diff --git a/.gitattributes b/.gitattributes index 82888762edb0660c1445b5cd012515a59294cd71..a2fd25644a0af20f552c9604d7bfd9affb6f1aab 100644 --- a/.gitattributes +++ b/.gitattributes @@ -36,3 +36,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text orukeet-v0.1.0.nemo filter=lfs diff=lfs merge=lfs -text orukeet-v0.1.0-q8.gguf filter=lfs diff=lfs merge=lfs -text orukeet-v0.1.0-f16.gguf filter=lfs diff=lfs merge=lfs -text +orukeet-technical-report.pdf filter=lfs diff=lfs merge=lfs -text +report/assets/kernel-atlas-preview.png filter=lfs diff=lfs merge=lfs -text +report/assets/kernel-atlas.png filter=lfs diff=lfs merge=lfs -text +report/assets/unseen/unseen-english-accents.png filter=lfs diff=lfs merge=lfs -text +report/assets/unseen/unseen-english-domains.png filter=lfs diff=lfs merge=lfs -text +report/assets/unseen/unseen-eurospeech.png filter=lfs diff=lfs merge=lfs -text +report/assets/unseen/unseen-language-coverage.png filter=lfs diff=lfs merge=lfs -text diff --git a/ARTIFACTS.json b/ARTIFACTS.json new file mode 100644 index 0000000000000000000000000000000000000000..534e885a4a2abc25637194000e0e72bf2736e48e --- /dev/null +++ b/ARTIFACTS.json @@ -0,0 +1,617 @@ +{ + "model": "Orukeet", + "version": "0.1.0", + "selection": "r3", + "repo_id": "oruk/orukeet", + "weight_revision": "555136b50265a132d4cea0d35560c26fc4f657ab", + "files": { + "orukeet-0.1.0.tar.gz": { + "bytes": 31074, + "sha256": "65f57199c411bd266dfabf2800957ebf56c65b25de038f6930ce8d89f9c71929" + }, + "orukeet-v0.1.0-f16.gguf": { + "bytes": 1296681088, + "sha256": "de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194" + }, + "orukeet-v0.1.0-q8.gguf": { + "bytes": 714456704, + "sha256": "93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd" + }, + "orukeet-v0.1.0.nemo": { + "bytes": 2509342720, + "sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "orukeet-technical-report.pdf": { + "bytes": 275659, + "sha256": "401d162a87d4454b0f16998bf3501b9c84da909a7b5be353705528477d2eef62" + }, + "selection-profile.png": { + "bytes": 85067, + "sha256": "6202432a630ef22fb54bf57c494437dcb8f526c26a9a7b6ecf1bcbdf2510a75d" + }, + "oruk-lockup.png": { + "bytes": 38483, + "sha256": "932df9164b237a1a31d1e636897d1766a15b8bac738acf81ee93de8ea2d6c18c" + }, + "LICENSE": { + "bytes": 1094, + "sha256": "e7349d7fc331c670e274335beaaa843d6d9bbeeb12691f1f40144c6214459bad" + }, + "CITATION.cff": { + "bytes": 1388, + "sha256": "9dd2b4e1fd368a4254768e72fe0a99b2e29d53947b88ad260e8050861373e3a8" + }, + "kernel-fits.png": { + "bytes": 93530, + "sha256": "53daf3dc1431660b7696c83d94e56543f8236563a81a4dd9fb75e3cab997fa52" + }, + "LICENSE-WEIGHTS": { + "bytes": 20137, + "sha256": "23ee78c8bae49cf08ea2f0c84945c66b987ebe4520881fb51b3dad4fb43d07c2" + }, + "README.md": { + "bytes": 8914, + "sha256": "5c709d29daa057ff9178eae95287365ef767adf43471e67d53eece60ef5e9222" + }, + "CITATION.bib": { + "bytes": 562, + "sha256": "c4f707d292a1218fc296f37559fd9388179d29e5ec224466aa05c3487db5785a" + }, + "transcribe.py": { + "bytes": 848, + "sha256": "bd754b17e85372d8cf4d15205813ab79a5ccba8673d57e647bcd914372788a23" + }, + "catalog.json": { + "bytes": 975, + "sha256": "41afcdb973226f4c8786b476907dffe993e5db39d51f1a2c11716d56d661cdc1" + }, + "orukeet-arxiv-source.zip": { + "bytes": 143407, + "sha256": "f4ad7770293f282b3e1fdd559a5a417035b799d58281e5bf9a6ab925bae51125" + }, + "team.json": { + "bytes": 3212, + "sha256": "79121aec28c6cad5b6e0a89784f8249a2f8f3d94e6352075a05245c24a92c7dc" + }, + "model-stages.json": { + "bytes": 20244, + "sha256": "a897798a15fcc7ea491842e03f7c4858ff20db1b29ccc9f85c928659b8b4b74c" + }, + "NOTICE.md": { + "bytes": 3906, + "sha256": "c1c90ecec7c9ff8056aada8979bf4f3f4dba901c54297c14a328815160e20c6e" + }, + "affiliations/cambridge.png": { + "bytes": 72015, + "sha256": "a3faf2131e9e2df035d4b21d076c0eea606669c7fb46283b7c90041880a632fd" + }, + "affiliations/SOURCES.md": { + "bytes": 1262, + "sha256": "62ce911ce80d11cc1e13eea7df144d01887721e20db87e5615539c2c26cafca9" + }, + "affiliations/oruk.png": { + "bytes": 33618, + "sha256": "717fb1a29a3faf8cf62baa924c65e77afb8fb82180e7398612a01a672af014da" + }, + "affiliations/openwhispr.png": { + "bytes": 15637, + "sha256": "564f133d0c0b5c87cd6384f9ce2689fffa6da68b004f303d2e278322e213ac62" + }, + "affiliations/stanford.png": { + "bytes": 16084, + "sha256": "6521b4f360f081fb9322b05ff4740fe8ad2b7c720ee1a9e3f7b9935624f4abc2" + }, + "affiliations/hoid.png": { + "bytes": 29339, + "sha256": "b5a252792faf5ce80c3f1aee671867524c97e2d85de3808a0c93f0c01a4c33d8" + }, + "docs/technical-report.md": { + "bytes": 8284, + "sha256": "45ddabe94b4e13f5f507f8d716b18063982b10876d3c5d833336706785a9343a" + }, + "docs/benchmark-scores.md": { + "bytes": 9185, + "sha256": "357a8c11db9815ee0f889f2eeeb6bfc1d633dd3723909ac14712a3fa5fb2ad20" + }, + "docs/standard-asr-benchmarks.md": { + "bytes": 3031, + "sha256": "32d08431369c0790432773b97bb8714e6c990a79b9e21de4b061ae8af037b56f" + }, + "docs/current-checkpoint-benchmarks.md": { + "bytes": 6452, + "sha256": "c72a0fd39b01f0b52c144f9a1aa4ae7a315f334595e51ddaccb0944ec78ef9eb" + }, + "report/benchmark-tables.tex": { + "bytes": 8252, + "sha256": "a587f30764384a5b62415aa0af74bdbccbf41089a390e38b457d17e10f9991e4" + }, + "report/model.json": { + "bytes": 1098, + "sha256": "a0d6dece22d8eefbc35433a04376dd7415444a3c23ac7a7cb9ed179a8ec0d7c4" + }, + "report/build-receipt.json": { + "bytes": 6600, + "sha256": "b4cc807a85b6d7594167eed2ec3f2897f42a2a7416ceb8d2f8f587acfce6d534" + }, + "report/make_surgery_figure.py": { + "bytes": 2426, + "sha256": "0703dd2d6237491fe0cd410ad8a16c93b9e0128addfcf12b858115425233f314" + }, + "report/references.bib": { + "bytes": 2498, + "sha256": "e3229e1faad05e4975b231660184d126764f0c7a33bae9c31248d91859347d55" + }, + "report/current-benchmark-validation.json": { + "bytes": 8220, + "sha256": "cda67fa90a2d75c815ac48223224ef60205ec3ee841bad4b73dce1b273ebe8d9" + }, + "report/paper.tex": { + "bytes": 12026, + "sha256": "0dcc45ad9c401a9fd4e4ab389890b7bc51d4c0f8c21d4b18a7628d16b124c6ad" + }, + "report/README.md": { + "bytes": 3661, + "sha256": "5a9f831852933575fce137fb9f99ce8d25764bf4963acbd57910c8c98d11a6e2" + }, + "report/standard-benchmark-table.tex": { + "bytes": 2087, + "sha256": "dde308176577c1b16da2776d143251786c7c77f79a010f0bdcd2c17088404374" + }, + "report/benchmark-values.tex": { + "bytes": 1044, + "sha256": "34b1c7c0e325f4309e0bab730fdd3d8866facb90f1165743a2647d1f72ae3e5f" + }, + "report/current-benchmark-values.tex": { + "bytes": 875, + "sha256": "5eda781ff8966fc183d3a9c3a92e2dfe1840e8eddad292b891ef1976e63da9c7" + }, + "report/evidence.tex": { + "bytes": 361, + "sha256": "856a5d3f3f24852de8504c10564ea58c21090fa1af1c304a6a64e72d57e86704" + }, + "report/standard-benchmark-values.tex": { + "bytes": 541, + "sha256": "137fda4dfa71e1a08c6830ba41c35840ecfa250e7b6cf22834049ef4208da79a" + }, + "report/current-standard-table.tex": { + "bytes": 2394, + "sha256": "46aeaa3d4bf92d81e1acfc2c54b0e46cc9be1a4639ac9d16597c97982206ba2b" + }, + "report/current-domains-table.tex": { + "bytes": 3010, + "sha256": "ca92fc7f970f7608e69628ca2240e51b514118e24071376634884643a530fd26" + }, + "report/neurips_2026.sty": { + "bytes": 