74610f0aab9246a499a9bed28c8992530685", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family' to 'roommates' while preserving the app, time frame, and action. The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/revision/revision_04_counterparty_group/fd8dfb3886f8b286/speech.wav", "wer": 0.08, "wav_sha256": "69279b878c442384078623dea7f4113879f9917e715a39f794a6b90add490060", "seconds": 11.6, "spoken_utterances": [ "Like all the um... venmo transactions of the ongoing year uh, let me think... to and from my family. ", "That should be roommates, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately corrects themselves to 'roommates' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to like Venmo transactions to and from roommates, which matches the original instruction. The correction is clear and final, so the task should be executed as revised.", "heard_transcript": "Like all the um Venmo transactions of the ongoing year, uh let me think, to and from my family. That should be roommates, not family." }, "candidate_id": "wav_ffb4c09772c90062", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker includes a filler ('uh let me think') and a clear correction from 'family' to 'roommates'. The correction is a believable change-of-mind, but the initial request is slightly less fluent, and the filler is more noticeable. The overall delivery is realistic but feels a bit more hesitant than the top two." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Like all the PayPal transactions of the ongoing year to and from my roommates. Change the app to Venmo, not PayPal.", "long_pause_offsets": [ 79 ], "source_sha256": "7501da1c3bd40af2d3a2feca8ba242adeefd50489007ad454127539a4b480731", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, preserving the action, time scope, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/revision/revision_01_app/b1ca63599030fb4d/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "e4c2b0397672b3ebd8417604eec9160956ccf01002188cc55cdf863964c90bf2", "seconds": 11.88, "spoken_utterances": [ "Like all the PayPal transactions of the um... ongoing year let's see... to and from my roommates. ", "Change the app to Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' with 'Change the app to Venmo, not PayPal.' This is a clear revision where the final intent is to use Venmo, matching the original instruction. The correction is immediate and unambiguous, satisfying the revision type criterion.", "heard_transcript": "Like all the PayPal transactions of the um ongoing year, let's see, to and from my roommates. Change the app to Venmo, not PayPal." }, "candidate_id": "wav_0441f562393880e1", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker makes a revision from 'PayPal' to 'Venmo' but the delivery is less natural. The initial request is slightly rushed, and the correction is stated as a separate, almost robotic command ('Change the app to Venmo, not PayPal'). The prosody lacks the natural flow of a spontaneous thought process, sounding more like a direct instruction than a self-corrected utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "13efead27c3e722ce3ee4fbae51d25bef351e9fe82c25c766d358ca116d9438b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/cancellation/cancel_023/fa3345bff617e376/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "90958a68ea42b5a7e0b677077bfcb90fc2027356dfde292ade26f0252d45bad4", "seconds": 12.88, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates.", "Hold on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins by stating the original request, then immediately follows with 'Hold on. Never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the task. The final interjection explicitly instructs not to execute the request, satisfying the cancellation type criterion.", "heard_transcript": "Like all the um Venmo transactions of the ongoing year uh to and from my roommates. Hold on. Never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "e55edf826bc990612fa96812cf2f14ef2f1304df4a213a3c9bee7d12fc1cb094", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/ambiguous_halt/halt_032/58b73fea5dfef3ec/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "0a2954168ca6b65cfdad60eab7884d4f53bee4e27a8ea5742cab7d516d524312", "seconds": 11.24, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates.", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Wait, before you do that,' which is a clear request to hold or pause the task. This is a final interjection that does not cancel the task but asks for a temporary halt. The initial request is fully audible and complete, and the final control intent is a pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Like all the um Venmo transactions of the ongoing year uh to and from my roommates. Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1c7eb58a1f9a9fcf29b23269eacd1ac11cb47284dcf24a71d0f4bb9578062df8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_1/audio_natural/filler/original/ea5f5165903f4376/speech.wav", "wer": 0.1875, "wav_sha256": "ee917735df8366f2b2309968f24ac6f3b248fb9e3263c8a612db373cdc93a1e2", "seconds": 7.24, "spoken_utterances": [ "Like all the um... Venmo transactions of the ongoing year uh... to and from my roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses filler words ('um', 'uh') and pauses but does not change the core request. The final utterance clearly retains the original intent to retrieve Venmo transactions with roommates for the ongoing year. The filler does not alter the goal, so the task should be executed as originally intended.", "heard_transcript": "Like all the um, Venmo transactions of the ongoing year uh, to and from my roommates." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing month to and from my family—I meant friends.", "long_pause_offsets": [], "source_sha256": "f139f4b90f1cb21bcf5ca2fff393261f16000ff41b8bbd9c9f51fd988dee5bff", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends' within the same utterance, replacing the group of people. The app, time frame, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/false_start/false_start_04_counterparty_group/646b8d3ad00a95c3/speech.wav", "wer": 0.15, "wav_sha256": "090e7dad07aeaaf936c5e81e64a0099d66715b3357884502e4d2a38d68c990bd", "seconds": 10.68, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing month um... to and from my family—hmmm... I meant friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects with 'I meant friends.' This is a clear false start where the final intended target is 'friends.' The correction is immediate and unambiguous, so the final intent matches the original instruction. The control intent is to execute the corrected task.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing month um to and from my family. I meant friends." }, "candidate_id": "wav_0514e22fc6e29b18", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent false-start pattern. The speaker begins with 'like all the uh Venmo transactions of the ongoing month,' then hesitates with 'um,' and after a brief pause, self-corrects with 'I meant friends' in a way that sounds like a genuine, spontaneous recollection. The prosody is consistent, the correction is immediate and integrated into the flow, and the intonation clearly marks the shift from the initial mistake to the intended meaning." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Like all the PayPal—I should say venmo transactions of the ongoing month to and from my friends.", "long_pause_offsets": [], "source_sha256": "5821abde778a56c0b18d0799cc7a132aa6dd4667abb08d913128d924652d4fc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects PayPal to Venmo within the same utterance, preserving the action, time scope, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/false_start/false_start_01_app/26ebcb1c80701b1a/speech.wav", "wer": 0.05, "wav_sha256": "677407110949811d054164995c3edc72998eecbdedef375c50585044be7c93cf", "seconds": 10.0, "spoken_utterances": [ "Like all the PayPal—um... I should say venmo transactions of the ongoing month ah... to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I should say Venmo'. This is a clear false start where the corrected value replaces the mistaken one. The rest of the instruction, including 'transactions of the ongoing month to and from my friends', remains unchanged and is executed as intended.", "heard_transcript": "Like all the PayPal um, I should say Venmo transactions of the ongoing month uh to and from my friends." }, "candidate_id": "wav_1919448724613893", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'like all the PayPal um... I should say Venmo transactions...' The correction is clear and the speaker smoothly transitions to the correct term. However, the initial 'PayPal' is stated as a mistaken word rather than a filler, and the correction is delivered as a separate clause ('I should say...') rather than being embedded in the original sentence. This makes the correction feel slightly more deliberate and less like a spontaneous slip, reducing its realism compared to audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the past week—no, that should be ongoing month to and from my friends.", "long_pause_offsets": [], "source_sha256": "e7970f5191de76ebf4ac72724ed3e85b92667659739e803a11bf784cc710936e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'past week' to 'ongoing month' within the same utterance, preserving the app, action, direction, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/false_start/false_start_02_time_reference/9fa19ff35d66eb28/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "8d427b42829c5688ce35f2bc00deebd60180ce49a1aa1b78cd9764a305f12bc1", "seconds": 11.72, "spoken_utterances": [ "Like all the venmo transactions of the past week—uh... no, that should be ongoing month let me think... to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'past week' but immediately self-corrects to 'ongoing month'. This is a false start where the final intent is the corrected version. The rest of the instruction, 'to and from my friends', remains unchanged and is preserved. The correction is immediate and clear, satisfying the false_start type criterion.", "heard_transcript": "Like all the Venmo transactions of the past week. No, that should be ongoing month. Let me think. To and from my friends." }, "candidate_id": "wav_980bc5425324b80d", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a false start ('past week') and a revision ('no, that should be ongoing month'), followed by a filler ('let me think'). While it includes multiple realistic elements, the revision is less natural because the speaker explicitly states the correction ('no, that should be...') rather than integrating it into the sentence. The 'let me think' pause is also slightly long and feels more like a deliberate pause than a spontaneous hesitation, making the overall flow less fluid than audio_02 or audio_01." } ] }, "is_default": false }, { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing month only to—or rather, to and from my friends.", "long_pause_offsets": [], "source_sha256": "b594279b9f83812b773d3ab66a092723adea29c78e3c7a4356c3a26291dc5a6f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially restricts the scope to 'only to' friends, then corrects to 'to and from' friends. The final intended scope includes both sending and receiving transactions with friends, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/false_start/false_start_03_transaction_direction/7c40ef5aa31f216c/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "9ca98d9f78b73a61cbc8d155664eb4129d32010b9961d18f8a3c20c2ab39af4e", "seconds": 11.48, "spoken_utterances": [ "Like all the uh... venmo transactions of the let me think... ongoing month only to—uh... or rather, to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to' but immediately self-corrects with 'or rather to and from my friends', which is the original instruction. The false start is corrected, and the final intent matches the original instruction exactly.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing month only to uh or rather to and from my friends." }, "candidate_id": "wav_430392c11e6f8f39", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the most complex and least natural-sounding structure. It begins with 'like all the uh Venmo transactions of the let me think ongoing month only to uh or rather to and from my friends.' The phrase 'let me think' is awkwardly placed mid-sentence, and the speaker then uses 'or rather' to correct the preposition 'to' to 'to and from.' The prosody is inconsistent, with unnatural pauses and a sense of forced correction. The flow is disrupted, and the self-correction feels more like a mechanical insertion than a spontaneous thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Like all the venmo transactions to and from my friends. Oh, I forgot to mention, this should only apply to the ongoing month.", "long_pause_offsets": [ 56 ], "source_sha256": "552a291866bb1c8bf7d89b6935aaf76dc650e952dc026cb2b9e9f601f866ec2e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a time restriction that narrows the scope of the initial request. The final goal combines the original action and target with the new temporal constraint." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/addition/addition_01/e262f2270ff6b935/speech.wav", "wer": 0.2, "wav_sha256": "fdd66b82f897478f801141bf31208143cf5454f36e4642e993d5251f96ed5d91", "seconds": 13.12, "spoken_utterances": [ "Like all the ah... venmo transactions to and from my friends. ", "Oh, I forgot to mention, this should only apply to the um... ongoing month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction matches the original request. The speaker then adds a clarifying constraint ('only apply to the ongoing months') after a pause, which is a valid addition. The final intent combines the original request with the new condition, satisfying the addition type criterion.", "heard_transcript": "Like all the Venmo transactions to and from my friends. Oh, I forgot to mention, this should only apply to the um ongoing months." }, "candidate_id": "wav_33061a8d20b04e68", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent initial request with natural prosody and a believable, smoothly integrated addition. The filler 'uh' and the phrase 'Oh, I forgot to mention' are placed at a realistic point in the utterance, and the correction is seamlessly woven into the flow without awkwardness. The pacing and intonation suggest spontaneous thought rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Like all the venmo transactions of the ongoing month. Oh, and one more thing: only include transactions to and from my friends.", "long_pause_offsets": [ 54 ], "source_sha256": "28ce44be1c3234a5026615f5703561fe980c2299bfd8c3f26f19d94908a849b4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing restriction that narrows the scope of the initial request to only transactions involving friends, which is consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/addition/addition_02/521d2db2fd96b6f5/speech.wav", "wer": 0.0, "wav_sha256": "3051f9c6d42ed05ddf5d5655d7f1b22050646137058fb5cf8b089fd87d75c829", "seconds": 17.92, "spoken_utterances": [ "Like all the uh, let's see... venmo transactions of the ... ongoing month. ", "Oh, and one more thing: only include transactions ... to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a general request for Venmo transactions of the ongoing month, then adds a constraint to only include transactions to and from friends. This matches the 'addition' type, where the final goal combines the initial instruction with the added information. The final intent is clear and executable.", "heard_transcript": "Like all the uh let's see Venmo transactions of the ongoing month. Oh, and one more thing, only include transactions to and from my friends." }, "candidate_id": "wav_bff79ab2456e171b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker's delivery is generally natural, the addition of 'Oh, and one more thing' feels slightly more abrupt and less integrated than in audio_01. The initial hesitation ('uh, let's see') is plausible, but the overall structure is more segmented, with the second part of the request sounding like a separate thought rather than a fluid continuation. The prosody is slightly less continuous, making it feel less spontaneous compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "Like all the PayPal transactions of the ongoing month to and from my friends. Wait, the transactions should be on Venmo, not PayPal.", "long_pause_offsets": [ 78 ], "source_sha256": "581ec569270ee8d19c0704433dc51834ff415b458ff456304776cb1fda9bce29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from PayPal to Venmo, while retaining the rest of the instruction including the transaction scope, time frame, and recipient group." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/revision/revision_01_app/7987f0197a344937/speech.wav", "wer": 0.0, "wav_sha256": "51c84033e7e42851568c834754a5768c7866af99b002a9a70f5b8bb94a64536a", "seconds": 16.96, "spoken_utterances": [ "Like all the uh, let me think... PayPal transactions of the ongoing month ... to and from my friends. ", "Wait, the transactions should be on ... Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal transactions' but immediately corrects themselves to 'Venmo not PayPal' after a pause. This is a revision where the final intent is to use Venmo, which matches the original instruction. The correction is clear and the final goal aligns with the reference.", "heard_transcript": "Like all the uh let me think. PayPal transactions of the ongoing month to and from my friends. Wait, the transactions should be on Venmo not PayPal." }, "candidate_id": "wav_fd0e5543d75b7020", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker starts with a clear initial request, uses a believable filler ('uh let me think'), then makes a clear mistake ('PayPal'), immediately self-corrects with a plausible reason ('Wait, the transactions should be on Venmo not PayPal'), and maintains consistent prosody and pacing throughout. The correction is integrated smoothly into the utterance, and the intonation reflects genuine self-correction rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing month only to my friends. That should be to and from, not only to.", "long_pause_offsets": [ 73 ], "source_sha256": "a64fa64369014f802b82ae270ec9ea955b9ae05fff82a50b9084a528170a8fb6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the direction from 'only to' to 'to and from', which matches the original instruction. The app, time scope, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/revision/revision_03_transaction_direction/b34d72291b443603/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "42d7aaa5de89ab0f3f73538412fc772603244918d0418dcea0d8bc928a28d5cd", "seconds": 15.4, "spoken_utterances": [ "Like all the venmo transactions of the um... ongoing month only to my friends. ", "That should be uh, let me think... to and from, not only to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to my friends' but then corrects themselves with 'That should be uh let me think, to and from, not only to.' This constitutes a revision where the final intent is to include both 'to and from' my friends, which matches the original instruction. The correction is clear and applied to the earlier instruction, preserving all other conditions.", "heard_transcript": "Like all the Venmo transactions of the um ongoing month only to my friends. That should be uh let me think, to and from, not only to." }, "candidate_id": "wav_cc958266cea0d9d5", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a believable revision with a self-correction ('only to' changed to 'to and from'), but the correction is less integrated than in audio_02. The speaker uses 'let me think' as a filler, which is natural, but the correction is delivered as a separate sentence rather than being embedded in the original request. The prosody is slightly less fluid, and the pause before the correction feels a bit longer than in the top-ranked clip, making the flow slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the last month to and from my friends. Make that the ongoing month, not last month.", "long_pause_offsets": [ 74 ], "source_sha256": "628074ee93ea8cb2039c3c689dc461692bb4d1b118772e061a1a84275eabcaae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the time reference from 'last month' to 'ongoing month', preserving the app, action, direction, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/revision/revision_02_time_reference/a6c4e7c4d60d6ac8/speech.wav", "wer": 0.0, "wav_sha256": "3f75e5d37b793e51ac888be2d9b6b03c26bee42d9da72223fa3069a8d716c376", "seconds": 13.64, "spoken_utterances": [ "Like all the ... venmo transactions of the last month um... to and from my friends. ", "Make that the ... ongoing month, not last month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last month' but immediately corrects it to 'ongoing month' with the phrase 'Make that the ongoing month, not last month.' This constitutes a revision where the final intent is to use 'ongoing month' instead of 'last month.' The rest of the instruction, including the app (Venmo), the action (transactions), and the target (to and from my friends), remains consistent. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Like all the Venmo transactions of the last month, um, to and from my friends. Make that the ongoing month, not last month." }, "candidate_id": "wav_dd5260a5007db527", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker first says 'last month' and then corrects to 'ongoing month'. The correction is clear and the speaker uses 'make that' to signal the change, which is pragmatic. However, the initial request is slightly less fluent, and the correction feels a bit more like a post-hoc addition than a natural thought process. The prosody is somewhat flat, and the overall delivery is less dynamic compared to the top two." } ] }, "is_default": false }, { "variant_id": "revision_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing month to and from my family. Use friends instead of family.", "long_pause_offsets": [ 76 ], "source_sha256": "ba3f2e8851e868a7f8b0bbe1ae5b30d2c01c23e17cebcfe9aa5da17d7833448f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the group from 'family' to 'friends', preserving the app, time scope, and action. The original instruction's 'friends' is restored." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/revision/revision_04_counterparty_group/0a402c9894c85793/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "5df6a28417f3709141751871b8f8aa168c79a07c15a25b837426d62820dd96ce", "seconds": 12.6, "spoken_utterances": [ "Like all the venmo transactions of the ah... ongoing month to and from my family. ", "Use uh... friends instead of family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'ongoing months' and 'my family', then corrects both to 'ongoing month' and 'my friends' after a pause. This is a revision where the final intent is the corrected version. The correction is clear and applies to the relevant parts of the instruction, so the final goal matches the original reference instruction.", "heard_transcript": "Like all the Venmo transactions of the ongoing months to and from my family. Use friends instead of family." }, "candidate_id": "wav_7e6097f79bcfe8c8", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has the least realistic delivery. It begins with a clear request but then abruptly cuts off mid-sentence ('...to and from my family. Use it...'). The second part sounds like a separate, disconnected instruction rather than a continuation of the original thought. The prosody is inconsistent, and the transition between the two parts is jarring, suggesting a possible edit or splice rather than a natural self-correction. The lack of a coherent flow significantly reduces its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_007", "hide_variant_original": true, "source_sha256": "a0a5eb55f2599b984a82cffa7e600b75b9828f2b606f32410ff7ef27f4173009", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/cancellation/cancel_007/bf95c09fcd64bc36/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "d123b87627a4160b68c68a9e24c3e621701e392b4966c5884b8859370e49095a", "seconds": 14.2, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends.", "Oh, wait, cancel my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first issues the original task and then immediately follows with a clear cancellation command: 'Oh wait, cancel my last request.' This constitutes a valid cancellation as per the type-specific criterion, which requires a clear revocation of the task. The final interjection is not ambiguous and explicitly cancels the prior instruction.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends. Oh wait, cancel my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_027", "hide_variant_original": true, "source_sha256": "8483fb3e209d3f859370b562238a25ec072d6df71fdc13089a72144dc8faf993", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/ambiguous_halt/halt_027/3deaf7443081a92b/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "525b2902b34ce71c963ef3ee65080ff9a4c4ac44d81473747b22eaa06cb7fd8a", "seconds": 13.68, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends.", "Hang on, I need a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hang on, I need a moment,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker does not resume or correct it. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends. Hang on, I need a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "27476db2a76a9a9f8027f3dfbd27bdcee09856fd2915c992bf9c408a6b1cdcca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_2/audio_natural/filler/original/91e07ba076101837/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "36f90ecada7096246bd6f5220da9456922e89f3e2223c72cc7b5ddb3b6da6af3", "seconds": 9.56, "spoken_utterances": [ "Like all the ... Venmo transactions of the hmmm... ongoing month to and from my friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker delivers the full instruction clearly and without interruption. The final intent matches the original instruction, and the filler type allows for minor hesitations or pauses that do not change the goal. Therefore, the task should be executed as stated.", "heard_transcript": "Like all the Venmo transactions of the ongoing month to and from my friends." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family—I should say coworkers.", "long_pause_offsets": [], "source_sha256": "32b9cdfc91e8655a5312ede11cbe97225e833615c5e2af8ca9c9d5d9870eb4dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'coworkers' within the same utterance, preserving the app, time scope, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/false_start/false_start_04_counterparty_group/3e3053a8be8a8438/speech.wav", "wer": 0.2, "wav_sha256": "3598b1396345321be46694bdc299f204bc06c3be5329f9132b23daea954c34c9", "seconds": 11.28, "spoken_utterances": [ "Like all the venmo transactions of the uh... ongoing year to and from my ... family—ah... I should say coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'familia' but immediately self-corrects to 'coworkers'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original instruction is discarded. The final audible intent matches the original instruction after the correction.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my famlia. I should say coworkers." }, "candidate_id": "wav_e37e21e8fc7363b4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects 'family' to 'coworkers' with a smooth, spontaneous-sounding false start. The prosody is consistent, and the correction is integrated into the flow without awkwardness or hesitation." } ] }, "is_default": true }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the last year—no, that should be ongoing year to and from my coworkers.", "long_pause_offsets": [], "source_sha256": "00105971845a90483148b7c9cecc165e41e5fac779babfafee94e446f696d091", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'last year' to 'ongoing year' within the same utterance, preserving the app, action, direction, and recipient scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/false_start/false_start_02_time_reference/53dcf93e992987f5/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "ae4fe3274b73a34df8a2cf6416c8cd804a173b8e1cb0bebdf08e8d812ce0ae1a", "seconds": 8.64, "spoken_utterances": [ "Like all the venmo transactions of the last year—uh... no, that should be ... ongoing year to and from my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'ongoing year'. According to the 'false_start' type, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the reference exactly, so the task should be executed as stated.", "heard_transcript": "Like all the Venmo transactions of the last year. No, that should be ongoing year to and from my coworkers." }, "candidate_id": "wav_fd0681fa201b0fe1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a plausible request, then corrects 'last year' to 'ongoing year' with a natural-sounding revision. The pause and intonation before the correction sound like genuine reconsideration, and the delivery is fluent and continuous." } ] }, "is_default": false }, { "variant_id": "false_start_01_app", "proposed_instruction": "Like all the PayPal—I meant venmo transactions of the ongoing year to and from my coworkers.", "long_pause_offsets": [], "source_sha256": "d9b96fcb6d2700680cd9cce2d2fae31faed43649d04d9e418ae2f1ff2627f6da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'PayPal' to 'Venmo' within the same utterance, preserving the rest of the instruction including the app, time frame, and direction of transactions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/false_start/false_start_01_app/7242a0a1cf1c757a/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "b79482f1fc5a947f7e49c8971d81c006eaf79f42a17561fc2223d1384977c6a8", "seconds": 8.08, "spoken_utterances": [ "Like all the PayPal—uh... I meant venmo transactions of the ongoing year um... to and from ... my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately self-corrects to 'Venmo' with 'I meant'. This is a false start where the final intent is the corrected version. The rest of the instruction matches the original, so the final audible task is to like all Venmo transactions of the ongoing year to and from coworkers. The correction is immediate and clear, satisfying the false_start criterion.", "heard_transcript": "Like all the PayPal uh, I meant Venmo transactions of the ongoing year um to and from my coworkers." }, "candidate_id": "wav_23a848b3e841f25a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with 'PayPal', then corrects to 'Venmo' with a believable self-repair. However, the correction is slightly less fluid than in audio_01 and audio_02, with a longer pause and a more hesitant tone, making it sound less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to—or rather, to and from my coworkers.", "long_pause_offsets": [], "source_sha256": "78941d47d4b79c913b0e211d739c9fb12b69bf1bc1b1723a7dc32de868eaf69d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially restricts the scope to 'only to' but immediately corrects to 'to and from' using 'or rather', restoring the original bidirectional intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/false_start/false_start_03_transaction_direction/dd773d5e7ec11b32/speech.wav", "wer": 0.25, "wav_sha256": "2ff065031e81fc3aea31016982769332f8ddbd76017c718f3e573bbe161157e0", "seconds": 7.28, "spoken_utterances": [ "Like all the venmo transactions of the uh... ongoing year only to—ah... or rather, to and from ... my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to' but immediately self-corrects with 'rather to and from'. The final audible intent matches the original instruction, and the correction is immediate and clear. The 'false_start' type requires applying the correction, which is done here.", "heard_transcript": "Like all the Venmo transactions of the uh ongoing year only to uh uh rather to and from my coworkers." }, "candidate_id": "wav_65ee428b8e6a4bd9", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request but then makes multiple, slightly disjointed corrections ('only to' to 'rather to and from'), with unnatural phrasing and a stutter-like repetition ('uh uh'). The flow is disrupted, and the corrections sound less like spontaneous speech and more like a rehearsed or awkward edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_03", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from. I left out a detail: only like transactions with my coworkers.", "long_pause_offsets": [ 65 ], "source_sha256": "99090b6a24eeff27af4b0e5e574bea513dfd48a562ba143a4fc8e5d0dfa4d526", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction 'with my coworkers', which narrows the initial broad request to the exact scope specified in the original instruction. The action, app, time frame, and direction are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/addition/addition_03/72b7845420453d70/speech.wav", "wer": 0.12, "wav_sha256": "128da53e13b651f2c115b306652dfd176783c54ca38891c16fbacdec63a1b4d5", "seconds": 11.48, "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing year ah... to and from. ", "I left out a detail: only like transactions with my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Like all the Venmo transactions of the ongoing year to and from my coworkers.' The speaker then adds a clarifying detail: 'Only like transactions with my coworkers.' This is a clear addition that narrows the scope of the original request. The final intent combines the original instruction with the added constraint, resulting in a more specific but coherent task. The addition is spoken after a pause and is not a revision or cancellation. Therefore, the task is executed with the combined intent.", "heard_transcript": "Like all the uh Venmo transactions of the ongoing year uh to and from. I left out a detail. Only like transactions with my coworkers." }, "candidate_id": "wav_83dfeea6bbdd2fbc", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Like all the venmo transactions to and from my coworkers. Oh, I forgot to mention, this should apply to all transactions of the ongoing year.", "long_pause_offsets": [ 58 ], "source_sha256": "77d9f4f4acd5195c109d45e97b44da671ff39f84062f6523afebd9ad383b2a03", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'of the ongoing year' to the initial request, which was otherwise complete. The final goal combines the original scope (Venmo, coworkers, transactions) with the new temporal restriction." }, "audio": { "status": "pending", "spoken_utterances": [ "Like all the venmo transactions uh... to and from hmmm... my coworkers. ", "Oh, I forgot to mention, this should apply to all transactions of the ... ongoing year." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "Like all the venmo transactions of the ongoing year my coworkers. Oh, and one more thing: include transactions going both to and from my coworkers.", "long_pause_offsets": [ 66 ], "source_sha256": "750f131d19e3755a1865565908d1537da37693258c329fae230442e04bf30b5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing directionality ('to and from') that was implied but not stated in the initial utterance, completing the original intent. The app, time frame, and subject remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "Like all the uh... venmo transactions of the ongoing year let's see... my coworkers. ", "Oh, and one more thing: include transactions going both to and from my coworkers." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_transaction_direction", "proposed_instruction": "Like all the venmo transactions of the ongoing year only to my coworkers. Wait, the transactions should be to and from, not only to.", "long_pause_offsets": [ 74 ], "source_sha256": "a5ea011d16d6beaf2c3fed61433f44ac34ac2cc66f650a3957f19fc680c5be13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the restriction from 'only to' to 'to and from', restoring the original bidirectional scope. The app, time frame, and recipient group remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/revision/revision_03_transaction_direction/c8bd037d3e525679/speech.wav", "wer": 0.04, "wav_sha256": "8892f5eee83e7ff7a0411b3593a1f00bdc74148e0bf1fee638e4352668071060", "seconds": 14.16, "spoken_utterances": [ "Like all the venmo transactions of the ongoing year uh... only to ... my coworkers. ", "Wait, the transactions should be ... to and from, not only to." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'only to my coworkers' but then corrects themselves with 'Wait, the transactions should be to and from not only to.' This is a clear revision where the final intent is to include both 'to and from' as specified in the original instruction. The revision applies to the earlier instruction, and the final goal matches the original instruction exactly.", "heard_transcript": "Like all the Venmo transactions of the ongoing year, uh, only to my coworkers. Wait, the transactions should be to and from not only to." }, "candidate_id": "wav_9ab1bfcc747812de", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then hesitates with 'uh' and 'wait,' followed by a self-correction that smoothly revises the scope from 'only to' to 'to and from.' The intonation and pacing reflect genuine reconsideration, and the correction is integrated into the same utterance without awkwardness. The prosody is consistent and the flow feels spontaneous." } ] }, "is_default": true }, { "variant_id": "revision_04_counterparty_group", "proposed_instruction": "Like all the venmo transactions of the ongoing year to and from my family. That should be transactions with coworkers, not family.", "long_pause_offsets": [ 75 ], "source_sha256": "965c37eb8b2cb720f626bc113bef50037813d8764047b3ab9c38c1f2203aefd9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the group from 'family' to 'coworkers', preserving the app (Venmo), time frame (ongoing year), and action (like transactions)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/revision/revision_04_counterparty_group/408f35b6ba70424c/speech.wav", "wer": 0.08, "wav_sha256": "ece96f89be544daabc59d4a6fcc2e626d52f220c38fab3d8e8f4b7d08cc9badc", "seconds": 12.6, "spoken_utterances": [ "Like all the let's see... venmo transactions of the ongoing year hmmm... to and from ... my family. ", "That should be transactions with coworkers, not family." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'my family' but immediately corrects themselves with 'That should be transactions with coworkers, not family.' This is a clear revision where the final intent is to target 'coworkers' instead of 'family'. The correction is direct and unambiguous, satisfying the revision type criterion. The rest of the instruction remains unchanged and is executed as intended.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my family. That should be transactions with coworkers, not family." }, "candidate_id": "wav_94234a844eae02f6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant 'let's see,' then makes a request, pauses with 'hmm,' and corrects the recipient from 'family' to 'co-workers.' The correction is delivered with a natural shift in intonation and a plausible change-of-mind tone. The pacing and fillers are realistic, though the initial hesitation is slightly more pronounced than in audio_01, making it feel slightly less fluid." } ] }, "is_default": false }, { "variant_id": "revision_02_time_reference", "proposed_instruction": "Like all the venmo transactions of the last year to and from my coworkers. Make that transactions from the ongoing year, not last year.", "long_pause_offsets": [ 75 ], "source_sha256": "0d09a71a95ae89c74eed12d6d160b820084d4e54485172ceaf7bec02abcd1c56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the time reference from 'last year' to 'ongoing year', while preserving the app (Venmo), action (like), and scope (transactions to and from coworkers)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/revision/revision_02_time_reference/b9ab87c4182e6519/speech.wav", "wer": 0.04, "wav_sha256": "aac8684adcaa79948e36d6fd5a9752159d68412074bc6183da01d9c0461ce2c2", "seconds": 14.84, "spoken_utterances": [ "Like all the ... venmo transactions of the um... last year to and from ... my coworkers. ", "Make that transactions from the ongoing year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'ongoing year' with 'Make that transactions from the ongoing year, not last year.' This is a clear revision where the final intent is to use 'ongoing year' instead of 'last year'. The rest of the instruction remains consistent. The revision type criterion is satisfied.", "heard_transcript": "Like all the Venmo transactions of the um, last year to and from my coworkers. Make that transactions from the ongoing year, not last year." }, "candidate_id": "wav_b74a5b744bf8729b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but includes a filler 'um' and a self-correction from 'last year' to 'ongoing year.' The correction is made in a separate sentence, which is less seamless than in audio_01. The prosody is natural, but the revision feels more like a post-hoc edit than a spontaneous thought process, reducing its realism slightly." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Like all the PayPal transactions of the ongoing year to and from my coworkers. Change the app to Venmo, not PayPal.", "long_pause_offsets": [ 79 ], "source_sha256": "8d1fd5602556f30db26c251711e52b553bcc68b57b32f0d6b8c513feabd56b42", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from PayPal to Venmo, preserving the transaction scope, time frame, and parties involved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/revision/revision_01_app/d68e0cff0f702434/speech.wav", "wer": 0.0, "wav_sha256": "12cb3c80e23c73c43740fdd395893366cc6016a7ce99e88b7224f91cc81a2f60", "seconds": 10.96, "spoken_utterances": [ "Like all the ... PayPal transactions of the ongoing year um... to and from my coworkers. ", "Change the app to Venmo, not PayPal." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'PayPal' but immediately corrects to 'Venmo' with the explicit phrase 'Change the app to Venmo, not PayPal.' This is a clear revision where the final intent is to apply the original instruction to Venmo instead of PayPal. The correction is unambiguous and directly modifies the target app, satisfying the revision type criterion.", "heard_transcript": "Like all the PayPal transactions of the ongoing year um to and from my coworkers. Change the app to Venmo, not PayPal." }, "candidate_id": "wav_8b620f8295f00e02", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker starts with a clear request but then makes a correction from 'PayPal' to 'Venmo.' The correction is delivered in a separate sentence, and the phrasing 'change the app to' sounds more like a command than a spontaneous realization. The prosody is consistent, but the delivery lacks the natural hesitation and flow of the others, making it sound more like a rehearsed or mechanical revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_002", "hide_variant_original": true, "source_sha256": "db6abbef4a8ae6a7dabae9063d80f97bb7a8d15b62636cd1ddcdf0c4361ce6c2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/cancellation/cancel_002/a8237ee8edf147a2/speech.wav", "wer": 0.0, "wav_sha256": "9a0cdf3dd6a5c2cfd0fb0dd7a6f1ccba8bbc813bae0e40933b6c1d15e2d06eb8", "seconds": 10.52, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers.", "Hold on, cancel the whole task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly ends with the speaker saying 'Hold on, cancel the whole task,' which is a direct and unambiguous cancellation of the entire preceding instruction. The initial request is fully revoked, satisfying the cancellation type criterion.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers. Hold on, cancel the whole task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_034", "hide_variant_original": true, "source_sha256": "a6b49f84292dc9e11bc9dac7aa0a1d8a9aafe262d24fd13018ffe55e33e979dd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/ambiguous_halt/halt_034/1193accd17f14299/speech.wav", "wer": 0.0, "wav_sha256": "ac637a3aa981ed12cb2e1ed3069695ddfbf7472f4a9306582c40eb230ea0d45f", "seconds": 10.0, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers.", "Wait, I was just thinking..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, I was just thinking...', which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection indicates a temporary halt without cancellation or execution. The initial request is not completed, but the final intent is to pause, not to execute or cancel.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers. Wait, I was just thinking..." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "df61dc5_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ceaa11393dd0b6a3e77c4f89d6faf65780de303749f981dffbabfa34c4a46695", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/df61dc5_3/audio_natural/filler/original/451df7f59f841b02/speech.wav", "wer": 0.0, "wav_sha256": "d85ac310211583fd4f0c8c1def91134c0a094ee54171ee954cb5c4520f145ede", "seconds": 6.04, "spoken_utterances": [ "Like all the ... Venmo transactions of the ... ongoing year to and from my coworkers." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains no hesitation, backchanneling, or self-correction that alters the task. The speaker completes the original instruction without any change in meaning or intent. The 'filler' type allows for such natural speech patterns, and the final audible task matches the original instruction.", "heard_transcript": "Like all the Venmo transactions of the ongoing year to and from my coworkers." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Add all spotify-recommended classical songs released in last year—no, that should be this year to a new \"Spotify Recommended Songs\" playlist.", "long_pause_offsets": [], "source_sha256": "62aad9cd8e00e90a806396f89a6bfaabe61a091fe53fef691d9610c6725ed8bf", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The immediate correction replaces 'last year' with 'this year'; all other conditions including genre, recommendation source, playlist creation, and action remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/false_start/false_start_03_time_reference/abb0e2da3cf1a23e/speech.wav", "wer": 0.12, "wav_sha256": "e3b6ec6d5bd43f1b31dc6d804e79c826bbf70d9062f8f17c3efbeaf80180b12d", "seconds": 12.8, "spoken_utterances": [ "Add all hmmm... spotify-recommended classical songs released in last year—um... no, that should be ... this year to a new \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference instruction in action, target, and constraint.", "heard_transcript": "Add all um, Spotify recommended classical songs released in last year. Um, no, that should be this year to a new Spotify recommended songs playlist." }, "candidate_id": "wav_d07d9c31f192cb88", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Spotify recommended classical songs released in last year', then self-corrects with 'No, that should be this year'—a believable and smoothly integrated revision. The intonation and pacing of the correction sound spontaneous, and the pause before the correction is appropriately placed. The overall prosody is consistent and conversational, with no signs of robotic delivery or unnatural pauses." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Add all Apple Music—I should say spotify-recommended classical songs released in this year to a new \"Spotify Recommended Songs\" playlist.", "long_pause_offsets": [], "source_sha256": "7b01ea91d3edc5d90332b5c0b90aaca45bcc1ed3ed56c090f63bad1ee5468392", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'spotify-recommended' within the same utterance, preserving the app, genre, release year, and playlist creation intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/false_start/false_start_01_app/951ecb7720daf3a8/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "e097c26edc3127f82964e511db768e58df9a5d33608edb4bed56ed983ad78dd5", "seconds": 13.84, "spoken_utterances": [ "Add all uh... Apple Music—um... I should say spotify-recommended classical songs released in ... this year to a new \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('Apple Music') is discarded. The rest of the instruction matches the reference exactly, including the target app, genre, release year, and playlist name. Therefore, the final audible intent is correct and matches the original instruction.", "heard_transcript": "Add all uh, Apple Music um, I should say Spotify recommended classical songs released in this year to a new Spotify recommended songs playlist." }, "candidate_id": "wav_aaf1d0b20e6ee274", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a false start with 'Apple Music' followed by a self-correction to 'Spotify recommended'. The repair is natural, but the initial error ('Apple Music') is more jarring than in audio_02, as it changes the platform entirely rather than just a minor detail like 'last year' vs 'this year'. The correction is spoken with a slightly rushed intonation, and the filler 'um' is used more frequently, making the delivery feel slightly less fluid than audio_02." } ] }, "is_default": false }, { "variant_id": "false_start_02_genre", "proposed_instruction": "Add all spotify-recommended jazz—or rather, classical songs released in this year to a new \"Spotify Recommended Songs\" playlist.", "long_pause_offsets": [], "source_sha256": "caf6f49f441eecaea8a679831bf9c1c39b4ab7ff6b352c2f98e7af9e662e10ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately corrects to 'classical' within the same utterance. The final genre is 'classical', and all other conditions (Spotify-recommended, released this year, new playlist) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/false_start/false_start_02_genre/6eb07589ef90f297/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "c926e9d5ea41ec90bd6010200ef92aec7bd9c1726762c9e9293bdb3d11fff0a6", "seconds": 14.0, "spoken_utterances": [ "Add all spotify-recommended ah... jazz—uh... or rather, classical songs released in this year to a new um... \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz' but immediately self-corrects to 'classical' with 'or rather'. This is a false start, and the final intent is the corrected version. The rest of the instruction matches the original, including the target playlist name and the condition 'released in this year'. The final control intent is to execute the corrected task.", "heard_transcript": "Add all Spotify recommended uh jazz uh or rather classical songs released in this year to a new um Spotify recommended songs playlist." }, "candidate_id": "wav_f4345fe7bdde9771", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has a false start with 'Spotify recommended jazz' followed by a correction to 'classical'. While the repair is present, the initial error is more significant (genre change), and the correction is delivered with a flatter, less expressive intonation. The filler 'um' is used awkwardly, and the overall pacing feels slightly hesitant and less natural compared to the other two. The prosody lacks the smoothness of a genuine self-correction, making it the least realistic of the three." } ] }, "is_default": false }, { "variant_id": "false_start_04_playlist_name", "proposed_instruction": "Add all spotify-recommended classical songs released in this year to a new \"My Favorite Tracks\"—scratch that, make it \"Spotify Recommended Songs\" playlist.", "long_pause_offsets": [], "source_sha256": "d668d4945a53d22e6d00768bf0ccc46b9668fef2564c775375e9cb5366496a53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests a playlist name \"My Favorite Tracks\" but immediately corrects it to \"Spotify Recommended Songs\". All other conditions—genre, recommendation source, release year, and action—are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Add all spotify-recommended classical songs released in this year to a new let me think... \"My Favorite Tracks\"—uh... scratch that, make it \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Add all spotify-recommended songs released in this year to a new \"Spotify Recommended Songs\" playlist. I left out a detail: only add classical songs.", "long_pause_offsets": [ 103 ], "source_sha256": "bc272aa82e0322b97fb53fd834ea2339b5a9f38242d05c0d5d9dba34b45c7576", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing genre restriction 'classical songs' to the initial broad request for 'spotify-recommended songs', preserving the app, release year, and playlist creation intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/addition/addition_02/7a0ea74067348125/speech.wav", "wer": 0.03125, "wav_sha256": "64cf5b20239935ac4d2b3dc6d16d1b3f62420bdf30555334f827a41554863a60", "seconds": 16.08, "spoken_utterances": [ "Add all uh, let's see... spotify-recommended songs released in this year to a new um... \"Spotify Recommended Songs\" playlist. ", "I left out a detail: only add let me think... classical songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is broad, requesting all Spotify-recommended songs from this year. After a pause, the speaker adds a clarifying constraint: 'only add classical songs'. This addition narrows the scope of the original request, and the final intent is to execute the original task with the added constraint. The addition is clear and directly modifies the initial instruction without introducing new conflicting goals.", "heard_transcript": "Add all uh let's see, Spotify recommended songs released in this year to a new um, Spotify recommended songs playlist. I left out a detail, only add, let me think, classical songs." }, "candidate_id": "wav_d2b1254f07694c75", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers ('uh', 'let's see', 'um') and a believable hesitation ('I left out a detail') before adding a specific restriction ('only add classical songs'). The self-correction is integrated smoothly into the flow, with a pause and 'let me think' that sounds like genuine recollection. The prosody is conversational and the pacing includes realistic pauses, making the addition feel spontaneous." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Add all classical songs released in this year to a new \"Spotify Recommended Songs\" playlist. Oh, I forgot to mention, these are Spotify-recommended songs.", "long_pause_offsets": [ 93 ], "source_sha256": "84012052a11b49a69017ddde8d216e7ed6bbeef943346bd2023dc367e318adea", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing Spotify-recommended condition, which refines the initial request. The action, target playlist, genre, and release year remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/addition/addition_01/1bfee30d5f8fa81d/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "05c31cffbbfc75e692a167dea10b171d7018ca92c470ec0317a592918fa20e0d", "seconds": 13.28, "spoken_utterances": [ "Add all ... classical songs released in ah... this year to a new \"Spotify Recommended Songs\" playlist. ", "Oh, I forgot to mention, these are Spotify-recommended songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is to add classical songs released this year to a new playlist. After a pause, the speaker adds a clarifying constraint: 'these are Spotify recommended songs.' This is a valid addition that narrows the scope of the initial request. The final intent combines both conditions, matching the reference instruction's goal. The addition is not a correction or revision but an extra constraint, so the 'addition' type criterion is satisfied.", "heard_transcript": "Add all classical songs released in this year to a new Spotify recommended songs playlist. Oh, I forgot to mention, these are Spotify recommended songs." }, "candidate_id": "wav_1b2f86a70805c999", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker delivers the initial request fluently but then adds a follow-up with 'Oh, I forgot to mention' and a clarification. While this is a plausible addition, the follow-up is delivered more abruptly and with less natural hesitation than in audio_01. The prosody is slightly more robotic, and the correction feels tacked on rather than emerging from a moment of recollection, reducing its realism compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_time_reference", "proposed_instruction": "Add all spotify-recommended classical songs released in last year to a new \"Spotify Recommended Songs\" playlist. Make that songs released this year, not last year.", "long_pause_offsets": [ 113 ], "source_sha256": "f904d2d8d30fbc4d60e0804cf53adc53fc6ee3ddf0097fd1dd353748d732be5e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the release year from 'last year' to 'this year', preserving the app, genre, action, and playlist name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/revision/revision_03_time_reference/b8c666c1545ace5c/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "fc62c3322a70b4d418e873f181889b7749004acdc456383179984a825609a533", "seconds": 14.72, "spoken_utterances": [ "Add all spotify-recommended classical songs released in ... last year to a new ... \"Spotify Recommended Songs\" playlist. ", "Make that songs released uh... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects to 'this year' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the playlist name and genre. Therefore, the task should be executed as revised.", "heard_transcript": "Add all Spotify recommended classical songs released in last year to a new Spotify recommended songs playlist. Make that songs released uh this year, not last year." }, "candidate_id": "wav_e0c939efc81909c3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear, slightly hesitant request, then corrects the year from 'last year' to 'this year' with a smooth, self-repairing intonation. The pause before the correction is brief and plausible, and the correction itself is integrated naturally into the sentence without awkwardness. The prosody is consistent, and the speaker sounds like they are formulating and adjusting a request in real time." } ] }, "is_default": true }, { "variant_id": "revision_02_genre", "proposed_instruction": "Add all spotify-recommended jazz songs released in this year to a new \"Spotify Recommended Songs\" playlist. I misspoke about the genre: add classical songs, not jazz.", "long_pause_offsets": [ 108 ], "source_sha256": "26eaaaa46633edad3e3f2540881bc4b747578cfbd8560b110aede7edf644a222", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the genre from jazz to classical, while retaining the app, recommendation source, release year, and playlist name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/revision/revision_02_genre/b33a2209763f2da0/speech.wav", "wer": 0.15625, "wav_sha256": "e17d2ceba10334c284931e062adc8283da63039812c3fd9e9653bf09ec7fb702", "seconds": 17.04, "spoken_utterances": [ "Add all um, let's see... spotify-recommended jazz songs released in uh... this year to a new um... \"Spotify Recommended Songs\" playlist. ", "I misspoke about the genre: add classical songs, not jazz." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'jazz songs' but immediately corrects themselves with 'I misspoke about the genre, add classical songs not jazz.' This is a clear revision where the final intent is to add classical songs, not jazz. The rest of the instruction, including the app (Spotify), the source ('recommended songs'), the condition ('released in this year'), and the target playlist name, remains consistent. The revision applies the correction to the earlier instruction, and the final goal matches the original reference instruction.", "heard_transcript": "Add all um let's see Spotify recommended jazz songs released in uh this year to a new um Spotify recommended songs playlist. I misspoke about the genre, add classical songs not jazz." }, "candidate_id": "wav_174a5b5a47573016", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision with a self-correction of the genre from 'jazz' to 'classical'. The initial request is delivered with natural hesitations ('um', 'uh') and a conversational tone. The correction is spoken immediately after the initial statement, with a plausible shift in intonation indicating a realization. The flow is mostly smooth, though the correction is slightly more abrupt than in audio_03, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Add all apple music-recommended classical songs released in this year to a new \"Spotify Recommended Songs\" playlist. Wait, use Spotify, not Apple Music.", "long_pause_offsets": [ 117 ], "source_sha256": "14c14dd040c39a086d25e18b0b75e31b74e26ccbf441f3fc0154a0ea5452c69f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the genre, release year, playlist name, and action. The correction is explicit and unambiguous." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/revision/revision_01_app/2e88af12355d50f4/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "6dd0bf16164cb97d6c4bfa06749ff8b09e76bc4a5da549677905b8698deb5cae", "seconds": 15.32, "spoken_utterances": [ "Add all apple music-recommended uh... classical songs released in um... this year to a new \"Spotify Recommended Songs\" playlist. ", "Wait, use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the source as 'Apple Music' but immediately self-corrects to 'Spotify' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the action, target, genre, release year, and playlist name. Therefore, the task is executed as intended.", "heard_transcript": "Add all Apple Music recommended uh classical songs released in um this year to a new Spotify recommended songs playlist. Wait, use Spotify not Apple Music." }, "candidate_id": "wav_90ecf2bdc4c80951", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker first mentions 'Apple Music' and then corrects to 'Spotify'. However, the correction is delivered with a noticeable pause and a more abrupt shift in tone, sounding less like a spontaneous self-correction and more like a deliberate rephrasing. The initial part of the sentence is slightly rushed, and the overall prosody feels less natural compared to the other two clips. The revision is clear but less seamlessly integrated." } ] }, "is_default": false }, { "variant_id": "revision_04_playlist_name", "proposed_instruction": "Add all spotify-recommended classical songs released in this year to a new \"My Favorite Tracks\" playlist. Change the playlist name to \"Spotify Recommended Songs\" instead of \"My Favorite Tracks\".", "long_pause_offsets": [ 106 ], "source_sha256": "07278a6020cbf39f66f2ef9ff0ebc02e361e882e2d12c55e1eb44130cfd75fbb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly renames the playlist from \"My Favorite Tracks\" to \"Spotify Recommended Songs\". All other requirements—genre, recommendation source, release year, and playlist creation—are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Add all um... spotify-recommended classical songs released in this year to a new ... \"My Favorite Tracks\" playlist. ", "Change the playlist name to \"... Spotify Recommended Songs\" instead of \"My Favorite Tracks\"." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_027", "hide_variant_original": true, "source_sha256": "3316b2d1803626121d417827294f29a695244c6f2c30abb01da93665f5727799", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/cancellation/cancel_027/bcfb2afde762ac40/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "c5443db00fc06cd2901ea48aa2ce4651a5ba2c0c100eb89ab04e5cee2444c1ae", "seconds": 17.04, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist.", "Oh, wait, please disregard my last request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first makes a request to add songs to a playlist, then immediately follows with 'Oh wait, please disregard my last request,' which is a clear and unambiguous cancellation of the prior instruction. The final interjection is a direct cancellation command, satisfying the criterion for the 'cancellation' type. The initial request is irrelevant once the cancellation is stated.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist. Oh wait, please disregard my last request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_019", "hide_variant_original": true, "source_sha256": "f4acd0b227fdb208b38d6b62a78cc71898581978a0eef0a8b782130fe39dcce8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/ambiguous_halt/halt_019/f0a3e881b8d5688b/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "2ae1a841617b9033ddf74288faf6af0c192c0e8300138b11838e2ea690ff0f53", "seconds": 16.0, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist.", "Can you give me a moment?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the initial request with 'Can you give me a moment?', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the speaker does not execute or modify the task. The final intent is to halt execution temporarily, not to cancel or proceed.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist. Can you give me a moment?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "65f40a852549fb7fbe8bcc5e95cd4cfa12e57e5b8d38976e06e641fab48e2235", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_1/audio_natural/filler/original/d184f4512e14b67d/speech.wav", "wer": 0.05, "wav_sha256": "f654b804a2fe68d48107a246f582fffa4cefa60a612136b8c711fd24d8df6b48", "seconds": 12.4, "spoken_utterances": [ "Add all uh... spotify-recommended ah... classical songs released in uh... this year to a new \"Spotify Recommended Songs\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh') but does not change the core request. The final intent matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "Add all uh Spotify recommended uh classical songs released in uh this year to a new Spotify recommended songs playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "Add all Apple Music—I should say spotify-recommended R&B songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist.", "long_pause_offsets": [], "source_sha256": "5622f340535a039de3ba65fdb3969ee2c7afcb64b4d3b8462a32b4d549bf898f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'spotify-recommended' within the same utterance, preserving the genre, release timeframe, playlist name, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/false_start/false_start_01_app/378fe263f5833ac0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "296fae5c8735280b87b292bd1e9c6e34415d146a84cbc3b9858f886ed1a70d4b", "seconds": 14.2, "spoken_utterances": [ "Add all ... Apple Music—um... I should say spotify-recommended R&B songs released in ... this or last year to a new \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('Apple Music') is discarded. The rest of the instruction, including the genre (R&B), release timeframe ('this or last year'), and playlist name ('Spotify R&B Recommendations'), is preserved and matches the original instruction. Therefore, the final audible intent is correct and complete.", "heard_transcript": "Add all Apple Music um, I should say Spotify recommended R&B songs released in this or last year to a new Spotify R&B Recommendations playlist." }, "candidate_id": "wav_7e4fcc6b9829ce60", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural false start with 'Apple Music' followed by a self-correction to 'Spotify recommended R&B songs'. The correction is smooth and the prosody remains consistent. The speaker's voice is clear, and the pacing feels like a genuine moment of recall. The flow from initial error to correction to completion is the most seamless and realistic among the four." } ] }, "is_default": true }, { "variant_id": "false_start_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this or last year to a new \"R&B Mix\"—scratch that, make it \"Spotify R&B Recommendations\" playlist.", "long_pause_offsets": [], "source_sha256": "5a76fd682c5452fd8fd4b0fb28f0c59c985953ae4da4882455cc6043ca9321d7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests a playlist name \"R&B Mix\" but immediately corrects it to \"Spotify R&B Recommendations\". All other conditions—genre (R&B), source (Spotify-recommended), release timeframe (this or last year), and action (add to a new playlist)—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/false_start/false_start_04_playlist_name/0e8cec0c34ab95f1/speech.wav", "wer": 0.0, "wav_sha256": "d08b761f2bc57cb7fcccc8f32110b2d68a92e7cddf3d2dc4411895989e5b38e7", "seconds": 14.8, "spoken_utterances": [ "Add all spotify-recommended R&B songs released in um... this or last year to a new \"R&B Mix\"—uh... scratch that, make it \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B mix' but immediately self-corrects with 'scratch that, make it Spotify R&B Recommendations playlist'. According to the 'false_start' type criterion, the final corrected target should be applied. The final target matches the original instruction's playlist name exactly, and all other conditions (Spotify-recommended, R&B, released in this or last year) are preserved. The correction is immediate and unambiguous, so the task is executed as intended.", "heard_transcript": "Add all Spotify recommended R&B songs released in um this or last year to a new R&B mix uh, scratch that, make it Spotify R&B Recommendations playlist." }, "candidate_id": "wav_9d283ba023a52573", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a clear false start with 'Spotify recommended R&B songs' and a self-correction to 'Spotify R&B recommendations playlist'. The correction is immediate and the intonation shift is natural. The speaker's voice is steady, and the delivery is fluent. The only minor issue is that the correction is slightly abrupt, but overall it sounds like a spontaneous correction." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Add all spotify-recommended R&B songs released in last year—no, that should be this or last year to a new \"Spotify R&B Recommendations\" playlist.", "long_pause_offsets": [], "source_sha256": "abc310645363a6e88dd3fbd226b0e1cd82f6b97aa8cfe7085147682933a03f2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this or last year'. The final intended time range matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/false_start/false_start_03_time_reference/502dcede2a67317a/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "fb4e5e873620f5321ede50daf028b13cc2a356c07b97cc1c3f5cf15cca78eb39", "seconds": 13.08, "spoken_utterances": [ "Add all ... spotify-recommended R&B songs released in last year—uh... no, that should be this or last year to a new \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this or last year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly, with only a minor grammatical paraphrase ('Spotify recommended' vs 'spotify-recommended'). No other constraints were altered.", "heard_transcript": "Add all Spotify recommended R&B songs released in last year, um, no, that should be this or last year to a new Spotify R&B Recommendations playlist." }, "candidate_id": "wav_c9a0d9ef376dba00", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio includes a false start with 'last year' followed by a correction to 'this or last year'. The correction is clear and the speaker's tone reflects a moment of reconsideration. However, the initial phrase 'last year' is spoken with a slightly unnatural emphasis, and the correction feels a bit rushed compared to the more organic flow in the top two." } ] }, "is_default": false }, { "variant_id": "false_start_02_genre", "proposed_instruction": "Add all spotify-recommended pop—or rather, R&B songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist.", "long_pause_offsets": [], "source_sha256": "7a008cb0c5ebd8da790676981963ef91d33830c066b25031a476cf5502b2dbb0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'pop' but immediately corrects to 'R&B' within the same utterance. The final genre is R&B, and all other conditions (source, release timeframe, playlist creation) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/false_start/false_start_02_genre/4ecd05bfeb10f006/speech.wav", "wer": 0.0, "wav_sha256": "8eeaa1d4d77b12c2aca7d7e001f4cfef2b71d211bacac52a05cd19981855c384", "seconds": 15.6, "spoken_utterances": [ "Add all uh... spotify-recommended pop—um... or rather, R&B songs released in this or last year to a new uh, let me think... \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pop' but immediately self-corrects to 'R and B'. The final intent is clear and matches the original instruction. The hesitation and filler words ('uh', 'let me think') do not change the goal. The task is executed as per the false_start type criterion, where the immediate correction overrides the initial mistake.", "heard_transcript": "Add all uh Spotify recommended pop um, or rather R and B songs released in this or last year to a new uh, let me think, Spotify R and B recommendations playlist." }, "candidate_id": "wav_43c568c949c2e97c", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains multiple hesitations ('uh', 'um', 'let me think') and a false start ('pop'), which makes it sound less fluent. The speaker seems to be struggling to formulate the request, and the pauses are longer and more awkward. The correction from 'pop' to 'R&B' is less smooth, and the overall delivery feels more hesitant and less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "Add all spotify-recommended R&B songs to a new \"Spotify R&B Recommendations\" playlist. I left out a detail: only include songs released in this or last year.", "long_pause_offsets": [ 87 ], "source_sha256": "a8e523b2683e86f71f2b95f4de12bed05e7cd0a450edb973418c6027380affec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing release-year constraint, which complements the initial instruction without contradicting any other requirement. The app, genre, playlist name, and action remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/addition/addition_02/b94a627fa9f31c80/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "fdd9fe1db9bc72ffa9b5696e89ebeccb6421643272ae955b25fb94e2b043f6df", "seconds": 14.32, "spoken_utterances": [ "Add all spotify-recommended ... R&B songs to a new uh... \"Spotify R&B Recommendations\" playlist. ", "I left out a detail: only include songs released in this or last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Add all Spotify recommended R&B songs to a new Spotify R&B recommendations playlist.' The speaker then adds a constraint: 'only include songs released in this or last year.' This is a clear addition of a new condition to the original task. The final intent combines the original action with the added constraint, matching the reference instruction. The type is 'addition', and the final goal is fully specified and executable.", "heard_transcript": "Add all Spotify recommended R&B songs to a new uh Spotify R&B recommendations playlist. I left out a detail, only include songs released in this or last year." }, "candidate_id": "wav_fab53e4643ac8464", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "Add all R&B songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist. Oh, I forgot to mention, these are Spotify-recommended songs.", "long_pause_offsets": [ 97 ], "source_sha256": "ef7abea9f18408388acd8bb0ae3568781532d6bce07650619fc6478ae26f1c3a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'Spotify-recommended' restriction, which was omitted in the initial utterance but is present in the original instruction. The genre, release timeframe, and playlist creation are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "Add all R&B songs released in ... this or last year to a new \"Spotify R&B Recommendations\" playlist. ", "Oh, I forgot to mention, these are um, let's see... Spotify-recommended songs." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this or last year to a new \"R&B Hits 2023\" playlist. Use \"Spotify R&B Recommendations\" instead of \"R&B Hits 2023\".", "long_pause_offsets": [ 103 ], "source_sha256": "77e32f51d5e748ec140d9401856970cb566cdb93dd4c8f438813247354696798", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision replaces the playlist name from \"R&B Hits two thousand twenty-three\" to \"Spotify R&B Recommendations\" while preserving all other conditions: genre (R&B), source (Spotify-recommended), release timeframe (this or last year), and the creation of a new playlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/revision/revision_04_playlist_name/3f4f401d2269d936/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "9a7e309d74755099d2669ec1e2d0dfb1b4ee11db0db9b4875c20bb09c2c8e13e", "seconds": 20.96, "spoken_utterances": [ "Add all spotify-recommended um... R&B songs released in this or last year to a new um, let's see... \"R&B Hits two thousand twenty-three\" playlist. ", "Use \"um, let me think... Spotify R&B Recommendations\" instead of \"R&B Hits two thousand twenty-three\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B hits 2023' as the playlist name but then corrects it to 'Spotify R&B Recommendations' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent matches the original instruction with the corrected playlist name.", "heard_transcript": "Add all Spotify recommended um R&B songs released in this or last year to a new um let's see, R&B hits 2023 playlist. Use um let me think, Spotify R&B Recommendations instead of R&B hits 2023." }, "candidate_id": "wav_675ccf432f756ab4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and continuous flow. The speaker uses multiple fillers ('um', 'let me think') at plausible points, with a clear self-correction from 'R&B hits 2023' to 'Spotify R&B recommendations'. The intonation and rhythm are consistent with someone formulating a request in real time, and the correction is smoothly integrated into the sentence without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Add all spotify-recommended R&B songs released in next year to a new \"Spotify R&B Recommendations\" playlist. Change 'this or last year' to 'next year'.", "long_pause_offsets": [ 109 ], "source_sha256": "351be8b6e384c5de8250d5b689601fb95162b4e4f671f0c6ab8d94b690c232b6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly revises the time condition from 'this or last year' to 'next year', replacing the original constraint. The app, genre, action, and playlist name remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/revision/revision_03_time_reference/ad74fffbb2e4cb4f/speech.wav", "wer": 0.05714285714285714, "wav_sha256": "5a2fe959304e2ad28c4e0abbebfd4b92bbb945d51262f73a5ca263e6f60ef447", "seconds": 18.32, "spoken_utterances": [ "Add all spotify-recommended R&B songs released in um, let me think... next year to a new um, let's see... \"Spotify R&B Recommendations\" playlist. ", "Change 'um... this or last year' to 'next year'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this or last year' but then explicitly corrects it to 'next year' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent is to add songs from 'next year', which matches the revised goal. All other constraints (action, target, playlist name) remain consistent with the original instruction.", "heard_transcript": "Add all Spotify recommended R&B songs released in um, let me think, next year to a new um, let's see, Spotify R&B recommendations playlist. Change um, this or last year to next year." }, "candidate_id": "wav_c7cd25c2bfbfaa17", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip has a natural hesitation pattern with 'um' and 'let me think', and includes a revision from 'this or last year' to 'next year'. However, the correction is less smoothly integrated than in audio_03, and the phrase 'change um this or last year to next year' sounds slightly more like a meta-instruction than a spontaneous thought process, reducing its realism." } ] }, "is_default": false }, { "variant_id": "revision_02_genre", "proposed_instruction": "Add all spotify-recommended pop songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist. Make that R&B, not pop.", "long_pause_offsets": [ 117 ], "source_sha256": "35327ec2becd03e5e7ba4c1e3b1368d1636fbe62f8a31144adb0742e8cf485e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the genre from 'pop' to 'R&B' while preserving the app, recommendation source, release timeframe, and playlist name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/revision/revision_02_genre/8574dab40c4c7ef8/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "f00005c1b093d69a2feaa3a0fa97edff5846e35659225de4d4a0b4e65b32264d", "seconds": 17.56, "spoken_utterances": [ "Add all ... spotify-recommended pop songs released in this or last year to a new uh... \"Spotify R&B Recommendations\" playlist. ", "Make that let's see... R&B, not pop." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pop songs' but immediately self-corrects to 'R&B, not pop' after a brief pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with the genre corrected from pop to R&B.", "heard_transcript": "Add all Spotify-recommended pop songs released in this or last year to a new uh Spotify R&B recommendations playlist. Make that, let's see, R&B, not pop." }, "candidate_id": "wav_8b2422381e724b54", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio has the most robotic and disjointed delivery. The speaker says 'make that let's see R&B not pop' as a separate clause, which sounds like a post-hoc correction rather than a natural self-repair. The prosody is flatter, and the transition between the initial request and the revision feels abrupt and less conversational compared to the other two." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Add all apple music-recommended R&B songs released in this or last year to a new \"Spotify R&B Recommendations\" playlist. Wait, use Spotify, not Apple Music.", "long_pause_offsets": [ 121 ], "source_sha256": "bd288345df9f230b71029bdd494665db59c3b5f0d664291b392671b6a891f18b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the genre, release timeframe, and playlist name." }, "audio": { "status": "pending", "spoken_utterances": [ "Add all apple music-recommended R&B songs released in ... this or last year to a new um... \"Spotify R&B Recommendations\" playlist. ", "Wait, use ... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "fc76128583693cc0d5a0327648ab349f24637601cd844c22c683a9997bc17af2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/cancellation/cancel_003/3df2e8f97093862c/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "ae4f9fad478395bc7cbfc8f5db8058435d9c4308e9f269846097e6eefc9b028d", "seconds": 15.88, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist.", "Oh, wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The recording ends with the explicit cancellation request 'Oh wait, please cancel that.' This clearly revokes the previously stated task. The cancellation is not ambiguous and applies to the entire task as originally described. The final intent is to cancel, satisfying the 'cancellation' type criterion.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist. Oh wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_012", "hide_variant_original": true, "source_sha256": "62e0bd8112f3a1944f4b534e60b3e35f3f69fdc2036c306318b82ffd547582d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/ambiguous_halt/halt_012/99e2eb93a153269e/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "37975659926a24570c1d1a9424bac1fec20c9ca90e906ff4c8f83ff41603f4c8", "seconds": 15.0, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist.", "Give me a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Give me a second,' which is a clear request to pause or wait. This is a final interjection that does not cancel the task or change its goal. The initial request is fully audible and complete, and the speaker does not attempt to revise, add to, or cancel it. Therefore, the final intent is to pause, matching the 'ambiguous_halt' type criterion.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist. Give me a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "06a859e365dca15a5d1bed550ba0c72734e39546586b571a8c458c6dd1e19b16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_2/audio_natural/filler/original/7220793e4131b5b9/speech.wav", "wer": 0.08, "wav_sha256": "a37c53db4e83ff3969753caa0e30f26cfc1d1eb2e64e0dad8fa0ef9dfb580486", "seconds": 11.96, "spoken_utterances": [ "Add all spotify-recommended uh, let's see... R&B songs released in um... this or last year to a new \"Spotify R&B Recommendations\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'let's see', 'um') but does not change the core request. The final intent matches the original instruction exactly, with no corrections or modifications. The fillers are typical hesitation markers and do not alter the task goal.", "heard_transcript": "Add all Spotify recommended uh let's see R&B songs released in um this or last year to a new Spotify R&B recommendations playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_genre", "proposed_instruction": "Add all spotify-recommended hip-hop—or rather, R&B songs released in this year to a new \"R&B Recommendation\" playlist.", "long_pause_offsets": [], "source_sha256": "2263d6dab590e6760cd2fe73c7cae5e0db4554a9e0cff9aca35ae73bab31e120", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'hip-hop' but immediately corrects to 'R&B' within the same utterance. The final genre is R&B, matching the original instruction. All other elements—source (Spotify-recommended), release year condition, and playlist creation—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/false_start/false_start_02_genre/677e7a3bcdd2f227/speech.wav", "wer": 0.12, "wav_sha256": "bf0be965fc3a28e73c2d5097ece0c36420e0f7ceb2458d3472a161e1e48bee9a", "seconds": 14.48, "spoken_utterances": [ "Add all spotify-recommended hip-hop—um... or rather, R&B songs released in this year hmmm... to a new uh... \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'hip hop' but immediately self-corrects to 'R&B' with 'or rather'. The final intent is clear and matches the original instruction. The correction is immediate and unambiguous, satisfying the false_start type criterion. All other constraints (Spotify-recommended, released in this year, new playlist named 'R&B Recommendation') are preserved.", "heard_transcript": "Add all Spotify recommended hip hop um, or rather R&B songs released in this year to a new uh R&B recommendation playlist." }, "candidate_id": "wav_f84c66affc58eb13", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'hip-hop', hesitates with 'um', then self-corrects to 'or rather, R&B' in a way that sounds spontaneous and conversational. The prosody is consistent, with a natural rise and fall in intonation, and the correction flows seamlessly into the rest of the sentence without awkward pauses or robotic phrasing. The filler 'um' is placed at a typical point of hesitation, and the overall rhythm feels like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "false_start_01_app", "proposed_instruction": "Add all Apple Music—I should say spotify-recommended R&B songs released in this year to a new \"R&B Recommendation\" playlist.", "long_pause_offsets": [], "source_sha256": "353f8948a98e3dcea6e989484a9cc76ac2703ecd1a6d2e4b1cce4c2398b87fac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify-recommended' within the same utterance, preserving the genre, release year, playlist name, and action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/false_start/false_start_01_app/bb4b0f8bf4bf41a4/speech.wav", "wer": 0.04, "wav_sha256": "9fd45b58805160bed71289640997ba822847498a05f1300c63594b9213fd1531", "seconds": 13.08, "spoken_utterances": [ "Add all Apple Music—um... I should say spotify-recommended R&B songs released in this year to a new ah... \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the action, genre, release year, and playlist name, remains consistent with the reference. Therefore, the final audible intent matches the original instruction after correction.", "heard_transcript": "Add all Apple Music um, I should say Spotify recommended R&B songs released in this year to a new uh R&B recommendation playlist." }, "candidate_id": "wav_536042a29853cc7d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear false start with 'Apple Music' corrected to 'Spotify'. The correction is immediate and the speaker continues smoothly. The 'um' filler is used naturally, and the overall delivery is fluent. However, the correction is slightly less organic than audio_03 because the initial error ('Apple Music') is more of a category mix-up than a minor slip, making the self-correction feel a bit more deliberate. The prosody is good, but the flow is slightly less spontaneous than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this year to a new \"R&B Hits\"—scratch that, make it \"R&B Recommendation\" playlist.", "long_pause_offsets": [], "source_sha256": "de2b672f40ab5cceddbe074133cfc17d1b15f15e39924342e420da43393affa7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially suggests \"R&B Hits\" but immediately corrects it to \"R&B Recommendation\". The rest of the instruction, including the genre, source, release year, and action, remains consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/false_start/false_start_04_playlist_name/212a35ec10eba820/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "c59d8f98da59596533b0c3768ea0d0d3c5a3f8b9714785e1e0b530087baecea8", "seconds": 12.72, "spoken_utterances": [ "Add all uh... spotify-recommended R&B songs released in this year to a new \"R&B Hits\"—hmmm... scratch that, make it um... \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B hits' but immediately self-corrects to 'R&B recommendation playlist'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value discarded. The final intent matches the original instruction exactly, with only a minor grammatical paraphrase ('recommendation playlist' vs 'Recommendation' playlist). The correction is immediate and unambiguous, satisfying the type-specific criterion.", "heard_transcript": "Add all uh Spotify recommended R and B songs released in this year to a new R and B hits. Scratch that, make it um R and B recommendation playlist." }, "candidate_id": "wav_6acaca1c865fe2dd", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision where the speaker starts with 'R&B hits', then says 'scratch that' and corrects to 'R&B recommendation playlist'. The 'scratch that' is a very natural way to cancel a previous idea, and the correction is clear. However, the pause before 'scratch that' is slightly long and the delivery of the correction feels a bit more rehearsed or less fluid than the top two. The prosody is consistent, but the overall pacing is slightly less natural, possibly due to the explicit cancellation phrase." } ] }, "is_default": false }, { "variant_id": "false_start_03_time_reference", "proposed_instruction": "Add all spotify-recommended R&B songs released in last year—no, that should be this year to a new \"R&B Recommendation\" playlist.", "long_pause_offsets": [], "source_sha256": "3cfa86255c207ee029077d293543ed61ca0b6adfd25b05abfe56127428623f15", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year'. The final intended time reference is 'this year', matching the original instruction. All other components—genre (R&B), source (Spotify-recommended), action (add), and destination (new playlist named \"R&B Recommendation\")—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/false_start/false_start_03_time_reference/c9d9034818500dd4/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "ced6f0b7cad150654a8835f518fc8094d71c74906ebd3222cbf49df29f5242c9", "seconds": 13.2, "spoken_utterances": [ "Add all um... spotify-recommended R&B songs released in last year—uh... no, that should be this year to a new let me think... \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately self-corrects to 'this year'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final instruction matches the original reference instruction in action, target, and constraint.", "heard_transcript": "Add all um Spotify recommended R&B songs released in last year. No, that should be this year to a new let me think R&B recommendation playlist." }, "candidate_id": "wav_7aa07cbf77d14507", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip contains a revision where the speaker initially says 'last year', then corrects to 'this year'. The correction is clear, but the 'no, that should be this year' feels more like a factual correction than a spontaneous thought process. The 'let me think' filler at the end is placed awkwardly and sounds like a forced pause rather than a genuine moment of hesitation. The prosody is somewhat flat, and the overall delivery feels less natural and more like a rehearsed or robotic correction compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Add all R&B songs released in this year to a new \"R&B Recommendation\" playlist. Oh, I forgot to mention, only add songs that are Spotify-recommended.", "long_pause_offsets": [ 80 ], "source_sha256": "a3def8469d1bb5518801bedebd63710f2fd2310d174a48f292b32e035c9a38ec", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a missing constraint that the songs must be Spotify-recommended, which is consistent with the original instruction and narrows the scope without contradiction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/addition/addition_01/aa537a28fabb596d/speech.wav", "wer": 0.03125, "wav_sha256": "c365f45fda70e4fb5dd1291462292cda91ea361977c1abd351f9155ee0da529e", "seconds": 18.08, "spoken_utterances": [ "Add all um... R&B songs released in this year ... to a new \"R&B Recommendation\" playlist. ", "Oh, I forgot to mention, only add songs that are let me think... Spotify-recommended." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Add all R&B songs released in this year to a new R&B recommendation playlist.' The speaker then adds a constraint: 'only add songs that are Spotify recommended.' This is a clear addition of a new condition to the original task. The final intent combines the original request with the added constraint, matching the reference instruction exactly. Therefore, the addition type criterion is satisfied.", "heard_transcript": "Add all um R&B songs released in this year to a new R&B recommendation playlist. Oh, I forgot to mention, only add songs that are let me think, Spotify recommended." }, "candidate_id": "wav_d6a85ee6536df991", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a coherent, multi-part request with natural hesitation and self-correction. The initial request is clear, followed by a believable 'oh I forgot to mention' revision that adds a constraint. The filler 'let me think' is placed naturally before the final detail, and the prosody maintains a conversational rhythm without robotic pauses or unnatural word boundaries." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Add all spotify-recommended songs released in this year to a new \"R&B Recommendation\" playlist. One more detail: only add R&B songs.", "long_pause_offsets": [ 96 ], "source_sha256": "4920dc7b6c0b12e56fddb816834695a100b4212a071ea3b7c3414e79650a9d57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds a genre restriction that narrows the initial broad request to only R&B songs, which matches the original instruction's genre requirement. The app, release year, playlist creation, and recommendation source remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/addition/addition_02/83f0775884fbf4fc/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "b45d05b2dbfe6f9f6757b9df0f129ad9a99add3b277538ee80584b26c9f5ee08", "seconds": 15.84, "spoken_utterances": [ "Add all spotify-recommended songs released in this year uh, let's see... to a new ... \"R&B Recommendation\" playlist. ", "One more detail: only add let's see... R&B songs." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'Add all Spotify recommended songs released in this year' to a new playlist. The addition 'only add, uh, let's see, R&B songs' refines the scope of the songs to be added, narrowing the initial broad request to only R&B songs. The final intent combines the original instruction with the added constraint, matching the reference instruction exactly. The 'uh, let's see' is a filler and does not change the meaning.", "heard_transcript": "Add all Spotify recommended songs released in this year, uh, let's see, to a new R&B recommendation playlist. One more detail, only add, uh, let's see, R&B songs." }, "candidate_id": "wav_7692d1c2e4a1411c", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker begins with a clear request but then adds a detail with a stutter ('uh let's see') and a repeated phrase ('uh let's see R&B songs'), which sounds less fluent and more hesitant than audio_02. The addition feels slightly awkward and less integrated into the flow, though the overall structure is still plausible as a spontaneous request." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_04_playlist_name", "proposed_instruction": "Add all spotify-recommended R&B songs released in this year to a new \"R&B Hits\" playlist. Change the playlist name to \"R&B Recommendation\" instead of \"R&B Hits\".", "long_pause_offsets": [ 90 ], "source_sha256": "0b64d6ec8ac035318819d97d2225b39e54f340097a4c6291f00668d90a7625ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the playlist name from \"R&B Hits\" to \"R&B Recommendation\"; all other requirements (genre, source, release year, and creation of a new playlist) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/revision/revision_04_playlist_name/5a300fcd9d22f94f/speech.wav", "wer": 0.1875, "wav_sha256": "a785d4f2d2f4878acc83407261c92e3b03fa41de4e2351e1945620931d563a25", "seconds": 17.48, "spoken_utterances": [ "Add all spotify-recommended ... R&B songs released in this year ... to a new um... \"R&B Hits\" playlist. ", "Change the playlist name to \"R&B Recommendation\" instead of \"R&B Hits\"." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'R&B hits playlist' but immediately revises it to 'R&B recommendation' in the same sentence. The revision is clear and final, and the rest of the instruction matches the original. The final intent is to create a playlist with the corrected name, so the revision criterion is satisfied.", "heard_transcript": "Add all Spotify recommended R&B songs released in this year to a new um R&B hits playlist. Change the playlist name to R&B recommendation instead of R&B hits." }, "candidate_id": "wav_25255deba81ebdb6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and continuous flow. The speaker begins with a clear request, then adds a follow-up revision ('Change the playlist name...') that is smoothly integrated into the same utterance. The prosody is consistent, with a believable 'um' filler and a natural-sounding correction. The revision is not abrupt but feels like a spontaneous addition to the original plan, making it sound like a single, coherent thought process." } ] }, "is_default": true }, { "variant_id": "revision_03_time_reference", "proposed_instruction": "Add all spotify-recommended R&B songs released in last year to a new \"R&B Recommendation\" playlist. Make that songs released this year, not last year.", "long_pause_offsets": [ 100 ], "source_sha256": "66e8230ae4e3a19539104d0b3856759719d6c96e618a5f219e4985cff346db21", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the release year from 'last year' to 'this year', preserving the genre, app, action, and playlist name." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/revision/revision_03_time_reference/21e24d3a41557a10/speech.wav", "wer": 0.034482758620689655, "wav_sha256": "dc3574c0f6fee22fe04ad95d1c5468cedeb9fa2c8afa7c95d17283060d2a2785", "seconds": 16.0, "spoken_utterances": [ "Add all let's see... spotify-recommended R&B songs released in last year to a new \"R&B Recommendation\" playlist. ", "Make that songs released ... this year, not last year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last year' but immediately corrects it to 'this year' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, including the action, target, genre, source, and playlist name. The correction is clear and unambiguous.", "heard_transcript": "Add all, let's see, Spotify recommended R&B songs released in last year to a new R&B recommendation playlist. Make that songs released this year, not last year." }, "candidate_id": "wav_631ff8160e6530d0", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a believable revision with a clear false start ('Let's see') and a pause before the initial request. The speaker then corrects 'last year' to 'this year' with a natural-sounding 'Make that...' construction. The prosody and pacing are consistent, and the correction feels like a genuine mid-sentence adjustment. The only minor drawback is the slightly longer pause before the correction, but it still sounds plausible." } ] }, "is_default": false }, { "variant_id": "revision_02_genre", "proposed_instruction": "Add all spotify-recommended pop songs released in this year to a new \"R&B Recommendation\" playlist. Sorry, I meant R&B, not pop.", "long_pause_offsets": [ 100 ], "source_sha256": "5b0258cd0dc0cbf34b6fd2a946f25ebc77418ef6f4bb3a4401939610f5d00374", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the genre from 'pop' to 'R&B', preserving the app recommendation source, release year condition, playlist name, and action of adding to a new playlist." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/revision/revision_02_genre/d1594cf1cef67fed/speech.wav", "wer": 0.08, "wav_sha256": "588868f6a0a0da92bc3bf42be968aa810f75f501658b63e1c206620e50c1c134", "seconds": 12.72, "spoken_utterances": [ "Add all spotify-recommended uh... pop songs released in ... this year to a new \"R&B Recommendation\" playlist. ", "Sorry, I meant R&B, not pop." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'pop songs' but immediately corrects to 'R&B, not pop' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Add all Spotify recommended uh pop songs released in this year to a new R&B recommendation playlist. Sorry, I meant R&B, not pop." }, "candidate_id": "wav_fed7e7dcd1fc4d4c", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision where the speaker first says 'pop songs' and then corrects to 'R&B'. The correction is clear and the prosody is natural. However, the initial request is slightly less fluent, with a noticeable pause and filler ('uh') before 'pop songs'. The correction is also a bit more abrupt compared to audio_03 and audio_04, making it sound slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_01_app", "proposed_instruction": "Add all apple music-recommended R&B songs released in this year to a new \"R&B Recommendation\" playlist. Wait, the app should be Spotify, not Apple Music.", "long_pause_offsets": [ 104 ], "source_sha256": "1a39a25169070b4edea46bda586c56b1f8eeaf297ac4237b3fdecafd2be18057", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, preserving the genre, release year condition, and playlist creation action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/revision/revision_01_app/c44d7aab770e4391/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "df4f7d72ddf7f85b8e7a7b0dda0d0aa6b810ae8088fc22ea31d626a784de3b78", "seconds": 17.64, "spoken_utterances": [ "Add all apple music-recommended R&B songs released in ... this year to a new um... \"R&B Recommendation\" playlist. ", "Wait, the app should be um, let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requests adding Apple Music-recommended songs but immediately corrects the app to Spotify after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, including the app, genre, release year, and playlist name.", "heard_transcript": "Add all Apple Music recommended R&B songs released in this year to a new um R&B Recommendation playlist. Wait, the app should be um, let me think, Spotify not Apple Music." }, "candidate_id": "wav_474be51eac7bf82f", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a revision where the speaker initially says 'Apple Music' and then corrects to 'Spotify'. However, the correction is delivered with a longer pause and a more hesitant 'um' and 'let me think' before the correction. The prosody is less fluid, and the correction feels more like a separate thought rather than a natural self-repair. The overall delivery is less continuous and more fragmented compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_003", "hide_variant_original": true, "source_sha256": "cfd684fd9c8d276274b5a0c7a66e31452604dfca5e3facd7b3d0f9138cf5932f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/cancellation/cancel_003/8a4e831e7dc006b6/speech.wav", "wer": 0.12, "wav_sha256": "5fd931289bff105538a34a96fad720aaedc3c45a63b5531460f2d297ba5cf2ff", "seconds": 14.28, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist.", "Oh, wait, please cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first gives a complete instruction to add songs to a playlist, then immediately corrects themselves by saying 'Oh wait, please cancel that.' This constitutes a clear cancellation of the entire task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "Add all Spotify recommended R&B songs released in um this year uh to a new R&B recommendation playlist. Oh wait, please cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_007", "hide_variant_original": true, "source_sha256": "c5ba41fe20eff523630636d242075aecb733f420f7266f8626239e86f24e45db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/ambiguous_halt/halt_007/34b38c9f0033bc9e/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "0710aac2792c53cd068bacc3456e71d87e1306100abbca3ab78ad768664372cb", "seconds": 13.6, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist.", "One moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit interjection 'One moment,' which is a clear request to pause or wait. This aligns with the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel. The initial request is not completed, but the final control intent is to halt, which satisfies the type-specific criterion.", "heard_transcript": "Add all Spotify-recommended R&B songs released in um this year uh to a new R&B Recommendation playlist. One moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e3d6c94_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6cc2832dd9873bebd3ff83fb6b51dc768db4630b324eea1ce33a285cd6e49a0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e3d6c94_3/audio_natural/filler/original/1b68a2ce9e7dcdca/speech.wav", "wer": 0.15, "wav_sha256": "e9671dde50e921e3e0848c03e526cf44fae9f6582470c00fae8692672bfe6188", "seconds": 10.6, "spoken_utterances": [ "Add all Spotify-recommended ... R&B songs released in um... this year uh... to a new \"R&B Recommendation\" playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'uh') but does not change the core task. The final intent matches the original instruction. The filler words do not alter the goal, so the task should be executed as intended.", "heard_transcript": "Add all Spotify recommended R&B songs released in um this year uh to a new R&B recommendation playlist." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_thread_priority", "proposed_instruction": "How many priority-2—I meant priority-1 unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "9e56e36bdc4f44db1aa8353a176f06a3232598dd9307d7d37c5e4d8471305806", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'priority-two' to 'priority-one' within the same utterance, preserving the app (Gmail), the object (unread email threads), and the action (counting)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/false_start/false_start_01_thread_priority/9d0a9d893323cd52/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "4bba2a67546648f5f0f89bc56446ce59dfd75600c8794ecd44b6e8603d30cb1f", "seconds": 7.48, "spoken_utterances": [ "How many priority-two—um... I meant ah... priority-one unread email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority two' but immediately self-corrects to 'priority one'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original instruction is executed with the correction. The rest of the instruction matches the original.", "heard_transcript": "How many priority two um, I meant uh priority one unread email threads are in my Gmail inbox?" }, "candidate_id": "wav_a19fe8f615a7f8f4", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker begins with a clear false start ('priority two'), immediately self-corrects to 'priority one' with a natural-sounding 'I meant', and smoothly continues the full request. The correction is concise, the prosody is consistent, and the flow feels spontaneous and unforced." } ] }, "is_default": true }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "How many priority-1 read—no, I mean unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "3b28433541afab5935d3daccd464b9bbb17384ee64a70b50a779a2fb8a057b0f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority-one read' but immediately corrects to 'unread'. The final state specifies the correct status (unread), and all other elements (priority level, app, action, and object) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/false_start/false_start_02_email_state/45a356a9cbc2cc80/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "3a54eb140c834ecc2b9f8c2927be78e97d52281042800bd1326f759e4bb3fb32", "seconds": 7.28, "spoken_utterances": [ "How many priority-one read—uh... no, I mean unread email threads are let's see... in my Gmail inbox?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_1cd72c09d25c0de3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker makes a false start ('priority one red'), self-corrects to 'unred', and continues. The correction is slightly less fluid than audio_03, with a minor stutter ('are let's see'), but the overall delivery is still credible and the repair is contextually appropriate." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "How many priority-1 unread email threads are in my Outlook—I should say Gmail inbox?", "long_pause_offsets": [], "source_sha256": "3ac216722e4889f0da2425dd6462b2819b963ec84323606cffe7679cd31b2ca8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail inbox'. The final app is Gmail, and all other conditions (priority-1, unread, email threads, inbox) match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/false_start/false_start_03_app/b749bcbb2da49f98/speech.wav", "wer": 0.2777777777777778, "wav_sha256": "b07fe66eace0ae580c9d46b3750fdb02e9f33f24a4b0f2226a4160c534836315", "seconds": 12.28, "spoken_utterances": [ "How many priority-one unread email threads are hmmm... in my Outlook—uh... I should say ... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value ('Gmail') should be applied, and the original incorrect value ('Outlook') should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many priority one unread email threads are in my Outlook uh, I should say Gmail inbox?" }, "candidate_id": "wav_38fd594d370ed3c1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker starts with a clear request, then pauses with 'hmm' before changing the email service from Outlook to Gmail. The correction is believable, but the initial pause and the phrase 'I should say' sound slightly rehearsed or less spontaneous than the other false starts, reducing naturalness." } ] }, "is_default": false }, { "variant_id": "false_start_04_email_folder", "proposed_instruction": "How many priority-1 unread email threads are in my Gmail drafts—scratch that, make it inbox?", "long_pause_offsets": [], "source_sha256": "5442d7b24b8ce2fdaa24007c4b39e550949874f16c9f602fd46df36a534d8f3e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'drafts' but immediately corrects to 'inbox' within the same utterance. The final target folder is 'inbox', matching the original instruction. All other elements—priority level, thread status, and app—remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/false_start/false_start_04_email_folder/ddf593229850fa1d/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "60d1c9970525a336230defab1e4d1ce4ecee9fff0e8de5304e34e2a52bb57883", "seconds": 12.96, "spoken_utterances": [ "How many priority-one unread email threads are in my Gmail uh, let's see... drafts—uh... scratch that, make it um... inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Drafts' but immediately self-corrects with 'scratch that, make it um inbox'. This is a false start where the final value is 'inbox', which matches the original instruction. The correction is immediate and unambiguous, so the final intent is to query the inbox, not drafts.", "heard_transcript": "How many priority one unread email threads are in my Gmail uh let's see. Drafts uh scratch that, make it um inbox." }, "candidate_id": "wav_bea7273bbfbc3382", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This speaker begins with a false start ('priority one unread email threads are in my Gmail'), then hesitates with 'uh let's see', and makes a correction to 'drafts'. The correction is abrupt and the phrase 'scratch that' sounds more like a command than a natural self-repair. The overall flow feels less organic and more like a scripted edit." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Oh, I forgot to mention, it should be priority-1 threads.", "long_pause_offsets": [ 53 ], "source_sha256": "91e8d809b7f4e8d4bdbe9252ffb6190af569f5664b52252cef0eb9f872a29dae", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'priority-one' restriction to the initial question about unread threads, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/addition/addition_01/25b333fd30f0a46c/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "9fd2ee46ae985d0ac1f2c59128bc7d8dbbd64eaa5717b323f647c79292b275da", "seconds": 12.32, "spoken_utterances": [ "How many unread email threads are ... in my Gmail inbox? ", "Oh, I forgot to mention, it should be um... priority-one threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How many unread email threads are in my Gmail inbox?' and was followed by an addition specifying 'priority one threads'. The addition clarifies the type of unread threads to count, combining with the original request to form the complete instruction. The final intent matches the reference instruction, and the addition type is correctly applied.", "heard_transcript": "How many unread email threads are in my Gmail inbox? Oh, I forgot to mention it should be um priority one threads." }, "candidate_id": "wav_7ac6e79ff48b8ec5", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. After a brief pause and filler ('oh'), they add a clarifying detail ('priority one threads') in a way that sounds like a spontaneous afterthought. The prosody is consistent, and the addition flows smoothly without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many priority-1 email threads are in my Gmail inbox? I forgot to add, only count unread threads.", "long_pause_offsets": [ 57 ], "source_sha256": "89ff77edef32409155575c1a4862411a6afc2fc6923406966eb3df79a9d017ad", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'unread' constraint, which refines the initial request without contradicting it. The final goal combines the original priority level and adds the unread condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/addition/addition_02/1859a57dacde77fd/speech.wav", "wer": 0.05, "wav_sha256": "cb3d492ab29d8c83ef7641891b7f991f17dda810e3c2e769abfa05883ef3de15", "seconds": 11.36, "spoken_utterances": [ "How many priority-one email threads are um... in my Gmail inbox? ", "I forgot to add, only count ... unread threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many priority one email threads are in my Gmail inbox?' which is a broad request. The speaker then adds the constraint 'only count unread threads' after a pause, which narrows the scope. The final intent combines the original request with the added constraint, matching the reference instruction. The addition is clear and directly modifies the initial request without introducing new or conflicting conditions.", "heard_transcript": "How many priority one email threads are um in my Gmail inbox? I forgot to add, only count unread threads." }, "candidate_id": "wav_c67d63a1ceec75d8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The initial request is clear, but the addition ('only count unread threads') is delivered with a slightly more abrupt and less fluid transition. The phrase 'I forgot to add' introduces a self-correction that feels a bit more rehearsed or less conversational than in audio_01. The prosody is slightly less natural, and the overall delivery is less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_thread_priority", "proposed_instruction": "How many priority-2 unread email threads are in my Gmail inbox? Make that priority-1 unread email threads, not priority-2.", "long_pause_offsets": [ 64 ], "source_sha256": "d7019d29e3a2d20bc636841f7005d814b75ed0fb17e4d837f572d7d7d11604e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the priority level from two to one, while retaining the other conditions: unread status, email threads, and Gmail inbox context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/revision/revision_01_thread_priority/c5b9bd4f6a5a2807/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "54e4a3aa6bfc5e130d1d50d2650ed5a6d3145c0bf9f5e2fc1c4854326008e91a", "seconds": 16.16, "spoken_utterances": [ "How many um, let me think... priority-two unread email threads are ... in my Gmail inbox? ", "Make that hmmm... priority-one unread email threads, not priority-two." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority two' but immediately corrects to 'priority one' after a pause, which matches the original instruction. The revision type criterion requires applying the correction spoken after a longer pause to the earlier instruction, which is satisfied here. The final audible intent is to count priority-1 unread threads, matching the original instruction.", "heard_transcript": "How many um, let me think, priority two unread email threads are in my Gmail inbox? Make that priority one unread email threads, not priority two." }, "candidate_id": "wav_db3e5fa4156b87dc", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent revision. The speaker begins with a realistic 'um' and 'let me think' filler, then delivers the initial request with a natural, slightly hesitant prosody. The correction ('make that...') is smoothly integrated into the same utterance, with a clear self-repair of the priority level. The intonation and pacing are consistent with spontaneous speech, and the correction feels like a genuine moment of recollection rather than a mechanical edit." } ] }, "is_default": true }, { "variant_id": "revision_04_email_folder", "proposed_instruction": "How many priority-1 unread email threads are in my Gmail sent? That should be inbox, not sent.", "long_pause_offsets": [ 63 ], "source_sha256": "40b187bda2b4e283758059b512d7cc0b1032719cc54625e86e756671af391ca6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from 'sent' to 'inbox', preserving the original intent of counting priority-1 unread threads in the inbox." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/revision/revision_04_email_folder/874b43161543ec59/speech.wav", "wer": 0.1, "wav_sha256": "01413de74c6a67e3f6a35b9c4e69e682c0e861ecfbfa6c7f157cd27d2ff23af2", "seconds": 12.16, "spoken_utterances": [ "How many uh... priority-one unread email threads are ... in my Gmail sent? ", "That should be ah... inbox, not sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately corrects to 'inbox' after a pause, which is a clear revision. The final intent matches the original instruction exactly, so the task should be executed as revised.", "heard_transcript": "How many uh priority one unread email threads are in my Gmail sent? That should be uh inbox, not sent." }, "candidate_id": "wav_9fb95e6bec571429", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision with a believable 'uh' filler and a natural-sounding correction of the mailbox ('sent' to 'inbox'). The speaker's intonation shifts appropriately when correcting the error, and the pause before the revision is plausible. However, the initial request is slightly more robotic in delivery compared to audio_01, and the correction is less fluidly embedded within the same thought process, making it feel slightly more like a separate statement." } ] }, "is_default": false }, { "variant_id": "revision_02_email_state", "proposed_instruction": "How many priority-1 read email threads are in my Gmail inbox? Wait, the email threads should be unread, not read.", "long_pause_offsets": [ 62 ], "source_sha256": "173c2164553fa7ee294cd05bf1013cc434429b82359359a6dba5a38fd9b29b00", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the status from 'read' to 'unread', preserving the priority level, app, and count action." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/revision/revision_02_email_state/4a950bfc35e63a59/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "08d37ceace5f684985b13f71b325886634fe33ce4c9aa3b3253550678a8807e8", "seconds": 13.72, "spoken_utterances": [ "How many um... priority-one read email threads are let me think... in my Gmail inbox? ", "Wait, the email threads should be uh... unread, not read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'read email threads' but immediately corrects themselves to 'unread, not read' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "How many um priority one read email threads are, let me think, in my Gmail inbox? Wait, the email threads should be uh unread, not read." }, "candidate_id": "wav_75f8cacd43bb3168", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision but includes two distinct errors: 'read' instead of 'unread' and a filler ('let me think') mid-sentence. The correction is delivered with a 'wait' and a clear self-repair, but the initial request is less coherent due to the filler interrupting the flow. The prosody is less natural, and the correction feels more like a post-hoc realization rather than a seamless thought process, reducing its realism compared to the others." } ] }, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "How many priority-1 unread email threads are in my Outlook inbox? Let me correct the app: check my Gmail inbox, not Outlook.", "long_pause_offsets": [ 66 ], "source_sha256": "72bb5b5125d6223cf196019ad0e685d5625060712dbe9cfa65f9044da6e2936c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email app from Outlook to Gmail, while retaining the original query about priority-1 unread threads. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/revision/revision_03_app/258f7a53e2765234/speech.wav", "wer": 0.0, "wav_sha256": "3b98a858b4f145f717d6eadae56df4ac8c7563c36e6246d14466d1d9630f604c", "seconds": 12.88, "spoken_utterances": [ "How many ... priority-one unread email threads are ... in my Outlook inbox? ", "Let me correct the app: check my Gmail inbox, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'Outlook inbox' but immediately corrects it to 'Gmail inbox not Outlook'. This is a revision where the final intent clearly specifies the correct app. The original instruction's app was wrong, but the revision corrects it to match the reference. All other constraints (priority-1, unread, email threads, inbox) are preserved and correctly applied to Gmail.", "heard_transcript": "How many priority one unread email threads are in my Outlook inbox? Let me correct the app, check my Gmail inbox not Outlook." }, "candidate_id": "wav_d421fc95c0e87b57", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. The speaker starts with a clear, unbroken request, then adds a correction ('let me correct the app...') that sounds like a separate, detached thought rather than a natural self-repair. The phrase 'check my Gmail inbox not Outlook' is delivered with a flat, declarative tone that lacks the intonation and hesitation of spontaneous speech. The correction feels tacked on, and the overall flow is less organic than the other clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "1c197057a0a2f8ddb5459f4b2b6623c85dacac569a9c08375d57f52f3ab34abb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/cancellation/cancel_021/2082ac9c7b9c26b1/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "e57c7fcc5748ef3df0a1c0c12f3b717ff523d2aac04a7fc75b82d2be35d1b428", "seconds": 11.32, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?", "Hold on, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question, then clearly revokes the entire task with 'Hold on, never mind, I don't need that done anymore.' This constitutes a complete cancellation of the original request. The final intent is to cancel, not to execute or pause, and the cancellation is unambiguous and applies to the whole task.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox? Hold on, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_037", "hide_variant_original": true, "source_sha256": "0506275714eca4d9d46cc278750db646a5dbfc710b6547887dd8b581b8d51d8b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/ambiguous_halt/halt_037/a855ff1a1b5cc330/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "56000963a728f408d7a92e13f1a1f327a7b1d8b855032a8b96dff7ec1261a6e9", "seconds": 9.28, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?", "Pause for a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit phrase 'Pause for a moment,' which constitutes a clear request to hold or wait. This matches the 'ambiguous_halt' type criterion, where the final interjection is a pause command. The initial request is not canceled or revised, but the final control intent is to pause, not execute.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox? Pause for a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "0e583ea672591dd803ec163d7837c19393ed289cdd82f82f2e027c68271193d8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_1/audio_natural/filler/original/b7ccd815ddeb9571/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "fd2a7db52b2aa7b43a5e3805470fcc458ad1e13ce58e669b693f198b2b67b960", "seconds": 6.08, "spoken_utterances": [ "How many um... priority one unread email threads are ... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'um' but continues with the complete and correct original instruction. The final intent matches the reference exactly, with no changes or omissions. The filler does not alter the task goal, so the execution criterion is satisfied.", "heard_transcript": "How many um priority one unread email threads are in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_thread_priority", "proposed_instruction": "How many priority-1—no, I meant priority-2 read email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "db78a20d3f6959745f5279d8fc3aff4624f0986e66da46c949982b7527fe1eb6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'priority-one' to 'priority-two' within the same utterance, preserving the rest of the query including the app (Gmail), status (read), and thread count request." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/false_start/false_start_01_thread_priority/cb99fe5cd14e35ab/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "44ba109269c9e4b86115ccb272bd428ad44779342f00d0c34e8a9d8d98904f0a", "seconds": 10.56, "spoken_utterances": [ "How many ... priority-one—um... no, I meant priority-two uh... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority one' but immediately self-corrects to 'priority two'. The final audible intent matches the original instruction exactly, with the correction applied. The 'false_start' type requires execution of the corrected version, which is fully present and clear.", "heard_transcript": "How many priority one um no I mean priority two uh read email threads are in my Gmail inbox?" }, "candidate_id": "wav_2dfa7f3d85e89974", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip has the most natural and spontaneous-sounding false start. The speaker begins with 'priority one', immediately self-corrects to 'priority two' with a clear, audible hesitation ('uh'), and continues smoothly. The prosody is consistent, and the correction feels like a genuine moment of recall. The pacing and intonation are very human, with no signs of artificiality or forced delivery." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "How many priority-2 read email threads are in my Outlook—I should say Gmail inbox?", "long_pause_offsets": [], "source_sha256": "8fd792cdfce63d5739ba31e46ec656165be2638ff131f2b5c8d3aaf528b7bf16", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail inbox'. The final goal matches the original instruction in all aspects: app (Gmail), thread status (read), priority level (2), and the query type (count)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/false_start/false_start_03_app/c24d33dcd70c6ecf/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "bd492c09c394c4849a1a928adab3c30be8cb9d18448b8eeee0210dc99b098532", "seconds": 9.6, "spoken_utterances": [ "How many um... priority-two read email threads are in my Outlook—hmmm... I should say Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail inbox'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value ('Outlook') is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "How many um priority two read email threads are in my Outlook? I should say Gmail inbox." }, "candidate_id": "wav_302c62e620f69d9d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a clear false start with 'Outlook' corrected to 'Gmail inbox'. The speaker uses a natural 'um' and a thoughtful pause before the correction. The intonation and rhythm are consistent with spontaneous speech. The correction is immediate and flows well, though the initial hesitation is slightly longer than in audio_04, making it feel a bit more deliberate." } ] }, "is_default": false }, { "variant_id": "false_start_04_email_folder", "proposed_instruction": "How many priority-2 read email threads are in my Gmail drafts—sorry, that should be inbox?", "long_pause_offsets": [], "source_sha256": "a20d3bec891acb70566adfd7e2971aa9c5b91adb8bbef939cbbbdc6d1b78a43f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'drafts' but immediately corrects it to 'inbox'. The final intended folder is the inbox, matching the original instruction. All other elements—priority level, read status, thread count, and app—remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/false_start/false_start_04_email_folder/e20710e7a15ea2d6/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "c4cee9cf1c6a961b9878cf7618e69a4dd5eb43c04d60cb24d36767cf91ff42c3", "seconds": 12.4, "spoken_utterances": [ "How many priority-two um, let me think... read email threads are in my Gmail ... drafts—um... sorry, that should be inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'drafts' but immediately self-corrects to 'inbox' after a pause and an explicit 'sorry, that should be inbox'. According to the 'false_start' type criterion, the final corrected value ('inbox') overrides the initial mistaken value ('drafts'), and the rest of the instruction remains unchanged. The final audible intent matches the original instruction.", "heard_transcript": "How many priority two, um, let me think, read email threads are in my Gmail drafts, um, sorry, that should be inbox." }, "candidate_id": "wav_a12aed148b72a2a1", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip includes a false start with 'drafts' corrected to 'inbox'. The speaker uses 'let me think' as a filler, which is a plausible hesitation strategy. However, the correction is slightly abrupt and the 'sorry' feels a bit more like a scripted acknowledgment than a natural self-correction. The prosody is slightly less fluid, and the overall delivery feels a bit more rehearsed compared to the top two." } ] }, "is_default": false }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "How many priority-2 unread—wait, I mean read email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "a0dc18b1eb8dbccbb617e42ea42b1c76d66f4bc420094fb7b4a8d733c3110ade", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately corrects it to 'read' within the same utterance. The final intended state is 'read', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/false_start/false_start_02_email_state/613a964417da711d/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "53a3dbdc4c6f120858b90e4087857e7183aef6754522ab26c6a03102a4cd49b6", "seconds": 11.04, "spoken_utterances": [ "How many uh... priority-two unread—um... wait, I mean ... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately self-corrects to 'read' after a pause and 'wait, I mean'. The final audible intent is to count 'read' email threads, which is the corrected version of the original instruction. The 'false_start' type requires applying the immediate self-correction, which is done here. The rest of the instruction matches the original.", "heard_transcript": "How many um priority two unread um wait, I mean read email threads are in my Gmail inbox?" }, "candidate_id": "wav_c4dd05ceacc06acb", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start with 'unread' corrected to 'read'. The correction is clear, but the initial word 'unread' is spoken with a slightly unnatural emphasis and the pause before 'wait' feels a bit too long and abrupt. The overall flow is less smooth, and the prosody seems slightly off, making it sound less spontaneous than the others. The correction feels more like a mechanical fix than a natural thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many read email threads are in my Gmail inbox? Oh, I forgot to mention, it's only for priority-2 threads.", "long_pause_offsets": [ 51 ], "source_sha256": "400e11651121606f88b856a2041645b9a4d678ce575489cad567fa40d78f6751", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'priority-two' restriction, which narrows the initial broad request to the exact scope specified in the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/addition/addition_01/8dca31276efcfb00/speech.wav", "wer": 0.15384615384615385, "wav_sha256": "fec28abb296618a6b12f7e688b5189b04fd6ff8708c9c0fdc12937f4d15cb851", "seconds": 14.84, "spoken_utterances": [ "How many um... read email threads are in my uh... Gmail inbox? ", "Oh, I forgot to mention, it's only for uh, let's see... priority-two threads." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "candidate_id": "wav_19301dc20a34de6f", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker's delivery is highly natural, with a believable hesitation ('uh') and a smooth, self-correcting flow. The initial request is slightly mumbled but understandable, followed by a plausible 'Oh I forgot to mention' that introduces a revision. The correction ('only for priority two threads') is clearly a follow-up to the earlier request, with appropriate intonation and pacing. The speaker sounds like they are genuinely recalling a detail mid-sentence, making the revision feel spontaneous and pragmatic." } ] }, "is_default": false }, { "variant_id": "addition_02", "proposed_instruction": "How many priority-2 email threads are in my Gmail inbox? I left out a detail: only count read email threads.", "long_pause_offsets": [ 57 ], "source_sha256": "5de72d55971910941e8ffd462d39a81aeef3fba2177d2a0f176dfa33f59e66d0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count read email threads', which refines the initial broad request. The priority level and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/addition/addition_02/5b72987e28a17505/speech.wav", "wer": 0.12, "wav_sha256": "d977dd9af9fd39d74b9c76461bdc45a945fa89cb043da3547bcdb668224812ce", "seconds": 15.84, "spoken_utterances": [ "How many um... priority-two email threads are in my ... Gmail inbox? ", "I left out a detail: only count um, let's see... read email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many priority-2 read email threads are in my Gmail inbox?'. The speaker then adds a clarifying condition 'Only count... read email threads' after a pause, which refines the original request by adding a constraint. This matches the 'addition' type, where the final goal is the combination of the initial instruction and the added information. The final audible intent is to count priority-2 threads that are read, which is a valid and complete task.", "heard_transcript": "How many um priority two email threads are in my Gmail inbox? I left out a detail. Only count um let's see read email threads." }, "candidate_id": "wav_a58e32add5395e94", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a natural-sounding addition, but the flow is less smooth. The speaker starts with a hesitant 'How many um' and then makes a self-correction ('I left out a detail') before adding the restriction. However, the addition ('only count um let's see read email threads') is less coherent and the phrase 'read email threads' is awkwardly phrased, possibly due to a slip of the tongue or unclear thought. The prosody is slightly less fluid, and the final phrase sounds like a mumbled afterthought rather than a fully formed addition, reducing its realism compared to audio_02." } ] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_thread_priority", "proposed_instruction": "How many priority-1 read email threads are in my Gmail inbox? Wait, the priority should be priority-2, not priority-1.", "long_pause_offsets": [ 62 ], "source_sha256": "07f4217469d36d3d22dad7226a9dc3ee8569ecfecb0322f2b949331204399e53", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the priority level from 'priority-one' to 'priority-two', while retaining the other conditions: 'read email threads' and 'in my Gmail inbox'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/revision/revision_01_thread_priority/3b922129fec1e7de/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "ff9f6d426e7222ff8e53ccbc0dcf3592fb6cbf52577c1dfcc1cf6d874e468596", "seconds": 15.16, "spoken_utterances": [ "How many uh, let's see... priority-one read email threads are in my Gmail inbox? ", "Wait, the priority should be uh... priority-two, not priority-one." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority one' but immediately corrects themselves to 'priority two, not priority one' after a pause. This is a clear revision where the final audible intent is to request priority-2 threads. The original instruction matches this final intent, so the revision is correctly applied.", "heard_transcript": "How many uh let's see, priority one read email threads are in my Gmail inbox? Wait, the priority should be uh priority two, not priority one." }, "candidate_id": "wav_f77d006ba2c14377", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent revision. The speaker begins with a plausible hesitation ('uh'), formulates the initial request, then self-corrects the priority level with a believable change-of-mind intonation ('Wait, the priority should be...'). The correction is smoothly integrated into the same utterance, and the prosody reflects genuine thought processing rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "revision_03_app", "proposed_instruction": "How many priority-2 read email threads are in my Outlook inbox? Let me correct the app: use Gmail, not Outlook.", "long_pause_offsets": [ 64 ], "source_sha256": "09f7e2b056ce6b0b80d3be7020b8ddbc2e6e1293b9f18114d53da11693be2a1c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email app from Outlook to Gmail, while retaining the original query about priority-2 read threads. The final goal matches the reference instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/revision/revision_03_app/eeb5e715e2084a75/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "e06e07b326debfc6fae32adc267ec084b36e924ebe0e177d88a820b5519d8469", "seconds": 14.84, "spoken_utterances": [ "How many uh, let me think... priority-two ... read email threads are in my um... Outlook inbox? ", "Let me correct the app: use Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Outlook but immediately corrects the app to Gmail. The revision type criterion requires applying the correction spoken after a pause to the earlier instruction. The final audible intent matches the original reference instruction exactly, so the task is executed as intended.", "heard_transcript": "How many uh, let me think, priority two read email threads are in my um Outlook inbox? Let me correct the app, use Gmail, not Outlook." }, "candidate_id": "wav_3a6b741704d9578a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a realistic revision with a clear self-correction. The speaker uses a natural filler ('let me think') and makes a mistake in the email client ('Outlook'), then corrects it to 'Gmail'. The correction is delivered as a follow-up clause ('Let me correct the app...'), which is a plausible way to handle a realization mid-sentence. The pacing and intonation are consistent with spontaneous speech." } ] }, "is_default": false }, { "variant_id": "revision_02_email_state", "proposed_instruction": "How many priority-2 unread email threads are in my Gmail inbox? Change 'read' to 'unread' for the email state.", "long_pause_offsets": [ 64 ], "source_sha256": "57da4d9f90f482134043d4d0e5da25660a2d96d852acb0f7b632687092c9b329", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the email state from 'read' to 'unread', aligning with the final spoken instruction. All other elements—priority level, thread count, app, and scope—remain consistent with the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/revision/revision_02_email_state/0bc7810a4b33c12e/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "66f25482a898325f75b079c0b5b719b44a07c8cc49211e62b65bf875006aa3d1", "seconds": 13.64, "spoken_utterances": [ "How many um... priority-two unread email threads are in my Gmail inbox? ", "Change 'read' to '... unread' for the email state." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'read' but immediately corrects it to 'unread' in the same sentence. This is a revision where the final audible intent is to count unread threads, which matches the revised instruction. The correction is clear and applied to the same context, so the final goal is correctly interpreted.", "heard_transcript": "How many um priority two read email threads are in my Gmail inbox? Change read to unread for the email state." }, "candidate_id": "wav_602d8d52854db4fc", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision where the speaker changes 'read' to 'unread'. The correction is made as a separate sentence after the initial request, which is a plausible way to revise a task. However, the delivery is slightly less fluid than audio_02 and audio_03, with a more abrupt transition between the two sentences and less natural intonation on the correction." } ] }, "is_default": false }, { "variant_id": "revision_04_email_folder", "proposed_instruction": "How many priority-2 read email threads are in my Gmail sent? That should be 'inbox', not 'sent'.", "long_pause_offsets": [ 61 ], "source_sha256": "1534b12919d26013724570ff6d4c8872e6f5268312a49c04e3ea49a9ac0ce7b8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from 'sent' to 'inbox', preserving the original request for priority-2 read threads in Gmail. The final goal matches the reference." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/revision/revision_04_email_folder/e7aafd8e08777dbb/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "33bc799e8c1b20f2211855cf2ced1eeb8b4d582eca31bd8882d4d73fd9b87943", "seconds": 12.88, "spoken_utterances": [ "How many priority-two ... read email threads are in my ... Gmail sent? ", "That should be 'uh... inbox', not 'sent'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'Gmail sent' but immediately corrects themselves to 'inbox, not sent' after a pause. This is a clear revision where the final intent is to query the inbox, matching the original instruction. The correction is self-contained and does not introduce new constraints or remove existing ones.", "heard_transcript": "How many priority two read email threads are in my Gmail sent? That should be uh inbox, not sent." }, "candidate_id": "wav_070c61123dfb7f8c", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This audio has the least realistic delivery. The speaker makes a correction ('That should be inbox, not sent') but the phrasing is awkward and the intonation is flat. The correction feels tacked on rather than naturally integrated, and the overall prosody lacks the fluidity and hesitation patterns of genuine spontaneous speech. The delivery sounds more like a rehearsed or robotic correction than a natural thought process." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "de9930a3bb9b897f2aa2745668cf8ee8bf7bbbfcc31a2db3eb7408534e390f22", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/cancellation/cancel_014/473af8b288f9ad50/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "f74eaaf4a59d77acddcaf178f7affc0147c2e6aae66dedd09494e628c9a5aae0", "seconds": 13.12, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?", "Oh, wait, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to ask a question but then explicitly says 'cancel that' after a brief pause and self-correction. This constitutes a clear cancellation of the initial request. The final audible intent is to cancel, not to execute the original task.", "heard_transcript": "How many uh let's see priority two read email threads are in my Gmail inbox? Oh wait, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_024", "hide_variant_original": true, "source_sha256": "02ec18a2e714016d63ba8aaf6f2c4b5e6ca0d5afb8fc0af9795866b0eeca2d62", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/ambiguous_halt/halt_024/0ef88e27673472dc/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "cb00f20782ac76fa9c4f2365cb791f0fdc634b07ca164f9a496bd78eda539229", "seconds": 11.56, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?", "Give me a moment to think." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then explicitly says 'Give me a moment to think,' which is a clear request to pause. The audio ends with this interjection, and there is no cancellation or further instruction. This matches the 'ambiguous_halt' type, where the final intent is to pause, not to execute or cancel.", "heard_transcript": "How many uh, let's see, priority two, read email threads are in my Gmail inbox? Give me a moment to think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "e3d755dfb8ce8a7626032d924a2c9707b373d9e919f320892eb0f60070a131e0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_2/audio_natural/filler/original/ba1ed4c7dec7d9a7/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "e073e60c5819bd864a707553cab0a611db5fbe26ca3a5ae1bb89fcd4be709b0d", "seconds": 7.92, "spoken_utterances": [ "How many uh, let's see... priority two ... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with hesitation ('uh', 'let's see') but completes the original instruction without changing the action, target, or value. The final intent matches the reference, and the filler words do not alter the task goal.", "heard_transcript": "How many uh let's see priority two read email threads are in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_03_app", "proposed_instruction": "How many priority-3 unread email threads are in my Outlook—wait, I meant Gmail inbox?", "long_pause_offsets": [], "source_sha256": "f002198d6c5862ded7974e426bc06b20a442c7c98e903c96f911bbd60a6c2429", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately corrects to 'Gmail inbox'. The rest of the query, including the priority level and unread status, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/false_start/false_start_03_app/69c5d63c41cc9b83/speech.wav", "wer": 0.058823529411764705, "wav_sha256": "b8739db5a80aeeda18598f5c25d51508e79251fe0d2b85a6fd71061ff117d90d", "seconds": 8.16, "spoken_utterances": [ "How many priority-three ... unread email threads are in my Outlook—ah... wait, I meant Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook app' but immediately self-corrects to 'Gmail inbox'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many priority three unread email threads are in my Outlook app? Wait, I meant Gmail inbox?" }, "candidate_id": "wav_abebe5d3ae0e3c7c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and spontaneous self-correction. The speaker starts with 'Outlook', then immediately self-corrects with 'wait, I meant Gmail inbox?' in a way that sounds like a genuine, real-time thought process. The prosody is fluid, with a believable shift in intonation and pacing during the correction. The pause before 'wait' is short and natural, fitting a moment of realization. The correction is integrated smoothly into the utterance without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "false_start_01_thread_priority", "proposed_instruction": "How many priority-2—sorry, I mean priority-3 unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "a61685a400e4575d9758c2788e8b9fa79aee2e2fbbb05895d2f307087f61b6ac", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority-two' but immediately corrects to 'priority-three'. The final value is the same as the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/false_start/false_start_01_thread_priority/20457dfc8376900f/speech.wav", "wer": 0.05, "wav_sha256": "9a4f543d626ce67744644032babf711cc535ca3bb979bae54ac24037512da4bd", "seconds": 11.56, "spoken_utterances": [ "How many priority-two—um... sorry, I mean priority-three hmmm... unread email threads are um... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority two' but immediately self-corrects to 'priority three'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final audible intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many priority two um sorry I mean priority three unread email threads are um in my Gmail inbox?" }, "candidate_id": "wav_1914b38ab05c91da", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a clear false start with 'priority two' corrected to 'priority three', followed by a filler 'um' and a pause. The correction is natural, and the speaker's hesitation and filler use are consistent with someone formulating a request. However, the correction is less integrated than in audio_03, as the speaker repeats the full phrase 'priority three unread email threads' after the correction, which slightly reduces the spontaneity compared to the more fluid correction in audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_email_state", "proposed_instruction": "How many priority-3 read—no, make that unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "267dda1975db3575377dcd7bc00c3554ef5ad6500f07605a66a6a8e6291d13a3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority-three read' but immediately corrects it to 'unread' within the same utterance. The final value for the read/unread status is 'unread', matching the original instruction. All other elements, including priority level and app, remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/false_start/false_start_02_email_state/707cc307c705dc9a/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "d642bbd5e8f0a951955cb59fc26cd6e3f32b2d254688cabd93f1c6836c0d805e", "seconds": 14.04, "spoken_utterances": [ "How many priority-three uh... read—um... no, make that unread email threads are uh, let me think... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'priority three of read' but immediately self-corrects to 'make that unread'. The final audible intent matches the original instruction. The hesitation and repair are typical of a false start, and the correction is applied to the initial value. The rest of the instruction remains unchanged.", "heard_transcript": "How many priority three of read um no make that unread email threads are uh let me think in my Gmail inbox?" }, "candidate_id": "wav_e4e00f4e2f01db22", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'priority three' followed by 'read' and then 'no, make that unread'. While the correction is present, the initial phrase 'priority three of read' is awkward and unclear, and the speaker seems to struggle to rephrase the entire clause. The 'let me think' pause is longer and less integrated than in the others, and the overall flow feels more hesitant and less coherent, making it the least realistic of the three." } ] }, "is_default": false }, { "variant_id": "false_start_04_email_folder", "proposed_instruction": "How many priority-3 unread email threads are in my Gmail drafts—scratch that, make it inbox?", "long_pause_offsets": [], "source_sha256": "651711f24d9a5e20d55657f444a77e235b87708d5c5f31ed8f078c30c63dadc5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'drafts' but immediately corrects to 'inbox' within the same utterance. The final target folder is 'inbox', matching the original instruction. All other elements—priority level, thread status, and app—remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "How many priority-three unread email threads are ... in my Gmail drafts—uh... scratch that, make it inbox?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How many priority-3 email threads are in my Gmail inbox? One more detail: only count unread email threads.", "long_pause_offsets": [ 57 ], "source_sha256": "c473dbc940b3bfd66fdee059dd1bb155a9fb9b62a10f09e141adddd4b20cf41c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing 'unread' constraint, which is a refinement of the initial request. The app (Gmail), priority level (priority-3), and action (count) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/addition/addition_02/dba0b02c2274e073/speech.wav", "wer": 0.0, "wav_sha256": "729d8b85e41c1e39a0ce5ffd140e932a54522f7288c6e6dfff5758b92af0144a", "seconds": 11.76, "spoken_utterances": [ "How many ... priority-three email threads are ... in my Gmail inbox? ", "One more detail: only count unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction is 'How many priority-3 unread email threads are in my Gmail inbox?'. The speaker first asks about 'priority three email threads' and then adds the constraint 'only count unread email threads'. This matches the 'addition' type, where the final goal combines the initial request with the added detail. The final intent is fully specified and executable, matching the original instruction's intent.", "heard_transcript": "How many priority three email threads are in my Gmail inbox? One more detail, only count unread email threads." }, "candidate_id": "wav_1f80c0fd89a8cbae", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Oh, I forgot to mention, it should be priority-3 threads.", "long_pause_offsets": [ 53 ], "source_sha256": "128cfcc5a7463494392446b7075145dde80635baf297c9e3fea7fec93824ef96", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing priority constraint, refining the initial broad request to include only priority-three threads. The app (Gmail), action (count), and state (unread) remain consistent." }, "audio": { "status": "pending", "spoken_utterances": [ "How many ah... unread email threads are hmmm... in my Gmail inbox? ", "Oh, I forgot to mention, it should be um... priority-three threads." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e775c78_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_thread_priority", "proposed_instruction": "How many priority-2 unread email threads are in my Gmail inbox? Wait, make that priority-3, not priority-2.", "long_pause_offsets": [ 64 ], "source_sha256": "4024d829522beaeb9fbd2391e2b824382de90cb6c64704ee6021fdd752375b60", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the priority level from two to three, while retaining the other conditions: unread status, email threads, and Gmail inbox context." }, "audio": { "status": "pending", "spoken_utterances": [ "How many priority-two uh, let's see... unread email threads are in my Gmail inbox? ", "Wait, make that uh... priority-three, not priority-two." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_02_email_state", "proposed_instruction": "How many priority-3 read email threads are in my Gmail inbox? I misspoke about the state: it should be unread, not read.", "long_pause_offsets": [ 62 ], "source_sha256": "d6d89b2934001726d4b82e7ab037c439484f0f41314da03d63fc5321bf488a1b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the state from 'read' to 'unread', preserving the priority level, app, and count action." }, "audio": { "status": "pending", "spoken_utterances": [ "How many let's see... priority-three read email threads are in my Gmail inbox? ", "I misspoke about the state: it should be um... unread, not read." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_03_app", "proposed_instruction": "How many priority-3 unread email threads are in my Outlook inbox? Let me correct the app: use Gmail, not Outlook.", "long_pause_offsets": [ 66 ], "source_sha256": "84cfd787a7975b813f6df5cb83e67f3c4694468717e72b49a30e0469899ec90c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email app from Outlook to Gmail, preserving the priority level, unread status, and thread count request." }, "audio": { "status": "pending", "spoken_utterances": [ "How many um... priority-three unread email threads are uh... in my Outlook inbox? ", "Let me correct the app: use um, let me think... Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_04_email_folder", "proposed_instruction": "How many priority-3 unread email threads are in my Gmail drafts? Change inbox to drafts instead.", "long_pause_offsets": [ 65 ], "source_sha256": "4e5cfa138db953cae655f99634eacc190a1c888bc67146a060f9ebdd1706bfda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the folder from 'inbox' to 'drafts', replacing the original location. The action, priority level, and unread status remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "How many hmmm... priority-three unread email threads are in my Gmail drafts? ", "Change ... inbox to ... drafts instead." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "no_eligible_candidate" }, { "task_id": "e775c78_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "794f3d928dd27ff07c0bdb79e7e4397399fa04c5946640a6cb919a372b82e46f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/cancellation/cancel_008/128bcf6b3e481e1e/speech.wav", "wer": 0.0, "wav_sha256": "c945a41cb8c4936657c0b3ceddc1a9423aea873c9be71900f6d16a5bba14c58b", "seconds": 11.64, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?", "Wait, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins to state the original task but immediately corrects themselves with 'Wait, I want to cancel that task,' which is a clear and final cancellation of the initial request. The audio ends with this cancellation, satisfying the 'cancellation' type criterion.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox. Wait, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_015", "hide_variant_original": true, "source_sha256": "cb80ffafbb4a51edba2546b879c0ab9b88dd6b8af4ee43ae0f97bfdad259ed5d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/ambiguous_halt/halt_015/1d851698578ceadd/speech.wav", "wer": 0.0, "wav_sha256": "ec06d1d073b44789a6e453f34e807ecaba034f8b8270faf9f1f461872f93d474", "seconds": 10.4, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?", "Please wait a moment." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Please wait a moment,' which is a clear instruction to pause. The initial part of the audio is a request for information, but the final interjection is a pause command. Since the type is 'ambiguous_halt,' the final intent to pause is what matters, and it is clearly audible and unambiguous.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox. Please wait a moment." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e775c78_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "4b5a7d51bd87095a3ed30d5d553d174b39d84c37aab26740715b07c1951921da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e775c78_3/audio_natural/filler/original/66ae3b296ee9465f/speech.wav", "wer": 0.0, "wav_sha256": "d2cff44138988eeaf66960d21ad2e19432e3ebd637f3379cf6837a3fb944c2ef", "seconds": 6.88, "spoken_utterances": [ "How many let's see... priority three unread email threads are ... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with 'How many let's see' as a filler, then proceeds to state the full original instruction without any correction or change in intent. The final audible task matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "How many let's see, priority three unread email threads are in my Gmail inbox." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my shortest—I meant longest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "5fe05e48808eac6e7806cfae31d41cc72991330d5a70cc217da7b4aca658c4bc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shortest' to 'longest' within the same utterance, preserving all other conditions including the app (Spotify), metric (minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/false_start/false_start_01_playlist_selection/3057ec531b689866/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "d7c703f6cc81b3d1ad3aa360080cc1bbc001ac23136f8365ac26cf0c2d790d9d", "seconds": 7.84, "spoken_utterances": [ "How long is my shortest—um... I meant uh... longest Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shortest' but immediately self-corrects to 'longest' with 'I meant uh longest'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value ('shortest') is discarded. The rest of the constraints ('Spotify playlist', 'in minutes', 'rounded to the nearest number') are preserved and match the reference. Therefore, the final audible intent matches the original instruction.", "heard_transcript": "How long is my shortest um, I meant uh longest Spotify playlist, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_3b06b48da417bf57", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with 'shortest' and immediately self-corrects to 'longest' with a natural-sounding 'I meant' repair. The correction is fluent and integrated into the sentence, and the rest of the utterance is delivered at a steady, conversational pace with a single, brief filler ('um') at the start. The prosody is consistent and the phrasing sounds like a spontaneous correction during a real task." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How long is my longest Apple Music—sorry, make that Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "ce6c9dda2db75963cd1de8d250ce5339d758aa9dd496bdc9d0924a45fd871ecb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance. The rest of the query, including the action (determine length), the entity (longest playlist), the app (Spotify), the unit (minutes), and the rounding instruction, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/false_start/false_start_02_app/25e8ce0e09a73d06/speech.wav", "wer": 0.23809523809523808, "wav_sha256": "307bbf411b51d2d8e1aafe3d279c20aa5654fba8caf0ea4772f09393f7960a26", "seconds": 8.56, "spoken_utterances": [ "How long is my longest Apple Music—um... sorry, make that uh... Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The rest of the instruction, including the request to round to the nearest number, is preserved. The final audible intent matches the original instruction.", "heard_transcript": "How long is my longest Apple Music um, sorry, make that uh Spotify playlist in minutes, round it to the nearest number?" }, "candidate_id": "wav_f1f62944656bb498", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with 'Apple Music' and corrects to 'Spotify' with a clear 'sorry, make that' self-repair. The correction is natural and the sentence flows well. However, the filler 'um' before 'Apple Music' is slightly longer and less integrated than in audio_03, and the overall delivery has a slightly more hesitant rhythm, making it feel a bit less fluid despite being a valid false start." } ] }, "is_default": false }, { "variant_id": "false_start_04_rounding_rule", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, rounded to the nearest whole number—I should say rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "70ac793377f9c6450fc609610af2f592f8574d39eb0333d3cc84d861bbd73e2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but then self-corrects to 'rounded to the nearest number'. Since 'whole number' and 'number' are semantically equivalent in this context, the final instruction retains the same rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/false_start/false_start_04_rounding_rule/bee6ead206585caf/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "21f5653a7f8135cf5e2cd63d0354c4982dc3520e0df42615d9115e5f8d3e4618", "seconds": 10.56, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, ah... in minutes, rounded to the nearest whole number—uh... I should say rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but immediately self-corrects to 'rounded to the nearest number'. Since this is a false start, the final intent is the corrected version, which matches the original instruction. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How long is my longest Spotify playlist, uh, in minutes, rounded to the nearest whole number? I should say rounded to the nearest number." }, "candidate_id": "wav_b103c81a49084abe", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with 'my who' instead of 'my longest', which is a more significant and less plausible error than a simple app name mix-up. The correction is delivered with 'I should say' but the phrasing 'rounded to the nearest whole number' is awkwardly repeated and then corrected to 'rounded to the nearest number', which sounds like a forced edit rather than a natural thought process. The delivery is less fluent, with multiple pauses and a slightly robotic intonation, reducing its realism." } ] }, "is_default": false }, { "variant_id": "false_start_03_duration_unit", "proposed_instruction": "How long is my longest Spotify playlist, in seconds—no, I mean minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "6ba3d89ddeebf3b4ad3e4f1c0c133a412ef807809140cd540869864ed5f2ba7e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'seconds' but immediately corrects to 'minutes' within the same utterance. The final unit is 'minutes', matching the original instruction. All other components, including the app (Spotify), metric (longest playlist), and rounding instruction, are preserved." }, "audio": { "status": "pending", "spoken_utterances": [ "How long is my uh... longest Spotify playlist, in hmmm... seconds—uh... no, I mean minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How long is my Spotify playlist, in minutes, rounded to the nearest number? Oh, and one more thing: find the longest playlist.", "long_pause_offsets": [ 76 ], "source_sha256": "126e927ea7926ee6428d12197b7a7a62c94c1ec8b6c44d13d8fb60fff60c3109", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the explicit condition to find the longest playlist, which clarifies the ambiguous 'my Spotify playlist' in the initial utterance. The units, rounding, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/addition/addition_01/d4556b77122394e0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "f05288b44702bec1b5f20a1945f16a5282d82c6c09be34a3ab760daf895398ab", "seconds": 15.2, "spoken_utterances": [ "How long is my Spotify playlist, hmmm... in minutes, um... rounded to the nearest number? ", "Oh, and one more thing: find the um, let's see... longest playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How long is my Spotify playlist?' with constraints 'in minutes' and 'rounded to the nearest number'. After a pause, the user adds 'find the... longest playlist'. This addition modifies the initial request by specifying which playlist to measure, changing the scope from 'my playlist' to 'my longest playlist'. The final goal combines the original constraints with the new target, matching the reference instruction. The addition is clear and unambiguous, satisfying the addition type criterion.", "heard_transcript": "How long is my Spotify playlist? Hmm. In minutes. Um. Round it to the nearest number. Oh, and one more thing, find the um, let's see, longest playlist." }, "candidate_id": "wav_4dd5c87149565572", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request ('How long is my Spotify playlist?') followed by a believable hesitation ('Hmm...') and a pause. The addition of 'Oh, and one more thing' is a pragmatic, conversational way to introduce a new request. The second part ('Find the... longest playlist') is delivered with a natural, slightly searching intonation ('let's see'), suggesting spontaneous thought. The prosody is continuous and the pacing realistic, with pauses that feel like genuine thinking rather than robotic breaks." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How long is my longest playlist, in minutes, rounded to the nearest number? I left out a detail: check my Spotify playlist.", "long_pause_offsets": [ 76 ], "source_sha256": "08d146021ee646aec88d8aaac8f49fe53dbe714291757866c9e0985847f24ac3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Spotify' to the initial playlist query, preserving the original intent of measuring the longest playlist's duration in minutes with rounding." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/addition/addition_02/24923256174e0bdc/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "12cc627704fb683cce533c1036b4e89bf7b897d20cd6a2a00243ce02465c3314", "seconds": 14.0, "spoken_utterances": [ "How long is my ... longest playlist, in minutes, hmmm... rounded to the nearest number? ", "I left out a detail: check my um... Spotify playlist." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is 'How long is my longest playlist in minutes, rounded to the nearest number?'. The speaker then adds 'check my um, Spotify playlist' as a clarification. Since the addition specifies the source of the playlist, the final intent is to execute the original task but with the added constraint that the playlist must be from Spotify. The addition is a clarification that narrows the scope, and the final control intent is to execute the combined task.", "heard_transcript": "How long is my longest playlist in minutes, rounded to the nearest number? I left out a detail, check my um, Spotify playlist." }, "candidate_id": "wav_c224c8ed32708960", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio starts with a similar structure but the second part ('I left out a detail') feels slightly less natural. The phrase 'I left out a detail' is a more formal way to introduce a correction than the conversational 'Oh, and one more thing' used in audio_01. The delivery of the second request ('Check my... Spotify playlist') is abrupt and lacks the same searching intonation as audio_01. While the pauses and fillers are present, the overall flow feels a bit more like a rehearsed or less spontaneous addition compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_playlist_selection", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes, rounded to the nearest number? Wait, I meant the longest Spotify playlist, not the shortest.", "long_pause_offsets": [ 85 ], "source_sha256": "ab4febbc98728ae4abcd40d2b01281b950c150148adf2e54b2ee5d5fba3c5d88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial request from 'shortest' to 'longest', preserving the app (Spotify), metric (duration in minutes), and rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/revision/revision_01_playlist_selection/979de3d511f76e5c/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "9924fcff1a4df1733cdee395a73a9771c44399a86b9d754686e272aae031c4c7", "seconds": 13.28, "spoken_utterances": [ "How long is my uh... shortest Spotify playlist, in minutes, rounded to the nearest number? ", "Wait, I meant the um... longest Spotify playlist, not the shortest." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the shortest playlist but immediately corrects themselves to request the longest playlist. This is a revision where the final intent is the corrected version. The final audible request matches the original instruction in meaning, with only the target (longest vs. shortest) being revised. The rest of the constraints (Spotify playlist, minutes, rounded) are preserved.", "heard_transcript": "How long is my uh shortest Spotify playlist in minutes, rounded to the nearest number? Wait, I meant the um longest Spotify playlist, not the shortest." }, "candidate_id": "wav_1c447440039101ee", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural revision flow. The speaker starts with a clear request, hesitates with 'uh' and 'wait', then self-corrects 'shortest' to 'longest' with a believable change-of-mind intonation and a smooth, continuous correction. The prosody is conversational, with appropriate pauses and filler usage that reflect genuine thought processing." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "How long is my longest Apple Music playlist, in minutes, rounded to the nearest number? Let me correct the app: it should be Spotify, not Apple Music.", "long_pause_offsets": [ 88 ], "source_sha256": "bdd78e0aa1c6c06971d36e7419d39846ea4b8ae912b3c13f75a3e09197df3efb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Apple Music to Spotify, while retaining the rest of the query structure including the metric (duration in minutes), the scope (longest playlist), and the rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/revision/revision_02_app/3d8d0c88edf6ec27/speech.wav", "wer": 0.030303030303030304, "wav_sha256": "98bf8226df11f70b7d0f35246bcc812829b56aceb2659fdae1192cdc84969fa3", "seconds": 15.36, "spoken_utterances": [ "How long is my let me think... longest Apple Music playlist, let's see... in minutes, rounded to the nearest number? ", "Let me correct the app: it should be uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause, which matches the revision type criterion. The final audible intent is to query the longest playlist duration in Spotify, consistent with the original instruction. All other constraints (in minutes, rounded) are preserved.", "heard_transcript": "How long is my let me think longest Apple Music playlist, let's see, in minutes, rounded to the nearest number. Let me correct the app, it should be uh Spotify, not Apple Music." }, "candidate_id": "wav_f306b4b52aaab0a3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision but with slightly less natural flow. The speaker uses 'let me think' and 'let's see' as fillers, which are plausible, but the correction 'it should be Spotify not Apple Music' is delivered more abruptly and with less of a conversational backtracking feel compared to audio_01. The prosody is slightly more robotic, and the correction feels more like a mechanical update than a spontaneous recollection." } ] }, "is_default": false }, { "variant_id": "revision_04_rounding_rule", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, not rounded? That should be rounded to the nearest number, not left unrounded.", "long_pause_offsets": [ 66 ], "source_sha256": "3b63332155aaf3a12114881801193fb7f308c65d11cb08f3376b24b73abc0b73", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the initial instruction by replacing 'not rounded' with 'rounded to the nearest number', reversing the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/revision/revision_04_rounding_rule/25f1bfd6a046f7cc/speech.wav", "wer": 0.0, "wav_sha256": "6f2cff7a3e526eca4c5071ca0dab98a13269e6f898d162a02474df259e1af7ca", "seconds": 12.24, "spoken_utterances": [ "How long is my ... longest Spotify playlist, in minutes, let's see... not rounded? ", "That should be rounded to the nearest number, not left unrounded." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'not rounded' but immediately corrects themselves with 'That should be rounded to the nearest number, not left unrounded.' This is a revision where the final audible intent is to have the result rounded. The correction is clear and applies to the original request, satisfying the revision type criterion.", "heard_transcript": "How long is my longest Spotify playlist in minutes, let's see. Not rounded. That should be rounded to the nearest number, not left unrounded." }, "candidate_id": "wav_adac52d79c3255c6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a revision, but the second part sounds like a post-hoc addition rather than a natural correction. The speaker first says 'in minutes let's see', then immediately corrects it to 'rounded to the nearest number'. The phrase 'not left unrounded' is awkward and unnatural, and the correction lacks the typical intonation of a spontaneous realization. The flow feels more like a task instruction being edited than a person thinking aloud." } ] }, "is_default": false }, { "variant_id": "revision_03_duration_unit", "proposed_instruction": "How long is my longest Spotify playlist, in seconds, rounded to the nearest number? Change the unit to minutes, not seconds.", "long_pause_offsets": [ 84 ], "source_sha256": "ac8b1148ec27ea1f33b10fa62ffc368da767e6a2832a2c19f899a00fccf41de2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the unit from seconds to minutes, preserving the rest of the request including the app, the object (longest playlist), and the rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/revision/revision_03_duration_unit/8cbf44e7bf5c2ebe/speech.wav", "wer": 0.0, "wav_sha256": "532ac2d69370feb7322c609b051e9208e101a4157ba4bba9c9161de7b6df1286", "seconds": 13.4, "spoken_utterances": [ "How long is my ... longest Spotify playlist, in seconds, ... rounded to the nearest number? ", "Change the unit to ... minutes, not seconds." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the duration in seconds but immediately corrects the unit to minutes. This is a revision where the final audible intent matches the original instruction. The correction is clear and unambiguous, and the rest of the constraints (longest playlist, rounded to nearest number) are preserved.", "heard_transcript": "How long is my longest Spotify playlist in seconds, rounded to the nearest number? Changed the unit to minutes, not seconds." }, "candidate_id": "wav_324278a8d159f6d5", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has the least realistic delivery. The speaker makes a request, then immediately corrects the unit from seconds to minutes. However, the correction is delivered as a flat, declarative statement without any hesitation, filler, or change-of-mind intonation. The prosody is monotonous and lacks the natural rhythm of someone formulating a request and then adjusting it. It sounds more like a pre-written prompt being read than a spontaneous utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "bebcc38a6d6e809b47da2ba323f5a9973e1216861dbbed4540d981bf108c5491", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/cancellation/cancel_021/7b6b25b942a0d622/speech.wav", "wer": 0.08, "wav_sha256": "3b831792cf8c48a73b21c26245bdd62feea1b4efb1f5842e9fa48e6728b8d162", "seconds": 12.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?", "Wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the length of their longest Spotify playlist, but immediately follows up with 'Wait, never mind, I don't need that done anymore,' which is a clear and final cancellation of the task. The audio ends with this cancellation, so the final control intent is to cancel, matching the 'cancellation' type criterion.", "heard_transcript": "How long is my hmm longest Spotify playlist in minutes, um rounded to the nearest number. Wait, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "e91f23a0a7f8c8972d7e399c446bc6fc60d3857e539d94f619f9cd9e1b9dfab2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/ambiguous_halt/halt_022/854dd8c0bf2a3378/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "bc366daa19bf877a61b434f17c5480e52c0a7bd1ae82d8c69b71fcc24f3846e6", "seconds": 11.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the user does not provide a new instruction. Therefore, the final control intent is to pause, and the task is not executed.", "heard_transcript": "How long is my longest Spotify playlist in minutes, um, round it to the nearest number. Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "3467c579f5e2259aefae8776a0516f63a67f9fb1809b7e9dd22df556c875637f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_1/audio_natural/filler/original/96845a344b805d03/speech.wav", "wer": 0.125, "wav_sha256": "8c49cf52e2b0b5f3c8d7cbbdaac48aa4021c9e1fdf55aaceb96091d5b52fbfea", "seconds": 7.64, "spoken_utterances": [ "How long is my hmmm... longest Spotify playlist, in minutes, um... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes filler words ('um') but does not change the core request. The final intent matches the original instruction, and the filler does not alter the task goal. The instruction is executed as intended.", "heard_transcript": "How long is my longest Spotify playlist in minutes, um, rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my longest—I meant shortest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "317305f2eacaeda6eba583f213c4ab70b8e2be21fc866f8c81e1b8e73e1eb9da", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'longest' to 'shortest' within the same utterance, preserving all other conditions including the app (Spotify), metric (duration in minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/false_start/false_start_01_playlist_selection/6cc777ed5e25843f/speech.wav", "wer": 0.05, "wav_sha256": "8db40a99269c394024ae7508cf51253447f669c83c90f7c116b36b7298aaa3ce", "seconds": 11.0, "spoken_utterances": [ "How long is my let's see... longest—ah... I meant shortest Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'longest' but immediately self-corrects to 'shortest'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final audible intent matches the original instruction exactly, with only a minor grammatical repair ('I meant' instead of 'I meant to'). All constraints (Spotify playlist, duration in minutes, rounded to nearest number) are preserved.", "heard_transcript": "How long is my, let's see, longest uh, I meant shortest Spotify playlist, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_be91bdc0c9f6a2c1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This clip exhibits the most natural and fluent self-correction. The speaker begins with 'How long is my... let's see...', which sounds like genuine hesitation. The false start 'longest' is immediately and smoothly corrected to 'shortest' with a believable 'I meant' phrase. The rest of the sentence flows naturally without further disruption, and the prosody is consistent and conversational." } ] }, "is_default": true }, { "variant_id": "false_start_03_duration_unit", "proposed_instruction": "How long is my shortest Spotify playlist, in seconds—sorry, I mean minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "ec55c00f6da7c36af715aa664018b59b6a41c2e85d894a30232f2f012ded7a08", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'seconds' but immediately corrects to 'minutes' within the same utterance. The final intended unit is minutes, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/false_start/false_start_03_duration_unit/ae5f98edad56c142/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "ac6c0b0e1fb2493df105a79ef2340e46dda4a7814b0aa07b9fb16f88105be8a1", "seconds": 10.04, "spoken_utterances": [ "How long is my shortest Spotify playlist, ... in seconds—um... sorry, I mean minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'seconds' but immediately self-corrects to 'minutes' with 'um, sorry, I mean minutes'. This is a clear false start where the final, corrected value is 'minutes'. The rest of the instruction, including 'rounded to the nearest number', is preserved. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "How long is my shortest Spotify playlist in seconds? um, sorry, I mean minutes, rounded to the nearest number?" }, "candidate_id": "wav_5a854894f38c2bae", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear self-correction from 'seconds' to 'minutes', which is a plausible error. The speaker says 'um' before the correction, indicating a moment of realization. The correction is delivered smoothly and the sentence continues without awkwardness. The prosody is natural, and the correction feels spontaneous rather than forced." } ] }, "is_default": false }, { "variant_id": "false_start_04_rounding_rule", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes, rounded to the nearest whole number—no, that should be rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "2ee09b737816a8497a5a986eeb56915420af95a2af9aa387b3c3371d56ef1943", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but immediately corrects it to 'rounded to the nearest number'. Since 'whole number' and 'number' are semantically equivalent in this context, the final instruction remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/false_start/false_start_04_rounding_rule/5eeb06f5d2123b01/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "4674174a7b45af34370cb06867cf0d5df77ec6184c846765ca1d3e8df9efe75d", "seconds": 12.8, "spoken_utterances": [ "How long is my shortest Spotify playlist, um, let me think... in minutes, rounded to the nearest whole number—um... no, that should be rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but immediately self-corrects to 'rounded to the nearest number'. Since this is a false_start type, the final corrected instruction is the one to execute. The meaning of 'rounded to the nearest number' is equivalent to the original instruction's 'rounded to the nearest number', so the final intent matches the original instruction.", "heard_transcript": "How long is my shortest Spotify playlist, um, let me think, in minutes, rounded to the nearest whole number, um, no, that should be rounded to the nearest number." }, "candidate_id": "wav_f6b1eeafdca30d42", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This clip contains a self-correction, but the flow is less natural. After the initial request, the speaker says 'um, let me think', which is a plausible filler. However, the correction 'no, that should be rounded to the nearest number' is delivered with a slightly abrupt and less conversational tone. The phrase 'rounded to the nearest whole number' is then replaced with 'rounded to the nearest number', which sounds like a mechanical edit rather than a natural rephrasing." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "How long is my shortest Apple Music—make that Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "f8f4472959b1a1c601b02a38a042c310f7ac48379173d3797b6c2955555834db", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance. The rest of the query, including the action (determine duration), the entity (shortest playlist), the app (Spotify), and the formatting (in minutes, rounded), remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/false_start/false_start_02_app/558fef921d0243b2/speech.wav", "wer": 0.2, "wav_sha256": "bc2699a365e7e93e459ba35e7932f5619645cf07e8b91a224f1d2fcf5733d802", "seconds": 8.96, "spoken_utterances": [ "How long is my shortest Apple Music—uh... make that Spotify playlist, um... in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify' with 'make that Spotify playlist'. This is a false start where the final corrected target matches the original instruction. The rest of the constraints (duration in minutes, rounded to nearest number) are preserved. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "How long is my shortest Apple Music uh, make that Spotify playlist um, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_d3d8fcda6104ee86", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "This clip has a false start with 'Apple Music' followed by a correction to 'Spotify playlist'. However, the correction is abrupt and the sentence structure feels disjointed. The speaker says 'uh' and 'um' in a way that sounds like they are searching for the right words, but the overall delivery lacks the smoothness of the other clips. The prosody is somewhat flat, and the correction feels less integrated into the flow of the sentence." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes,? Oh, I forgot to mention, round the result to the nearest number.", "long_pause_offsets": [ 55 ], "source_sha256": "abe2d8d096d10314f626dcb37c87f831219fdb60abcc5bd3211de6fc707d12cb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rounding instruction, which was not in the initial utterance but is now included in the final goal. The original instruction already specified the playlist and unit, and the addition completes the requirement without changing any other aspect." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/addition/addition_01/e3762ebae063370b/speech.wav", "wer": 0.0, "wav_sha256": "879200178d082e6483af1c3e85f399372025fbabeaac80bc45ea76ae3ca61263", "seconds": 14.96, "spoken_utterances": [ "How long is my let me think... shortest Spotify playlist, ... in minutes,? ", "Oh, I forgot to mention, round the result um... to the nearest number." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is a slightly hesitant version of the original instruction. The speaker then adds the constraint 'round the result to the nearest number' after a pause, which is an addition to the original request. The final intent combines the original task with the added rounding requirement, matching the expected behavior for an 'addition' type. The final control intent is to execute the combined task.", "heard_transcript": "How long is my, let me think, shortest Spotify playlist, in minutes? Oh, I forgot to mention, round the result, um, to the nearest number." }, "candidate_id": "wav_1e4e70ccc435cab6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think') and then adds a missing detail ('round the result') after a brief pause, which is followed by a second, more specific addition ('to the nearest number'). The flow is continuous and the speaker's intonation reflects genuine recollection and correction, making the addition feel spontaneous and realistic." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How long is my shortest Spotify playlist,, rounded to the nearest number? I left out a detail: express the duration in minutes.", "long_pause_offsets": [ 74 ], "source_sha256": "2a31412e17f81fb3ad91945e31e765ff2e1d5e44b1cb48f9c186da47bb83c2c0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing unit requirement 'in minutes', which was implied but not stated in the initial utterance. The final goal now includes both the original intent (shortest playlist duration) and the added constraint (expressed in minutes)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/addition/addition_02/e5d9d6776e10d5f0/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "02a095168b94a5fee480824ab232ed50c896d54a4c9b8b4b0716ca8899ce82ee", "seconds": 16.16, "spoken_utterances": [ "How long is my um, let me think... shortest Spotify playlist,, rounded to the nearest number? ", "I left out a detail: express the duration uh... in minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is a filler-filled version of the original instruction. The speaker then adds a clarifying detail ('Express the duration in minutes') after a pause, which is a valid addition. The final goal combines the original request with the added constraint, matching the reference instruction exactly.", "heard_transcript": "How long is my um, let me think, shortest Spotify playlist, rounded to the nearest number? I left out a detail. Express the duration in minutes." }, "candidate_id": "wav_b67166ec2c0fcffd", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker also adds a missing detail ('express the duration in minutes'), but the phrasing is slightly more awkward and less fluid. The initial hesitation ('um') and 'let me think' are natural, but the addition comes after a longer, more abrupt pause and the speaker's tone sounds slightly more like a rehearsed correction than a spontaneous thought. The overall delivery is less seamless than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_playlist_selection", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, rounded to the nearest number? Wait, I meant the shortest playlist, not the longest.", "long_pause_offsets": [ 84 ], "source_sha256": "01e0807656c22d2494456cb37b63dbc88e2ac6c398ab67db112f78b06a4c3048", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects 'longest' to 'shortest', preserving the app (Spotify), unit (minutes), and rounding instruction. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/revision/revision_01_playlist_selection/fc8bce102fcc81b7/speech.wav", "wer": 0.08, "wav_sha256": "3aa3cc2bbeeca6e3772e6830d8bc4065f998eb0f31d102f34537a53b78e84489", "seconds": 13.48, "spoken_utterances": [ "How long is my let's see... longest Spotify playlist, ... in minutes, rounded to the nearest number? ", "Wait, I meant the shortest playlist, not the longest." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'longest' playlist but immediately corrects themselves to 'shortest' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the initial instruction. The final audible intent matches the original reference instruction exactly, so the task should be executed as requested.", "heard_transcript": "How long is my let's see, longest Spotify playlist in minutes, rounded to the nearest number? Wait, I meant the shortest playlist, not the longest." }, "candidate_id": "wav_829c682bec6dfbd6", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects with a plausible 'Wait, I meant...' followed by a smooth revision. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pause before the correction is brief and realistic, fitting a moment of realization." } ] }, "is_default": true }, { "variant_id": "revision_04_rounding_rule", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes, rounded to the nearest ten? That should be rounded to the nearest number, not rounded to the nearest ten.", "long_pause_offsets": [ 82 ], "source_sha256": "f0af0b5017e720d973c5a9033c21e52be16675597741322b938dace8fe047983", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the rounding rule from 'nearest ten' to 'nearest number', reversing the intended precision." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/revision/revision_04_rounding_rule/d9061ffe3ce4ddcf/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "fd143b168446f5edac8b09c8e80c428b2f3c642691cff1937918d44b1df5ffc7", "seconds": 13.84, "spoken_utterances": [ "How long is my shortest ... Spotify playlist, um... in minutes, rounded to the nearest ten? ", "That should be rounded to the nearest number, not rounded to the nearest ten." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest ten' but immediately corrects themselves to 'rounded to the nearest number'. This is a revision where the final, corrected instruction matches the original reference. The correction is clear and unambiguous, so the task should be executed as revised.", "heard_transcript": "How long is my shortest Spotify playlist, um, in minutes, rounded to the nearest ten? That should be rounded to the nearest number, not rounded to the nearest ten." }, "candidate_id": "wav_5aa51be1ad3b8e77", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then hesitates with 'um' before stating a rounding instruction. After a short pause, they correct themselves with a believable 'That should be rounded to the nearest number, not rounded to the nearest ten.' The correction is integrated into the same utterance, and the intonation reflects a genuine change of mind. The delivery is slightly less fluid than audio_01 but still very natural." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How long is my shortest Apple Music playlist, in minutes, rounded to the nearest number? Let me correct the app: use Spotify, not Apple Music.", "long_pause_offsets": [ 89 ], "source_sha256": "e778c1ace41415f360daee4e821c165f6aaaa2f53d01d827611cd7a7029f3a30", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music app from Apple Music to Spotify, while retaining the original query about the shortest playlist duration in minutes, rounded to the nearest number." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/revision/revision_02_app/f4e788dc13d814e5/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "9a9d0a512abf8392424d7aa41322e6cf357ac8fa9915010dc3dd439249f7c7e1", "seconds": 14.12, "spoken_utterances": [ "How long is my ah... shortest Apple Music playlist, in minutes, rounded to the nearest number? ", "Let me correct the app: use um... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially requested the duration of the shortest Apple Music playlist but immediately corrected the app to Spotify. The revision is clear and unambiguous, and the final intent is to execute the original task with the corrected app. All other constraints (duration in minutes, rounded to nearest number) remain unchanged.", "heard_transcript": "How long is my um shortest Apple Music playlist in minutes, rounded to the nearest number? Let me correct the app. Use um Spotify, not Apple Music." }, "candidate_id": "wav_34517a4cebbb5a42", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with a clear request but makes a minor error ('arm' instead of 'shortest') and then corrects it with 'Let me correct the app.' The correction is somewhat abrupt and less integrated into the flow. The prosody is slightly off, and the phrase 'Let me correct the app' sounds a bit mechanical, possibly due to a slight mispronunciation or awkward phrasing. Still, the overall structure is realistic." } ] }, "is_default": false }, { "variant_id": "revision_03_duration_unit", "proposed_instruction": "How long is my shortest Spotify playlist, in seconds, rounded to the nearest number? Change the unit to minutes, not seconds.", "long_pause_offsets": [ 85 ], "source_sha256": "90018f854b2d7c0fef44767c89c352d9fde6e03173ea3b802f173c8d13fcfe58", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the unit from seconds to minutes, preserving the rest of the request including the app, the metric (shortest playlist), and the rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/revision/revision_03_duration_unit/5abb6e859e28ee51/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "d3d294001314fb72e3315db14c373e16ec7b59035fd30cdd3a9e1d47aba33e09", "seconds": 13.92, "spoken_utterances": [ "How long is my um... shortest uh... Spotify playlist, in seconds, rounded to the nearest number? ", "Change the unit to minutes, not seconds." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the duration in seconds but immediately corrects the unit to minutes with 'Change the unit to minutes, not seconds.' This constitutes a revision where the final instruction overrides the initial incorrect unit. The final audible intent matches the original reference instruction.", "heard_transcript": "How long is my own shortest uh Spotify playlist in seconds, rounded to the nearest number? Change the unit to minutes, not seconds." }, "candidate_id": "wav_dc02cff894af6c60", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker begins with a clear request but makes a significant error ('my own' instead of 'my shortest') and then corrects it. The correction is abrupt and the phrase 'Change the unit to minutes, not seconds' sounds more like a command than a natural self-correction. The prosody is inconsistent, and the delivery feels less spontaneous, possibly due to a stutter or mispronunciation. The overall flow is less natural compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_008", "hide_variant_original": true, "source_sha256": "8d5367d3ce28babfa5dc203cc0d1ca850c83b1b12f3c15dc2115ef69ddf51008", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/cancellation/cancel_008/95a8806dcae486d9/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "d786c2dff34119bcbbb2792668a1127491666afbab6e7159f7ba008a6f2c68a1", "seconds": 12.96, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?", "Hang on, I want to cancel that task." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, I want to cancel that task.' This is a clear and direct cancellation of the entire task. The final interjection is a cancellation command, not a pause or a revision. Therefore, the control intent is correctly identified as 'cancel'.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes rounded to the nearest number? Hang on, I want to cancel that task." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "b3a9d638d33bb07a4e6dcaf9221d1dcbd35ef02aa042099a0e882de0a79f095d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/ambiguous_halt/halt_033/5bbf39d0b8b4bafa/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "5e75c414c361598d651d5b0a1b43a4a9cbf2d073f05313852b7d0a0b9cadc525", "seconds": 12.16, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then audibly pauses with 'Hold on, there's something', which is a clear request to hold/wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, and the speaker intends to resume after a brief interruption.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes, rounded to the nearest number? Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "594bd3005e1345aba27cb4f9bb8eeac28e54fc240cea02fe6d792e4b65e2e66e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_2/audio_natural/filler/original/3e2e7e3a4cf9f1a0/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "b70ba9b48abc18d2418ca3fd9331a2549d3dcdbd28eae26d059b87eaf04b169f", "seconds": 8.28, "spoken_utterances": [ "How long is my um, let's see... shortest Spotify playlist, uh... in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('um', 'let's see', 'uh') but does not change the core request. The final intent is clear and matches the original instruction. The filler words do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How long is my um let's see shortest Spotify playlist uh in minutes rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_playlist_selection", "proposed_instruction": "How long is my shortest—no, I mean longest Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "0588c694f1fee35923ac5d7b49363571e30c274811b8559009d0db91e58a3783", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'shortest' to 'longest' within the same utterance, preserving all other conditions including the app (Spotify), metric (duration in minutes), and rounding requirement." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/false_start/false_start_01_playlist_selection/6f7b78e672bd2e33/speech.wav", "wer": 0.0, "wav_sha256": "4b3f082a2ce7434258487b79c36ebd12a4616aa6392feda2fef821347a40d939", "seconds": 7.4, "spoken_utterances": [ "How long is my shortest—um... no, I mean longest Spotify playlist, in minutes, ... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'shortest' but immediately self-corrects to 'longest'. According to the 'false_start' type criterion, the final corrected value ('longest') should be applied, and the original incorrect value ('shortest') should be discarded. The rest of the instruction, including 'Spotify playlist', 'in minutes', and 'rounded to the nearest number', remains unchanged and is preserved. Therefore, the final intent matches the original instruction.", "heard_transcript": "How long is my shortest um, no, I mean longest Spotify playlist in minutes, rounded to the nearest number?" }, "candidate_id": "wav_babf050632e2b64d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with 'How long is my shortest' and immediately self-corrects to 'longest Spotify playlist'. The correction is smooth, with a natural pause and intonation shift. The phrasing 'rounded to the nearest number' is delivered without hesitation, suggesting a spontaneous, realistic thought process rather than a rehearsed or robotic delivery." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How long is my longest Apple Music—I should say Spotify playlist, in minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "81c047aa0bfc8be744f48d42cb19ce4a378fe4158fb89be92b3dfdefffb11d2a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Apple Music' to 'Spotify' within the same utterance, preserving the rest of the query including the metric (minutes), rounding instruction, and scope (longest playlist)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/false_start/false_start_02_app/78f5f323e1d2dc7d/speech.wav", "wer": 0.0, "wav_sha256": "ec7edb63dd29f98e61dda642524fac34622c90b45210060447f70ada2ec51d36", "seconds": 10.0, "spoken_utterances": [ "How long is my longest Apple Music—um... I should say uh... Spotify playlist, in minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final intent is the corrected version, which matches the original instruction exactly. The correction is immediate and unambiguous, so the task should be executed as requested.", "heard_transcript": "How long is my longest Apple Music um, I should say uh Spotify playlist, in minutes, rounded to the nearest number?" }, "candidate_id": "wav_5262a2b4db1344ee", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with 'How long is my longest Apple Music' and corrects to 'Spotify playlist'. The correction is clear and the pause before 'I should say' sounds like a genuine moment of reconsideration. The delivery is slightly more hesitant than audio_01, but still very natural and credible." } ] }, "is_default": false }, { "variant_id": "false_start_03_duration_unit", "proposed_instruction": "How long is my longest Spotify playlist, in seconds—sorry, make that minutes, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "6a92a30b920cb40940b93b3d70e07d04453bc8f0aeeface398e1541e8ac1962f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'seconds' but immediately corrects to 'minutes' using 'make that minutes'. The rest of the instruction, including the rounding requirement, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/false_start/false_start_03_duration_unit/2313b8465faa6fde/speech.wav", "wer": 0.1, "wav_sha256": "7fb14fcde1322fdb26a3faa4785ea8b5d2e17d3bf7bb7eb98080437c78bb0cb5", "seconds": 11.16, "spoken_utterances": [ "How long is my uh... longest Spotify playlist, in seconds—hmmm... sorry, make that minutes, rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the duration in seconds but immediately self-corrects to minutes. According to the 'false_start' type criterion, the final corrected value (minutes) should be applied, and the original incorrect value (seconds) should be discarded. The final intent matches the original instruction, with the only difference being the correction of the unit from seconds to minutes.", "heard_transcript": "How long is my uh longest Spotify playlist in seconds? Sorry, make that minutes, rounded to the nearest number." }, "candidate_id": "wav_c0a0ba2afd0e0988", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with 'How long is my uh longest Spotify playlist in seconds' and corrects to 'minutes'. The 'uh' and the self-correction are realistic, but the correction is abrupt and the intonation shift feels slightly less fluid than in the top two. The 'sorry' is a bit too formal for a spontaneous correction, making it slightly less natural." } ] }, "is_default": false }, { "variant_id": "false_start_04_rounding_rule", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, rounded to the nearest whole number—let me correct that, rounded to the nearest number?", "long_pause_offsets": [], "source_sha256": "3fb90df1a43459d3119598f4e4ade59a8770a03339a49ebf7d33fd15628349a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but immediately corrects it to 'rounded to the nearest number'. The final phrase is identical to the original instruction, and 'whole number' and 'number' are semantically equivalent in this context when referring to rounding." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/false_start/false_start_04_rounding_rule/a70b78e8d3d8abf6/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "c713692cc5263bbbcbbfbb0a312dc3ceee58bcca54eac790e81d7a314765480e", "seconds": 13.2, "spoken_utterances": [ "How long is my longest Spotify playlist, um... in minutes, rounded to the nearest whole number—ah... let me correct that, ... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'rounded to the nearest whole number' but immediately self-corrects to 'rounded to the nearest number'. Since this is a false_start type, the final intent is determined by the immediate correction. The phrase 'rounded to the nearest number' is semantically equivalent to 'rounded to the nearest whole number' in this context, as 'number' implies an integer value. Therefore, the final intent matches the original instruction with no meaningful discrepancy.", "heard_transcript": "How long is my longest Spotify playlist, um, in minutes, rounded to the nearest whole number? Let me correct that. Rounded to the nearest number?" }, "candidate_id": "wav_5ee013b95b052c09", "naturalness": { "rank": 4, "mean_rank": 4.0, "singleton": false, "reasons": [ { "round": 1, "rank": 4, "reason": "The speaker says 'How long is my longest Spotify playlist um in minutes, rounded to the nearest whole number' and then says 'Let me correct that' before repeating 'rounded to the nearest number'. The correction is redundant and the phrase 'let me correct that' sounds rehearsed rather than spontaneous. The pause and delivery feel more like a scripted edit than a natural self-correction." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How long is my longest Spotify playlist,, rounded to the nearest number? Oh, and one more thing: return the duration in minutes.", "long_pause_offsets": [ 73 ], "source_sha256": "2e72a95757dad187a9abdd09e53180d9bfa2ba4b7376a57625282a9c3d5e4704", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing unit 'in minutes' to the initial request, which was otherwise complete. The original instruction already specified rounding to the nearest number, so the final goal matches exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/addition/addition_02/ad8cc0501faff1e5/speech.wav", "wer": 0.0, "wav_sha256": "d4a05d882f8f090a05ce1f6e883af047f77c6e0b5effd8b68336ef36a2bef25c", "seconds": 14.8, "spoken_utterances": [ "How long is my let me think... longest Spotify playlist,, um... rounded to the nearest number? ", "Oh, and one more thing: return the duration ... in minutes." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial utterance contains hesitation ('let me think') and filler ('um'), but the core request 'How long is my longest Spotify playlist' is clear. The addition 'return the duration in minutes' after a pause refines the output format. The final intent combines the original request with the added constraint, matching the reference instruction. The 'rounded to the nearest number' is preserved as part of the original request.", "heard_transcript": "How long is my let me think longest Spotify playlist um rounded to the nearest number. Oh, and one more thing, return the duration in minutes." }, "candidate_id": "wav_fe2ffd227d572d03", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker uses natural fillers like 'let me think' and 'um' with plausible timing and intonation. The addition of 'Oh, and one more thing' is a realistic, conversational way to append a detail. The prosody is continuous and the pacing reflects genuine thought, making the delivery highly spontaneous and believable." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How long is my longest Spotify playlist, in minutes,? Oh, I forgot to mention, round the result to the nearest number.", "long_pause_offsets": [ 54 ], "source_sha256": "d9df06152cd61be90fe7283752ff92d51a604a1ad7a83613d5cfd3ab5adca90d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing rounding instruction, which is a constraint on the result format. The original instruction already specifies the entity (longest Spotify playlist), unit (minutes), and the action (determine length). The addition completes the requirement without altering any other aspect." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/addition/addition_01/4fb796b78b02e965/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "a4a18d9c20dcc04c0ae643b88a2a768c3937e88dfc2fe66a1876fc5b0548e986", "seconds": 14.12, "spoken_utterances": [ "How long is my ah... longest Spotify playlist, ... in minutes,? ", "Oh, I forgot to mention, round the result ... to the nearest number." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks for the duration of the longest Spotify playlist in minutes. After a pause, they add the constraint to round the result to the nearest number. This matches the 'addition' type, where the final goal combines the original instruction with the added constraint. The final audible intent is clear and executable.", "heard_transcript": "How long is my uh longest Spotify playlist in minutes? Oh, I forgot to mention, round the result to the nearest number." }, "candidate_id": "wav_dbb1a11c66a5e497", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker uses a filler ('uh') and a self-correction ('Oh, I forgot to mention'), which is a plausible repair. However, the addition of the rounding instruction feels slightly tacked on and less smoothly integrated than in audio_02. The prosody is natural, but the flow of the addition is less organic, making it slightly less realistic overall." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How long is my longest Apple Music playlist, in minutes, rounded to the nearest number? Make that Spotify, not Apple Music.", "long_pause_offsets": [ 88 ], "source_sha256": "54787e8d5f87dfd45247e45a9835db5d3698bfa33e9e5670aecbf7ff2b96ff9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music service from Apple Music to Spotify, preserving the rest of the query including the metric (duration in minutes), the scope (longest playlist), and the rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/revision/revision_02_app/629a26a5328b80d9/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "8d2ff85d864d0163809f83f48e8cee9d18e48b314b4babbfa0a0a373bb4e8c71", "seconds": 14.88, "spoken_utterances": [ "How long is my uh... longest Apple Music playlist, ah... in minutes, rounded to the nearest number? ", "Make that uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Apple Music but immediately corrects to Spotify. According to the revision type criterion, the correction spoken after the initial instruction must be applied. The final audible intent matches the original reference instruction exactly, with only a minor grammatical difference ('Make that' vs. 'How long is') that does not alter the meaning.", "heard_transcript": "How long is my uh longest Apple Music playlist uh in minutes, rounded to the nearest number. Make that uh Spotify, not Apple Music." }, "candidate_id": "wav_0cd8f7ea3fe22564", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant request ('How long is my...') and then self-corrects the platform from 'Apple Music' to 'Spotify' with a natural-sounding 'Make that's uh...'. The correction is integrated smoothly into the same utterance, maintaining a consistent rhythm and prosody. The pause before the correction is brief and plausible, fitting a moment of realization. The overall delivery sounds like a spontaneous, real-time correction without any robotic or disjointed qualities." } ] }, "is_default": true }, { "variant_id": "revision_01_playlist_selection", "proposed_instruction": "How long is my shortest Spotify playlist, in minutes, rounded to the nearest number? Wait, I meant the longest Spotify playlist, not the shortest.", "long_pause_offsets": [ 85 ], "source_sha256": "ab4febbc98728ae4abcd40d2b01281b950c150148adf2e54b2ee5d5fba3c5d88", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial request from 'shortest' to 'longest', preserving the app (Spotify), metric (duration in minutes), and rounding instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/revision/revision_01_playlist_selection/d84282be428647ce/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "06633aa98afad6b06c5b365c7d171fe8366d3622653a165eff003a6b0e3f4c8d", "seconds": 15.6, "spoken_utterances": [ "How long is my um... shortest Spotify playlist, uh... in minutes, rounded to the nearest number? ", "Wait, I meant the ... longest Spotify playlist, not the shortest." ] }, "audio_semantic": { "status": "equivalent", "reason": "The user initially asks for the shortest playlist but immediately corrects themselves to the longest playlist. This is a revision where the final intent is the corrected version. The final audible request matches the original instruction in meaning, with only the target (longest vs shortest) being revised. The rest of the constraints (in minutes, rounded to the nearest number) are preserved.", "heard_transcript": "How long is my um shortest Spotify playlist uh in minutes, rounded to the nearest number? Wait, I meant the longest Spotify playlist, not the shortest." }, "candidate_id": "wav_c57ecc4c19f30c4b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request but then corrects the playlist type from 'shortest' to 'longest' after a longer, more deliberate pause ('Wait, I meant the...'). While this is a plausible revision, the pause is longer and more pronounced than in audio_01, and the correction feels slightly more like a separate thought than a fluid self-correction. The prosody remains natural, but the overall flow is less continuous, making it slightly less realistic than audio_01." } ] }, "is_default": false }, { "variant_id": "revision_03_duration_unit", "proposed_instruction": "How long is my longest Spotify playlist, in seconds, rounded to the nearest number? Change the unit to minutes, not seconds.", "long_pause_offsets": [ 84 ], "source_sha256": "ac8b1148ec27ea1f33b10fa62ffc368da767e6a2832a2c19f899a00fccf41de2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the unit from seconds to minutes, preserving the rest of the request including the app, the object (longest playlist), and the rounding instruction." }, "audio": { "status": "pending", "spoken_utterances": [ "How long is my longest Spotify playlist, ... in seconds, rounded to the nearest number? ", "Change the unit to uh... minutes, not seconds." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "revision_04_rounding_rule", "proposed_instruction": "How long is my longest Spotify playlist, in minutes, rounded to the nearest ten? That should be rounded to the nearest number, not rounded to the nearest ten.", "long_pause_offsets": [ 81 ], "source_sha256": "8b1a6d4149719b031cf3dbc9293f4a8e727e0434c32ed1c0df012494119e0603", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects the rounding rule from 'nearest ten' to 'nearest number', reversing the original specification." }, "audio": { "status": "pending", "spoken_utterances": [ "How long is my uh... longest Spotify playlist, in minutes, ... rounded to the nearest ten? ", "That should be rounded to the nearest number, not rounded to the nearest ten." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_021", "hide_variant_original": true, "source_sha256": "c13f69783c13ef2d2b02fd0dc4cc491e7ccf03f886c1dac5832afab4b6d43a11", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/cancellation/cancel_021/a3db7f6f7ea14a26/speech.wav", "wer": 0.0, "wav_sha256": "8cab0c9fbe731962fd50bcce52d8ba44fca622e1f73aa9214924de586528cc71", "seconds": 13.08, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?", "Wait, never mind. I don't need that done anymore." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Wait, never mind, I don't need that done anymore.' This constitutes a direct cancellation of the entire task as per the type-specific criterion. The final intent is to cancel, not to execute.", "heard_transcript": "How long is my uh, let me think. longest Spotify playlist in minutes, rounded to the nearest number. Wait, never mind, I don't need that done anymore." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "c73522420e129c08c49acbc20cb6d5cff2834712a5b27f921f23edae84e79a68", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/ambiguous_halt/halt_004/3599c3c22bcf6f79/speech.wav", "wer": 0.0, "wav_sha256": "982f5077b01afe0bdb401b48a95ca417f0b41eec0c4c392f9dcb67224a3119ca", "seconds": 11.48, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains a filler phrase 'let me think' followed by a clear request to pause, 'Just a second.' This constitutes an unambiguous halt request. The original instruction is not canceled or revised, and the user's intent is to delay execution, not to execute or cancel the task. Therefore, the final control intent is 'pause', which matches the type-specific criterion.", "heard_transcript": "How long is my uh, let me think. longest Spotify playlist in minutes, rounded to the nearest number. Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e7a10f8_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d5e86970559f550ec887d5ef5a7dee439369aabdecf7d6213c9f00784afc63a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e7a10f8_3/audio_natural/filler/original/38998782e0c0acc0/speech.wav", "wer": 0.0, "wav_sha256": "eb6498a4ccee2fd7ac0b1813a37d0b30ac8005b8203a63f25c71b95075da7334", "seconds": 7.96, "spoken_utterances": [ "How long is my uh, let me think... longest Spotify playlist, in minutes, ... rounded to the nearest number?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes the filler 'uh' and 'let me think' but does not change the core request. The final, complete instruction matches the original reference in both action and constraints. The 'filler' type allows for such hesitations and backchannels without altering the task goal.", "heard_transcript": "How long is my uh, let me think, longest Spotify playlist in minutes, rounded to the nearest number?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the least played—no, I mean most played song by Velvet Echo on Spotify.", "long_pause_offsets": [], "source_sha256": "785912f5d65d7b64d343eec054393633cd5cd5ebb9f2b636016b1ab75d4d0a0c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately corrects to 'most played' within the same utterance. The final intended action is to identify the most played song, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/false_start/false_start_01_song_selection/0d6aa09a4b0b927f/speech.wav", "wer": 0.125, "wav_sha256": "382742177e7122f05a27daabbf400bae5acf5403cd7f788b9584dfa6f0285db3", "seconds": 11.56, "spoken_utterances": [ "What is the title of the least played—uh... no, I mean hmmm... most played song by um, let's see... Velvet Echo on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately self-corrects to 'most played' with 'no, I mean'. This is a false start where the final intent is the corrected version. The rest of the instruction, including the artist 'Velvet Echo' and platform 'Spotify', remains unchanged and clear. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "What is the title of the least played uh no, I mean most played song by um let's see Velvet Echo on Spotify." }, "candidate_id": "wav_a6e5ab623649f821", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and continuous flow. The false start ('least played') is immediately and smoothly corrected with 'no, I mean, most played', followed by a brief, believable hesitation ('hmm') and a natural recovery ('let's see') before completing the request. The prosody is consistent, and the self-correction feels spontaneous rather than rehearsed." } ] }, "is_default": true }, { "variant_id": "false_start_02_artist_name", "proposed_instruction": "What is the title of the most played song by Velvet Light—sorry, I meant Velvet Echo on Spotify.", "long_pause_offsets": [], "source_sha256": "9817df0569f30f7621c4857cca7cd2ad084b133448c90a5001d8a12d3ff1d285", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'Velvet Light' but immediately corrects to 'Velvet Echo' within the same utterance. The final intended artist is 'Velvet Echo', matching the original instruction. All other elements—action, app, and query type—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/false_start/false_start_02_artist_name/8728240df51d6c75/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "77e7aa17a7b2420e75d26adaa337adedbd1755bd9ad0670dcfc10af1700fae40", "seconds": 9.92, "spoken_utterances": [ "What is the title of the most played song by ah... Velvet Light—um... sorry, I meant Velvet Echo on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mispronounces 'Velvet Echo' as 'Velvet Light' but immediately self-corrects to 'Velvet Echo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "What is the title of the most played song by uh Velvet Light um sorry, I meant Velvet Echo on Spotify." }, "candidate_id": "wav_ffc32fd5e19a2ab1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker shows a clear false start ('Velvet Light') and a self-correction ('sorry, I meant Velvet Echo'). The delivery is natural, with a slight stumble and a genuine-sounding apology. However, the correction is slightly less fluid than in audio_03, and the overall pace is a bit more hesitant, making it feel slightly less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "What is the title of the most played song by Velvet Echo on YouTube Music—scratch that, make it Spotify.", "long_pause_offsets": [], "source_sha256": "616b34d455d02775d8f004ef34a6699fb6b0ef36d00f66299834e417c72497f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions YouTube Music but immediately corrects to Spotify, preserving the artist, action, and the concept of 'most played song'." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/false_start/false_start_03_app/c663a85950084b76/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "ac1f3c9f3dac9627879259da656f5c9be18ec58344f71af0f8e449641387ccc1", "seconds": 9.44, "spoken_utterances": [ "What is the title of the most played song by ah... Velvet Echo on YouTube Music—um... scratch that, make it ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'YouTube Music' but immediately self-corrects with 'scratch that, make it Spotify'. This is a clear false start where the final value replaces the initial incorrect one. The final intent matches the original instruction exactly, so the task should be executed as requested.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on YouTube Music? Um, scratch that, make it Spotify." }, "candidate_id": "wav_48a9d8753d5ddd09", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a revision ('YouTube Music' to 'Spotify') rather than a self-correction within the same utterance. The 'scratch that' is a common phrase, but the subsequent 'make it Spotify' sounds more like a direct command to the system than a natural thought process. The prosody is slightly more robotic, and the revision feels less integrated into the speaker's thought flow compared to the other two." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Velvet Echo on Spotify. Oh, I forgot to mention, it should be the most played song.", "long_pause_offsets": [ 57 ], "source_sha256": "fae52e994e107531e02d3d2469ee67b012d0997501cbf45bf557819f7e457957", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most played' to the initial request about a song by Velvet Echo, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/addition/addition_01/d2f34a2530d3d6a0/speech.wav", "wer": 0.04, "wav_sha256": "a802935440d8f661ad100fd54b86f84db56536bc1385dcf751115842589b0499", "seconds": 13.04, "spoken_utterances": [ "What is the title of the song by ... Velvet Echo on um... Spotify. ", "Oh, I forgot to mention, it should be the ... most played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, asking for 'the title of the song' without specifying 'most played'. After a pause, the speaker adds the constraint 'it should be the most played song'. This is a clear addition that completes the original instruction. The final intent matches the reference instruction exactly, so the task passes.", "heard_transcript": "What is the title of the song by Velvet Echo on um Spotify. Oh, I forgot to mention, it should be the most played song." }, "candidate_id": "wav_20af90e86e049d91", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. After a brief pause and filler ('um'), they add a crucial detail ('most played song') in a way that sounds like a spontaneous memory. The intonation and pacing of the addition feel organic, not forced or rehearsed. The overall prosody is consistent and conversational." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "What is the title of the most played song by on Spotify. I left out a detail: the artist is Velvet Echo.", "long_pause_offsets": [ 57 ], "source_sha256": "075a2ab60a9f42fe929a93907f71b4a2e73340727962291bd2ab332cfee103a1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing artist name 'Velvet Echo' to the initial incomplete request, completing the full query as intended." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/addition/addition_02/55f22106b894a279/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "0879a0e2274bb5a08ce65a3eab07a811d7807d33e6da11493fa2437105163674", "seconds": 12.32, "spoken_utterances": [ "What is the title of the most played song by on um... Spotify. ", "I left out a detail: the artist is ah... Velvet Echo." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, missing the artist name. After a pause and self-correction, the speaker adds the missing detail 'Velvet Echo'. The final intent combines the original request with the added information, matching the reference instruction. The addition type criterion is satisfied.", "heard_transcript": "What is the title of the most played song by an um Spotify. I left out a detail. The artist is uh Velvet Echo." }, "candidate_id": "wav_a695bb7dbf699c52", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly more hesitant and fragmented phrasing ('most played song by an um Spotify'). The addition of the artist name ('Velvet Echo') comes after a longer, more deliberate pause and a self-corrective phrase ('I left out a detail'). While this is a plausible revision, the delivery feels slightly more rehearsed or constructed compared to audio_01, and the initial phrasing is less fluid." } ] }, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "What is the title of the most played song by Velvet Echo. One more thing: do this on Spotify.", "long_pause_offsets": [ 58 ], "source_sha256": "ffd9e2a46e40995a24ced6eb7f9f2c12af3e45f23be9da487a519edb83bc935c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing platform requirement, preserving the artist and the query about the most played song." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the most played song by um... Velvet Echo. ", "One more thing: do this ... on Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the top rated song by Velvet Echo on Spotify. Wait, the song selection should be 'most played', not 'top rated'.", "long_pause_offsets": [ 67 ], "source_sha256": "2759bc425a5abcb51224213c588870d40bb09aed6f8a130410ac95e9e3250b0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the selection criterion from 'top rated' to 'most played', while preserving the artist, app, and query type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/revision/revision_01_song_selection/c4b3ebe0d9f51ae6/speech.wav", "wer": 0.03333333333333333, "wav_sha256": "eecfb31988e10ff39c024dbf1204b64dc9d7746dd4a83c9b005c67ce09adb0bd", "seconds": 16.72, "spoken_utterances": [ "What is the title of the top rated song by ... Velvet Echo on hmmm... Spotify. ", "Wait, the song selection should be 'uh, let me think... most played', not 'top rated'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'top rated' song but immediately corrects themselves to 'most played' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction, so the task should be executed as requested.", "heard_transcript": "What is the title of the top rated song by Velvet Echo on Spotify? Wait, the song selection should be, uh, let me think, most played, not top rated." }, "candidate_id": "wav_889c80957fff1c05", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits the most natural and plausible revision process. The initial request is slightly hesitant ('top-rated'), followed by a believable self-correction ('wait... let me think... most played') with a realistic filler and a pause. The prosody is varied, with a genuine-sounding shift in intonation when correcting the metric, and the overall rhythm mimics a person recalling and adjusting their request in real time." } ] }, "is_default": true }, { "variant_id": "revision_03_app", "proposed_instruction": "What is the title of the most played song by Velvet Echo on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 73 ], "source_sha256": "2d09b7e9e5888cf788dde116340efe76fafc798a6749287b9e6e7ff4014f944d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, while retaining the artist name and the request for the most played song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/revision/revision_03_app/737b4131594715ca/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "f123cd1eec879a33a8fd8a96bac6a47721615b4de082521fd3fc437325905214", "seconds": 14.36, "spoken_utterances": [ "What is the title of the most played song by ... Velvet Echo on uh... Apple Music. ", "Make that hmmm... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately corrects it to 'Spotify' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "What is the title of the most played song by Velvet Echo on uh Apple Music. Make that hmm Spotify, not Apple Music." }, "candidate_id": "wav_15753e779f6277e6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This speaker makes a revision from 'Apple Music' to 'Spotify' after a brief pause and filler ('make that...'). The correction is clear and the intonation shift is natural. However, the initial hesitation ('on uh...') is less pronounced than in audio_03, and the correction is more abrupt, lacking the 'let me think' hesitation that makes the revision feel more spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_artist_name", "proposed_instruction": "What is the title of the most played song by Velvet Noise on Spotify. Sorry, I meant Velvet Echo, not Velvet Noise.", "long_pause_offsets": [ 70 ], "source_sha256": "160507b46069c2f8c18da8b80c39b8c6d337af1211393e4e7919e960e3a79816", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the artist name from Velvet Noise to Velvet Echo, preserving the rest of the query including the platform and the request for the most played song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/revision/revision_02_artist_name/e46306bab2b03660/speech.wav", "wer": 0.0, "wav_sha256": "78b2acda8c8129f01dc7d7f7f57673309d7a6707192fd71f6433595793102593", "seconds": 11.48, "spoken_utterances": [ "What is the title of the most played song by ... Velvet Noise on um... Spotify. ", "Sorry, I meant Velvet Echo, not Velvet Noise." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Velvet Noise' but immediately corrects to 'Velvet Echo' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the most played song by Velvet Echo on Spotify, which matches the original instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "What is the title of the most played song by Velvet Noise on um Spotify. Sorry, I meant Velvet Echo, not Velvet Noise." }, "candidate_id": "wav_355c3174b3931bf3", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker makes a self-correction from 'Velvet Noise' to 'Velvet Echo' but does so with a flat, robotic intonation and no natural hesitation or filler before the correction. The pause is long and unnatural, and the correction is delivered as a flat statement rather than a spontaneous realization. The overall delivery lacks the prosodic variation and believable cognitive process seen in the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_014", "hide_variant_original": true, "source_sha256": "9bd75fbd248f054cd7000ffebbfae70e528f75a1e5e452e89e6290f1c8062a0d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/cancellation/cancel_014/60035d3410603b93/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "934e70afdf0935952b1dbac49cb4839127da7dda7a4bfa44cc7185333aa9dd6a", "seconds": 12.08, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify.", "Hold on, on second thought, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Hold on, on second thought, cancel that,' which is a clear and unambiguous cancellation of the task. The final interjection is a direct command to cancel, satisfying the cancellation type criterion.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify? Hold on, on second thought, cancel that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_014", "hide_variant_original": true, "source_sha256": "431abf1b574623604c221e5533b78aa2a5e74630e5eb361d6b6f84e1e6584679", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/ambiguous_halt/halt_014/013eb87e1d4506f4/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "01499f8cfe23176442268cdbc2f11f0fc199933beee041dff9696b10d62684df", "seconds": 10.96, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify.", "Can you hold on a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Can you hold on a second?', which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause request, not a cancellation or a new instruction. The original task is not revoked, but the speaker is asking for a momentary halt.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify. Can you hold on a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_1", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "ae084364de2983b6668598e8c2327d7b17bb03bed4ba6fc8215aeac6e22f9490", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_1/audio_natural/filler/original/81d8ba65ea175606/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "dd58735fa4ab3963b0d2dfb2bea284d65567891d968e173dd5bb3a7d3cf58a53", "seconds": 7.28, "spoken_utterances": [ "What is the title of the most played song by uh... Velvet Echo on ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent is to retrieve the title of the most played song by Velvet Echo on Spotify, which matches the original instruction. The filler does not alter the task goal.", "heard_transcript": "What is the title of the most played song by uh Velvet Echo on Spotify?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_artist_name", "proposed_instruction": "What is the title of the least played song by Zoey Jones—sorry, I meant Zoey James on Spotify.", "long_pause_offsets": [], "source_sha256": "d7935fda17b89ae88455f21e2ec241e3a51a9c92ecf8cbd9bacbd91d431afbc1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Zoey Jones' to 'Zoey James' within the same utterance. The final artist name matches the original instruction, and all other conditions (least played, song title, Spotify) are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/false_start/false_start_02_artist_name/e7d0dd7a358ea9b2/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "fff4378aa11dfcdef2e33e78002f5328b8955564d9483f29efa87c604b2e294f", "seconds": 10.92, "spoken_utterances": [ "What is the title of the let me think... least played song by Zoey Jones—um... sorry, I meant Zoey James on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Zoey Jones' but immediately self-corrects to 'Zoey James' after a pause and 'sorry'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original incorrect value is discarded. The final goal matches the original instruction exactly.", "heard_transcript": "What is the title of the least played song by Zoey Jones, um, sorry, I meant Zoey James on Spotify." }, "candidate_id": "wav_91c25eb7d3ebdd71", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start pattern. The speaker begins with 'What is the title of the' and then hesitates with 'let me think'. After a brief pause, they incorrectly say 'least played song by Zoe Jones' and immediately self-correct to 'Zoe James' with a clear, spontaneous-sounding 'sorry, I meant'. The correction is integrated smoothly into the sentence, and the prosody reflects genuine hesitation and recollection. The pacing and intonation are consistent with a person formulating a request in real time." } ] }, "is_default": true }, { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the most played—no, I mean least played song by Zoey James on Spotify.", "long_pause_offsets": [], "source_sha256": "dc3347243c625b9185ff8040f79f9bfc252c2937be2f5e5713d2a354e66303e8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately corrects to 'least played' within the same utterance. The final intended action is to identify the least played song, matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/false_start/false_start_01_song_selection/346d4cdc20560478/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "9d38bf7baf000d78c51dd3e022f7adb308913a677789737eceb6f64bb935b0c4", "seconds": 13.72, "spoken_utterances": [ "What is the title of the most played—hmmm... no, I mean ... least played song by um, let me think... Zoey James on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'most played' but immediately self-corrects to 'least played'. According to the 'false_start' type criterion, the final corrected value should be applied, and the initial mistake should be discarded. The final intent matches the original instruction exactly, with only a minor hesitation and filler ('um', 'let me think') that does not alter the core task.", "heard_transcript": "What is the title of the most played um, no, I mean, least played song by um, let me think, Zoey James on Spotify." }, "candidate_id": "wav_dee24a515fc5cf12", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a false start and self-correction, but the flow is less natural. The speaker starts with 'most played', then corrects to 'least played', but the initial error is not a spontaneous slip—it sounds more like a deliberate switch. The 'um' and 'let me think' fillers are placed awkwardly, and the overall rhythm feels slightly rehearsed or unnatural. The correction is less integrated and more abrupt compared to audio_01, making it sound less like a genuine moment of recollection." } ] }, "is_default": false }, { "variant_id": "false_start_03_app", "proposed_instruction": "What is the title of the least played song by Zoey James on Apple Music—make that Spotify.", "long_pause_offsets": [], "source_sha256": "724ea62790a5a824cc171f55b59280cc5b97dcbef2bfcabea2d636e053c09f48", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify. The rest of the query, including the artist name and the intent to find the least played song, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/false_start/false_start_03_app/4718aff0ede1767f/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "78c423aec7932637c6c8fe8c4774a5cfdc736c90f0feabb7b19f361be410be25", "seconds": 10.24, "spoken_utterances": [ "What is the title of the least played song by Zoey James on Apple Music—uh... make that ... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Apple Music' but immediately self-corrects to 'Spotify'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "What is the title of the least played song by Zoey James on Apple Music, uh, make that Spotify." }, "candidate_id": "wav_d9f562a160cdcc21", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio features a revision rather than a false start. The speaker begins with 'Apple Music' and then corrects to 'Spotify'. However, the correction is delivered with a flat, robotic intonation and lacks the natural prosody of a spontaneous change of mind. The 'uh' filler is too short and doesn't convey hesitation. The overall delivery sounds rehearsed or synthesized, with an unnatural rhythm and lack of genuine vocal variation, making it the least realistic of the three." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Zoey James on Spotify. Oh, I forgot to mention, it should be the least played song.", "long_pause_offsets": [ 56 ], "source_sha256": "2087918e6149c92fd745a1c982713e2f332bcbb1e0ed20fed02fcf15147b96c7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'least played' to the initial request about Zoey James's songs, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/addition/addition_01/bafda86b631304c6/speech.wav", "wer": 0.10344827586206896, "wav_sha256": "d0025b70644f18ce063c789678cf0c9b2be1573bda96c0d566f7d5536738b66d", "seconds": 16.44, "spoken_utterances": [ "What is the title of the song by Zoey James on um, let me think... Spotify. ", "Oh, I forgot to mention, it should be the hmmm... least played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete, missing the 'least played' constraint. After a pause and self-correction, the speaker adds the missing constraint. The final intent combines the original request with the added condition, matching the reference instruction. The addition is clear and directly modifies the initial request without introducing new errors.", "heard_transcript": "What is the title of the song by Zoey James on um, let me think, Spotify. Oh, I forgot to mention it should be the least played song." }, "candidate_id": "wav_9b6ab0b0dcbe5886", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "What is the title of the least played song by on Spotify. I left out a detail: the artist is Zoey James.", "long_pause_offsets": [ 58 ], "source_sha256": "6655451df7c9ade52bdceee91f5bffe5e1c7b8d8ae2fa9f38e4730e713c8239d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing artist identity, completing the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the ... least played song by on uh... Spotify. ", "I left out a detail: the artist is um... Zoey James." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false }, { "variant_id": "addition_03", "proposed_instruction": "What is the title of the least played song by Zoey James on. One more thing: this should be on Spotify.", "long_pause_offsets": [ 61 ], "source_sha256": "a66949354f524cf481df31b97230e338b73142b75dad34cde656c0a2e0f1e15e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing platform requirement, completing the initial incomplete instruction. The artist and the query type (least played song) remain unchanged." }, "audio": { "status": "pending", "spoken_utterances": [ "What is the title of the ... least played song by uh... Zoey James on. ", "One more thing: this should be on Spotify." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_app", "proposed_instruction": "What is the title of the least played song by Zoey James on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 73 ], "source_sha256": "dfc22e3a58989c5a1bb60875964a09a92c710f52922e1ccd93d3e72faa2d4a70", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, preserving the artist, the query type (least played song), and the intent to retrieve the title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/revision/revision_03_app/5cfef3b25f43049c/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "a72f534d5d1df12ce4a406e3749f4181ef44fa95068b7358150cca7181892823", "seconds": 13.64, "spoken_utterances": [ "What is the title of the least played song by Zoey James on ... Apple Music. ", "Make that let me think... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the least played song on Apple Music, then immediately corrects the platform to Spotify. This is a clear revision where the final intent is to query Spotify, not Apple Music. The rest of the instruction (artist, metric, action) remains unchanged. The revision is spoken with a natural pause and self-correction, satisfying the revision type criterion.", "heard_transcript": "What is the title of the least played song by Zoey James on Apple Music. Make that, let me think, Spotify, not Apple Music." }, "candidate_id": "wav_c0b83910f83e90d9", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with a natural pause and filler ('let me think') before a smooth, self-corrected revision. The correction ('Spotify not Apple Music') is integrated fluidly into the same utterance, maintaining a consistent rhythm and intonation. The prosody is continuous and the correction feels spontaneous, not forced." } ] }, "is_default": true }, { "variant_id": "revision_02_artist_name", "proposed_instruction": "What is the title of the least played song by Zoey Jones on Spotify. Sorry, I meant Zoey James, not Zoey Jones.", "long_pause_offsets": [ 69 ], "source_sha256": "c77479b80a06c61933630c8dc5a80fcb36df67f40511c70090c32a1b21829bca", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the artist name from Zoey Jones to Zoey James, preserving the rest of the query including the app (Spotify), the action (finding the least played song), and the scope (by the artist)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/revision/revision_02_artist_name/fc2e76535516ac08/speech.wav", "wer": 0.11538461538461539, "wav_sha256": "46cff7a701cf2b935ad88ccb4a4d1033878810eb17763d84091988c5099a6317", "seconds": 15.2, "spoken_utterances": [ "What is the title of the least played song by um, let's see... Zoey Jones on Spotify. ", "Sorry, I meant um... Zoey James, not Zoey Jones." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Zoey Jones' but immediately corrects themselves to 'Zoey James' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the least played song by Zoey James, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were changed.", "heard_transcript": "What is the title of the least played song by um, let's see, Zoey Jones on Spotify? Sorry, I meant um, Zoey James, not Zoey Jones." }, "candidate_id": "wav_0de450d5d2649124", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then corrects the artist's name ('Zoe Jones' to 'Zoe James'). The correction is delivered with a natural pause and filler ('sorry, I meant'), and the intonation reflects a genuine self-correction. However, the initial part of the sentence is slightly more hesitant and less fluid than audio_03, making it slightly less realistic overall." } ] }, "is_default": false }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the most played song by Zoey James on Spotify. Wait, that should be the least played song, not the most played.", "long_pause_offsets": [ 68 ], "source_sha256": "521c209dc3829b8580b1dbd80ac5279ee9ab554ae0d28adbfaa3d6cf766b62bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the comparison operator from 'most played' to 'least played', while retaining the artist, app, and query type." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/revision/revision_01_song_selection/2659d0228a0e439e/speech.wav", "wer": 0.06451612903225806, "wav_sha256": "fd09ed677d17c0a07c5734cc549253ace8d1cd9a20008d00b1dd6c808ab7f830", "seconds": 17.24, "spoken_utterances": [ "What is the title of the um... most played song by um, let's see... Zoey James on Spotify. ", "Wait, that should be the um... least played song, not the most played." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the 'most played' song but immediately corrects themselves to 'least played' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the least played song, which matches the original reference instruction. All other conditions (artist: Zoey James, platform: Spotify) are preserved.", "heard_transcript": "What is the title of the um most played song by um let's see Zoey James on Spotify. Wait, that should be the um least played song, not the most played." }, "candidate_id": "wav_99c116cf36216dd6", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then corrects the metric from 'most played' to 'least played'. The correction is delivered with a noticeable pause and a slightly awkward, hesitant tone ('wait, that should be...'). The prosody is less continuous, and the correction feels more like a disjointed afterthought than a natural part of the thought process, reducing its realism compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "ad51cfb912059a15e9f595f4513dbac46996e3c3b03b7304e92072e7f1789f13", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/cancellation/cancel_039/06e6649211f4d747/speech.wav", "wer": 0.07142857142857142, "wav_sha256": "605386ac1d0b83252d78699c5083c64a7a61db27a90bf2bf15c17850a97c4136", "seconds": 14.24, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify.", "Hold on, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the requested task but then clearly and unambiguously cancels it with 'Hold on, don't go ahead with that. Cancel it.' This constitutes a final cancellation of the entire task as per the type-specific criterion.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify. Hold on, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_005", "hide_variant_original": true, "source_sha256": "5d1de2c34df55c60637f8543c99851fb2328fdf0d872b8d4fc7cddf6838b9d25", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/ambiguous_halt/halt_005/1009dfdd2f22431e/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "1a5a46837c1e960b87651ad2087f1026ae92e474ac5ea9fcd298af936d041bce", "seconds": 13.2, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify.", "One second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request with 'What is the title of the least played song by Zoey James on Spotify' and then immediately says 'One second,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is fully audible and unaltered, and the final control intent is to pause, not execute or cancel.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify. One second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_2", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b782130edda27a3e3537b1551d0912bacf296366717e1c61b607f5860c3d3877", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_2/audio_natural/filler/original/2e0d4004558f299c/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a040c1635bb69c0249c857f8b0406fc16225b5e494e8dea957a9a51397f6670a", "seconds": 9.56, "spoken_utterances": [ "What is the title of the hmmm... least played song by um, let me think... Zoey James on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes hesitation ('um', 'let me think') but does not change the core request. The final audible intent matches the original instruction. The filler words do not alter the goal, so the task should be executed.", "heard_transcript": "What is the title of the least played song by um, let me think, Zoey James on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_song_selection", "proposed_instruction": "What is the title of the least played—no, I meant most played song by Jasper Skye on Spotify.", "long_pause_offsets": [], "source_sha256": "40e5b72a29ff88ff3a02eefd1405b3bde8c6134d9859e05017e0b5b7d9c984b9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'least played' to 'most played' within the same utterance, preserving the artist, app, and the intent to retrieve the song title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/false_start/false_start_01_song_selection/b74bda724bc41df6/speech.wav", "wer": 0.05, "wav_sha256": "d9c4d714e73e494e0f3ed2e732a5dfc9ac04a550ddc6ad2337fe16eb64faa4c8", "seconds": 10.88, "spoken_utterances": [ "What is the title of the least played—um... no, I meant most played song by ... Jasper Skye on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'least played' but immediately self-corrects to 'most played' with 'no, I meant'. This is a clear false start where the final intended value is 'most played'. The rest of the instruction, including the artist 'Jasper Skye' and the platform 'Spotify', remains unchanged and correct. The final intent matches the original instruction, so the task should be executed.", "heard_transcript": "What is the title of the least played um, no, I meant most played song by Jasper Skye on Spotify." }, "candidate_id": "wav_a4a0ba001d4a70d1", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a very natural-sounding false start. The initial request 'least played' is clearly spoken, followed by a believable self-correction to 'most played' with a smooth intonation shift and a pause that sounds like genuine reconsideration. The phrasing 'no, I meant' is a common, spontaneous way to correct oneself, and the overall pace and prosody are consistent with someone thinking aloud." } ] }, "is_default": true }, { "variant_id": "false_start_03_app", "proposed_instruction": "What is the title of the most played song by Jasper Skye on Apple Music—scratch that, make it Spotify.", "long_pause_offsets": [], "source_sha256": "30074a36870643102a76b4eb2b06bd6172b373c154ff40f0fc3e9a89878c4dd1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially mentions Apple Music but immediately corrects to Spotify, which is the app specified in the original instruction. The rest of the query, including the artist and the request for the most played song, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/false_start/false_start_03_app/0bf526c9a7053e25/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "d55e4d5bce1bf86d8d4dac34a2e719ca03cdc1f00ea17a2d360a062972547f7b", "seconds": 10.16, "spoken_utterances": [ "What is the title of the most played song by Jasper Skye on Apple Music—um... scratch that, make it uh... Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested the task for Apple Music but immediately self-corrected to Spotify. According to the 'false_start' type criterion, the final intent after the immediate correction should be executed. The correction is clear and unambiguous, changing the platform from Apple Music to Spotify while preserving the rest of the request. Therefore, the final intent matches the original instruction.", "heard_transcript": "What is the title of the most played song by Jasper Skye on Apple Music? Um, scratch that, make it uh Spotify." }, "candidate_id": "wav_33bc77e6fa13f6f1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a believable revision. The speaker starts with a correct request for Apple Music, then hesitates with 'um' before deciding to change it to Spotify. The 'scratch that' is a natural, informal way to cancel and replace a part of a sentence. The correction is integrated smoothly, and the prosody reflects a moment of decision-making rather than a robotic edit." } ] }, "is_default": false }, { "variant_id": "false_start_02_artist_name", "proposed_instruction": "What is the title of the most played song by Jasper Light—sorry, I should have said Jasper Skye on Spotify.", "long_pause_offsets": [], "source_sha256": "6f89ded17abc71566e4022fcda22513f974532aa11c80821c01e2bf8f2818287", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Jasper Light' to 'Jasper Skye' within the same utterance, preserving the action, app, and all other conditions." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/false_start/false_start_02_artist_name/0eb1158843f2d8ff/speech.wav", "wer": 0.19230769230769232, "wav_sha256": "7d0e6b90dfe0739f40c19f77a1943fdf7f9aee3507ca5104f9a0a054b2714ccd", "seconds": 12.48, "spoken_utterances": [ "What is the title of the uh... most played song by uh, let's see... Jasper Light—uh... sorry, I should have said Jasper Skye on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially misstates the artist as 'Jasper Lida' but immediately self-corrects to 'Jasper Skye' with an apology. According to the 'false_start' type criterion, the final corrected value must be applied, and the original mistaken value is discarded. The final intent matches the original instruction exactly, with only a minor grammatical paraphrase ('most played' vs 'most played song') that does not alter the meaning.", "heard_transcript": "What is the title of the uh most played song by uh let's see, Jasper Lida. Sorry, I should have said Jasper Skye on Spotify." }, "candidate_id": "wav_a67f02b63cd799ef", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "While this clip contains a false start ('uh' before 'most played'), the correction is less natural. The speaker says 'Jasper Lida' and then immediately corrects to 'Jasper Sky' without a pause or hesitation, making the correction feel abrupt. The initial 'uh' and 'let's see' are plausible fillers, but the lack of a natural pause before the correction reduces the realism of the self-correction flow." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "What is the title of the song by Jasper Skye on Spotify. Oh, I forgot to mention, it should be the most played song.", "long_pause_offsets": [ 57 ], "source_sha256": "466e958a76fbf028d599571f9e2bc51ae261c87df10d3fc247c2b4fdbb0d065f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'most played' to the initial request about a song by Jasper Skye, refining the scope to match the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/addition/addition_01/d9bcd3104cd1876f/speech.wav", "wer": 0.08, "wav_sha256": "420c73d63ddb1f7b268453f80f19a1193eea7f371fceba159d74486075051f1c", "seconds": 11.4, "spoken_utterances": [ "What is the title of the song by um... Jasper Skye on ... Spotify. ", "Oh, I forgot to mention, it should be the most played song." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was to find the title of a song by Jasper Skye on Spotify. The speaker then added the constraint 'it should be the most played song' after a pause, which is a clear addition to the original instruction. The final intent combines the original request with the added constraint, matching the reference instruction. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "What is the title of the song by um Jasper Skye on Spotify? Oh, I forgot to mention, it should be the most played song." }, "candidate_id": "wav_d563c96f73c98756", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear initial request with natural prosody and a believable, additive follow-up. The phrase 'Oh, I forgot to mention' is a pragmatic and common way to add a detail, and the intonation and pacing of the entire utterance sound spontaneous and conversational. The addition is smoothly integrated without awkwardness or hesitation." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "What is the title of the most played song by on Spotify. One more detail: the artist is Jasper Skye.", "long_pause_offsets": [ 57 ], "source_sha256": "a27ed882ff322ee68c26cbe78e0d867e10072d0f51045735668d0f182f645f71", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup supplies the missing artist identity, completing the initial incomplete request. The final goal matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/addition/addition_02/db0c63078da66db2/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "0a7e2e1aecc74c8a2d7e92a9c5f891dc202d515a89c7b97b6ec3742698621932", "seconds": 14.76, "spoken_utterances": [ "What is the title of the um... most played song by on hmmm... Spotify. ", "One more detail: the artist is ... Jasper Skye." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('by on Spotify'), but the speaker adds the missing detail 'the artist is Jasper Skye' after a pause. This addition completes the original instruction, specifying the artist. The final intent is clear and matches the reference instruction. The addition type criterion is satisfied.", "heard_transcript": "What is the title of the um most played song by on Spotify. One more detail, the artist is Jasper Skye." }, "candidate_id": "wav_64ef67cd62f63810", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker's delivery is less fluid, with multiple filled pauses ('um', 'uh') and a halting rhythm that suggests more effortful formulation. The initial request is fragmented and unclear, and the addition of the artist's name comes after a long pause, sounding less like a spontaneous thought and more like a rehearsed or corrected statement. The overall prosody is less natural and more hesitant than audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_artist_name", "proposed_instruction": "What is the title of the most played song by Luna Moon on Spotify. Sorry, I meant Jasper Skye, not Luna Moon.", "long_pause_offsets": [ 67 ], "source_sha256": "2b63c526cc16822886e6c42bcbc8fbf00983b6f37fe40c2bf273cabe5b7bbdc9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the artist name from Luna Moon to Jasper Skye, preserving the rest of the query including the action, app, and the 'most played song' condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/revision/revision_02_artist_name/7748499ac1d69228/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "9a01da0739d076f138fc7aa915fbf6fa695205d7435fc91817b2f6f413a42eca", "seconds": 15.0, "spoken_utterances": [ "What is the title of the most played song by ... Luna Moon on uh... Spotify. ", "Sorry, I meant ... Jasper Skye, not Luna Moon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Luna Moon' but immediately corrects to 'Jasper Skye' after a pause and an apology. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to find the most played song by Jasper Skye on Spotify, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "What is the title of the most played song by Luna Moon on uh Spotify. Sorry, I meant Jasper Skye, not Luna Moon." }, "candidate_id": "wav_735a5af5f1cef5eb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then self-corrects from 'Luna Moon' to 'Jasper Skye' with a believable hesitation and intonation shift. The correction is smoothly integrated into the same utterance, and the overall prosody is consistent and conversational, suggesting a spontaneous thought process." } ] }, "is_default": true }, { "variant_id": "revision_03_app", "proposed_instruction": "What is the title of the most played song by Jasper Skye on Apple Music. Make that Spotify, not Apple Music.", "long_pause_offsets": [ 73 ], "source_sha256": "ec93cfeffe336de96a5c838be875ae38146c0214bfd2bf675e0003cb39ac3276", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the music platform from Apple Music to Spotify, preserving the artist, the action, and the intent to find the most played song." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/revision/revision_03_app/a6ff38f2565868b7/speech.wav", "wer": 0.13636363636363635, "wav_sha256": "2dbf119b527ac0f3a8021e690ed4a285e2024447a9a100fbc936b4221d9950cf", "seconds": 14.12, "spoken_utterances": [ "What is the title of the ... most played song by Jasper Skye on ... Apple Music. ", "Make that uh... Spotify, not Apple Music." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for the most played song by Jasper Skye on Apple Music, then immediately corrects the platform to Spotify. This is a revision where the final intent is to query Spotify, which matches the original instruction. The correction is clear and applied to the relevant part of the task.", "heard_transcript": "What is the title of the most played song by Jasper Skye on Apple Music? Make that uh Spotify not Apple Music." }, "candidate_id": "wav_c60b73c533229db1", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then corrects 'Apple Music' to 'Spotify' with a natural-sounding 'uh' and a slight pause. The correction is integrated into the flow, though the initial phrasing is slightly more hesitant than audio_02. The delivery is realistic but less fluid in its self-correction." } ] }, "is_default": false }, { "variant_id": "revision_01_song_selection", "proposed_instruction": "What is the title of the top tracks song by Jasper Skye on Spotify. Wait, the selection should be 'most played', not 'top tracks'.", "long_pause_offsets": [ 68 ], "source_sha256": "c01e2fc7156fc5c0fb3817549e0bf12008416548c32441b6a31f700c279fcd29", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the selection criterion from 'top tracks' to 'most played', preserving the artist, app, and the intent to retrieve a song title." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/revision/revision_01_song_selection/3c7ac7f6d3b0c8f4/speech.wav", "wer": 0.2692307692307692, "wav_sha256": "2bd818e29784847445f90ea5979d94e0b57fac1c939d38a924572c1fd92de4f1", "seconds": 16.08, "spoken_utterances": [ "What is the title of the um... top tracks song by hmmm... Jasper Skye on ... Spotify. ", "Wait, the selection should be 'most played', not 'top tracks'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'top tracks' but immediately corrects themselves to 'most played' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent is to find the most played song, which matches the original reference instruction. No other constraints were changed or lost.", "heard_transcript": "What is the title of the top tracks song by Jasper Skye on Spotify? Wait, the selection should be most played, not top tracks." }, "candidate_id": "wav_1bd8074323e84483", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker uses multiple fillers ('um', 'hmm') and hesitates significantly before stating the request. The correction from 'top tracks' to 'most played' is made after a longer pause, and the phrasing sounds more like a recollection than a spontaneous correction. The overall delivery is less fluid and more fragmented compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_036", "hide_variant_original": true, "source_sha256": "7aa6f2c27450d68a3e1c5c892c1ebb53075aa8e60982022a6ae75d7e51b5ad5a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/cancellation/cancel_036/6115fefc3eb3cbc1/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "37d5ae8607ac1ae110c4b12d1ea604dc73c8971a21481cf738be8ca408c79cb1", "seconds": 13.16, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify.", "Oh, wait, don't continue with that request. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio clearly contains a final directive to cancel the previously stated task. The speaker first formulates the request, then immediately interrupts with 'Oh wait, don't continue with that request. Cancel it.', which constitutes a clear cancellation of the entire task. The final control intent is a cancellation, matching the expected behavior for this type.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify. Oh wait, don't continue with that request. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_021", "hide_variant_original": true, "source_sha256": "a34dfe1f284ec07c3f3ad43339168b541c33e06aba6f7d2d9229a163eeeae40b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/ambiguous_halt/halt_021/efc92fa2a3eace61/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "f80c092a080565005c479fb903265903e23ef61a77ec84f204082e493317adb9", "seconds": 11.76, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify.", "Wait, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Wait, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled or revised, and the speaker does not provide a corrected or final instruction. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify. Wait, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "e85d92a_3", "split": "train", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1fb5e8b3f63feb1cd0cc37696c8dc589dea08c4b58b8e34e8e2aec21a05b9d35", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/e85d92a_3/audio_natural/filler/original/7f5f5a842e0f72d9/speech.wav", "wer": 0.2222222222222222, "wav_sha256": "397c076cec6365e36701a6830a21170bbcfc89e53eb17398f3f533bd9724e748", "seconds": 7.8, "spoken_utterances": [ "What is the title of the ah... most played song by uh, let's see... Jasper Skye on Spotify." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'let's see') and pauses but does not change the core request. The final intent is clear and matches the original instruction. The filler type allows for such hesitations without altering the task goal.", "heard_transcript": "What is the title of the uh most played song by uh let's see Jasper Skye on Spotify." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in this calendar year?", "long_pause_offsets": [], "source_sha256": "7b8af3f643590b6bef72d83eb0fcfe79eb8f3bbc813f91e667c80749bdd12be2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the app and time frame." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/false_start/false_start_01_app/f45de312f5a4855a/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "91a13f5083d6fcb218d1cd4d3aedd301ea7dd74a049c426feaa1e57052c3bf8c", "seconds": 7.84, "spoken_utterances": [ "How much did I spend on ebay—um... sorry, I meant amazon in uh... this calendar year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. This is a clear false start where the final intended target is Amazon. The time frame 'this calendar year' remains unchanged. The correction is immediate and unambiguous, satisfying the false_start type criterion.", "heard_transcript": "How much did I spend on eBay um sorry, I meant Amazon in uh this calendar year?" }, "candidate_id": "wav_c2a43e89df78a8ef", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('eBay') followed by a smooth self-correction ('I meant Amazon') with a believable hesitation ('uh') and a pause. The intonation and rhythm are consistent with spontaneous speech, and the correction flows naturally within the same utterance without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "How much did I spend on amazon in last calendar year—no, I mean this calendar year?", "long_pause_offsets": [], "source_sha256": "d10bb090701313f9e69c12e0bc9cf60dc3d32a3fbf8e0552765fd8f178876296", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar year' but immediately corrects to 'this calendar year'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/false_start/false_start_02_time_reference/a24fea85dc903884/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "2c7b976f977ccc7fa2569766aab6451fa0e9075210027445c380c94f4542f6b9", "seconds": 10.92, "spoken_utterances": [ "How much did I spend on uh... amazon in last calendar year—um... no, I mean ... this calendar year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'last calendar year' but immediately self-corrects to 'this calendar year'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original mistaken value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How much did I spend on uh Amazon in last calendar year? Um, no, I mean this calendar year." }, "candidate_id": "wav_4fd49f67fe7375b2", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker also makes a false start ('Amazon') and corrects it ('this calendar year'), but the correction is less fluid. The initial 'uh' and 'um' fillers are present, but the correction feels slightly more abrupt and less integrated into the sentence structure. The prosody is slightly less natural, with a flatter intonation and a less convincing transition between the mistaken and corrected phrases." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How much did I spend on amazon? I left out a detail: it's for this calendar year.", "long_pause_offsets": [ 32 ], "source_sha256": "8493fc56f91ff7c2f74dacd3734366fea6a626bbab77c32e917b03d5145742f8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'this calendar year' to the initial question about Amazon spending, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/addition/addition_02/bf9c67e0f8472785/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "837bf8fbe1df9d1b1089d8342e211ca23ecddda44bf1ba9e627da676024c1640", "seconds": 11.16, "spoken_utterances": [ "How much did I spend on um... amazon? ", "I left out a detail: it's for uh... this calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is incomplete ('How much did I spend on um Amazon?'), but the speaker immediately adds the missing constraint ('it's for uh this calendar year'). The final intent combines the original request with the added detail, matching the reference instruction exactly. The addition type criterion is satisfied.", "heard_transcript": "How much did I spend on um Amazon? I left out a detail, it's for uh this calendar year." }, "candidate_id": "wav_b5455d953936dee7", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant start ('How much did I spend on um...'), followed by a plausible self-correction ('I left out a detail') and a smooth, additive continuation ('It's for this calendar year'). The intonation and pacing reflect genuine thought process, and the addition feels like a remembered clarification rather than a forced insertion." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How much did I spend in this calendar year? Oh, I forgot to mention, it's specifically on Amazon.", "long_pause_offsets": [ 44 ], "source_sha256": "200c02468a7d34b3bf9d595715c3c0eb329557ddd2b3e74caca8e2cef729685e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing restriction to the initial broad question, specifying that the spending inquiry is limited to Amazon. The time frame 'this calendar year' remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/addition/addition_01/623c60f0af4e4ffd/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "db01d0aa1f0223aa89d52efdf4c1a160d2d11f3d38d891611f67ecf1af5af405", "seconds": 11.64, "spoken_utterances": [ "How much did I spend in hmmm... this calendar year? ", "Oh, I forgot to mention, it's specifically ... on Amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How much did I spend in this calendar year?' which is incomplete without the spending context. The speaker then adds 'it's specifically on Amazon' after a pause, which completes the original instruction by specifying the target. This matches the addition type criterion: the final goal combines the initial instruction with the added information, resulting in the complete and correct task as per the reference.", "heard_transcript": "How much did I spend in this calendar year? Oh, I forgot to mention, it's specifically on Amazon." }, "candidate_id": "wav_c2375f164014cd19", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar hesitant tone ('How much did I spend in...') and then adds a detail ('Oh, I forgot to mention it's specifically on Amazon'). While the addition is natural, the initial phrasing is slightly less fluid than audio_01, and the correction ('Oh, I forgot') is more abrupt, making the overall flow feel a bit less spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_time_reference", "proposed_instruction": "How much did I spend on amazon in last calendar year? Make that this calendar year, not last calendar year.", "long_pause_offsets": [ 54 ], "source_sha256": "449a5aa7a5f8154429d4d6166240a092ca926515e5e88a18b8c75e3deeb09b8b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'last calendar year' to 'this calendar year', aligning with the original instruction. The app (amazon) and action (query spending) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/revision/revision_02_time_reference/10043b9e91a82e51/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "e34f1411caad16cef333f261c305fbe5db4acbaeb9243430b760b65409354d4a", "seconds": 13.44, "spoken_utterances": [ "How much did I spend on ah... amazon in um... last calendar year? ", "Make that ah... this calendar year, not last calendar year." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'last calendar year' but immediately corrects it to 'this calendar year' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to inquire about spending on Amazon in 'this calendar year', which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "How much did I spend on uh Amazon in um last calendar year? Make that uh this calendar year, not last calendar year." }, "candidate_id": "wav_a8fabce4a4dec45d", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, slightly hesitant question ('How much did I spend on...') with typical fillers ('uh', 'um') and a smooth, self-corrected revision ('Make that uh this calendar year not last calendar year'). The correction is integrated into the same utterance with a believable shift in intonation and pacing, showing a spontaneous change of mind without awkwardness or repetition. The prosody is consistent and the flow is continuous, making it sound like a real-time thought process." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "How much did I spend on ebay in this calendar year? Wait, I meant Amazon, not eBay.", "long_pause_offsets": [ 52 ], "source_sha256": "601d5df2cf8faa94a7f91d3c1a95ca2dfc3ecead96ed9a13a39e8913a874abf0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the platform from eBay to Amazon, preserving the action, time frame, and intent to query spending." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/revision/revision_01_app/da9380e6f4c2bd59/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "12b57e18175259ad4e78dc10af5bb4732416ebb36241f3334b5e3f77553e7889", "seconds": 11.2, "spoken_utterances": [ "How much did I spend on uh... ebay in this calendar year? ", "Wait, I meant ... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'eBay' but immediately corrects to 'Amazon' after a pause and the phrase 'Wait, I meant'. This constitutes a revision where the final intent is to query Amazon, matching the original instruction. The correction is clear and the final goal is fully specified.", "heard_transcript": "How much did I spend on uh eBay in this calendar year? Wait, I meant Amazon, not eBay." }, "candidate_id": "wav_67c9065824f113a6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also features a natural-sounding request with fillers ('uh') and a clear revision ('Wait, I meant Amazon not eBay'). However, the correction is delivered as a separate, slightly abrupt utterance after a short pause, which makes it sound more like a post-hoc correction than a fluid self-repair. While still realistic, the break between the initial request and the correction is more noticeable than in audio_01, reducing the sense of continuous, spontaneous speech." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_019", "hide_variant_original": true, "source_sha256": "a44c560059e32350370e750e51179abe112d97928161869ad6cce43a37fa5cd4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/cancellation/cancel_019/4b700486d345f2e3/speech.wav", "wer": 0.05555555555555555, "wav_sha256": "bc05fd8412851297a149f40d6f14375bf1ba2f3c52a3035535e9c4abab34c4c7", "seconds": 11.48, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?", "Wait, never mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original request but immediately cancels it with 'Wait, never mind, don't do that.' This constitutes a clear cancellation of the task. The final audible intent is to cancel, not to execute the original query.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year? Wait, never mind, don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_036", "hide_variant_original": true, "source_sha256": "45046fbfcc8b860f05d886b6dcf71de9af1f299faa6411011387039eac77f5e5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/ambiguous_halt/halt_036/6e0d49a08743f4f5/speech.wav", "wer": 0.05263157894736842, "wav_sha256": "bea447a01b2dc224c3950fd0a992734c304920133755f0b8e8ae5c4132f69491", "seconds": 11.12, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?", "Wait a second before you go ahead." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the explicit request 'Wait a second before you go ahead,' which is a clear instruction to pause. The initial query is not canceled, and the final interjection is a directive to hold, not to execute or cancel. This matches the 'ambiguous_halt' type criterion, where the final audible intent is to pause.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year? Wait a second before you go ahead." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "b48011f66c2d9e28e92b6b81166ebaf96107effb881d4703bdfb4bcad4d967bd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_1/audio_natural/filler/original/a48de506a2e1b1ad/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "c15b5ed629cf30b761d54fa396e8af419d0e314cc393a44673e9b7b0a357e8fa", "seconds": 6.84, "spoken_utterances": [ "How much did I spend on uh... amazon in ... this calendar year?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'uh' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the task goal. Therefore, the task should be executed.", "heard_transcript": "How much did I spend on uh Amazon in this calendar year?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in the last calendar month?", "long_pause_offsets": [], "source_sha256": "b043561fce559bd065f666cd563baf679857a1074d2a96566202341e26d7e780", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the time frame and intent to retrieve spending information." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/false_start/false_start_01_app/88ae93a4f89bda10/speech.wav", "wer": 0.0, "wav_sha256": "2d9b09354655b442b4c0ae4ebd5de06cc427726c22721a03bac01cb78833d0ab", "seconds": 7.52, "spoken_utterances": [ "How much did I spend on ebay—um... sorry, I meant ... amazon in the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How much did I spend on eBay, um, sorry, I meant Amazon in the last calendar month?" }, "candidate_id": "wav_48a2473ff4cd0a9e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear false start ('eBay') and immediately self-corrects to 'Amazon' with a natural-sounding 'sorry, I meant'. The prosody is smooth, the correction is integrated into the flow without awkwardness, and the intonation reflects genuine recall. The phrase 'in the last calendar month' is delivered with appropriate emphasis and clarity, suggesting a spontaneous, realistic formulation." } ] }, "is_default": true }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "How much did I spend on amazon in the last three months—no, that should be the last calendar month?", "long_pause_offsets": [], "source_sha256": "4de4aa8ede04ee4848314ad7421fa30e04d6d7a7b8c97a171523d867ef46bdb2", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'last three months' but immediately corrects it to 'last calendar month'. The final intended time reference matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/false_start/false_start_02_time_reference/fa78557e7f1c43d2/speech.wav", "wer": 0.0, "wav_sha256": "3e64ea55179f653f046c877e060e844fda824d93e61707e2c8459be15425a3e3", "seconds": 10.92, "spoken_utterances": [ "How much did I spend on let's see... amazon in the last three months—um... no, that should be uh... the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states 'last three months' but immediately self-corrects to 'last calendar month'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How much did I spend on let's see, Amazon in the last three months? Um, no, that should be uh the last calendar month." }, "candidate_id": "wav_3301bc958d09dbff", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip contains a revision rather than a false start. The speaker initially says 'last three months' but then corrects to 'last calendar month'. However, the correction is less fluid—the pause before 'no' feels slightly longer and more hesitant than in audio_01, and the phrase 'that should be' introduces a more deliberative tone. While still realistic, the delivery is less seamless and slightly more self-conscious compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_02", "proposed_instruction": "How much did I spend on amazon? I left out a detail: this should apply only to the last calendar month.", "long_pause_offsets": [ 32 ], "source_sha256": "bd80c61503f78c8508259ea9d789d1f8abe1bd9eb3a8422315abf9085c4e5f37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint, refining the initial broad question to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/addition/addition_02/e34e55636bcd2778/speech.wav", "wer": 0.0, "wav_sha256": "98c2a0597b167f010cf206a14d4670249800b1ce7df5fc989e6b3361180f7272", "seconds": 13.84, "spoken_utterances": [ "How much did I spend on let me think... amazon? ", "I left out a detail: this should apply only to um, let's see... the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about spending on Amazon, then adds the constraint 'the last calendar month' after a pause and self-correction. This matches the 'addition' type criterion, where the final goal combines the original request with the added information. The final intent is clear and matches the reference instruction.", "heard_transcript": "How much did I spend on, let me think, Amazon? I left out a detail. This should apply only to, um, let's see, the last calendar month." }, "candidate_id": "wav_ff912c90ac9071bd", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('let me think'), formulates the initial request, then adds a clarifying detail ('this should apply only to...') with a plausible follow-up pause ('let's see') before stating the time frame. The prosody is continuous and the flow reflects a genuine attempt to recall and refine a request, making it sound like a spontaneous, real-time thought process." } ] }, "is_default": true }, { "variant_id": "addition_01", "proposed_instruction": "How much did I spend in the last calendar month? Oh, I forgot to mention, it's specifically on Amazon.", "long_pause_offsets": [ 49 ], "source_sha256": "6a016b1dc3557204499baa743d64a77a538002a84f51360f39fe6cfe6dcce71d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context 'on Amazon' to the initial question, refining the scope without changing the time frame or action. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/addition/addition_01/9b93293bac5b8054/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "9700bdb57fef851b6e10ecf12f6639fff95bab21a3e9d99c4f653ec379935253", "seconds": 13.2, "spoken_utterances": [ "How much did I spend uh, let me think... in the last calendar month? ", "Oh, I forgot to mention, it's specifically ... on Amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about spending in the last calendar month, then adds the constraint 'specifically on Amazon' after a pause. This matches the 'addition' type, where the final goal combines the original instruction with the added detail. The final audible intent is complete and matches the reference instruction.", "heard_transcript": "How much did I spend uh let me think in the last calendar month? Oh, I forgot to mention it's specifically on Amazon." }, "candidate_id": "wav_a98e0feeb0aeea75", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker also uses natural fillers ('uh', 'let me think') and adds a restriction ('specifically on Amazon') after an initial request. However, the addition feels slightly more abrupt and less integrated into the original sentence than in audio_01. The phrase 'Oh, I forgot to mention' introduces a minor meta-commentary that slightly breaks the flow, making the overall delivery feel a bit more rehearsed or less seamlessly spontaneous." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_app", "proposed_instruction": "How much did I spend on ebay in the last calendar month? Wait, I misspoke about the app: use Amazon, not eBay.", "long_pause_offsets": [ 57 ], "source_sha256": "a42469761ef535af69c4cdf43ca62e38553bae24c4b183dce72f773df5786c6d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from eBay to Amazon, while retaining the time frame and the action of querying spending. The original instruction's app and the final spoken instruction's app are the only differing elements, and the correction resolves the mismatch." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/revision/revision_01_app/ea8f72b958eb77a8/speech.wav", "wer": 0.043478260869565216, "wav_sha256": "81a1674bb1b94201d6463062b040873a21a277289b2693b457c30a4c116bb3ef", "seconds": 14.2, "spoken_utterances": [ "How much did I spend on ... ebay ... in the last calendar month? ", "Wait, I misspoke about the app: use uh... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially requested information for eBay but immediately corrected themselves to Amazon after a pause and self-correction. The revision clearly replaces the original app name with the correct one, and the rest of the instruction remains unchanged. The final intent matches the original reference instruction, so the task should be executed as revised.", "heard_transcript": "How much did I spend on eBay in the last calendar month? Wait, I misspoke about the app. Use uh Amazon, not eBay." }, "candidate_id": "wav_8b24888a8fd3727d", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true }, { "variant_id": "revision_02_time_reference", "proposed_instruction": "How much did I spend on amazon in the last three weeks? Change the time reference to 'the last calendar month' instead of 'the last three weeks'.", "long_pause_offsets": [ 56 ], "source_sha256": "6f04f73eb0b6f76bc88ed64ec728b2a6cb2a496080d0a2c261414aec1dc59810", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time reference from 'the last three weeks' to 'the last calendar month', preserving the app (amazon) and action (query spending)." }, "audio": { "status": "pending", "spoken_utterances": [ "How much did I spend on amazon ... in the last three weeks? ", "Change the time reference to 'uh... the last calendar month' instead of 'the last three weeks'." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_030", "hide_variant_original": true, "source_sha256": "7080236019fe5cadc45e9c9d24429f0e992ad898494e9f1c81422379500444ef", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/cancellation/cancel_030/827a476ee9ac3835/speech.wav", "wer": 0.04, "wav_sha256": "d93282c3978e321e5425a2b6a2506cb3d01501fd3bac55361b9d20cdfdb41d7d", "seconds": 12.4, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?", "Oh, wait, you can drop that task. I no longer need it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately follows with 'Oh wait, you can drop that task. I no longer need it.' This constitutes a clear and final cancellation of the initial request. The cancellation is not ambiguous and is not followed by any resumption or modification of the original task. Therefore, the final control intent is cancellation, which matches the expected behavior for the 'cancellation' type.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month? Oh wait, you can drop that task. I no longer need it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_004", "hide_variant_original": true, "source_sha256": "d07a35530b12d78d9e07678cd1126cfbfd2a9f1a47e6a889324bb6cd969fc3af", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/ambiguous_halt/halt_004/e94252f66c75f83e/speech.wav", "wer": 0.0625, "wav_sha256": "c839c8b171ad6194b5c9ca4fe39feff242786e0d69030934cadd800c4d27423b", "seconds": 10.52, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?", "Just a second..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker completes the original request and then says 'Just a second,' which is a clear audible request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original task is not canceled or revised, and the speaker does not provide a new instruction. The final intent is to halt execution temporarily.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month? Just a second." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "a58076a6d629c67866efe7d02e22e800321f9137a1d6f04956f4b37df708f258", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_2/audio_natural/filler/original/2415a5576c36d8d7/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "442fea92f36b40b0b23b37bbd2d53cc08f7bb059122962d1d77b4b9a2bf800dc", "seconds": 7.2, "spoken_utterances": [ "How much did I spend on um... amazon ... in the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent matches the original instruction, and the filler does not alter the goal. Therefore, the task should be executed.", "heard_transcript": "How much did I spend on um Amazon in the last calendar month?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_app", "proposed_instruction": "How much did I spend on ebay—sorry, I meant amazon in this or the last calendar month?", "long_pause_offsets": [], "source_sha256": "58da42790d0dea596b793d0061a8eb04d361c69f64135e13258b7d0d2df81504", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'ebay' to 'amazon' within the same utterance, preserving the rest of the query including the time frame and intent to retrieve spending information." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/false_start/false_start_01_app/4f1091d827561915/speech.wav", "wer": 0.1, "wav_sha256": "c445ca605c336339f173529e617e76b6daabf52263e7dc0f06761934252ca8a7", "seconds": 10.0, "spoken_utterances": [ "How much did I spend on ebay—uh... sorry, I meant amazon in um... this or the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'eBay' but immediately self-corrects to 'Amazon' with 'sorry, I meant Amazon'. This is a clear false start where the final intent is the corrected version. The time frame 'this or the last calendar month' is preserved. The correction is immediate and unambiguous, satisfying the false_start criterion.", "heard_transcript": "How much did I spend on eBay up, sorry, I meant Amazon in um, this or the last calendar month?" }, "candidate_id": "wav_bc8b72f0680db47b", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding false start ('How much did I spend on eBay...'), immediately self-corrects with a believable 'sorry, I meant Amazon' that matches the intended target, and continues with a plausible, slightly hesitant phrasing ('in um...') before the clip ends. The flow from error to correction to continuation is smooth and mimics genuine spontaneous speech." } ] }, "is_default": true }, { "variant_id": "false_start_02_time_reference", "proposed_instruction": "How much did I spend on amazon in this calendar month—no, that should be this or the last calendar month?", "long_pause_offsets": [], "source_sha256": "41958a9f1cee46e816c2285534dd6917e5c8c4a39e686783b858d8e606874fc0", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'this calendar month' but immediately corrects it to 'this or the last calendar month'. The final intended time range is broader than the original instruction, but the correction is explicit and complete, replacing the initial value with the intended one." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/false_start/false_start_02_time_reference/ae419d14f723a5e3/speech.wav", "wer": 0.08695652173913043, "wav_sha256": "462c2025bc74d157588f1abd016f43623c99f01a76fa546b3e54db06cf0eb075", "seconds": 9.36, "spoken_utterances": [ "How much did I spend on amazon in this calendar month—um... no, that should be uh... this or the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'this calendar month' but immediately self-corrects to 'this or the last calendar month'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial incorrect value should be discarded. The final audible intent matches the original instruction exactly, so the task is executed as intended.", "heard_transcript": "How much did I spend on Amazon in this calendar month? Um, no, that should be uh this or the last calendar month." }, "candidate_id": "wav_76c8ed88e0acc61b", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then self-corrects with 'No, that should be uh this or the last calendar month.' While the correction is present, the initial error ('on Amazon') is not a false start but a misstatement of the platform, and the correction is less fluid and more of a rephrasing than a natural self-repair. The prosody is slightly less continuous, and the correction feels more like a restart than a seamless fix." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How much did I spend in this or the last calendar month? Oh, I forgot to mention, I want to know about Amazon.", "long_pause_offsets": [ 57 ], "source_sha256": "53937311c88be638b840e83213a83aa367a87c987a5c6d202d8165a35300b9f3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing context about Amazon, which was omitted in the initial utterance. The time frame and the action of querying spending remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/addition/addition_01/7d1e86012a835be2/speech.wav", "wer": 0.037037037037037035, "wav_sha256": "1832fbe4be3b1fc058b2435558b04a2c4c7bf2b5a6981960f120d601cab3a9bf", "seconds": 14.68, "spoken_utterances": [ "How much did I spend in uh... this or the last calendar month? ", "Oh, I forgot to mention, I want to know about uh, let's see... Amazon." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about spending in 'this or the last calendar month' without specifying a merchant. After a pause and an interjection, they add 'Amazon' as a new constraint. According to the 'addition' type criterion, the final goal is the combination of the initial instruction with the added information. The final audible intent correctly combines the time frame with the merchant, matching the original instruction's intent.", "heard_transcript": "How much did I spend in the this or the last calendar month? Oh, I forgot to mention, I want to know about uh, let's see, Amazon." }, "candidate_id": "wav_4ab416ebf0b05b38", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request, then pauses and adds a specific detail ('Amazon') with a believable 'let's see' filler. The intonation and rhythm are consistent with spontaneous speech, and the addition flows smoothly without awkwardness or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How much did I spend on amazon in? I left out a detail: just for this or the last calendar month.", "long_pause_offsets": [ 35 ], "source_sha256": "40c19b10fc128fed9914fe3f60264d265de4111d565788683baea51d105f1e59", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing time constraint 'just for this or the last calendar month', which clarifies the original ambiguous 'in' clause. The final goal is the same as the original instruction but with a more precise time scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/addition/addition_02/7e077adabfa9888a/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "abc4a0a586291afa475f746e88af51cba65fbd711de13c31743d2043e63e16db", "seconds": 11.16, "spoken_utterances": [ "How much did I spend on um... amazon in? ", "I left out a detail: just for ... this or the last calendar month." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially hesitates with 'um' and 'I left I be detail', which are filler and self-corrections. The final, clear request is 'How much did I spend on Amazon in this or the last calendar month?', which matches the original instruction. The addition type criterion is satisfied as the final intent is the same as the original instruction, with no contradictory or additional constraints.", "heard_transcript": "How much did I spend on um Amazon in I left I be detail just for this or the last calendar month?" }, "candidate_id": "wav_27d8bf6ff9a7025a", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a plausible request but includes a noticeable stutter ('on um') and a slightly unnatural pause before 'Amazon'. The phrase 'I left I be detailed' is awkward and unclear, suggesting a possible self-correction that doesn't sound fully integrated. The overall delivery is less fluent and more hesitant than audio_02." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_time_reference", "proposed_instruction": "How much did I spend on amazon in this calendar month? Make that 'this or the last calendar month', not just 'this calendar month'.", "long_pause_offsets": [ 55 ], "source_sha256": "0b0889d78a33422bfb9d8b884fd80a37e3d158e9fd8d3f4cfa996f4efa49a17b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the time range from 'this calendar month' to 'this or the last calendar month', which matches the original instruction. The app (amazon) and action (query spending) remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/revision/revision_02_time_reference/29a60fd3dcb78a15/speech.wav", "wer": 0.06896551724137931, "wav_sha256": "a9bf4c798100ebd5ee8f0ff7ce28da5da361a5b45928213b98f36fdd3b306040", "seconds": 14.04, "spoken_utterances": [ "How much did I spend on um... amazon in uh, let me think... this calendar month? ", "Make that 'this or the last calendar month', not just 'this calendar month'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks for spending in 'this calendar month', then immediately revises the request to 'this or the last calendar month, not just this calendar month'. This is a clear revision of the time constraint. The final intent matches the original instruction, so the task should be executed as revised.", "heard_transcript": "How much did I spend on um Amazon in uh let me think this calendar month. Make that this or the last calendar month, not just this calendar month." }, "candidate_id": "wav_73320c2d77ac3466", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural, hesitant formulation ('How much did I spend on...') and includes a plausible 'let me think' pause. The revision ('Make that this or the last calendar month') is a smooth, self-corrected rephrasing of the time frame, delivered with a believable change-of-mind intonation and a clear reference back to the original request. The prosody is continuous and the pacing realistic, with no signs of artificiality or disjointedness." } ] }, "is_default": true }, { "variant_id": "revision_01_app", "proposed_instruction": "How much did I spend on ebay in this or the last calendar month? Sorry, I meant Amazon, not eBay.", "long_pause_offsets": [ 65 ], "source_sha256": "7be14870602a6bd0b12df2a983a0ab4c4c36eb40df4ff896c7613cb7db1a8f94", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the platform from eBay to Amazon, preserving the time range and the action of querying spending." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/revision/revision_01_app/7967ece4b169a217/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "d89f8d796f81f7173e84d21da81cb9e58201fee95b5cac6d83c5c68fb7bd7e48", "seconds": 14.4, "spoken_utterances": [ "How much did I spend on uh... ebay in this or the last calendar month? ", "Sorry, I meant um, let's see... Amazon, not eBay." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially mentions 'eBay' but immediately corrects to 'Amazon' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to inquire about spending on Amazon, which matches the original reference instruction. The correction is clear and unambiguous, and no other constraints were altered.", "heard_transcript": "How much did I spend on uh eBay in this or the last calendar month? Sorry, I meant um let's see, Amazon, not eBay." }, "candidate_id": "wav_dea325d2a1b62224", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a similar hesitant tone but makes a more abrupt and less integrated correction. The initial mention of 'eBay' is followed by a 'sorry' and a self-correction to 'Amazon', which feels slightly rushed and less naturally embedded in the flow. The 'let me see' filler is less effective as a thought process indicator compared to the 'let me think' in audio_01, and the overall delivery has a slightly more robotic cadence, possibly due to the quicker correction and less organic-sounding revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_039", "hide_variant_original": true, "source_sha256": "1d609c7dff15e4b1eea3dd5f31ecf3f472c8c8351b20867b1a60d641a2c43289", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/cancellation/cancel_039/c5905ad28ea6cc31/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "e7a8920551d25a314656c63c07943b6460809368a8b58ae78f7978b1c803ff67", "seconds": 11.8, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?", "Hold on, don't go ahead with that. Cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hold on, don't go ahead with that. Cancel it.' This is a clear and final cancellation of the initial request. The control intent matches the 'cancellation' type criterion, as the task is revoked and not executed.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month? Hold on, don't go ahead with that. Cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_022", "hide_variant_original": true, "source_sha256": "2ce9579b3970a2f4b393b0541c05b0d41e4ed8bd51f172a9f3cb3aca7aca0d0b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/ambiguous_halt/halt_022/1eff909463e91470/speech.wav", "wer": 0.05, "wav_sha256": "5be2147c83524352fed19fa6f1b1cb056da867b9c12d3a52d7c07351398f3443", "seconds": 11.36, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?", "Hang on, let me think..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hang on, let me think,' which is a clear request to pause or hold the current task. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. No cancellation or execution is requested, and the initial query is not revoked. The final intent is to pause, which satisfies the type-specific criterion.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month? Hang on, let me think." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "eb5ad85_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f29f41f585ae800b7a48c3bb184ca22030257ef25e5f232b75d40af19e0b2c56", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/eb5ad85_3/audio_natural/filler/original/11f1aae34a827ba4/speech.wav", "wer": 0.06666666666666667, "wav_sha256": "44601832083d0003c644ba9757a2042fce280e04f26bd2ef10dcca2da4ace655", "seconds": 6.88, "spoken_utterances": [ "How much did I spend on ... amazon in uh... this or the last calendar month?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker repeats 'in in' as a filler, which is a common hesitation pattern. The core request remains unchanged and fully intact. The repetition does not alter the goal, so the final intent matches the original instruction. The type is 'filler', and the final control intent is to execute the task.", "heard_transcript": "How much did I spend on Amazon in in this or the last calendar month?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "How many read email threads are in my Outlook—I should say Gmail inbox?", "long_pause_offsets": [], "source_sha256": "73bee7e6a6d722b0cae3d90dcd7b7c0faa472ebaa0c6aae76facaa8c8858c534", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail inbox' within the same utterance, preserving the rest of the query including the metric (read email threads), the scope (inbox), and the intent (count)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/false_start/false_start_02_app/9fc8e4b1f5f0fa7c/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "587b6b06b557f9fd7c2aebf32a3b6338a82b3a157bb81e40c7eae100b509e421", "seconds": 9.12, "spoken_utterances": [ "How many read email threads are in my um... Outlook—uh... I should say ah... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many read email threads are in my um Outlook uh of I should say uh Gmail inbox?" }, "candidate_id": "wav_f93c9b3a4fd2efaf", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker delivers a clear, continuous request with a natural self-correction from 'Outlook' to 'Gmail'. The false start is brief and the correction is smoothly integrated into the sentence. The prosody is consistent, and the speaker sounds like they are formulating the request in real time without robotic pauses or unnatural word boundaries." } ] }, "is_default": true }, { "variant_id": "false_start_01_email_state", "proposed_instruction": "How many unread—sorry, I meant read email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "b160516f18785ea722e90858d70c92cea88680524d26f02bdc8fced2b3fdfe2c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'unread' to 'read' within the same utterance, preserving the rest of the query including the app (Gmail), the object (email threads), and the context (inbox)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/false_start/false_start_01_email_state/16ca25a267cca286/speech.wav", "wer": 0.3125, "wav_sha256": "18055ed7dd738ffae140fe6a576637a9de7984e676fcce6e75e6986272e0ef73", "seconds": 8.0, "spoken_utterances": [ "How many unread—uh... sorry, I meant um... read email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'unread' but immediately self-corrects to 'read' after a pause and filler words. According to the 'false_start' type criterion, the final corrected value ('read') should be applied, and the original incorrect value ('unread') should be discarded. The final audible intent is to count 'read' email threads, which matches the corrected instruction.", "heard_transcript": "How many unread uh sorry, I meant um read email threads are in my Gmail inbox?" }, "candidate_id": "wav_bf0771234e96afc8", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip features a false start with 'unread' followed by a clear self-correction to 'read'. The repair is immediate and the sentence flows naturally. The speaker's hesitation and correction are believable, and the overall delivery is fluent with appropriate pauses. It is slightly less natural than audio_03 due to the initial word being more jarring to correct." } ] }, "is_default": false }, { "variant_id": "false_start_03_email_folder", "proposed_instruction": "How many read email threads are in my Gmail sent—no, make that inbox?", "long_pause_offsets": [], "source_sha256": "26b1829c808e2a50f56f4e54f7ab262bbd1e5d23c92b56386924359186642c69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately corrects to 'inbox'. The final value is 'inbox', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/false_start/false_start_03_email_folder/71c80e34c9dcdc43/speech.wav", "wer": 0.0, "wav_sha256": "bc21f88f74a6512d1c2260a568c0c55eb53efa61f59836612b6a96978986cbdf", "seconds": 11.8, "spoken_utterances": [ "How many um... read email threads are um, let me think... in my Gmail sent—um... no, make that inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects to 'inbox'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many um read email threads are um, let me think. In my Gmail sent um, no, make that inbox." }, "candidate_id": "wav_e2a9b1cd3dd1a936", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker includes multiple fillers ('um', 'let me think') and a revision ('sent' to 'inbox'). However, the pauses are longer and more disruptive, and the revision feels less integrated. The speech is less fluid, with a more hesitant and fragmented delivery, making it sound less spontaneous and more like a rehearsed or awkward attempt." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail inbox? Oh, I forgot to mention, only count the read email threads.", "long_pause_offsets": [ 46 ], "source_sha256": "efebe0baecd57180d3eb6d68785bc4afa75869bedab7e84500e3d705146e0e1c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count the read email threads', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/addition/addition_01/3623ed455790b455/speech.wav", "wer": 0.21739130434782608, "wav_sha256": "aa0cfede64baee23f28726cdae18b588bd7b613812064fa1f35fac64756b4c83", "seconds": 13.36, "spoken_utterances": [ "How many email threads are um... in my Gmail inbox? ", "Oh, I forgot to mention, only count the let's see... read email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial instruction is a general count of email threads. The speaker then adds a constraint to only count 'read email threads'. This addition refines the original request, and the final audible intent is to count only the read threads. The addition is clear and unambiguous, satisfying the 'addition' type criterion.", "heard_transcript": "How many email threads are um in my Gmail inbox? Oh, I forgot to mention, only count the, let's see, read email threads." }, "candidate_id": "wav_07d9fa3960e78000", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding question about email threads. The addition of 'Oh, I forgot to mention' followed by a pause and a specific clarification ('only count the... let's see... read email threads') is a highly plausible spontaneous memory recall. The filler 'let's see' and the slight hesitation before specifying the condition sound authentic, and the overall prosody is smooth and conversational." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many read email threads are in my inbox? One more thing: make sure it's in my Gmail inbox.", "long_pause_offsets": [ 45 ], "source_sha256": "3a396371852128b02df6472537f96ab2dbb113be7e8248004052cb3969fb0afd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Gmail' to the initial inbox query, preserving the original intent of counting read email threads." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/addition/addition_02/28fd60e926976f97/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "cd7cb25e22ebd49204db5fb910afb87f42c5483d6ae67bdd69e42d5536faa3ad", "seconds": 14.32, "spoken_utterances": [ "How many hmmm... read email threads are uh... in my inbox? ", "One more thing: make sure it's ... in my Gmail inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many read email threads are uh in my inbox?' which is a slightly garbled version of the original instruction. The user then adds the clarification 'One more thing, make sure it's in my Gmail inbox.' This addition specifies the email service, which is a necessary constraint. The final intent combines the original request with the added constraint, matching the reference instruction. The addition type requires combining the earlier instruction with the added information, which is done correctly here.", "heard_transcript": "How many read email threads are uh in my inbox? One more thing, make sure it's in my Gmail inbox." }, "candidate_id": "wav_a86d7d1417d4505e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "While the speaker also adds a clarification, the initial phrasing 'How many mum read email threads are uh in my inbox' contains a noticeable stutter ('mum') and a filler ('uh') that disrupt the flow more than in audio_01. The addition ('One more thing. Make sure it's in my Gmail inbox') is grammatically awkward and sounds like a forced, less natural afterthought compared to the more integrated addition in audio_01. The prosody is less fluid overall." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_email_state", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Wait, I meant read email threads, not unread.", "long_pause_offsets": [ 53 ], "source_sha256": "00fd9293a03981be05e9f241a89fff9ecaaec257a17b963ca63f7f56adb4a5a7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "different", "reason": "The followup explicitly corrects 'unread' to 'read', reversing the original condition. The final goal is the opposite of the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/revision/revision_01_email_state/0f2e8915c94f5280/speech.wav", "wer": 0.1, "wav_sha256": "8eaa8efa1fbb245fd594959a14acd046143afda48beecca58f4bd620fee054ad", "seconds": 14.04, "spoken_utterances": [ "How many ... unread email threads are um... in my Gmail inbox? ", "Wait, I meant uh... read email threads, not unread." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'unread email threads' but immediately corrects themselves to 'read email threads, not unread'. This is a clear revision where the final intent is to count read threads. The correction is made after a pause and with a self-repair, matching the revision type criterion. The final goal aligns with the original instruction.", "heard_transcript": "How many unread email threads are um in my Gmail inbox? Wait, I meant uh read email threads, not unread." }, "candidate_id": "wav_56c37afb4c2f6dc2", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, slightly hesitant question about unread email threads, then self-corrects to 'read email threads' with a natural 'wait, I meant...' transition. The correction is immediate and fluent, with a believable shift in intonation and pacing. The pause and filler 'um' at the start sound like genuine hesitation, and the overall prosody is smooth and continuous, making the revision feel spontaneous and realistic." } ] }, "is_default": false }, { "variant_id": "revision_03_email_folder", "proposed_instruction": "How many read email threads are in my Gmail sent? Change the folder to inbox instead of sent.", "long_pause_offsets": [ 50 ], "source_sha256": "be6caf8340cb3dcca5d4888557219fc7af18cc94736734380af358ca867393a5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from 'sent' to 'inbox', preserving the action, the entity (read email threads), and the app (Gmail)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/revision/revision_03_email_folder/073ff67cd7325d7d/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "9645b7a2784d724999525848fd2f6717d69498a232b1cae523fd0990f4ceea53", "seconds": 13.44, "spoken_utterances": [ "How many uh... read email threads are ... in my Gmail sent? ", "Change the folder to let's see... inbox instead of sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'Gmail sent' but then corrects to 'inbox' after a pause, which is the final intent. The revision type requires applying the correction spoken after a longer pause to the earlier instruction. The final goal matches the original instruction.", "heard_transcript": "How many uh read email threads are in my Gmail sent? Change the folder to let's see inbox instead of sent." }, "candidate_id": "wav_8561a32fa6e7fe13", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a hesitant 'how many uh' and a filler 'uh', then asks about read email threads in 'sent'. After a pause, they revise the folder to 'inbox'. While the revision is present, the second part of the utterance sounds more like a follow-up thought than a natural self-correction. The intonation and pacing of the revision are slightly abrupt, and the phrase 'change the folder to let's see inbox' feels more like a task instruction than a spontaneous correction, reducing its realism compared to audio_01." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "How many read email threads are in my Outlook inbox? Let me correct the app: check for read email threads in Gmail, not Outlook.", "long_pause_offsets": [ 53 ], "source_sha256": "ea674c995bcd9aa9e7f57a4ffc90df732adcfd4dcf4d44ab98030c2eb7fb4c84", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Outlook to Gmail, preserving the original query about read email threads in the inbox." }, "audio": { "status": "pending", "spoken_utterances": [ "How many read email threads are uh... in my Outlook inbox? ", "Let me correct the app: check for read email threads in let me think... Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_012", "hide_variant_original": true, "source_sha256": "10ac01bda22e844ca3aac72f20a988afb1acf2265b725917e5ffcc602422c27b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/cancellation/cancel_012/4d0f30fc362a458f/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "b5016d7acf976b3abed9af905938918183f2689d195522300cbde78a4a90ac88", "seconds": 11.6, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?", "Wait, actually, I've changed my mind. Don't do that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks a question but then clearly revokes the task with 'Wait, actually, I've changed my mind. Don't do that.' This constitutes a direct cancellation of the original request. The final intent is to cancel, not to execute the counting task.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox? Wait, actually, I've changed my mind. Don't do that." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_032", "hide_variant_original": true, "source_sha256": "95b8ee26be82e6b67f5e4300f66a72fca38e8af6dd8d9f145da8174045cae19a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/ambiguous_halt/halt_032/4f4d983280dd4a38/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "0b5ad73e4638718aed89cf90e81e77499bdcaa57e0bbbc744d5f622ef3dcbb3a", "seconds": 9.72, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?", "Wait, before you do that..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then immediately says 'Wait, before you do that,' which is a clear request to pause the current action. The final interjection is a pause command, not a cancellation or a new instruction. The audio ends with this request, satisfying the 'ambiguous_halt' type criterion.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox? Wait, before you do that" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "f5e7b573e2889c590e349806155632c9efa15d21ec2b96d49adcdce99e271140", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_1/audio_natural/filler/original/ce4b9d3d35e01740/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "03f2169ac672bbf93b9a39b7ea0b0832d1e53c863b4d7161b8af836cb4536b36", "seconds": 5.6, "spoken_utterances": [ "How many uh... read email threads are um... in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker uses fillers ('uh', 'um') but does not change the core request. The final intent is to count read email threads in the Gmail inbox, matching the original instruction. The fillers do not alter the task goal, so the execution intent is preserved.", "heard_transcript": "How many uh read email threads are um in my Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_email_state", "proposed_instruction": "How many read—no, I mean unread email threads are in my Gmail outbox?", "long_pause_offsets": [], "source_sha256": "8ebdc4424bd3a7c227fbf08e49eab0eb74f9cabd524f869cb08dcaf3af6fc899", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker corrects 'read' to 'unread' within the same utterance, preserving the app (Gmail), the object (email threads), and the location (outbox)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/false_start/false_start_01_email_state/4148469929c78f11/speech.wav", "wer": 0.13333333333333333, "wav_sha256": "a1a0d22b31c3441541e0eea0ede0926a827cc82924a9724bb1ebb58ad5948bcb", "seconds": 5.32, "spoken_utterances": [ "How many read—uh... no, I mean unread email threads are in my Gmail ... outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'read' but immediately self-corrects to 'unread' with 'no, I mean unread'. This is a false start where the final, corrected value is the intended goal. The rest of the instruction matches the original, so the final intent is clear and executable.", "heard_transcript": "How many read uh, no, I mean unread email threads are in my Gmail outbox?" }, "candidate_id": "wav_e735cb0d2112226e", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio features a very natural false start with 'read' followed by an immediate self-correction to 'unread'. The correction is smooth and spoken with a realistic, slightly hesitant tone. The rest of the sentence flows without further issues, and the speaker maintains a consistent, believable pace and prosody throughout." } ] }, "is_default": true }, { "variant_id": "false_start_03_email_folder", "proposed_instruction": "How many unread email threads are in my Gmail inbox—scratch that, make it outbox?", "long_pause_offsets": [], "source_sha256": "65cc414db6ce2e7f9988e8dfdb9a79c32f8ef179711562488a2e7c671a424884", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'inbox' but immediately corrects it to 'outbox' within the same utterance. The final intended folder is 'outbox', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/false_start/false_start_03_email_folder/12f260f744901a6b/speech.wav", "wer": 0.10526315789473684, "wav_sha256": "a3f49792a40f64737eaaa479c4c6c5c89473d0113415d0537211de1b06c6d83b", "seconds": 10.0, "spoken_utterances": [ "How many um, let's see... unread email threads are in my Gmail inbox—hmmm... scratch that, make it outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the inbox but immediately self-corrects with 'Scratch that. Make it outbox.' This constitutes a false start where the final audible intent is the corrected version. The final instruction matches the original reference, so the task is executed as intended.", "heard_transcript": "How many um, let's see, unread email threads are in my Gmail inbox? Hmm. Scratch that. Make it outbox." }, "candidate_id": "wav_d2edae9a0516ff74", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'unred' and a self-correction to 'unread', which is realistic. However, the speaker then says 'scratch that' and changes the email service to 'outbox', which is a more abrupt and less natural shift. The prosody becomes slightly more robotic after the correction, and the final phrase lacks the smoothness of audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "How many unread email threads are in my Outlook—sorry, I meant Gmail outbox?", "long_pause_offsets": [], "source_sha256": "2aa010c52533c6c0c2be3ec40faa2f4020031b4198d2ce7fdf42c50aa50c7fc4", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the rest of the query including the mailbox ('outbox') and the count type ('unread email threads')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/false_start/false_start_02_app/8c78d0210f5e58ba/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "667aea07d3b7e8479ea7a882cd7a30379b1828fea81ad30304d3feca75a4d2ab", "seconds": 10.32, "spoken_utterances": [ "How many unread email threads are in my hmmm... Outlook—uh... sorry, I meant hmmm... Gmail outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail' with 'sorry, I meant'. This is a clear false start where the final value replaces the initial incorrect one. The rest of the instruction matches the original, so the final intent is correct.", "heard_transcript": "How many unread email threads are in my Outlook uh sorry, I meant Gmail outbox?" }, "candidate_id": "wav_6e8a1466a4237fe7", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains multiple hesitations and a more complex revision process. The speaker starts with 'Gmail', then hesitates and says 'Outlook', then corrects to 'Gmail'. The back-and-forth is less fluid and sounds more like a forced or rehearsed correction rather than a spontaneous thought process. The pauses and fillers feel slightly overused, reducing the overall naturalness." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail outbox? Oh, I forgot to mention, only count the unread email threads.", "long_pause_offsets": [ 47 ], "source_sha256": "bc811b322114d20b7a55c42347db9d6c0836b116f0617fd612e11d9df3a43200", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'unread' to the initial question, refining the scope without contradicting the original intent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/addition/addition_01/2ce6dcbad4e92582/speech.wav", "wer": 0.047619047619047616, "wav_sha256": "015a5d554100ee7dfb7492538089671a4cc34a7e8e434729ebf4500d860391d6", "seconds": 12.0, "spoken_utterances": [ "How many email threads are in my Gmail um... outbox? ", "Oh, I forgot to mention, only count the ... unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was to count email threads in the Gmail outbox, which is a broad request. The speaker then adds a constraint to only count unread email threads. This matches the 'addition' type, where the final goal combines the original instruction with the added information. The final audible intent is to count unread email threads, which is a narrower version of the original request and fully satisfies the addition type criterion.", "heard_transcript": "How many email threads are in my Gmail um, outbox? Oh, I forgot to mention, only count the unread email threads." }, "candidate_id": "wav_779e150fcc901d1a", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a clear, natural-sounding request. The pause and 'um' are brief and plausible, followed by a self-correcting addition ('Oh, I forgot to mention...') that smoothly integrates the new constraint ('only count the unread email threads'). The intonation and pacing reflect genuine thought processing, and the correction is grammatically and semantically coherent within the context." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many unread email threads are in my outbox? One more detail: this is specifically in my Gmail.", "long_pause_offsets": [ 48 ], "source_sha256": "3c43806408f8db041214d4c79dc0aadf8210bc97f76ae93e2b768b7cf7873a9d", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Gmail' to the initial outbox query, preserving the original intent of counting unread threads in the outbox." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/addition/addition_02/019a5043b1471366/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "c5a5302296efd150c077bca3d18362194d0853c5bb09bd64d87f117a95ecd8ac", "seconds": 11.04, "spoken_utterances": [ "How many ... unread email threads are in my let me think... outbox? ", "One more detail: this is specifically in my Gmail." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How many unread email threads are in my let me think outbox?', which contains a mistaken value ('let me think') for the mailbox. After a pause, the speaker adds 'One more detail, this is specifically in my Gmail.' This addition clarifies the email service. The final intent combines the original request with the added constraint, resulting in the correct goal: 'How many unread email threads are in my Gmail outbox?'. The addition type criterion is satisfied as the final goal matches the reference instruction.", "heard_transcript": "How many unread email threads are in my let me think outbox? One more detail, this is specifically in my Gmail." }, "candidate_id": "wav_53fb6f6922b7afaa", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a slightly hesitant 'let me think' filler, which is realistic, but the subsequent addition ('one more detail...') feels less integrated. The phrase 'this is specifically in my Gmail' is delivered as a separate, slightly abrupt afterthought rather than a smooth revision. The prosody is flatter, and the overall flow is less cohesive compared to audio_01, making it sound more like a rehearsed or less spontaneous addition." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_email_folder", "proposed_instruction": "How many unread email threads are in my Gmail inbox? Change the folder to outbox, not inbox.", "long_pause_offsets": [ 53 ], "source_sha256": "21b7ab4cd522b67232122920cfb9054d4696547be305d0a35a102625505dfabc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from inbox to outbox, preserving the action, metric, and app context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/revision/revision_03_email_folder/182b1569fe59aa93/speech.wav", "wer": 0.09523809523809523, "wav_sha256": "0baabc2e9bb61c83c38509309e5eabac53131d137124bb46520218724425285b", "seconds": 11.84, "spoken_utterances": [ "How many um, let's see... unread email threads are in my uh... Gmail inbox? ", "Change the folder to outbox, not inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'inbox' but immediately corrects the folder to 'outbox' with the phrase 'Change the folder to outbox, not inbox.' This constitutes a revision where the final intent is to query the outbox, matching the original instruction. The correction is clear and unambiguous, satisfying the revision type criterion.", "heard_transcript": "How many um, let's see, unread email threads are in my uh Gmail inbox? Change the folder to outbox, not inbox." }, "candidate_id": "wav_5d813af545ed71cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker begins with a natural hesitation ('um') and a self-pacing phrase ('let's see'), followed by a clear initial request. The correction ('change the folder to outbox, not inbox') is delivered as a smooth, immediate follow-up with a natural tone shift, indicating a spontaneous realization. The prosody is consistent, and the correction is integrated into the flow without awkwardness or repetition." } ] }, "is_default": true }, { "variant_id": "revision_01_email_state", "proposed_instruction": "How many read email threads are in my Gmail outbox? Wait, it should be unread email threads, not read.", "long_pause_offsets": [ 52 ], "source_sha256": "0360be0168e11742e18a19802346a263a8ee07bc60e28310d3437827fcbe63b5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the initial condition from 'read' to 'unread', preserving the app (Gmail), location (outbox), and the count query." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/revision/revision_01_email_state/6fb0d1c0cfee8cb0/speech.wav", "wer": 0.05, "wav_sha256": "1c2abae098c64309cba9dd992a37d9122cde3cd9d9ccaeb747fdf2297972adbf", "seconds": 11.6, "spoken_utterances": [ "How many read email threads are in my Gmail ... outbox? ", "Wait, it should be um... unread email threads, not read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'read email threads' but immediately corrects themselves to 'unread email threads' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to count unread email threads, which matches the original instruction.", "heard_transcript": "How many read email threads are in my Gmail outbox? Wait, it should be um unread email threads, not read." }, "candidate_id": "wav_98589596ee6ff484", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear but slightly rushed request, then pauses and corrects 'read' to 'unread' with a hesitant 'wait, it should be um...'. While the correction is plausible, the initial request is delivered more flatly, and the correction feels slightly more abrupt and less fluid than in audio_01. The prosody lacks the same natural pacing and intonation variation, making it sound less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "How many unread email threads are in my Outlook outbox? Let me correct the app: check Gmail, not Outlook.", "long_pause_offsets": [ 56 ], "source_sha256": "53bf7d8a31cc03c3b83fbe40e321c3dd068d0e34522f1c547d5c0dfa946f2896", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Outlook to Gmail, preserving the rest of the query including the metric (unread email threads), the folder (outbox), and the user context (my)." }, "audio": { "status": "pending", "spoken_utterances": [ "How many unread email threads are in my uh, let me think... Outlook um... outbox? ", "Let me correct the app: check Gmail, not Outlook." ] }, "audio_semantic": { "status": "unchecked", "reason": "현재 WAV 바이트와 원본 목표에 일치하는 Audio semantic 결과가 없습니다." }, "naturalness": { "rank": null }, "candidate_id": null, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_001", "hide_variant_original": true, "source_sha256": "c576ddc884f349fa266fc058911395191812450d2aecad641ebaaa5ad8af7248", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/cancellation/cancel_001/0e788788c57e0f83/speech.wav", "wer": 0.0, "wav_sha256": "d5e75ec667b7a2d320b7c278287a86481e66be4e66aa96006c4bbf3fe391b3f7", "seconds": 9.88, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?", "Wait, cancel that." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Wait, cancel that,' which is a clear and direct cancellation of the task. The final interjection 'Um' does not introduce a new task or modify the cancellation. The audio ends with a cancellation intent, satisfying the criterion for the 'cancellation' type.", "heard_transcript": "How many unread email threads are in my Gmail outbox? Wait, cancel that. Um" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_016", "hide_variant_original": true, "source_sha256": "007ad2cc2b2b1bb4ec73f36817492983eb751816001eb035da61351af403a005", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/ambiguous_halt/halt_016/1f708f1fe403b408/speech.wav", "wer": 0.0, "wav_sha256": "dcf1a6d5add9de98b9880d60cf5cdacce7d24cdc17f5e79552a9fea711427d38", "seconds": 9.32, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?", "Just a second, please." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Just a second, please,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The original instruction is not canceled or revised, and the user does not resume or modify the task. Therefore, the final intent is to pause, and the control intent is 'pause'.", "heard_transcript": "How many unread email threads are in my Gmail outbox? Just a second, please." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "9eb3bb845076bf27aa79a24bdee7f5187cec03777ef6e5e5540e499708a0ab96", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_2/audio_natural/filler/original/939e53f0303b4858/speech.wav", "wer": 0.0, "wav_sha256": "1962ff2efecbed21d13e3c9382629facdb2131a89e9b7fb2f624a88091a9580a", "seconds": 5.84, "spoken_utterances": [ "How many ... unread email threads are in my Gmail ... outbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio contains only the original instruction without any hesitation, backchannel, or repair. The speaker delivers the full request clearly and directly, matching the reference instruction. As this is a 'filler' type task, any initial hesitation or pause within the instruction is ignored, and the final goal is to execute the complete request as heard.", "heard_transcript": "How many unread email threads are in my Gmail outbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_email_state", "proposed_instruction": "How many read—no, I meant unread email threads are in my Gmail inbox?", "long_pause_offsets": [], "source_sha256": "97567802033a201be991ad656e9df2ac1e23c7821a472e20ddfe20b3c2c77932", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'read' to 'unread' within the same utterance, preserving the rest of the query including the app (Gmail), the object (email threads), and the context (inbox)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/false_start/false_start_01_email_state/a70a024834fcd4df/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "aa61bcb834f93f5b9ced42b0fac837fbb17b5fe6008b8a0650bc8e35cc154214", "seconds": 11.24, "spoken_utterances": [ "How many um... read—uh... no, I meant hmmm... unread email threads are in my Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'red' but immediately self-corrects to 'unread' after a pause and a laugh. The final, corrected value 'unread' is the intended target. The rest of the instruction matches the original. This is a clear false start where the correction is applied to the initial request.", "heard_transcript": "How many um red uh no, I meant unread email threads are in my Gmail inbox?" }, "candidate_id": "wav_195fad98d7e37b32", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This speaker exhibits a very natural false start with 'red' followed by a self-correction to 'unread' and a brief, genuine laugh. The prosody is fluid, the pause before the correction is realistic, and the intonation suggests spontaneous thought. The correction is immediate and integrated smoothly into the sentence, making it sound like a real-time mental correction." } ] }, "is_default": true }, { "variant_id": "false_start_02_app", "proposed_instruction": "How many unread email threads are in my Outlook—sorry, that should be Gmail inbox?", "long_pause_offsets": [], "source_sha256": "977dbab81e7f699ab0359b9d993451c2800bd8b5266b3b197b281d791b3b7d39", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'Outlook' to 'Gmail' within the same utterance, preserving the rest of the query including the app, metric, and context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/false_start/false_start_02_app/b9231353bca14572/speech.wav", "wer": 0.16666666666666666, "wav_sha256": "b39f76a5eae2e6d53f8f2e14690d8ed9156cf5e04ea38b7e0cd4ded4fcaa446b", "seconds": 9.28, "spoken_utterances": [ "How many um... unread email threads are in my Outlook—uh... sorry, that should be um... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Outlook' but immediately self-corrects to 'Gmail'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "How many um unread email threads are in my Outlook uh sorry, that should be um Gmail inbox?" }, "candidate_id": "wav_137b1020708d9877", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a false start with 'Outlook' and a correction to 'Gmail'. The speaker uses 'sorry' and 'um' naturally, and the correction is clear. However, the initial error ('Outlook') is more jarring than 'red' in audio_01, and the correction is slightly less fluid, with a longer pause and a more hesitant tone, making it feel a bit less spontaneous." } ] }, "is_default": false }, { "variant_id": "false_start_03_email_folder", "proposed_instruction": "How many unread email threads are in my Gmail sent—scratch that, make it inbox?", "long_pause_offsets": [], "source_sha256": "64d8ee7cfed4b102d360eacde2bcac6b66b1f3bd4cb90cada1e32c58f23c7504", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately corrects to 'inbox' within the same utterance. The final intended folder is 'inbox', matching the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/false_start/false_start_03_email_folder/8fc224a91db55de7/speech.wav", "wer": 0.125, "wav_sha256": "d018318ee15c6370dd32ab2093588eb59fda46c399cd53a9e89692511bb4215e", "seconds": 8.8, "spoken_utterances": [ "How many ... unread email threads are in my Gmail sent—um... scratch that, make it inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'sent' but immediately self-corrects with 'scratch that, make it inbox.' This is a clear false start where the final value replaces the mistaken one. The final intent matches the original instruction exactly, so the task should be executed.", "heard_transcript": "How many unread email threads are in my Gmail sent um, scratch that, make it inbox." }, "candidate_id": "wav_56f1b463f690c8c9", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This speaker says 'sent' and then immediately corrects to 'inbox' with 'scratch that'. While this is a valid revision, the correction is abrupt and lacks the natural hesitation or filler words seen in the others. The delivery is more robotic, and the correction feels like a mechanical edit rather than a genuine moment of self-correction, reducing its realism." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "How many email threads are in my Gmail inbox? Oh, I forgot to mention, only count the unread email threads.", "long_pause_offsets": [ 46 ], "source_sha256": "fd954453fa265184322c30ebc134385fc9ff5bae3932b05e0b7f5f6402440ac8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing constraint 'only count the unread email threads', which refines the initial broad request to match the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/addition/addition_01/f7e97292b60e68d9/speech.wav", "wer": 0.041666666666666664, "wav_sha256": "22887b74377a28ab7aa77cf8d1fcf75bd88b637a256e8cd661f9c6e3d99f26ac", "seconds": 12.84, "spoken_utterances": [ "How many email threads are in my let me think... Gmail inbox? ", "Oh, I forgot to mention, only count the um... unread email threads." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request was 'How many email threads are in my Gmail inbox?' and was followed by an addition specifying 'only count the unread email threads.' This addition narrows the scope of the original request to only unread threads, which matches the final intent of the reference instruction. The addition is clear and directly modifies the original task without introducing conflicting or incorrect constraints.", "heard_transcript": "How many email threads are in my let me think Gmail inbox? Oh, I forgot to mention, only count the um unread email threads." }, "candidate_id": "wav_d92125f77741f2ce", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays a highly natural and plausible sequence of self-correction and addition. The speaker begins with a clear request, hesitates with a believable 'let me think' pause, then adds a specific restriction ('only count the unread email threads') after a second pause. The prosody includes natural intonation shifts, filler words ('um'), and a smooth, continuous flow that mimics genuine spontaneous speech. The addition feels like a remembered detail rather than a rehearsed command." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "How many unread email threads are in my inbox? One more detail: check my Gmail inbox.", "long_pause_offsets": [ 47 ], "source_sha256": "f7ee3839a54df59b69a8b80fcea920acbfb68863b32de3cbe83e12874e7ffbb8", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing app context 'Gmail' to the initial inbox query, preserving the original intent of counting unread email threads in the specified inbox." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/addition/addition_02/38b96f1be488d999/speech.wav", "wer": 0.1111111111111111, "wav_sha256": "4575b919b043665e87e3eb6e4392ca6ec6370e30167867ef1a00afc65399ac36", "seconds": 12.36, "spoken_utterances": [ "How many uh... unread email threads are in my inbox? ", "One more detail: check my hmmm... Gmail inbox." ] }, "audio_semantic": { "status": "equivalent", "reason": "The initial request is 'How many unread email threads are in my inbox?' and is followed by an addition 'check my Gmail inbox.' The addition specifies the email service as Gmail, which refines the original request. The final intent combines the original query with the added detail, resulting in a complete and executable task. The addition is not a revision or cancellation, and the final control intent is to execute the combined task.", "heard_transcript": "How many uh unread email threads are in my inbox? One more detail, check my hmm Gmail inbox." }, "candidate_id": "wav_e8c27a98348bfba6", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio also contains a natural-sounding addition, but the flow is less smooth. The speaker starts with a slight false start ('hominy'), then corrects to 'how many', which is a minor but noticeable self-repair. The addition ('check my Gmail inbox') is less integrated into the original request and sounds more like a separate afterthought. The 'hmm' pause is longer and less connected to the prior sentence, making the overall delivery feel slightly more disjointed compared to audio_01." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_02_app", "proposed_instruction": "How many unread email threads are in my Outlook inbox? Let me correct the app: check Gmail, not Outlook.", "long_pause_offsets": [ 55 ], "source_sha256": "5aea73f5b031c59cd73b14b83712398707ed3d62b6eaabb05038386013034f4e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the app from Outlook to Gmail, preserving the original query about unread email threads and the inbox context." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/revision/revision_02_app/441633135a2edf38/speech.wav", "wer": 0.05, "wav_sha256": "31a08928ede55cb955a29e0dd823671d94188a8955d8c4ca6327ad10ecb8cd2c", "seconds": 12.16, "spoken_utterances": [ "How many um... unread email threads are in my Outlook inbox? ", "Let me correct the app: check ... Gmail, not Outlook." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about Outlook but immediately corrects the target application to Gmail. The revision is clear and final, and the rest of the instruction (counting unread email threads in the inbox) remains unchanged. The final intent matches the original reference instruction.", "heard_transcript": "How many um unread email threads are in my Outlook inbox? Let me correct that. Check Gmail, not Outlook." }, "candidate_id": "wav_acf4e4c6a10cacaa", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and plausible revision flow. The speaker begins with a clear but slightly hesitant request ('How many um unread email threads are in my Outlook inbox'), then self-corrects with a believable 'Let me correct the app' followed by a smooth, contextually appropriate correction to 'Gmail not Outlook'. The prosody is consistent, the pause before the correction feels like genuine recollection, and the intonation reflects a spontaneous change of mind rather than a rehearsed or robotic edit." } ] }, "is_default": true }, { "variant_id": "revision_03_email_folder", "proposed_instruction": "How many unread email threads are in my Gmail sent? Change the folder to inbox instead of sent.", "long_pause_offsets": [ 52 ], "source_sha256": "76b2a46b3da3a8368ab4e32c821bbe52762f3db73349f75683a04a8b91ce45eb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the folder from 'sent' to 'inbox', preserving the action, metric, and app. The final goal matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/revision/revision_03_email_folder/85c693bb259c9f05/speech.wav", "wer": 0.15, "wav_sha256": "874a880420db3fe54a27d20e80d17e5751c383d6dbb71e280398d308ae5bc564", "seconds": 13.0, "spoken_utterances": [ "How many um... unread email threads are in my ... Gmail sent? ", "Change the folder to uh... inbox instead of sent." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about the 'sent' folder but immediately corrects the folder to 'inbox' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final goal matches the original reference instruction exactly.", "heard_transcript": "How many um unread email threads are in my Gmail sent? Change the folder to uh inbox instead of sent." }, "candidate_id": "wav_ee6514095f22c7d3", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio shows a revision where the speaker initially asks about 'Gmail sent' and then corrects to 'inbox'. The correction is clear and the prosody is natural, but the initial request is slightly more awkward ('change the folder to uh inbox instead of sent') compared to audio_02's more fluent phrasing. The pause before 'uh inbox' is plausible, but the overall flow feels a bit more like a direct command than a spontaneous thought process, making it less realistic than audio_02." } ] }, "is_default": false }, { "variant_id": "revision_01_email_state", "proposed_instruction": "How many read email threads are in my Gmail inbox? Wait, the email state should be unread, not read.", "long_pause_offsets": [ 51 ], "source_sha256": "07f10c7ec95602319ae67f970150f3dcdad49fc4c3a1cefd04bfed4acb0a9175", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the email state from 'read' to 'unread', preserving the app (Gmail), the object (email threads), and the action (counting)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/revision/revision_01_email_state/f7ac36ebc99bc61c/speech.wav", "wer": 0.07407407407407407, "wav_sha256": "6903278c773057b2541d7ac96f7e9b5a03bc36def3413181b0f2fe9683943e75", "seconds": 18.2, "spoken_utterances": [ "How many uh... read email threads are in my uh, let me think... Gmail inbox? ", "Wait, the email state should be uh, let's see... unread, not read." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially asks about 'read email threads' but immediately corrects themselves to 'unread, not read' after a pause. This constitutes a revision where the final audible intent is to count unread threads, matching the original instruction. The correction is clear and applied to the earlier request.", "heard_transcript": "How many uh um read email threads are in my uh let me think Gmail inbox? Wait, the email state should be uh let's see unread, not read." }, "candidate_id": "wav_33a8da043a3cfd96", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains multiple fillers ('uh', 'um', 'let me think') and a more complex revision process. The speaker starts with 'read email threads', then corrects to 'unread', but the correction is less smooth and more fragmented. The phrase 'let's see' after 'wait' sounds like a filler rather than a genuine thought process, and the overall delivery feels more hesitant and less coherent than the others. The multiple pauses and self-corrections make it sound less natural and more like a forced or rehearsed attempt at revision." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_023", "hide_variant_original": true, "source_sha256": "d76ca8f07e5b622a7fd028cacd1642d7793e1eec0b816fd36e5b37b58cf79696", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/cancellation/cancel_023/fc4c85a32f68f87c/speech.wav", "wer": 0.045454545454545456, "wav_sha256": "e82c599c031f202f6a3f34b30e11402bad16fd914b9764ffb6a553b4e0ca3e21", "seconds": 12.76, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?", "Hang on, never mind about that. Don't carry out the request." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker first asks a question, then immediately says 'Hang on, never mind about that. Don't carry out the request,' which is a clear and unambiguous cancellation of the initial task. The final interjection explicitly instructs not to execute the request, satisfying the cancellation type criterion.", "heard_transcript": "How many unread email threads are in my um Gmail inbox? Hang on, never mind about that. Don't carry out the request." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_002", "hide_variant_original": true, "source_sha256": "4ede491384041ee352659aaeed5b56489e1c7a27a94f83e0fdd560ef0c28407e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/ambiguous_halt/halt_002/0eeb718341e14428/speech.wav", "wer": 0.07692307692307693, "wav_sha256": "9f26e7add47cf5c0141bc2c0a8914989f052f4336a55c7c68fbf72272fc22d98", "seconds": 9.72, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?", "Hold on..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with the speaker saying 'Hold on,' which is a clear request to pause or wait. This matches the 'ambiguous_halt' type, where the final interjection is a pause command. The initial request is not canceled, but the speaker explicitly halts further action.", "heard_transcript": "How many unread email threads are in my um Gmail inbox? Hold on." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "f691597_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "d6b556f2851b8e5d9f1b91403e652e733d6ece66f46024cc67a0895d69b67ddc", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/f691597_3/audio_natural/filler/original/abc49c8643fa99bb/speech.wav", "wer": 0.09090909090909091, "wav_sha256": "229018add82654cdafc7694cce3c8258c95a24079bcaeda767da0189d7f8c3e3", "seconds": 6.64, "spoken_utterances": [ "How many ... unread email threads are in my um... Gmail inbox?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler word 'um' but does not change the core request. The final intent is to count unread email threads in the Gmail inbox, matching the original instruction. The filler does not alter the task goal.", "heard_transcript": "How many unread email threads are in my um Gmail inbox?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_01_contact_group", "proposed_instruction": "Add all my family—I meant friends and roommates as friends on venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "2453711e96b061a415a5e75894e1b3679a22f3835f1df1caf0b311d906c8d437", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family' to 'friends and roommates' within the same utterance, preserving the app, action, and conditional scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/false_start/false_start_01_contact_group/a59836a908e5f589/speech.wav", "wer": 0.1, "wav_sha256": "cdf454030d9baa29af4b8e93a373a1c44190579ab532190d747a3d09571d4545", "seconds": 10.0, "spoken_utterances": [ "Add all my uh... family—hmmm... I meant friends and roommates as friends on venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family' but immediately self-corrects to 'friends and roommates'. According to the 'false_start' type criterion, the final corrected intent overrides the initial mistake. The corrected instruction matches the original reference exactly, including all conditions and targets. Therefore, the task should be executed as stated in the correction.", "heard_transcript": "Add all my uh family, I meant friends and roommates as friends on Venmo, if they are not already." }, "candidate_id": "wav_0c8c5d4c08ecd103", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent self-correction. The speaker begins with 'family' and immediately self-corrects to 'friends and roommates' with a smooth, conversational tone and a believable 'I meant' phrase. The prosody is consistent, and the correction feels spontaneous rather than rehearsed. The pacing and intonation suggest genuine thought-in-progress, making it the most realistic." } ] }, "is_default": true }, { "variant_id": "false_start_03_deduplication_condition", "proposed_instruction": "Add all my friends and roommates as friends on venmo, if they are already—no, make that if they are not already.", "long_pause_offsets": [], "source_sha256": "ea46e384f21f5e034e8515c397f41d07c5792f900fd6465fdd0047c48aabf7a6", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately corrects it to 'if they are not already'. The final condition matches the original instruction." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/false_start/false_start_03_deduplication_condition/eb2577cfd38e15ca/speech.wav", "wer": 0.16, "wav_sha256": "8fbe880600a687d2cd5f402b3a0dd85b56efc94e1cc5a3dcab968877e2e1429e", "seconds": 12.8, "spoken_utterances": [ "Add um... all my friends and roommates as friends on venmo, uh... if they are already—hmmm... no, make that if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately self-corrects with 'no, make that, if they are not already.' This is a clear false start where the final, corrected condition matches the original instruction. The final intent is to add friends only if they are not already added, which aligns with the original task.", "heard_transcript": "Add all my friends and roommates as friends on Venmo. Uh, if they are already, uh, no, make that, if they are not already." }, "candidate_id": "wav_47b2caf272337259", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a clear revision after an initial statement. The speaker starts with 'add all my friends and roommates as friends on Venmo' and then corrects 'if they are already' to 'if they are not already'. The correction is natural, with a slight hesitation and a 'make that' phrase that sounds like a genuine rephrasing. The prosody is slightly more hesitant than audio_02, but the flow remains credible." } ] }, "is_default": false }, { "variant_id": "false_start_02_app", "proposed_instruction": "Add all my friends and roommates as friends on cashapp—sorry, make that venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "31b75096c936e051ff5668099e84f8abd47dbf68441f8c37f8d7ef839ef7f543", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'cashapp' but immediately corrects it to 'venmo' within the same utterance. The rest of the instruction, including the list of people and the conditional, remains unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/false_start/false_start_02_app/df47b3c21796a507/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "2626d2e96a7a509d7788571a8c501bfb53524332098a0bd3a17b2f05237ffdda", "seconds": 12.68, "spoken_utterances": [ "Add all my friends and roommates as friends um... on cashapp—uh... sorry, make that um... venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' after a pause and 'sorry'. According to the 'false_start' type criterion, the final corrected value should be applied. The rest of the instruction matches the original, and the correction is clear and immediate. Therefore, the final intent is to execute the task on Venmo.", "heard_transcript": "Add all my friends and roommates as friends, um, on Cash App, uh, sorry, make that, um, Venmo, if they are not already." }, "candidate_id": "wav_4792cef5d45585fa", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start with 'add all my friends and roommates' friends' followed by a pause and an unclear word ('uncash up'), which is then corrected to 'Venmo'. The initial phrase is awkward and the correction feels less natural. The speaker's hesitation and the unclear word suggest either a genuine slip or a recording issue, making the delivery less realistic compared to the others." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Add all my friends and roommates as friends on venmo,. And make sure to add them only if they are not already friends.", "long_pause_offsets": [ 55 ], "source_sha256": "93b756985fe3381884b05c03a20507a291f3a9cafabdc010782f9a7983c2389e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the condition 'only if they are not already friends', which clarifies the original instruction's implicit intent. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/addition/addition_01/ae635dc5b1333161/speech.wav", "wer": 0.08333333333333333, "wav_sha256": "2f695809eb981de26b9ad92c14feb32632cfe43039b4f67ea1e2c5a1f15af1fe", "seconds": 13.68, "spoken_utterances": [ "Add uh... all my friends and roommates as friends ... on venmo,. ", "And make sure to add them ... only if they are not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The final instruction combines the original request with the added constraint 'only if they are not already friends', which is a clear addition that narrows the scope of the original task. The final audible intent matches the reference instruction with the added condition preserved. No other issues are present.", "heard_transcript": "Add uh all my friends and roommates as friends on Venmo, and make sure to add them only if they are not already friends." }, "candidate_id": "wav_91eda236961d50f4", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_deduplication_condition", "proposed_instruction": "Add all my friends and roommates as friends on venmo, if they are already. That should be 'if they are not already', not 'if they are already'.", "long_pause_offsets": [ 75 ], "source_sha256": "7a709374228f447d9205691da65e9eac43d7bf38c2efb330785b40868d06039a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'if they are already' to 'if they are not already', preserving the action, target group, and app." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/revision/revision_03_deduplication_condition/0e6f36cf6124ccbd/speech.wav", "wer": 0.0967741935483871, "wav_sha256": "142ec5f4ab1a4f9fb07ea774d1e5a39ab1afc7eda568729af0e39e16b7c9ddac", "seconds": 19.56, "spoken_utterances": [ "Add hmmm... all my friends and roommates as friends on venmo, um, let's see... if they are already. ", "That should be 'if they are not already', not 'if they are already'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately self-corrects to 'if they are not already' after a pause and explicit revision. The final audible intent matches the original instruction, so the revision criterion is satisfied.", "heard_transcript": "Add all my friends and roommates as friends on Venmo, um, let's see, if they are already. That should be, if they are not already. Not if they are already." }, "candidate_id": "wav_dc68891ba38842cb", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker exhibits a natural, multi-stage thought process with a clear false start ('At'), a thoughtful pause ('Hmm'), and a self-correction of the condition from 'if they are already' to 'if they are not already'. The intonation and pacing reflect genuine deliberation, and the correction is smoothly integrated into the ongoing utterance, making it highly realistic." } ] }, "is_default": true }, { "variant_id": "revision_02_app", "proposed_instruction": "Add all my friends and roommates as friends on cashapp, if they are not already. Change the app to Venmo, not Cash App.", "long_pause_offsets": [ 81 ], "source_sha256": "b1ecc4bc10760f1505612cff3b1c99f483151faf43f7f68aa7244918b59a6532", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Cash App to Venmo, preserving the action, target group, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/revision/revision_02_app/435c7654dcc29959/speech.wav", "wer": 0.2916666666666667, "wav_sha256": "edd3eadf1d31c498e4cc82cc21985a02ce762f43c6df3b99eeba44347dfc2618", "seconds": 12.4, "spoken_utterances": [ "Add all my friends and roommates as friends on cashapp, um... if they are not already. ", "Change the app to ... Venmo, not Cash App." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo, not Cash App' after a pause. This is a revision where the final instruction overrides the initial incorrect app name. The rest of the instruction, including the condition 'if they are not already', remains unchanged and is preserved in the final intent. The correction is clear and final, satisfying the revision type criterion.", "heard_transcript": "Add all my friends and roommates as friends on Cash App, um, if they are not already. Change the app to Venmo, not Cash App." }, "candidate_id": "wav_260bb413a56e909f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision where the speaker first says 'Cash App' and then corrects it to 'Venmo'. The correction is clear and the pause before it sounds like a moment of realization. However, the initial part of the sentence is delivered in a slightly more robotic, read-aloud manner compared to audio_03, and the correction is less integrated into a continuous thought flow, making it less spontaneous." } ] }, "is_default": false }, { "variant_id": "revision_01_contact_group", "proposed_instruction": "Add all my family as friends on venmo, if they are not already. Wait, the contact group should be 'friends and roommates', not 'family'.", "long_pause_offsets": [ 64 ], "source_sha256": "6536770ede9863c76c31cdbcce27c26d56680b17688fc1f558775fb43ab60acd", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision corrects the contact group from 'family' to 'friends and roommates', while preserving the app, action, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/revision/revision_01_contact_group/2cf5409b27d3d0d2/speech.wav", "wer": 0.08, "wav_sha256": "8cecf2468afbe2ce62438c7acd9eb36f292cc7d8dec62a2e4b631dcd51090dad", "seconds": 14.6, "spoken_utterances": [ "Add uh... all my family as friends on venmo, ... if they are not already. ", "Wait, the contact group should be '... friends and roommates', not 'family'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'all my family' but immediately corrects themselves to 'friends and roommates' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, so the task should be executed as revised.", "heard_transcript": "Add uh all my family as friends on Venmo, if they are not already. Wait, the contact group should be friends and roommates, not family." }, "candidate_id": "wav_69256773748ff7f4", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker starts with 'Add uh all my family as friends on Venmo' and then corrects to 'friends and roommates'. While it includes a self-correction, the initial 'uh' and the slightly flat delivery make it sound less natural. The correction is abrupt and lacks the organic flow seen in the other clips, suggesting a less spontaneous, more rehearsed delivery." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_028", "hide_variant_original": true, "source_sha256": "bc1e35177b87a9306d1d0815296610788b9ef9c8f49cafddb80d51801e0b5ca7", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/cancellation/cancel_028/168e0e8b18caff03/speech.wav", "wer": 0.16129032258064516, "wav_sha256": "c71d17fa1223a2037dee94d64051033a8b69f814fc3f014817b9f5d7ffb55d2a", "seconds": 16.92, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already.", "Hang on, disregard that instruction. Don't carry it out." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins the original task but then clearly revokes it with 'Hang on disregard that instruction don't carry it out.' This constitutes a direct cancellation of the entire task. The final intent is unambiguous and matches the cancellation type criterion.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already. Hang on disregard that instruction don't carry it out." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_006", "hide_variant_original": true, "source_sha256": "ef255626f04caf5774acf3b2513447110b2b82f3a71caadbf5e87aa78e630fb1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/ambiguous_halt/halt_006/062fd55da5865003/speech.wav", "wer": 0.16, "wav_sha256": "d73282d603c8595b72e6489ba51e97b6269a895aa00f38cd9278e9131344c111", "seconds": 15.16, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already.", "Just a moment..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'just a moment', which is a clear request to pause or wait. The speaker does not cancel the task or provide a new instruction. The initial request is still present but not completed, and the final interjection is a pause command. This matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already just a moment" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_1", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "7beee4c9012eb26ba680eb67e8ff0262e539bf726d52cd27425fbb0990816a57", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_1/audio_natural/filler/original/0f573c75a3fa3111/speech.wav", "wer": 0.18181818181818182, "wav_sha256": "2042639f35e8d1cb9bae8cddf719fcc70d8f2845aa62c9b82b1075d9b46812ca", "seconds": 11.8, "spoken_utterances": [ "Add uh... all my friends and roommates as friends um, let me think... on Venmo, let's see... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes fillers ('uh', 'um', 'let me think') and a brief pause but does not change the core request. The final audible intent matches the original instruction. The hesitation and backchanneling are typical of a filler type, and the task goal remains consistent.", "heard_transcript": "Add uh all my friends and roommates as friends um let me think on Venmo let's see if they are not already" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Add all my coworkers and friends as friends on cashapp—make that venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "434dc5550100a7c428212be6573841c6414b93a631b0c057c219fbb6f1aa5898", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'cashapp' to 'venmo' within the same utterance, preserving the action, target group, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/false_start/false_start_02_app/4fb31271172ca763/speech.wav", "wer": 0.21052631578947367, "wav_sha256": "6b4b678c96b9698e60d01f2cfa15121274a100c7497eada15dab729617e41426", "seconds": 9.48, "spoken_utterances": [ "Add all my coworkers and friends as friends on cashapp—ah... make that ... venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'Make that, Venmo'. This is a false start where the final value is the corrected one. The rest of the instruction matches the original, including the conditional 'if they are not already'. The final intent is clear and executable.", "heard_transcript": "Add all my coworkers and friends as friends on Cash App. Make that, Venmo, if they are not already." }, "candidate_id": "wav_3e360c404ccfd26c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Cash App', then self-corrects to 'Venmo' with a clear, spontaneous-sounding 'Make that' transition. The prosody is consistent, the pace is conversational, and the correction feels like a genuine moment of recall. The final phrase 'if they are not already' is delivered smoothly without hesitation, completing the thought naturally." } ] }, "is_default": true }, { "variant_id": "false_start_03_deduplication_condition", "proposed_instruction": "Add all my coworkers and friends as friends on venmo, if they are already—no, make that if they are not already.", "long_pause_offsets": [], "source_sha256": "a98d357b60c2c45ec3d6b3f7f694be76d354ecd9726800dd60b5b7720d6982b1", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately corrects it to 'if they are not already'. The final condition matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/false_start/false_start_03_deduplication_condition/f3b3e414464d9bea/speech.wav", "wer": 0.16, "wav_sha256": "6ca91900f8dbb799a926c231909e5c73301f2e00f70ce11927dddf428efb74fe", "seconds": 13.4, "spoken_utterances": [ "Add all hmmm... my coworkers and friends as friends on venmo, um... if they are already—ah... no, make that if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately self-corrects with 'no, make that if they are not already.' This is a clear false start where the final, corrected condition is 'if they are not already.' The final intent matches the original instruction, so the task should be executed as corrected.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are already, uh, no, make that if they are not already." }, "candidate_id": "wav_881b2147318f7b4e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip also contains a false-start repair, but the correction is less fluid. The speaker says 'if they are already' and then corrects to 'no, make that if they are not already'. The 'no' is abrupt and slightly unnatural, breaking the flow. The prosody shifts noticeably, and the correction feels more like a mechanical insertion than a spontaneous thought correction, reducing its realism compared to audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_01_contact_group", "proposed_instruction": "Add all my family and neighbors—I meant coworkers and friends as friends on venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "0f32be4617c3f605d9d3874f7f660a4e3c1d3ba99917f20d630757ade418f55e", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker immediately corrects 'family and neighbors' to 'coworkers and friends' within the same utterance, preserving the app, action, and conditional scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/false_start/false_start_01_contact_group/cdbb2ddac7474784/speech.wav", "wer": 0.125, "wav_sha256": "3105d1353e2a774f8207f4ff0beb3e847c831efcbdce2dc5b9fd61dd7d4d1ffb", "seconds": 9.64, "spoken_utterances": [ "Add all my family and neighbors—um... I meant coworkers and friends as friends on venmo, um, let's see... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family and neighbors' but immediately self-corrects to 'coworkers and friends' with 'I meant'. The correction is immediate and unambiguous, and the rest of the instruction ('as friends on Venmo, if they are not already') remains unchanged. This matches the 'false_start' type criterion, where the final intent is the corrected version.", "heard_transcript": "And all my family and neighbors, um, I meant coworkers and friends as friends on Venmo, um, let's see, if they are not already." }, "candidate_id": "wav_716f40d96e659e25", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio is the least realistic. It begins with a false start ('all my family and neighbors um'), but the speaker never completes the correction or continues the thought. Instead, they shift to a new, unrelated request ('I met coworkers and friends as friends on Venmo um let's see if they are not already'). The transition is jarring and lacks the natural flow of a self-correcting thought process. The 'um' fillers are placed awkwardly, and the overall delivery sounds disjointed and less like a single, coherent utterance." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Add all my coworkers and friends as friends on venmo,. Oh, I forgot to mention, only add them if they are not already friends.", "long_pause_offsets": [ 55 ], "source_sha256": "d90ed5aa2e636123367478dfd5ab708716c991767ec30da838f399fc390e6c48", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only if they are not already friends', which refines the initial broad request. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/addition/addition_01/dc55fc0b609ee287/speech.wav", "wer": 0.16, "wav_sha256": "0fd800ac4819e7ba225652fd9047d42db3d1d2d3aaef1c71fab1f8080b5c8c26", "seconds": 10.64, "spoken_utterances": [ "Add all ... my coworkers and friends as friends uh... on venmo,. ", "Oh, I forgot to mention, only add them if they are not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to add all coworkers and friends as friends on Venmo. After a pause, the speaker added the constraint 'only add them if they are not already friends.' This is a clear addition that narrows the scope of the original request. The final intent combines the original action with the new condition, matching the expected behavior for an 'addition' type. The constraint is explicitly stated and preserved.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo. Oh, I forgot to mention, only add them if they are not already friends." }, "candidate_id": "wav_0e1db0c1ddf6c33e", "naturalness": { "rank": 1, "mean_rank": null, "singleton": true, "reasons": [] }, "is_default": true } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_01_contact_group", "proposed_instruction": "Add all my family members as friends on venmo, if they are not already. Wait, I meant coworkers and friends, not family members.", "long_pause_offsets": [ 72 ], "source_sha256": "cc0d1bf02add5854c8fad1ff5ba0cfe1e1ab20198b39c4e55941290681b8be28", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the group from 'family members' to 'coworkers and friends', preserving the app, action, and conditional logic." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/revision/revision_01_contact_group/5c3b285964cbb478/speech.wav", "wer": 0.038461538461538464, "wav_sha256": "4d2bc5ec4497ff161c0794559f82670e36e0df7cf5ee1b8a9a1eb7ba191afbf3", "seconds": 12.52, "spoken_utterances": [ "Add all my family members as friends on venmo, um... if they are not already. ", "Wait, I meant let's see... coworkers and friends, not family members." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family members' but immediately corrects to 'coworkers and friends' after a pause and self-correction. According to the revision type criterion, the final corrected instruction overrides the initial one. The final intent matches the original reference instruction exactly, including the conditional 'if they are not already'. Therefore, the task should be executed as revised.", "heard_transcript": "Add all my family members as friends on Venmo, um, if they are not already. Wait, I meant, let's see, coworkers and friends, not family members." }, "candidate_id": "wav_5bb24c9b1cf400a3", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio displays the most natural and plausible revision flow. The speaker starts with a clear request, then after a brief pause and a self-correcting 'wait, I meant', smoothly revises the target group from 'family members' to 'co-workers and friends'. The intonation and pacing suggest genuine hesitation and correction, with a believable change-of-mind. The prosody is consistent and the delivery is continuous without robotic or disjointed segments." } ] }, "is_default": true }, { "variant_id": "revision_03_deduplication_condition", "proposed_instruction": "Add all my coworkers and friends as friends on venmo, if they are already. Change the condition to 'if they are not already', not 'if they are already'.", "long_pause_offsets": [ 75 ], "source_sha256": "8a9ace528b69816e024ce4a7a61b0450958d21a233dad075136b0335711a9e9b", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup explicitly corrects the condition from 'if they are already' to 'if they are not already', which matches the original instruction. The action, target group, and app remain unchanged." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/revision/revision_03_deduplication_condition/dbd6a874118e32cc/speech.wav", "wer": 0.11764705882352941, "wav_sha256": "2e9d82ae873acd1f983e20e5732edeeb08701255d3dc9d2ee9ae4d7546e96b4b", "seconds": 18.24, "spoken_utterances": [ "Add all uh... my coworkers and friends as friends on venmo, um, let me think... if they are already. ", "Change the condition to 'hmmm... if they are not already', not 'if they are already'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the condition 'if they are already' but then immediately self-corrects to 'if they are not already' after a pause and hesitation. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final audible intent matches the original reference instruction exactly.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo. Um, let me think. If they are already. Change the condition to, hmm, if they are not already. Not if they are already." }, "candidate_id": "wav_38a405559e49cb9f", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision but with less natural flow. The speaker begins with a request, then inserts a filler 'let me think' before revising the condition. While the revision is semantically clear, the phrasing 'change the condition to...' sounds slightly more like a meta-instruction than a spontaneous thought. The prosody is slightly less fluid, and the revision feels a bit more like a post-hoc correction than a natural thought process." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Add all my coworkers and friends as friends on CashApp, if they are not already. Make that on Venmo, not CashApp.", "long_pause_offsets": [ 81 ], "source_sha256": "423a750fe06ed2eb11c65cfe2e821e3160f4548409307135d211106cbc155c4f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from CashApp to Venmo, while retaining the list of people (coworkers and friends) and the conditional logic ('if they are not already')." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/revision/revision_02_app/fa0bc89b51cc789e/speech.wav", "wer": 0.36363636363636365, "wav_sha256": "3a859ade5d4fb986142d33583826f5abdfde78dc2b474f750ee400a9d2fa3a83", "seconds": 11.36, "spoken_utterances": [ "Add all um... my coworkers and friends as friends ... on CashApp, if they are not already. ", "Make that on Venmo, not CashApp." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' after a pause. According to the revision type criterion, the correction spoken after a longer pause overrides the earlier instruction. The final intent matches the original reference instruction exactly, with the correct app specified.", "heard_transcript": "Add all um my coworkers and friends as friends on Cash App if they are not already. Make that on Venmo, not Cash App." }, "candidate_id": "wav_14195850af4bdc2b", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a revision but with a less realistic delivery. The speaker starts with a request, then corrects 'Cash App' to 'Venmo'. However, the correction is abrupt and lacks the natural hesitation or self-repair seen in the other clips. The prosody is flatter, and the transition between the initial and corrected version feels more like a read-aloud correction than a spontaneous change of mind. The overall delivery is less conversational and more robotic." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_018", "hide_variant_original": true, "source_sha256": "fb70bd1d21f46a2bd2a2ee1b26ae46d2e7f78abd3e49a9c8acf440a4481fd3fb", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/cancellation/cancel_018/57ac4c9c206e0ca5/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "08497a6b3deef24e016a36c9f91568fb8d09ad11bd5234cb4c7dbef2f876e2c4", "seconds": 13.48, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already.", "Oh, wait, I'd rather not do that after all. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to add coworkers and friends on Venmo, then clearly revokes it by saying 'Oh wait, I'd rather not do that after all. Please cancel it.' This constitutes a direct cancellation of the entire task as per the 'cancellation' type criterion.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already. Oh wait, I'd rather not do that after all. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_033", "hide_variant_original": true, "source_sha256": "f4e6ce39cadd22f9f64565d7a52b0e8913bb6a58e988d9fe8bd3129405516b37", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/ambiguous_halt/halt_033/3eff73d1a736b74f/speech.wav", "wer": 0.2, "wav_sha256": "be4d1e8d89d97ef359db19d8f5c46f8e2ab2f5ea006b2a738709928dcee7edc9", "seconds": 11.6, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already.", "Hold on, there's something..." ] }, "audio_semantic": { "status": "equivalent", "reason": "The audio ends with 'Hold on, there's something', which is a clear request to pause or wait. There is no cancellation or instruction to resume. The initial task is not completed, but the final audible intent is to pause, which matches the 'ambiguous_halt' type criterion.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already. Hold on, there's something" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_2", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "6f85ffb8c3feea17ca36b5f63f4edddf8c52316b50e9ab4d690352fdd4caad3f", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_2/audio_natural/filler/original/789801a4ac67976a/speech.wav", "wer": 0.25, "wav_sha256": "a4b8d2ec1600975a20f87e5d2afbc3db67fa94603ace02c56de3050efd2b43d5", "seconds": 7.72, "spoken_utterances": [ "Add all ... my coworkers and friends as friends on Venmo, um... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker includes a filler 'um' but does not change the task or add any new conditions. The final intent matches the original instruction exactly, with no correction or modification. The filler does not alter the goal, so the task should be executed as stated.", "heard_transcript": "Add all my coworkers and friends as friends on Venmo, um, if they are not already." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "false_start", "criterion": "한 발화 안에서 잘못 말한 내용을 즉시 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "false_start_02_app", "proposed_instruction": "Add all my coworkers and roommates as friends on cashapp—sorry, make that venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "6e5e3c351f3e44a6f2d8790146db9a435a387825c63ac5f01cfe22eb50b56eda", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'cashapp' but immediately corrects to 'venmo'. The app is the only changed element, and the correction is clear and complete. All other conditions—target group (coworkers and roommates), action (add as friends), and condition (if not already)—are preserved." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/false_start/false_start_02_app/3dad92aa1ca4e834/speech.wav", "wer": 0.3, "wav_sha256": "1f012c2533f4f4b8221226d43a7b62d104d4a3973ff887609d569d8c2212bb3f", "seconds": 12.12, "spoken_utterances": [ "Add all ... my coworkers and roommates as friends on cashapp—hmmm... sorry, make that venmo, ... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately self-corrects to 'Venmo' with 'sorry, make that Venmo'. According to the 'false_start' type criterion, the final corrected value should be applied, and the original incorrect value should be discarded. The final intent matches the original instruction with the correct app name.", "heard_transcript": "Add all my coworkers and roommates as friends on Cash App. Um, sorry, make that Venmo, if they are not already." }, "candidate_id": "wav_c6dafdf0ec541e4c", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "This audio exhibits the most natural and fluent false-start repair. The speaker begins with 'Cash App', then self-corrects to 'Venmo' with a clear, spontaneous-sounding 'sorry, make that Venmo'. The intonation and pacing of the correction sound like a genuine moment of recall, and the overall delivery is continuous and conversational without awkward pauses or robotic phrasing." } ] }, "is_default": true }, { "variant_id": "false_start_03_deduplication_condition", "proposed_instruction": "Add all my coworkers and roommates as friends on venmo, if they are already—no, that should be if they are not already.", "long_pause_offsets": [], "source_sha256": "6cb00ee27b884f5e825c54db1186d1df92fa912bc71ad28930757b0d9d669551", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately corrects it to 'if they are not already'. The final condition matches the original instruction exactly." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/false_start/false_start_03_deduplication_condition/ab853ae19ba2f108/speech.wav", "wer": 0.14814814814814814, "wav_sha256": "fbfd7b8c3e7372e57716ac505f6cd9c5382eebdf725279144cd8a7e4f2f87e8a", "seconds": 13.48, "spoken_utterances": [ "Add all let me think... my coworkers and roommates as friends on venmo, if they are already—hmmm... no, that should be if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately self-corrects to 'if they are not already'. According to the 'false_start' type criterion, the final corrected instruction should be executed, and the initial mistake should be discarded. The final intent matches the original instruction exactly.", "heard_transcript": "Add all my coworkers and roommates as friends on Venmo, if they are already. No, that should be if they are not already." }, "candidate_id": "wav_461ff3d6fd34583d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This clip shows a revision where the speaker initially says 'if they are already' and then corrects to 'if they are not already'. The correction is clear and the intonation reflects a believable change of mind. However, the initial phrase 'add all let me think' is slightly less natural than audio_03's opening, and the overall flow is a bit more hesitant, making it less realistic than audio_03." } ] }, "is_default": false }, { "variant_id": "false_start_01_contact_group", "proposed_instruction": "Add all my classmates and neighbors—I should have said coworkers and roommates as friends on venmo, if they are not already.", "long_pause_offsets": [], "source_sha256": "78bf0ece0b3442f8d2e2c016e26fda6723f53a5f8bcb3cc88346fd7d3fc88f69", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The speaker explicitly corrects 'classmates and neighbors' to 'coworkers and roommates' within the same utterance, preserving the app, action, and conditional scope." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/false_start/false_start_01_contact_group/91eb7460c5b16803/speech.wav", "wer": 0.16, "wav_sha256": "c5f2f22be597916a073bafcec478769515ae061bf8db08151cd2a6546e99c5d2", "seconds": 12.68, "spoken_utterances": [ "Add all uh... my classmates and neighbors—hmmm... I should have said coworkers and roommates as friends on venmo, uh... if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'classmates and neighbors' but immediately self-corrects to 'coworkers and roommates'. According to the 'false_start' type criterion, the final intent is the corrected version, and the original mistaken value is discarded. The final instruction matches the original reference exactly, including the conditional 'if they are not already'. Therefore, the task is executed as intended.", "heard_transcript": "Add all uh my classmates and neighbors. Hmm. I should have said coworkers and roommates as friends on Venmo uh if they are not already." }, "candidate_id": "wav_09fc9e02245b1202", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "This audio contains a false start where the speaker says 'classmates and neighbors' and then corrects to 'co-workers and roommates'. While the repair is present, the initial phrase sounds less plausible as a spontaneous thought, and the correction is delivered with a slightly unnatural, clipped intonation. The 'uh' and 'hm' fillers are present but feel less integrated into the flow compared to the other two clips." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "addition", "criterion": "처음 빠진 정보를 후속 발화로 추가하고, 최종 목표는 원문과 같아야 합니다. 기존 조건을 바꾸는 revision과 구분하세요.", "status": "supported", "variants": [ { "variant_id": "addition_01", "proposed_instruction": "Add all my coworkers and roommates as friends on venmo,. Oh, I forgot to mention, only add them if they're not already friends.", "long_pause_offsets": [ 57 ], "source_sha256": "d1f9e6cca5c59d9db4240e59534771ddacfd2933fb57d5201a1bf786045e528c", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing condition 'only if they're not already friends', which refines the initial instruction without contradicting it. The action, target group, and app remain consistent." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/addition/addition_01/2f2aa619e0910de4/speech.wav", "wer": 0.125, "wav_sha256": "4c53da8405749a3292adcce71d95015bee3dda716aca2927bf74cb78e5b52003", "seconds": 12.64, "spoken_utterances": [ "Add all um... my coworkers and roommates as friends ... on venmo,. ", "Oh, I forgot to mention, only add them if they're not already friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The original instruction was to add coworkers and roommates as friends on Venmo with the condition 'if they are not already'. The speaker initially omitted this condition, then added it as an afterthought. According to the 'addition' type criterion, the final goal is the combination of the initial instruction and the added constraint. The final audible intent matches the original instruction exactly, including the conditional clause.", "heard_transcript": "Add all um my coworkers and roommates as friends on Venmo. Oh, I forgot to mention, only add them if they're not already friends." }, "candidate_id": "wav_fafe96a18aa63018", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a natural, continuous request with a plausible self-correction. The initial request is clear, followed by a believable 'Oh, I forgot to mention' filler that introduces a revision. The prosody is conversational, with appropriate pauses and intonation, and the correction flows smoothly without sounding rehearsed or robotic. The addition of the condition feels like a spontaneous afterthought." } ] }, "is_default": true }, { "variant_id": "addition_02", "proposed_instruction": "Add all as friends on venmo, if they are not already. One more detail: add all my coworkers and roommates.", "long_pause_offsets": [ 54 ], "source_sha256": "63b1e89088ca8848cd7736d92ef63118db362706a15353369590afa2d33607a9", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup adds the missing group specification 'my coworkers and roommates' to the initial broad 'all' request, refining the scope without contradicting the original condition." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/addition/addition_02/7c07ab396d468ff9/speech.wav", "wer": 0.0, "wav_sha256": "7daccfd1be3fc600bb89c312a681c418330c1c547b774120458297d3d0aca1c6", "seconds": 11.44, "spoken_utterances": [ "Add all as friends ... on venmo, if they are not already. ", "One more detail: add all my ... coworkers and roommates." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with the original instruction and then adds a clarifying detail that repeats the same group ('coworkers and roommates'). Since the addition does not change the scope or intent and is consistent with the original, the final goal remains to add all coworkers and roommates as friends on Venmo if they are not already. The addition is redundant but not contradictory, so the task is executed as intended.", "heard_transcript": "Add all my coworkers and roommates as friends on Venmo, if they are not already. One more detail, add all my coworkers and roommates." }, "candidate_id": "wav_cfb89bdeaee16c1d", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "This audio is less realistic due to its unnatural structure. The speaker first makes a complete request, then adds a new detail ('One more detail') before continuing. However, the second part ('Add all my coworkers and roommates') is delivered as a separate, abrupt statement rather than a natural continuation. The transition lacks the conversational flow seen in audio_02, and the overall delivery sounds more like a scripted addition than a spontaneous thought." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "revision", "criterion": "완결된 초기 지시 뒤에 잘못된 값을 정정하고, 최종 목표는 원문과 같아야 합니다.", "status": "supported", "variants": [ { "variant_id": "revision_03_deduplication_condition", "proposed_instruction": "Add all my coworkers and roommates as friends on venmo, if they are already. Make that 'if they are not already' instead of 'if they are already'.", "long_pause_offsets": [ 77 ], "source_sha256": "c92fb20fbf81c55299fa4c0d9fd0b79858033cbb3d2f15919d2ea40443dbf42a", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the condition from 'if they are already' to 'if they are not already', preserving the action, target group, and app. The original instruction's condition is fully corrected." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/revision/revision_03_deduplication_condition/256e7a54428fce65/speech.wav", "wer": 0.2, "wav_sha256": "667a8d44697557979eb3a9a0084d05ad159505f27c6db07bc7ca478478cbe77f", "seconds": 19.84, "spoken_utterances": [ "Add all um... my coworkers and roommates as friends uh... on venmo, hmmm... if they are already. ", "Make that 'if they are not already' instead of 'if they are already'." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'if they are already' but immediately corrects themselves to 'if they are not already' after a pause. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final audible intent is to add all coworkers and roommates as friends on Venmo only if they are not already friends, which matches the original instruction. The correction is clear and final, so the task is executed as revised.", "heard_transcript": "Add all um my coworkers and roommates as friends uh on Venmo who if they are already. Make that if they are not already instead of if they are already." }, "candidate_id": "wav_328cbe35715187c0", "naturalness": { "rank": 1, "mean_rank": 1.0, "singleton": false, "reasons": [ { "round": 1, "rank": 1, "reason": "The speaker delivers a clear, continuous request with natural hesitation ('um', 'uh') and a self-corrected phrase ('if they are not already instead of if they are already'). The prosody is consistent, the correction is smoothly integrated, and the pacing reflects genuine thought processing. The speaker sounds like a real person formulating and refining a task." } ] }, "is_default": true }, { "variant_id": "revision_01_contact_group", "proposed_instruction": "Add all my family and friends as friends on venmo, if they are not already. Wait, the group should be coworkers and roommates, not family and friends.", "long_pause_offsets": [ 76 ], "source_sha256": "8a898ae6c5789a113f6079719f657a75f2280b231227e556fd80b7baffe865f5", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup revision correctly replaces the group from 'family and friends' to 'coworkers and roommates', while preserving the app, action, and conditional clause." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/revision/revision_01_contact_group/db64d864cc021de7/speech.wav", "wer": 0.14285714285714285, "wav_sha256": "6c4dfaee4bacac0d40e565a7a1eb72cdd23db12fe05d15829bec2f5634dc791b", "seconds": 16.92, "spoken_utterances": [ "Add all ... my family and friends as friends ... on venmo, um... if they are not already. ", "Wait, the group should be coworkers and roommates, not family and friends." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'family and friends' but immediately corrects themselves to 'coworkers and roommates' after a pause and self-correction. According to the revision type criterion, the final corrected instruction should be applied. The final intent matches the original reference instruction exactly, so the task should be executed.", "heard_transcript": "Add all my family and friends as friends on Venmo, um, if they are not already. Wait, the group should be coworkers and roommates, not family and friends." }, "candidate_id": "wav_7debc256c18e846e", "naturalness": { "rank": 2, "mean_rank": 2.0, "singleton": false, "reasons": [ { "round": 1, "rank": 2, "reason": "The speaker starts with a clear request, then corrects the group type ('family and friends' to 'co-workers and roommates'). The correction is delivered with a natural 'wait' and a change in intonation, indicating a genuine realization. The prosody is mostly natural, though the correction is slightly abrupt compared to audio_01, making it feel a bit less fluid." } ] }, "is_default": false }, { "variant_id": "revision_02_app", "proposed_instruction": "Add all my coworkers and roommates as friends on cashapp, if they are not already. Change the app to Venmo, not Cash App.", "long_pause_offsets": [ 83 ], "source_sha256": "641c6211fb63107054e601b3c7409961dc9ef5e254b9a1c0626251c2d1ff6531", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "equivalent", "reason": "The followup corrects the app from Cash App to Venmo, preserving the target group (coworkers and roommates) and the conditional logic (if not already friends)." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/revision/revision_02_app/12e4f36d0223d027/speech.wav", "wer": 0.28, "wav_sha256": "26164b44969ac38accf710fb2133591e2a4ff07e7c21b2787385d0eccff7c54a", "seconds": 16.92, "spoken_utterances": [ "Add all my coworkers and roommates as friends ... on cashapp, hmmm... if they are not already. ", "Change the app to uh... Venmo, not Cash App." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially says 'Cash App' but immediately corrects to 'Venmo' after a pause and self-correction. According to the revision type criterion, the correction spoken after a longer pause must be applied to the earlier instruction. The final intent is to perform the action on Venmo, which matches the original reference instruction. All other conditions (adding coworkers and roommates, if not already friends) are preserved.", "heard_transcript": "Add all my coworkers and roommates as friends on Cash App, if they are not already. Change the app to uh, Venmo, not Cash App." }, "candidate_id": "wav_b04bd7861be7880a", "naturalness": { "rank": 3, "mean_rank": 3.0, "singleton": false, "reasons": [ { "round": 1, "rank": 3, "reason": "The speaker begins with a clear request but then makes a correction ('Cash App' to 'Venmo'). However, the correction is delivered with a noticeable pause and a hesitant 'uh' that sounds less spontaneous and more like a rehearsed edit. The overall delivery feels slightly less natural, possibly due to the awkward phrasing of the correction and the less seamless integration of the change." } ] }, "is_default": false } ], "reason": "", "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "cancellation", "criterion": "원래 작업 전체를 명확하게 취소하고 추가 실행을 중단하라는 뜻이면 Yes. 이미 한 작업의 복구는 요구하지 않습니다.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "cancel_040", "hide_variant_original": true, "source_sha256": "ecfdcf199f7dc2378a5bb3ff0b73a5ef02820e6c54220713e48573b381f4a0d3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/cancellation/cancel_040/be77c986b04c71f0/speech.wav", "wer": 0.1, "wav_sha256": "b8fdc81e795a58db41827950ab929102aaf0003cee2256de00fb90503d60a78b", "seconds": 16.2, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already.", "Oh, wait, I don't want to proceed with that task. Please cancel it." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker initially states the task to add coworkers and roommates as friends on Venmo. After a brief pause, they explicitly state, 'Oh wait, I don't want to proceed with that task. Please cancel it.' This constitutes a clear cancellation of the entire task. The final control intent is to cancel, which matches the 'cancellation' type criterion.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Oh wait, I don't want to proceed with that task. Please cancel it." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "ambiguous_halt", "criterion": "일단 멈추고 다음 지시를 기다리라는 뜻이면 Yes. 영구 취소·자동 재개·새 목표를 명시하면 No.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "halt_018", "hide_variant_original": true, "source_sha256": "92dde09775c597869782261128dbbd9a0d022e52964e24c5e51086fd24beb8b3", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원래 목표 수행이 아니라 취소/대기 동작을 검수하는 유형입니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/ambiguous_halt/halt_018/e19d89c4c12ca8ed/speech.wav", "wer": 0.13043478260869565, "wav_sha256": "81ae2cf2bb4cf5c8c6ec3042c448937a3ee0ec6f2ff375a6e7ad24dcf8e27193", "seconds": 13.6, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already.", "Could you pause for a second?" ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker finishes the original instruction and then clearly requests a pause with 'Could you pause for a second?'. This constitutes a valid ambiguous_halt as the final interjection is a direct request to hold/wait, with no cancellation or resumption permission. The original task intent is not revoked, but execution is requested to be paused.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Could you pause for a second?" }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" }, { "task_id": "ff58e36_3", "split": "test", "kind": "filler", "criterion": "망설임이나 filler가 들어가도 원문과 같은 작업 목표를 전달하면 Yes.", "status": "supported", "reason": "이 sample에 배정된 TTS 생성 계획이 아직 없습니다.", "variants": [ { "variant_id": "original", "hide_variant_original": true, "source_sha256": "1e4b91b55ab9224c840b103b659bedd36eed064bcc8e7986e4be74477931c850", "decision": null, "review_revision": null, "stale_review": false, "semantic": { "status": "not_applicable", "reason": "원문 기반 filler의 의미는 음성으로 검수합니다." }, "audio": { "status": "ready", "path": "/home/hbcho/agent/appworld/data/tasks_llm/ff58e36_3/audio_natural/filler/original/5d60f954d7092a19/speech.wav", "wer": 0.17647058823529413, "wav_sha256": "3fdce475dcf64ae4aaebaa36dda6d254f631c24678b405f1f8dc34b5a913a4f9", "seconds": 9.52, "spoken_utterances": [ "Add all uh... my coworkers and roommates as friends hmmm... on Venmo, if they are not already." ] }, "audio_semantic": { "status": "equivalent", "reason": "The speaker begins with a filler 'uh' and a pause ('Hmm'), but the final, complete instruction matches the original reference. The hesitation and backchannel do not alter the goal. The task is executed as intended.", "heard_transcript": "Add all uh my coworkers and roommates as friends on Venmo if they are not already. Hmm." }, "naturalness": { "rank": 1, "singleton": true }, "candidate_id": null, "is_default": true } ], "selection_status": "selected" } ], "splits": { "07b42fd_1": "train", "07b42fd_2": "train", "07b42fd_3": "train", "23cf851_1": "train", "23cf851_2": "train", "287e338_1": "train", "287e338_2": "train", "287e338_3": "train", "37a8675_1": "train", "37a8675_2": "train", "37a8675_3": "train", "3ab5b8b_1": "train", "3ab5b8b_2": "train", "3ab5b8b_3": "train", "57c3486_1": "train", "57c3486_2": "train", "57c3486_3": "train", "6bdbc26_1": "train", "6bdbc26_2": "train", "6bdbc26_3": "train", "6ea6792_1": "train", "6ea6792_2": "train", "6ea6792_3": "train", "771d8fc_1": "train", "771d8fc_2": "train", "771d8fc_3": "train", "82e2fac_1": "train", "82e2fac_2": "train", "82e2fac_3": "train", "b7a9ee9_1": "train", "b7a9ee9_2": "train", "b7a9ee9_3": "train", "c901732_1": "train", "c901732_2": "train", "c901732_3": "train", "cf6abd2_2": "train", "cf6abd2_3": "train", "d4e9306_1": "train", "d4e9306_2": "train", "d4e9306_3": "train", "df61dc5_1": "train", "df61dc5_2": "train", "df61dc5_3": "train", "e3d6c94_3": "train", "e7a10f8_1": "train", "e7a10f8_2": "train", "e7a10f8_3": "train", "e85d92a_1": "train", "e85d92a_2": "train", "e85d92a_3": "train", "23cf851_3": "train", "cf6abd2_1": "train", "e3d6c94_1": "train", "e3d6c94_2": "train", "024c982_1": "test", "07bb666_1": "test", "07bb666_2": "test", "07bb666_3": "test", "099174d_1": "test", "099174d_2": "test", "099174d_3": "test", "0d22252_1": "test", "0d22252_2": "test", "0d22252_3": "test", "1150ed6_1": "test", "1150ed6_2": "test", "1150ed6_3": "test", "13547f5_1": "test", "13547f5_3": "test", "1a79e37_1": "test", "1a79e37_2": "test", "1a79e37_3": "test", "1c4bd27_1": "test", "1c4bd27_2": "test", "1c4bd27_3": "test", "21abae1_1": "test", "21abae1_2": "test", "277d81d_1": "test", "277d81d_2": "test", "277d81d_3": "test", "2c544f9_1": "test", "2c544f9_2": "test", "2c544f9_3": "test", "31dc501_1": "test", "31dc501_2": "test", "31dc501_3": "test", "365e0a3_1": "test", "365e0a3_2": "test", "365e0a3_3": "test", "3f3c139_1": "test", "3f3c139_2": "test", "3f3c139_3": "test", "425a494_1": "test", "425a494_2": "test", "425a494_3": "test", "476b213_1": "test", "476b213_2": "test", "476b213_3": "test", "4d12842_1": "test", "4d12842_2": "test", "4d12842_3": "test", "522e5e5_1": "test", "522e5e5_2": "test", "522e5e5_3": "test", "5238afc_1": "test", "5238afc_2": "test", "5238afc_3": "test", "552869a_1": "test", "552869a_2": "test", "552869a_3": "test", "5a83b05_1": "test", "5a83b05_2": "test", "5a83b05_3": "test", "5e27cd7_1": "test", "5e27cd7_2": "test", "5e27cd7_3": "test", "6a5e690_1": "test", "6a5e690_2": "test", "6a5e690_3": "test", "7574325_1": "test", "7574325_2": "test", "7574325_3": "test", "77bcb81_1": "test", "77bcb81_2": "test", "77bcb81_3": "test", "7847649_1": "test", "7847649_3": "test", "7b97e6d_1": "test", "7b97e6d_2": "test", "7b97e6d_3": "test", "7e1be84_1": "test", "7e1be84_2": "test", "7e1be84_3": "test", "8749218_1": "test", "8749218_2": "test", "8749218_3": "test", "9871968_1": "test", "9871968_2": "test", "9871968_3": "test", "98d2608_1": "test", "98d2608_2": "test", "98d2608_3": "test", "9bf2c8a_1": "test", "9bf2c8a_2": "test", "9bf2c8a_3": "test", "c77c005_2": "test", "c77c005_3": "test", "cb4ee7a_1": "test", "cb4ee7a_2": "test", "cb4ee7a_3": "test", "cef9191_1": "test", "cef9191_2": "test", "cef9191_3": "test", "d18139b_1": "test", "d18139b_2": "test", "d18139b_3": "test", "dac78d9_1": "test", "dac78d9_2": "test", "dac78d9_3": "test", "e775c78_1": "test", "e775c78_2": "test", "e775c78_3": "test", "eb5ad85_1": "test", "eb5ad85_2": "test", "eb5ad85_3": "test", "f691597_1": "test", "f691597_2": "test", "f691597_3": "test", "ff58e36_1": "test", "ff58e36_2": "test", "ff58e36_3": "test", "32616b5_1": "test", "3aa1a22_1": "test", "3aa1a22_2": "test", "3aa1a22_3": "test", "986aa4e_1": "test", "986aa4e_2": "test", "986aa4e_3": "test", "988af8e_1": "test", "988af8e_2": "test", "988af8e_3": "test", "024c982_2": "test", "024c982_3": "test", "13547f5_2": "test", "21abae1_3": "test", "32616b5_2": "test", "32616b5_3": "test", "7847649_2": "test", "c77c005_1": "test" }, "review_code_sha256": { "review_store.py": "873e2a2f0afd532bf9ad0119934fb5607429dcb8df604bd018812f88d7c1d2e5", "app.py": "7b2ab7cded543a76e0b78de0bedb41498f85c43e81eba33ac7e91f1d1b9eaa80", "test_review_store.py": "27c3ac28bdec53fe0f09bfa9aadb90811af593ff31fa69eb27d8490bc258484e", "editor.py": "82d07eb5ca5e036ef832c9716a12b3961324a853fe7c48d7f834200ebce76665", "controller.py": "5546003c301440c0c7acfcff124b588f7f2df9646716dfe55102cb9774b40a88", "variant_review.py": "801f13ad8ea2d91589efa8a4c6daa696fa2aca9a5872de156bb1bcc520ea9915", "test_variant_review.py": "87fe82010f2ff0f451f59eea41d549f6e70c105c111ffd9888b911c18bc38e03", "results.py": "1f5936632cb4bf11163c420027c4ef163b4c0cc839b51bc753b5227d5352659e", "test_results.py": "f10e693fbeaafa92808c8bd7f80d0958f9052996e75e6d5430e010f6808e2b0d", "test_candidate_selection.py": "cbaeab118e1b050dfe833b08eb8c7277e0686e4bebb95302407a21f1f2080edf" } }