# File format and timing reference All JSON is UTF-8. `manifest.jsonl` has one JSON object per audio file. It is the canonical audio index. JSON object order is not meaningful. All consumer-facing `audio` and `metadata` paths, including those in `variants.json`, are relative to the dataset root, not to the sample folder. ## Manifest fields | Field | Meaning | |---|---| | `task_id` | Sample identifier, e.g. `024c982_1`. | | `family_id` | Shared task template identifier, e.g. `024c982`; `_1`, `_2`, `_3` are siblings. | | `split` | `train` or `test`; siblings are kept together. | | `type` | Export name: `clean`, `filler`, `false-start`, `addition`, `revision`, `cancel`, `halt`. | | `source_type` | Generation name: `clean`, `filler`, `false_start`, `addition`, `revision`, `cancellation`, `ambiguous_halt`. | | `variant_id` | Original selected variant identifier. | | `audio`, `metadata` | Dataset-relative paths to the WAV and its JSON metadata. | | `wav_sha256` | SHA-256 of the complete WAV file. | | `audio_info` | `sample_rate` (Hz), `frames`, `channels`, `sample_width_bytes`, `duration_seconds`. | | `expected_control_intent` | `execute` for clean and the first four perturbation types, `cancel` for cancel, `wait` for halt. | | `speaker_id` | Identifier recorded by the TTS generator; not a verified real-person identity. | ## Sample files `specs.json` is the authoritative task instruction and task setup. `annotation.json` annotates that instruction. Its character positions use Python Unicode string indices, zero based, with end offsets exclusive; they are not byte offsets or audio times. `pause_candidates`/`correction_candidates` refer to `instruction`; `spoken_pause_candidates` refer to normalized `spoken_text`. `dbs/` contains task database snapshots. `ground_truth/` contains evaluation code and answers. Use the task's AppWorld-compatible environment to execute/evaluate it; this export does not bundle that runtime. Ground-truth files are evaluation material and should not be exposed as agent input. Some inherited database rows retain historical instruction text: use `specs.json`, not such rows, as the instruction authority. `variants.json` has `available` (type → manifest record), `missing_types`. Absence means unavailable or ineligible under the selection rules; do not substitute a different type or silently synthesize replacement audio. `reports/missing.json` lists each skipped sample/type combination. There are no empty sample directories in this export. ## Audio metadata Each `metadata/.json` has schema version `1.0`: | Field | Meaning | |---|---| | `task_id`, `type`, `source_type` | Identity and type mapping. | | `audio_info` | Same physical audio properties as the manifest. | | `source_wav` | Historical absolute source path, for provenance only. | | `llm_selection` | Exact selected UI candidate, with text/audio judgments, naturalness rank/reasons and any recorded human-review status. Human-review status did not determine this export. | | `source_metadata` | Complete generation metadata: spoken utterances, speaker, settings, source dependencies, pause reports, ASR/WER, long boundaries and verification flags. | | `source_documents` | Embedded dependency documents keyed by source filename; each has original path, SHA-256 and JSON document. These include the source annotation/variant or shared sentence library as applicable. | | `inserted_pause_intervals` | Added silence intervals mapped into final exported WAV coordinates. | | `timing_notes` | Coordinate and alignment limitations. | Find the actual TTS input in `source_metadata.spoken_utterances`. It can differ from the normalized variant text because fillers were inserted. ASR transcripts are model outputs and are not guaranteed word-for-word references. `final_content_wer` is a generation check, not an independent task-success or listening-quality score. Correction spans and changed initial/final values are retained in the embedded source variant document; match `variant_id` within that document's `variants` array. Shared cancel/halt documents may contain unused library sentences; only the selected spoken utterances correspond to this WAV. LLM `equivalent` is an automatic judgment, not a human approval or proof of successful task execution. Other perturbation types may also contain filler words or hesitations: type labels identify the primary perturbation, not mutually exclusive acoustic features. ## Filler and pause coordinates There are three different coordinate systems: 1. Annotation character offsets locate candidate boundaries in the specified text. 2. Original ASR word times in `source_metadata.utterance_reports` refer to source utterances. Editing inserts silence, so do not use these timestamps directly as final-WAV word alignment. 3. `inserted_pause_intervals` uses final WAV frames and seconds. Frames are zero based; end frames are exclusive. Divide frame indices by `audio_info.sample_rate` to obtain seconds. For an inserted short pause, `utterance_index`, `event_id`, `filler`, `source_before`, `start_frame`, `end_frame`, `start_seconds` and `end_seconds` identify the edited event. `filler: null` means no lexical filler is attached to that event; the silence is still a pause. Long boundaries have `kind: explicit_long_pause` and use the same final frame/second coordinates. Timeline order need not be chronological; sort by `start_frame` for playback overlays. The complete original event is in `source_metadata.utterance_reports[utterance_index].timeline`, matched by `event_id`. This includes neighboring recognized words, acoustic search details, insertion positions and verification flags. Skipped edits remain in the original report. Short-pause output coordinates there are local to the edited utterance. The first utterance starts at frame zero; subsequent utterances start at the preceding `long_boundaries` entry's `end_frame`. **An inserted silence interval is not a filler speech interval.** Filler labels and neighboring ASR anchors preserve the estimated location, but this dataset does not supply human-verified exact start/end times for every spoken filler. See each original `human_listening_verified` and `boundary_acoustically_verified` flag. No listening or forced-alignment validation was added during export. ## Provenance and validation reports - `selection_snapshot.json`: all UI candidate rows, selected defaults, skipped types, original split assignments and hashes of the review backend code. - `task_source_overrides.json`: six existing TTS-aligned task overrides, with before/after documents and hashes. Three also synchronize the expected password answer. All other task assets come from `tasks`. - `missing.json`, `excluded_samples.json`: skipped types and samples with no selected audio; the latter is empty in this export. - `rebuild_appworld_vox.py`: historical export script. It requires the original internal source tree and review backend and is not a standalone TTS/regeneration pipeline. - `file_inventory.json`, `inventory_policy.json`: dataset-relative file hashes and explicit exclusions. The inventory excludes itself and changing validation/lifecycle reports to avoid self-referential hashes. - `validation.json`: original full export verification. `documentation_audit.json`: later structural/documentation verification using the portable tool. Neither certifies human listening quality or execution success. - `lifecycle.json`: replacement verification and requested deletion of the previous backup. This directory is sufficient to read and validate the exported audio and embedded metadata without the original absolute paths. Recreating LLM judgments, TTS audio or running AppWorld task evaluation requires external runtime/model dependencies. License and citation information were not supplied as part of this export; no new license is assigned here. ## Clean baseline addition All 189 samples include clean audio from the later `audio_natural/clean` generation, with the same per-sample TTS speaker as the perturbation variants. No generated fillers or inserted silence edits are present; natural pauses can still occur. Clean `llm_selection` is null and `import_selection` identifies the user-requested singleton import and its verification scope. ASR/WER and generation quality flags are retained without inventing an LLM semantic pass. `reports/clean_import.json` records the current full export verification; older audit reports describe earlier export stages.