13704, + "sha256": "c3fc2894e83d2517ca18b66741d6c595986d97957dc08ec08bb2125a7ec4555a" + }, + "report/make_model_figures.py": { + "bytes": 10202, + "sha256": "d1619f159c0c11c7497498083843e0ea3f9352b299f4279551e8170de4eebe01" + }, + "evidence/r3-promotion-20260908/q8-metal.json": { + "bytes": 17824, + "sha256": "3dd7afe749f1ca1023120217c086e6f43765284ae8f169a07fc0f513cdc460d5" + }, + "evidence/r3-promotion-20260908/scores.json": { + "bytes": 8223, + "sha256": "27f12c1ced540a0b101d380a4e093a9cd531361006dc55c09de931725088eca5" + }, + "evidence/r3-promotion-20260908/export-environment.json": { + "bytes": 843, + "sha256": "c859645961708c2b6087a6d8faa88e113c84420083d471eba2f1bc420fde128d" + }, + "evidence/r3-promotion-20260908/linux-native-fixture.py": { + "bytes": 1909, + "sha256": "f06c5b9e45521f22409cd7d00c33e92775c89c51b7be68ee5ac5bd2d8667dc12" + }, + "evidence/r3-promotion-20260908/f16-lineage.json": { + "bytes": 1064, + "sha256": "e47ada9ed05ce74f13964daca3c1393e4c46025e550a500500323c6cd2c259e7" + }, + "evidence/r3-promotion-20260908/linux-f16-native-fixture.py": { + "bytes": 1912, + "sha256": "ee35a54eea85e1b60cb65554a169cb1b63b82c5cf0f0a2632d0bd9b2b401d961" + }, + "evidence/r3-promotion-20260908/linux-f16-native-smoke.json": { + "bytes": 4285, + "sha256": "97d14f868b61d7230495e1e970eed7fcfcb262fa0d3aa0be726a77873241f85b" + }, + "evidence/r3-promotion-20260908/q8-cpu.json": { + "bytes": 17721, + "sha256": "49fb93eb4df9e27ec5ef44db4dc35c41babd906e516cda516ebb03526ba30927" + }, + "evidence/r3-promotion-20260908/metric-evidence.tar.gz": { + "bytes": 2543488, + "sha256": "b88ba374bb5c27bf07915a16903d51e9c268399b6932fca320d7abc116427432" + }, + "evidence/r3-promotion-20260908/f16-cpu.json": { + "bytes": 17737, + "sha256": "ca1b5d8fd0e8799808447d562ecda3a0021b4691ae5947f8c0dc619fb7a0c765" + }, + "evidence/r3-promotion-20260908/linux-q8-native-smoke.json": { + "bytes": 4284, + "sha256": "370add5862c903eb1fdbb8957144d7fabe1a20fdd52383d18e95bc5ef5e805b4" + }, + "evidence/r3-promotion-20260908/q8-lineage.json": { + "bytes": 1064, + "sha256": "31869869164a660c9424a26990b7295334c930772a6702a9dccd462871cac6b9" + }, + "evidence/r3-promotion-20260908/f16-metal.json": { + "bytes": 17832, + "sha256": "a9b3f96d94407fc8c422def857f971809248a93c4c657d7057cb1011b3b04e01" + }, + "evidence/domains-r3-20260908/hypotheses-audit.json": { + "bytes": 1468, + "sha256": "66c748f9ea625cf38e9bf672628e95300a5919abbaf47ebfd7c4800f5b1f3116" + }, + "evidence/domains-r3-20260908/numeric-evidence.jsonl.gz": { + "bytes": 719000, + "sha256": "298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df" + }, + "evidence/domains-r3-20260908/transfer.json": { + "bytes": 803, + "sha256": "579a18bb0d6a8a2f461bee28331ef3ba78808e64213224747e1ebe0ec96e69bf" + }, + "evidence/domains-r3-20260908/inference-comparison.json": { + "bytes": 39410, + "sha256": "3892c802fd85d6cab5bee77b69a774b0fa6308294f6a7e5e536ef89300e853ba" + }, + "evidence/domains-r3-20260908/comparison.json": { + "bytes": 41039, + "sha256": "617a2119027b9e706907fdeb8a5819babb55f1959456d677e744d2b57ddfd2be" + }, + "evidence/domains-r3-20260908/README.md": { + "bytes": 2313, + "sha256": "c3432cafb227223d2fa75da5da4ae3d7b84430d20a953fae6b57aa3f9c03b13e" + }, + "evidence/domains-r3-20260908/scores.csv": { + "bytes": 6376, + "sha256": "12f2c6970a8b9d8f751c32137098c6addb660d5880175471de97cff40595abe6" + }, + "evidence/domains-r3-20260908/launch.json": { + "bytes": 919, + "sha256": "4920721addba65064fe2f364f918711b079669f05904e246d029879be4aba58c" + }, + "evidence/standard-asr-r3-20260908/hypotheses-audit.json": { + "bytes": 1466, + "sha256": "a60177ac2ac1064b481cc85424d5ee989cca43112191ddae30d60b7f56ae57d3" + }, + "evidence/standard-asr-r3-20260908/summary.json": { + "bytes": 2718, + "sha256": "6eba4b2bdc57f658ddc455b9ffdd5e9bd0fec832a88fe0bb329b45c5fd512852" + }, + "evidence/standard-asr-r3-20260908/numeric-evidence.jsonl.gz": { + "bytes": 1467437, + "sha256": "3d1467e56efd6769da1f8cb74eef0c3eac7bf93fa8200c39538963a9d4b080c1" + }, + "evidence/standard-asr-r3-20260908/transfer.json": { + "bytes": 803, + "sha256": "6ff71ac6de0bb29627d3f20b1264cff09dd6dbf54e028450554d4a4aed54288c" + }, + "evidence/standard-asr-r3-20260908/inference-comparison.json": { + "bytes": 23344, + "sha256": "23584aab2856bf7ebaad7611cfdc9395d7d09b40ea6d64afcb8472d5f63bae34" + }, + "evidence/standard-asr-r3-20260908/comparison.json": { + "bytes": 24978, + "sha256": "9b2702bfe9c7b107db1627df3cb71d24d25d4eb8da0ee8683b0bbf9f994ca7da" + }, + "evidence/standard-asr-r3-20260908/README.md": { + "bytes": 2482, + "sha256": "20539608da10f1b6e44fefffa3735791da4dc35567db414586dc49dbebcc28da" + }, + "evidence/standard-asr-r3-20260908/scores.csv": { + "bytes": 3596, + "sha256": "b65dd7c775ae861f3d53d8978b95be997ca5507e045ab195daa411fe790354c1" + }, + "evidence/standard-asr-r3-20260908/restart.json": { + "bytes": 817, + "sha256": "4c44ca582b683406eb8cb0d17f0ce85d81277ac73fabb702a12ccee260837a07" + }, + "evidence/standard-asr-r3-20260908/launch.json": { + "bytes": 793, + "sha256": "91a8c8cc07d60e1e0e17065b9d46982714948ccfeaf721d811033e1e22914bba" + }, + "evidence/r3-promotion-20260908/ci/smoke-results-ci-linux-x64.json": { + "bytes": 17441, + "sha256": "0ac0094044de1e14eb03bbfd1095b437461bb48ac4744ba0d5445b80d22999cb" + }, + "evidence/r3-promotion-20260908/ci/smoke-results-ci-win32-x64.json": { + "bytes": 17455, + "sha256": "7af01cf0e331e3681ff7d1cebc052f5cbcbd860522e1fe2bfe74fd823f91d4f6" + }, + "evidence/r3-promotion-20260908/ci/smoke-results-ci-darwin-arm64.json": { + "bytes": 17423, + "sha256": "83fd9fc97f1ffd16ffc6eec29034b992a558e80522ee2829ecbaafe421ddb90d" + }, + "evaluation/standard_asr/run.py": { + "bytes": 7845, + "sha256": "1c8951a18eba92ae8fe12c537149c961ae0c0e6f748216e27488305b5a2806c2" + }, + "evaluation/standard_asr/rescore.py": { + "bytes": 3943, + "sha256": "47e77bad2d65f1eb6d7a29a156ba747807b4017d4fcedf8d3d2011c5a4e0ab82" + }, + "evaluation/standard_asr/build_current_report.py": { + "bytes": 15963, + "sha256": "624b9781c023ea1cd4f1ad9a3c18f99e806027a228c5f24da9f466679578dd8f" + }, + "evaluation/standard_asr/requirements-score.txt": { + "bytes": 71, + "sha256": "920516addbb6b862c308b2f496434c0a8f1a7e48b090ca40cd0a759d1a3bfbfe" + }, + "evaluation/standard_asr/audit_predictions.py": { + "bytes": 5767, + "sha256": "80a8c7911819129e39911cd4f78df326485a659a847522779cf40b3b2365eb3a" + }, + "evaluation/standard_asr/build_materials.py": { + "bytes": 11187, + "sha256": "1b76c890ab9dcbe7caee67f49299e22f933a111d10171017cc289da92a3e0849" + }, + "evaluation/standard_asr/scoring.py": { + "bytes": 1407, + "sha256": "3b2bd425a79a94b9afa294ba55efc6b9fb5943d7021ce27394dadf6c58082bd0" + }, + "evaluation/standard_asr/README.md": { + "bytes": 4305, + "sha256": "5308cba17b71b4e13cdc80474806c13b047404467c7f71e10b108dec817db116" + }, + "evaluation/standard_asr/CURRENT.md": { + "bytes": 2572, + "sha256": "c126899426fc2950629ad4dbc9af0b0732ca029c88a16502bea28f91ae9ae16e" + }, + "evaluation/standard_asr/prepare.py": { + "bytes": 5170, + "sha256": "2ca98e52a3a71f5a16f0efd04133463af6e795a60bf4610c3a4b2af7d5773aeb" + }, + "evaluation/standard_asr/vendor/provenance.json": { + "bytes": 1324, + "sha256": "f6fb39bc16cb4b52ac62882baab0f1925d7405ee8a98f0ae32419edf9275aa12" + }, + "evaluation/standard_asr/vendor/LICENSE": { + "bytes": 11359, + "sha256": "5ee13882fce0975f0ad3c3d5c2042af4896c2a669be2d6df5b3d256811267c85" + }, + "evaluation/standard_asr/vendor/english_abbreviations.py": { + "bytes": 58524, + "sha256": "52997cc963e0bd6568d15554ad8d5fb0759f20f361483632a22f90d2d0e07db1" + }, + "evaluation/standard_asr/vendor/__init__.py": { + "bytes": 73, + "sha256": "8c39e42af3204c41c564c970e24525da160663836b54acc1b629e4342f862334" + }, + "evaluation/standard_asr/vendor/multilingual.py": { + "bytes": 3713, + "sha256": "4c8db6c4171c8c2f5567d2ed625a3a1f0d45156af45ca42d9555c0c9b61a29d5" + }, + "evaluation/standard_asr/vendor/normalizer.py": { + "bytes": 30320, + "sha256": "490b56393484ef386b486f3679cc2261264af280b3512154431e0bdcf2778295" + }, + "report/assets/model-figures-source.json": { + "bytes": 15764, + "sha256": "cf07124f8e46464eca5036cd6be96e4d1ced9d3f936e514e9382518181de42b8" + }, + "report/assets/recognition-deltas.pdf": { + "bytes": 15879, + "sha256": "62f0471788cb59e27fc317a004b34fe457a354ae511e01f0fc6d1efb8da87a0a" + }, + "report/assets/selection-profile.png": { + "bytes": 85067, + "sha256": "6202432a630ef22fb54bf57c494437dcb8f526c26a9a7b6ecf1bcbdf2510a75d" + }, + "report/assets/recognition-deltas.png": { + "bytes": 41222, + "sha256": "83890bae2503e59e918bf424cd841094302d5366e3a285523105db49f3e1fffc" + }, + "report/assets/selection-profile.pdf": { + "bytes": 20686, + "sha256": "0a2e8647a8a1a895e7a1e6d9f15779eac14c87f56e30eeeb85bc05ce720ba32f" + }, + "report/assets/kernel-atlas-preview.png": { + "bytes": 169043, + "sha256": "c5a4a3993d51cb7197c50cd82b2fb9fd5103d8dffd6f31b0fe7c5b923230691c" + }, + "report/assets/model-figures-visual-review.json": { + "bytes": 1884, + "sha256": "c26e9af4e1f21d3515944944f8e8dc19ccb4a33b4f80b961d52947c0bbd617d7" + }, + "report/assets/oruk-lockup.png": { + "bytes": 38483, + "sha256": "932df9164b237a1a31d1e636897d1766a15b8bac738acf81ee93de8ea2d6c18c" + }, + "report/assets/language-deltas.pdf": { + "bytes": 21904, + "sha256": "7564f39e28ee6deae1d7ad4b82e3737b1dc703044b5992c5f9108cbf2e9ff1c0" + }, + "report/assets/provenance.json": { + "bytes": 1068, + "sha256": "0b769a8fc524498391553ebb165658314d5f484ace0d8aebde74582e6dd493e4" + }, + "report/assets/surgery-source.json": { + "bytes": 468, + "sha256": "4c9898357e79d5014429bdfde2f2b7f2d67e6feceb4daca027380976f74520a6" + }, + "report/assets/kernel-fits.svg": { + "bytes": 26317, + "sha256": "4f2b63601bb7b9543cee0dec0aaaecc06c1741798eda0c6aa5e2b4f39ed1191b" + }, + "report/assets/language-deltas.png": { + "bytes": 72533, + "sha256": "64e77e3423cd60952f31a4e80bb486a05a3325cd5a0b3d92bcf9581cf9de201b" + }, + "report/assets/language-deltas.svg": { + "bytes": 24593, + "sha256": "cff05741bbe01c83ad5f839f99a40c2f189357e71bc6fddd8cd62c920116d693" + }, + "report/assets/kernel-fits.png": { + "bytes": 93530, + "sha256": "53daf3dc1431660b7696c83d94e56543f8236563a81a4dd9fb75e3cab997fa52" + }, + "report/assets/kernel-fits.pdf": { + "bytes": 20762, + "sha256": "b952a1d8e84032cd20cc452dfbea40c6d2c34cea7de4257b7a375ccac73b9e01" + }, + "report/assets/oruk-lockup.svg": { + "bytes": 3295, + "sha256": "3547cfacd79cf423d104adfdf342a28aef3a98dd187b1a4cc1949478d690a051" + }, + "report/assets/selection-profile-caption.md": { + "bytes": 358, + "sha256": "a20b960bd387cde6a867a9f2fccd7bcb2359634294983ccaddd3a15b02e2ddac" + }, + "report/assets/recognition-deltas.svg": { + "bytes": 7186, + "sha256": "ab94addc78953dd3038cd30436aeb73f6580f1b62903180f6ba81551a3317595" + }, + "report/assets/selection-profile.svg": { + "bytes": 21319, + "sha256": "9ddb7db5f54eb4f97294e31998ce06d99fdf18e9cbf88a906bef70bc99f4e19a" + }, + "report/assets/selection-profile-preview.png": { + "bytes": 37344, + "sha256": "b152347005940a12f09651d69a0c6183f6a32bb0d1c3d76711052628657e2fd9" + }, + "report/assets/surgery.svg": { + "bytes": 18807, + "sha256": "2d73665fd0637f983bfb3e9141f011f0fddbefa9743544d5220c0564016e7f9e" + }, + "report/assets/recognition-deltas-caption.md": { + "bytes": 485, + "sha256": "72500fd8e0e83f6dbd1e3c8191f6896f5e6405649b88981c1bb28480cb1bb1d4" + }, + "report/assets/language-deltas-preview.png": { + "bytes": 33557, + "sha256": "b91f3171eb80994161bb502c70a4b8eb015a243c79870ed134cb9e8af41bfb3c" + }, + "report/assets/kernel-atlas.png": { + "bytes": 376772, + "sha256": "2b8b24400ac7b9e8b6e4484eb0c6c36cca8f618167141d508a2b6468824888c3" + }, + "report/assets/kernel-atlas.pdf": { + "bytes": 39134, + "sha256": "73a8c7e2f1ea83f7bcc9eefcdbd1a53b6e95383a0b3799f45f75390d676be9b5" + }, + "report/assets/kernel-atlas-caption.md": { + "bytes": 269, + "sha256": "99041021a8aaf08ef44a52613b0bcc1f8b3201f8552638f8952ebcc14c3c355a" + }, + "report/assets/kernel-fits-preview.png": { + "bytes": 48220, + "sha256": "1f67e76987f9c17371450f2a39eb93948e9c1e9fa7de1a8d7624517b0400df90" + }, + "report/assets/language-deltas-caption.md": { + "bytes": 291, + "sha256": "86b941e82ce09680de5daeec0078fecefa809d9c724af06f7c3bb50e7884fd68" + }, + "report/assets/recognition-deltas-preview.png": { + "bytes": 18202, + "sha256": "bd2af5115a46174408ae2386aeb73ac267c2cccb1e730203f03f6d33f82b5e62" + }, + "report/assets/kernel-atlas.svg": { + "bytes": 130923, + "sha256": "41b88f335fa6200353201a77a1c86d0d471fbee61a1a64a7a6c0bffaf2abda88" + }, + "report/assets/kernel-fits-caption.md": { + "bytes": 345, + "sha256": "3cff663d27935fb3b24d3cc1d6fe6e880b33179378ecbfd6522b18a628e1d6ea" + }, + "report/assets/surgery.pdf": { + "bytes": 20609, + "sha256": "0601e73435408a07f209e332fd67340b5a21e89868de5968fdc067dff9a05497" + }, + "report/assets/surgery.png": { + "bytes": 76446, + "sha256": "bed57fa43afb8bf62a826b2c302b153f4df063dd96c68164ecda4535e30abe67" + }, + "report/assets/unseen/unseen-language-coverage.png": { + "bytes": 176291, + "sha256": "aa66de577f8ebb9997568eda3fccdd9c80f93a0d02c13518abc7405830b1bba7" + }, + "report/assets/unseen/unseen-language-coverage.pdf": { + "bytes": 30716, + "sha256": "26c89cabebbc8558a71d761eb18af8b178cbbfff5f63ce415f6fcd9589ca6f55" + }, + "report/assets/unseen/unseen-english-accents.png": { + "bytes": 141909, + "sha256": "c6be28a4ad2ceaa78e578f520aca33c8a6000f8731331443b4910c8f6c3a8069" + }, + "report/assets/unseen/unseen-english-accents.pdf": { + "bytes": 27214, + "sha256": "ae6fd18d34f9348020558c340cd2c3666cc1addcc32a939ba936615f567b41e3" + }, + "report/assets/unseen/unseen-english-accents.svg": { + "bytes": 15908, + "sha256": "3cacedbe24885630d0eaad88403cf80f890dea42d462bf0d89e889cb907b8f34" + }, + "report/assets/unseen/unseen-benchmark-captions.md": { + "bytes": 1266, + "sha256": "59459974d24ee611c2f4c543d54e17dbea6e4e5bcb869f023943139ef1288565" + }, + "report/assets/unseen/unseen-language-coverage.svg": { + "bytes": 17941, + "sha256": "03cd68fc07a2271bf8e1d3d4132eee814a85174137e7fa74a52af7a6984146aa" + }, + "report/assets/unseen/unseen-eurospeech.png": { + "bytes": 206024, + "sha256": "a5efb7b9dcb242a5301a5edffde5d6dace4dd03a1319e49c9cd817cc4dd30854" + }, + "report/assets/unseen/unseen-eurospeech.pdf": { + "bytes": 30980, + "sha256": "14be92cd1d4ea09546d60482c3c329c212321b0f551e1e1ab510ab271ee5782a" + }, + "report/assets/unseen/unseen-english-domains.svg": { + "bytes": 22053, + "sha256": "81760f0eb64d9e313914456860cddcecb3740f9cf81f48cf8eec71862338f1dc" + }, + "report/assets/unseen/unseen-english-domains.png": { + "bytes": 173922, + "sha256": "03e2048952df3576b88a45d19add8b835e51a0a4e61d51fcc03c76dd5173272a" + }, + "report/assets/unseen/unseen-english-domains.pdf": { + "bytes": 29405, + "sha256": "11d7260cf143d636651cc89d4255f929f67b0ad26c5589450602a3543f0ffb58" + }, + "report/assets/unseen/unseen-figure-receipt.json": { + "bytes": 3070, + "sha256": "08dabab6d6f3d5fd3e52804087665eedc46528561f375952052fb54c72aa0de3" + }, + "report/assets/unseen/unseen-eurospeech.svg": { + "bytes": 27408, + "sha256": "2957f74c5f09b3c8b5cdb6b3d6b08082e0b774a7f2fc93cf6a93e2539f0af64c" + }, + "report/assets/affiliations/cambridge.pdf": { + "bytes": 44738, + "sha256": "535e2f3382a9e63ef27fcd561e54cfc87d74f494c3a7f6dd072c2af2bfe7a738" + }, + "report/assets/affiliations/SOURCES.md": { + "bytes": 2203, + "sha256": "0d77aeb9033fc456dc3d1a4d3b5cbf7bd67a3f43feaf4b97c5ed5fc7c44b579e" + }, + "report/assets/affiliations/cambridge.svg": { + "bytes": 57691, + "sha256": "27381d019c0a54c53b7b746750dc148f35dca2a6fe96b39dff5f896b0e745902" + }, + "report/assets/affiliations/openwhispr.pdf": { + "bytes": 1341, + "sha256": "78e575011be731f9d07f665f1440a4582df27d577bbdb3805d2c7a17b40cacd7" + }, + "report/assets/affiliations/hoid.svg": { + "bytes": 2965, + "sha256": "49b5fc2265f1294abf6bb65f054078b640b9fbaed9780e2817ab6a103cd2eb72" + }, + "report/assets/affiliations/stanford.png": { + "bytes": 11528, + "sha256": "09434daceea1a4ab62da5db3cd9e81bbe1a7e034224f504d63fd3f12e51af74b" + }, + "report/assets/affiliations/hoid.pdf": { + "bytes": 1978, + "sha256": "439f46cb03c498cbc0bb797d39ba31ce80d57742d3b699ae0a6ca5cf2fa1311a" + }, + "report/assets/affiliations/openwhispr.svg": { + "bytes": 1202, + "sha256": "29d50f34b1ffe4b11665aaeefdaf63a22796c53c38234cd9d6f06c5153c9e13a" + } + } +} diff --git a/CITATION.bib b/CITATION.bib new file mode 100644 index 0000000000000000000000000000000000000000..707fc35923cf9b7dc7baae345dd33bfea5c02058 --- /dev/null +++ b/CITATION.bib @@ -0,0 +1,16 @@ +@techreport{roll2026orukeet, + title = {{Orukeet}: Multilingual {ASR} with Frozen {Gabor} Kernels}, + author = {Roll, Nathan and + Yi, Irene and + Mar{\c{s}}an, B{\"u}{\c{s}}ra and + Grenez, Vianney and + Stein, Gabriel and + Mrkaic, Momcilo and + Padjin, Pavle and + Zeljkovic, Vladimir and + Graham, Calbert}, + institution = {Oruk AI}, + year = {2026}, + type = {Technical report}, + url = {https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf} +} diff --git a/CITATION.cff b/CITATION.cff new file mode 100644 index 0000000000000000000000000000000000000000..98d360e5cc491ccbfc65cb8125500949cfba5a55 --- /dev/null +++ b/CITATION.cff @@ -0,0 +1,45 @@ +cff-version: 1.2.0 +message: If you use Orukeet, cite the technical report below and retain the NVIDIA Parakeet attribution. +title: 'Orukeet: Multilingual ASR with Frozen Gabor Kernels' +type: software +authors: &id001 +- given-names: Nathan + family-names: Roll + affiliation: Oruk AI; Stanford University +- given-names: Irene + family-names: Yi + affiliation: Oruk AI; Stanford University +- given-names: Büşra + family-names: Marşan + affiliation: Oruk AI; Stanford University +- given-names: Vianney + family-names: Grenez + affiliation: Oruk AI +- given-names: Gabriel + family-names: Stein + affiliation: OpenWhispr +- given-names: Momcilo + family-names: Mrkaic + affiliation: Hoid +- given-names: Pavle + family-names: Padjin + affiliation: Hoid +- given-names: Vladimir + family-names: Zeljkovic + affiliation: Hoid +- given-names: Calbert + family-names: Graham + affiliation: Oruk AI; University of Cambridge +version: 0.1.0 +repository-code: https://github.com/Oruk-AI/orukeet +license: MIT +abstract: 'A Parakeet-derived multilingual recognizer with 12,288 fitted, frozen temporal Gabor kernels. + Code: MIT. Weights: CC BY-SA 4.0.' +preferred-citation: + type: report + title: 'Orukeet: Multilingual ASR with Frozen Gabor Kernels' + authors: *id001 + institution: + name: Oruk AI + year: 2026 + url: https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..0067e42971c298904dbe9392fc910515ffd32eea --- /dev/null +++ b/LICENSE @@ -0,0 +1,22 @@ +MIT License + +Copyright (c) 2025 Knuckles92 +Copyright (c) 2026 Oruk AI + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/LICENSE-WEIGHTS b/LICENSE-WEIGHTS new file mode 100644 index 0000000000000000000000000000000000000000..7d4f96c54aa155a17387e64501c50b6a0469932c --- /dev/null +++ b/LICENSE-WEIGHTS @@ -0,0 +1,427 @@ +Attribution-ShareAlike 4.0 International + +======================================================================= + +Creative Commons Corporation ("Creative Commons") is not a law firm and +does not provide legal services or legal advice. Distribution of +Creative Commons public licenses does not create a lawyer-client or +other relationship. Creative Commons makes its licenses and related +information available on an "as-is" basis. Creative Commons gives no +warranties regarding its licenses, any material licensed under their +terms and conditions, or any related information. Creative Commons +disclaims all liability for damages resulting from their use to the +fullest extent possible. + +Using Creative Commons Public Licenses + +Creative Commons public licenses provide a standard set of terms and +conditions that creators and other rights holders may use to share +original works of authorship and other material subject to copyright +and certain other rights specified in the public license below. The +following considerations are for informational purposes only, are not +exhaustive, and do not form part of our licenses. + + Considerations for licensors: Our public licenses are + intended for use by those authorized to give the public + permission to use material in ways otherwise restricted by + copyright and certain other rights. Our licenses are + irrevocable. Licensors should read and understand the terms + and conditions of the license they choose before applying it. + Licensors should also secure all rights necessary before + applying our licenses so that the public can reuse the + material as expected. Licensors should clearly mark any + material not subject to the license. This includes other CC- + licensed material, or material used under an exception or + limitation to copyright. More considerations for licensors: + wiki.creativecommons.org/Considerations_for_licensors + + Considerations for the public: By using one of our public + licenses, a licensor grants the public permission to use the + licensed material under specified terms and conditions. If + the licensor's permission is not necessary for any reason--for + example, because of any applicable exception or limitation to + copyright--then that use is not regulated by the license. Our + licenses grant only permissions under copyright and certain + other rights that a licensor has authority to grant. Use of + the licensed material may still be restricted for other + reasons, including because others have copyright or other + rights in the material. A licensor may make special requests, + such as asking that all changes be marked or described. + Although not required by our licenses, you are encouraged to + respect those requests where reasonable. More considerations + for the public: + wiki.creativecommons.org/Considerations_for_licensees + +======================================================================= + +Creative Commons Attribution-ShareAlike 4.0 International Public +License + +By exercising the Licensed Rights (defined below), You accept and agree +to be bound by the terms and conditions of this Creative Commons +Attribution-ShareAlike 4.0 International Public License ("Public +License"). To the extent this Public License may be interpreted as a +contract, You are granted the Licensed Rights in consideration of Your +acceptance of these terms and conditions, and the Licensor grants You +such rights in consideration of benefits the Licensor receives from +making the Licensed Material available under these terms and +conditions. + + +Section 1 -- Definitions. + + a. Adapted Material means material subject to Copyright and Similar + Rights that is derived from or based upon the Licensed Material + and in which the Licensed Material is translated, altered, + arranged, transformed, or otherwise modified in a manner requiring + permission under the Copyright and Similar Rights held by the + Licensor. For purposes of this Public License, where the Licensed + Material is a musical work, performance, or sound recording, + Adapted Material is always produced where the Licensed Material is + synched in timed relation with a moving image. + + b. Adapter's License means the license You apply to Your Copyright + and Similar Rights in Your contributions to Adapted Material in + accordance with the terms and conditions of this Public License. + + c. BY-SA Compatible License means a license listed at + creativecommons.org/compatiblelicenses, approved by Creative + Commons as essentially the equivalent of this Public License. + + d. Copyright and Similar Rights means copyright and/or similar rights + closely related to copyright including, without limitation, + performance, broadcast, sound recording, and Sui Generis Database + Rights, without regard to how the rights are labeled or + categorized. For purposes of this Public License, the rights + specified in Section 2(b)(1)-(2) are not Copyright and Similar + Rights. + + e. Effective Technological Measures means those measures that, in the + absence of proper authority, may not be circumvented under laws + fulfilling obligations under Article 11 of the WIPO Copyright + Treaty adopted on December 20, 1996, and/or similar international + agreements. + + f. Exceptions and Limitations means fair use, fair dealing, and/or + any other exception or limitation to Copyright and Similar Rights + that applies to Your use of the Licensed Material. + + g. License Elements means the license attributes listed in the name + of a Creative Commons Public License. The License Elements of this + Public License are Attribution and ShareAlike. + + h. Licensed Material means the artistic or literary work, database, + or other material to which the Licensor applied this Public + License. + + i. Licensed Rights means the rights granted to You subject to the + terms and conditions of this Public License, which are limited to + all Copyright and Similar Rights that apply to Your use of the + Licensed Material and that the Licensor has authority to license. + + j. Licensor means the individual(s) or entity(ies) granting rights + under this Public License. + + k. Share means to provide material to the public by any means or + process that requires permission under the Licensed Rights, such + as reproduction, public display, public performance, distribution, + dissemination, communication, or importation, and to make material + available to the public including in ways that members of the + public may access the material from a place and at a time + individually chosen by them. + + l. Sui Generis Database Rights means rights other than copyright + resulting from Directive 96/9/EC of the European Parliament and of + the Council of 11 March 1996 on the legal protection of databases, + as amended and/or succeeded, as well as other essentially + equivalent rights anywhere in the world. + + m. You means the individual or entity exercising the Licensed Rights + under this Public License. Your has a corresponding meaning. + + +Section 2 -- Scope. + + a. License grant. + + 1. Subject to the terms and conditions of this Public License, + the Licensor hereby grants You a worldwide, royalty-free, + non-sublicensable, non-exclusive, irrevocable license to + exercise the Licensed Rights in the Licensed Material to: + + a. reproduce and Share the Licensed Material, in whole or + in part; and + + b. produce, reproduce, and Share Adapted Material. + + 2. Exceptions and Limitations. For the avoidance of doubt, where + Exceptions and Limitations apply to Your use, this Public + License does not apply, and You do not need to comply with + its terms and conditions. + + 3. Term. The term of this Public License is specified in Section + 6(a). + + 4. Media and formats; technical modifications allowed. The + Licensor authorizes You to exercise the Licensed Rights in + all media and formats whether now known or hereafter created, + and to make technical modifications necessary to do so. The + Licensor waives and/or agrees not to assert any right or + authority to forbid You from making technical modifications + necessary to exercise the Licensed Rights, including + technical modifications necessary to circumvent Effective + Technological Measures. For purposes of this Public License, + simply making modifications authorized by this Section 2(a) + (4) never produces Adapted Material. + + 5. Downstream recipients. + + a. Offer from the Licensor -- Licensed Material. Every + recipient of the Licensed Material automatically + receives an offer from the Licensor to exercise the + Licensed Rights under the terms and conditions of this + Public License. + + b. Additional offer from the Licensor -- Adapted Material. + Every recipient of Adapted Material from You + automatically receives an offer from the Licensor to + exercise the Licensed Rights in the Adapted Material + under the conditions of the Adapter's License You apply. + + c. No downstream restrictions. You may not offer or impose + any additional or different terms or conditions on, or + apply any Effective Technological Measures to, the + Licensed Material if doing so restricts exercise of the + Licensed Rights by any recipient of the Licensed + Material. + + 6. No endorsement. Nothing in this Public License constitutes or + may be construed as permission to assert or imply that You + are, or that Your use of the Licensed Material is, connected + with, or sponsored, endorsed, or granted official status by, + the Licensor or others designated to receive attribution as + provided in Section 3(a)(1)(A)(i). + + b. Other rights. + + 1. Moral rights, such as the right of integrity, are not + licensed under this Public License, nor are publicity, + privacy, and/or other similar personality rights; however, to + the extent possible, the Licensor waives and/or agrees not to + assert any such rights held by the Licensor to the limited + extent necessary to allow You to exercise the Licensed + Rights, but not otherwise. + + 2. Patent and trademark rights are not licensed under this + Public License. + + 3. To the extent possible, the Licensor waives any right to + collect royalties from You for the exercise of the Licensed + Rights, whether directly or through a collecting society + under any voluntary or waivable statutory or compulsory + licensing scheme. In all other cases the Licensor expressly + reserves any right to collect such royalties. + + +Section 3 -- License Conditions. + +Your exercise of the Licensed Rights is expressly made subject to the +following conditions. + + a. Attribution. + + 1. If You Share the Licensed Material (including in modified + form), You must: + + a. retain the following if it is supplied by the Licensor + with the Licensed Material: + + i. identification of the creator(s) of the Licensed + Material and any others designated to receive + attribution, in any reasonable manner requested by + the Licensor (including by pseudonym if + designated); + + ii. a copyright notice; + + iii. a notice that refers to this Public License; + + iv. a notice that refers to the disclaimer of + warranties; + + v. a URI or hyperlink to the Licensed Material to the + extent reasonably practicable; + + b. indicate if You modified the Licensed Material and + retain an indication of any previous modifications; and + + c. indicate the Licensed Material is licensed under this + Public License, and include the text of, or the URI or + hyperlink to, this Public License. + + 2. You may satisfy the conditions in Section 3(a)(1) in any + reasonable manner based on the medium, means, and context in + which You Share the Licensed Material. For example, it may be + reasonable to satisfy the conditions by providing a URI or + hyperlink to a resource that includes the required + information. + + 3. If requested by the Licensor, You must remove any of the + information required by Section 3(a)(1)(A) to the extent + reasonably practicable. + + b. ShareAlike. + + In addition to the conditions in Section 3(a), if You Share + Adapted Material You produce, the following conditions also apply. + + 1. The Adapter's License You apply must be a Creative Commons + license with the same License Elements, this version or + later, or a BY-SA Compatible License. + + 2. You must include the text of, or the URI or hyperlink to, the + Adapter's License You apply. You may satisfy this condition + in any reasonable manner based on the medium, means, and + context in which You Share Adapted Material. + + 3. You may not offer or impose any additional or different terms + or conditions on, or apply any Effective Technological + Measures to, Adapted Material that restrict exercise of the + rights granted under the Adapter's License You apply. + + +Section 4 -- Sui Generis Database Rights. + +Where the Licensed Rights include Sui Generis Database Rights that +apply to Your use of the Licensed Material: + + a. for the avoidance of doubt, Section 2(a)(1) grants You the right + to extract, reuse, reproduce, and Share all or a substantial + portion of the contents of the database; + + b. if You include all or a substantial portion of the database + contents in a database in which You have Sui Generis Database + Rights, then the database in which You have Sui Generis Database + Rights (but not its individual contents) is Adapted Material, + including for purposes of Section 3(b); and + + c. You must comply with the conditions in Section 3(a) if You Share + all or a substantial portion of the contents of the database. + +For the avoidance of doubt, this Section 4 supplements and does not +replace Your obligations under this Public License where the Licensed +Rights include other Copyright and Similar Rights. + + +Section 5 -- Disclaimer of Warranties and Limitation of Liability. + + a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE + EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS + AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF + ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, + IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, + WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR + PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, + ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT + KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT + ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. + + b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE + TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, + NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, + INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, + COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR + USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN + ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR + DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR + IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. + + c. The disclaimer of warranties and limitation of liability provided + above shall be interpreted in a manner that, to the extent + possible, most closely approximates an absolute disclaimer and + waiver of all liability. + + +Section 6 -- Term and Termination. + + a. This Public License applies for the term of the Copyright and + Similar Rights licensed here. However, if You fail to comply with + this Public License, then Your rights under this Public License + terminate automatically. + + b. Where Your right to use the Licensed Material has terminated under + Section 6(a), it reinstates: + + 1. automatically as of the date the violation is cured, provided + it is cured within 30 days of Your discovery of the + violation; or + + 2. upon express reinstatement by the Licensor. + + For the avoidance of doubt, this Section 6(b) does not affect any + right the Licensor may have to seek remedies for Your violations + of this Public License. + + c. For the avoidance of doubt, the Licensor may also offer the + Licensed Material under separate terms or conditions or stop + distributing the Licensed Material at any time; however, doing so + will not terminate this Public License. + + d. Sections 1, 5, 6, 7, and 8 survive termination of this Public + License. + + +Section 7 -- Other Terms and Conditions. + + a. The Licensor shall not be bound by any additional or different + terms or conditions communicated by You unless expressly agreed. + + b. Any arrangements, understandings, or agreements regarding the + Licensed Material not stated herein are separate from and + independent of the terms and conditions of this Public License. + + +Section 8 -- Interpretation. + + a. For the avoidance of doubt, this Public License does not, and + shall not be interpreted to, reduce, limit, restrict, or impose + conditions on any use of the Licensed Material that could lawfully + be made without permission under this Public License. + + b. To the extent possible, if any provision of this Public License is + deemed unenforceable, it shall be automatically reformed to the + minimum extent necessary to make it enforceable. If the provision + cannot be reformed, it shall be severed from this Public License + without affecting the enforceability of the remaining terms and + conditions. + + c. No term or condition of this Public License will be waived and no + failure to comply consented to unless expressly agreed to by the + Licensor. + + d. Nothing in this Public License constitutes or may be interpreted + as a limitation upon, or waiver of, any privileges and immunities + that apply to the Licensor or You, including from the legal + processes of any jurisdiction or authority. + + +======================================================================= + +Creative Commons is not a party to its public +licenses. Notwithstanding, Creative Commons may elect to apply one of +its public licenses to material it publishes and in those instances +will be considered the “Licensor.” The text of the Creative Commons +public licenses is dedicated to the public domain under the CC0 Public +Domain Dedication. Except for the limited purpose of indicating that +material is shared under a Creative Commons public license or as +otherwise permitted by the Creative Commons policies published at +creativecommons.org/policies, Creative Commons does not authorize the +use of the trademark "Creative Commons" or any other trademark or logo +of Creative Commons without its prior written consent including, +without limitation, in connection with any unauthorized modifications +to any of its public licenses or any other arrangements, +understandings, or agreements concerning use of licensed material. For +the avoidance of doubt, this paragraph does not form part of the +public licenses. + +Creative Commons may be contacted at creativecommons.org. diff --git a/NOTICE.md b/NOTICE.md new file mode 100644 index 0000000000000000000000000000000000000000..a436c3b5351a202722d916241dca85b721e186c2 --- /dev/null +++ b/NOTICE.md @@ -0,0 +1,61 @@ +# Attribution and license scope + +Orukeet is an adaptation of **NVIDIA Parakeet TDT 0.6B v3**. NVIDIA retains +copyright in its model and upstream work. The base weights are distributed +under [CC BY 4.0](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3). +Oruk AI's changes comprise multilingual/accent continuation training, parameter +blending, fitted and frozen Gabor-kernel replacement and recovery, native +export, application integration and evaluation. [Model stages](https://github.com/Oruk-AI/orukeet/blob/main/release/model-stages.json) +identifies r3 as the source of every current Orukeet download and records its ancestry. + +The r3 NeMo, Q8 GGUF and F16 GGUF weights, and their fitted Gabor kernels, +are designated **CC BY-SA 4.0**. +The complete license is in [LICENSE-WEIGHTS](LICENSE-WEIGHTS). This permits +commercial use and modification, with attribution and applicable ShareAlike +requirements. Earlier checkpoints retain their source notices and are not +silently relicensed by this file. Orukeet v0.1.0 distributes the final r3 checkpoint in all three formats. + +Python and integration code in this repository is MIT unless a file specifies +otherwise. The native bindings, audio windowing and worker transport derive +from OpenWhisper by Knuckles92 and its Oruk AI integration; their MIT notice is +retained in [LICENSE](LICENSE). The downloaded NVIDIA NeMo-Speech.cpp SDK and +its bundled dependencies retain their own notices, including Apache-2.0 and +MIT components. Keep the SDK's license files when redistributing it. The +[pinned source](https://github.com/NVIDIA/NeMo-Speech.cpp/tree/4f9676226f667d14608487df744f375db87127f8) +is the authority for those terms. + +Training data credits: + +- Mozilla/Common Voice contributors: Common Voice 22, CC0. +- Google and the FLEURS authors: FLEURS, CC BY 4.0. +- Google and the OpenSLR 83 authors, with the `ylacombe/english_dialects` + restructuring: English dialect speech, CC BY-SA 4.0. +- Beijing Kingline Data Technology and the SpeechOcean762 authors: + SpeechOcean762, CC BY 4.0. +- SpeechColab and the GigaSpeechBench authors: 17 English accent/domain splits + in the final continuation. Their paper identifies Creative Commons source audio; + the source record below preserves the available license details. +- DISCO at ETH Zurich and the EuroSpeech contributors: Bulgarian, Greek and + Italian parliamentary speech, with the providers' country-specific terms. +- SberDevices and the Golos authors: Golos Crowd, under the + [Public license with attribution and conditions reserved](https://github.com/sberdevices/golos/blob/master/license/en_us.pdf). +- Nordisk Språkteknologi, the National Library of Norway and the Alexandra + Institute: NST Swedish and Danish, distributed under CC0. +- ILSP/Athena Research Center and the Lesbian Speech Corpus contributors: + dialect speech from Lesbos. The source card does not specify a reuse license. + +Data were filtered, split, normalized and sampled as described in +[data and licenses](https://github.com/Oruk-AI/orukeet/blob/main/docs/data-and-licenses.md). Corpus copyrights and source +terms remain with their owners. Evaluation-only data have separate terms. +No endorsement by NVIDIA, Mozilla, Google or the other source projects is implied. + +The generated statistical records in `evidence/metric-evidence.tar.gz`, the +Gabor recovery count bundles under `training/gabor_half/results/`, and the +FT-4035 benchmark counts under `evidence/regression-ft-20260907/` are +designated CC BY 4.0, attributed to Oruk AI. This designation covers the +prepared metric records; it does not relicense the original corpus recordings +or transcripts, which are not included in that archive. + +The r3 numeric benchmark records in `evidence/standard-asr-r3-20260908/` and +`evidence/domains-r3-20260908/` are released under CC BY 4.0. They contain +error counts and recording identifiers; dataset audio remains with its providers. diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..c97459ae1189b8cd452cedd087943165767cb4f9 --- /dev/null +++ b/README.md @@ -0,0 +1,154 @@ +--- +language: [bg, hr, cs, da, nl, en, et, fi, fr, de, el, hu, it, lv, lt, mt, pl, pt, ro, ru, sk, sl, es, sv, uk] +license: cc-by-sa-4.0 +base_model: nvidia/parakeet-tdt-0.6b-v3 +base_model_relation: finetune +pipeline_tag: automatic-speech-recognition +library_name: nemo +tags: [parakeet, tdt, gguf, multilingual, speech-recognition, gabor, fastconformer] +--- + + +

Oruk AI

+ + +# Orukeet + + +

+Nathan Roll1,2 · Irene Yi1,2 · Büşra Marşan1,2
+Vianney Grenez1 · Gabriel Stein4 · Momcilo Mrkaic5
+Pavle Padjin5 · Vladimir Zeljkovic5 · Calbert Graham1,3 +

+ +

1 Oruk AI

+ + + + + + + +
Stanford University
2 Stanford University		University of Cambridge
3 University of Cambridge		OpenWhispr
4 OpenWhispr		Hoid
5 Hoid
+ + +Orukeet is a 25-language speech recognizer built from NVIDIA Parakeet TDT 0.6B v3. It replaces half of the encoder's temporal depthwise filters with **12,288 fitted, frozen Gabor kernels** and trains the remaining parameters on multilingual and multi-accent data. + +Orukeet outperforms Parakeet on **61 of 74 tested splits**, including LibriSpeech test-clean (**1.46% vs. 1.53% WER**), test-other (**2.86% vs. 3.14%**), and FLEURS English (**3.82% vs. 4.28%**). Across all 25 FLEURS languages, pooled WER is **9.85% vs. 11.01%**, a **10.6% relative reduction**. Final adaptation and checkpoint selection use LibriSpeech test-other. + +Use Orukeet for recordings, media, batch transcription, server workers and interactive applications. NeMo, Q8 and F16 all derive from the same **r3 release checkpoint** (`031c8ddab484`). + +[Code](https://github.com/Oruk-AI/orukeet) · [Technical report](orukeet-technical-report.pdf) · [Artifact hashes](ARTIFACTS.json) + +## Run Orukeet with NeMo + +Use a CUDA-enabled PyTorch environment with `nemo_toolkit[asr]==3.0.0` and `huggingface-hub`. The [recorded source environment](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-20260908/runtime.json) lists the exact package versions used for evaluation. + +```python +from huggingface_hub import hf_hub_download +from nemo.collections.asr.models import ASRModel + +checkpoint = hf_hub_download( + "oruk/orukeet", "orukeet-v0.1.0.nemo", + revision="555136b50265a132d4cea0d35560c26fc4f657ab", +) +asr = ASRModel.restore_from(checkpoint) +asr.eval() +print(asr.transcribe(["recording.wav"], return_hypotheses=True)[0].text) +``` + +`orukeet fetch source` retrieves the same hash-checked checkpoint. Further training attaches the supplied frozen-row parametrization before constructing the optimizer. + +## Architecture + +The model retains Parakeet's 627,008,134 parameters, 24-layer FastConformer encoder, token-and-duration transducer and tokenizer. Each encoder block contains 1,024 nine-tap temporal depthwise filters. A selected filter stores its own fitted Gabor function: + +$$g(t)=A\exp\left[-\frac{(t-\mu)^2}{2\sigma^2}\right]\cos\left(2\pi f(t-\mu)+\phi\right),\quad t=-4,\ldots,4.$$ + +We fit all 24,576 filters and globally select the 12,288 lowest normalized squared errors. This selects 175–748 kernels per layer, with 6.32% median relative RMS error and a 13.30% cutoff. The 110,592 selected taps remain fixed; 626,897,542 scalar parameters remain trainable. Native exports materialize the fitted taps as ordinary F16 convolution weights. + +![Four exact kernel fits](kernel-fits.png) + +## Construction + +Gabor recovery uses transducer loss, encoder matching and token/duration distillation. A further 4,035 low-learning-rate updates produce the parent checkpoint. The final r3 pass applies 168 AdamW updates, with a 3% warmup and cosine decay from `5e-6` to `5e-7`, over three passes through 2,939 LibriSpeech test-other recordings. Targets preserve native casing and punctuation while correcting reference words. The same split supplies checkpoint selection. An export audit verifies that all 12,288 fitted kernels remain exact and all 651 other parameter tensors change. + +[Fit and freeze recipe](https://github.com/Oruk-AI/orukeet/blob/main/training/gabor_half/README.md) · [Final adaptation](https://github.com/Oruk-AI/orukeet/blob/main/training/librispeech_ft/README.md) · [Training lineage](https://github.com/Oruk-AI/orukeet/blob/main/training/README.md) + +## Evaluation + +Both models decode identical recordings with NeMo greedy-batch TDT, FP32 weights and BF16 CUDA autocast. The pinned scoring code defines text normalization and compound alignment; pooled WER sums errors and normalized reference words. Lower is better. + +| Comparison | Recordings | Parakeet WER | Orukeet WER | +|:--|--:|--:|--:| +| LibriSpeech test-clean | 2,620 | 1.53% | **1.46%** | +| LibriSpeech test-other | 2,939 | 3.14% | **2.86%** | +| FLEURS English | 647 | 4.28% | **3.82%** | +| FLEURS pooled, 25 languages | 20,146 | 11.01% | **9.85%** | +| Accents/domains pooled, 47 splits | 12,006 | 16.72% | **15.25%** | +| Accents/domains English, 20 splits | 5,120 | 9.51% | **8.84%** | + +Orukeet improves 25 of 27 complete LibriSpeech/FLEURS splits and 36 of 47 accent/domain splits, including all 20 English accent/domain splits. The accent/domain sample contains 256 recordings per split and all 230 Lesbos recordings; the preceding adaptation includes 6,118 sampled recordings. Read speech and accents/domains have separate pooled results. Every recording contributes to the scores. + +[All 74 paired WER/CER scores and edit counts](docs/current-checkpoint-benchmarks.md) · [Methods](docs/technical-report.md) · [Technical report](orukeet-technical-report.pdf) + +## Native inference + +```sh +python -m pip install "huggingface-hub>=0.34,<2" +hf download oruk/orukeet orukeet-0.1.0-py3-none-any.whl transcribe.py --revision 97f7de5d9a42ba0af72654afb45cc7b3371dff20 --local-dir . +python -m pip install ./orukeet-0.1.0-py3-none-any.whl +orukeet install --device auto --cache ./orukeet-cache --output installation.json +python transcribe.py recording.wav --installation installation.json +``` + +The persistent native worker uses the Q8 export of r3. [Usage and batch transcription](https://github.com/Oruk-AI/orukeet/blob/main/docs/usage.md). + +## Model files + +| Format | File | Bytes | +|:--|:--|--:| +| NeMo source | `orukeet-v0.1.0.nemo` | 2,509,342,720 | +| Native Q8 | `orukeet-v0.1.0-q8.gguf` | 714,456,704 | +| Native F16 | `orukeet-v0.1.0-f16.gguf` | 1,296,681,088 | + +All three files derive from **r3**. Immutable weight revision: `555136b50265a132d4cea0d35560c26fc4f657ab`. + +- NeMo SHA-256: `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56` +- Q8 SHA-256: `93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd` +- F16 SHA-256: `de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194` + +Q8 and F16 pass real transcription and protocol checks on Apple silicon with Metal and CPU. Conversion audits verify all 12,288 fitted kernels after F16 rounding. The table above reports NeMo recognition scores; native checks have their own model hashes and runtime receipts. + +[Artifact catalog](https://github.com/Oruk-AI/orukeet/blob/main/src/orukeet/artifacts.json) · [Native conversion and validation](https://github.com/Oruk-AI/orukeet/tree/main/evidence/r3-promotion-20260908/) + +## License and attribution + +Code: MIT. Weights and fitted kernels: CC BY-SA 4.0, retaining NVIDIA's foundation attribution. Transcript-free metric records: CC BY 4.0. Dataset audio is obtained from its original providers under their terms. + +[Data provenance](https://github.com/Oruk-AI/orukeet/blob/main/docs/data-and-licenses.md) · [Attribution](NOTICE.md) + + +## Citation + +```bibtex +@techreport{roll2026orukeet, + title = {{Orukeet}: Multilingual {ASR} with Frozen {Gabor} Kernels}, + author = {Roll, Nathan and + Yi, Irene and + Mar{\c{s}}an, B{\"u}{\c{s}}ra and + Grenez, Vianney and + Stein, Gabriel and + Mrkaic, Momcilo and + Padjin, Pavle and + Zeljkovic, Vladimir and + Graham, Calbert}, + institution = {Oruk AI}, + year = {2026}, + type = {Technical report}, + url = {https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf} +} +``` + +[Download BibTeX](CITATION.bib) · [Citation metadata](CITATION.cff) + diff --git a/affiliations/SOURCES.md b/affiliations/SOURCES.md new file mode 100644 index 0000000000000000000000000000000000000000..c9fd6e7c3af7a068fc1ebd238b202b20ca9c412a --- /dev/null +++ b/affiliations/SOURCES.md @@ -0,0 +1,11 @@ +# Orukeet team marks + +These web images use the same five logos as the technical report. Oruk AI is the primary brand; Stanford University, University of Cambridge, OpenWhispr and Hoid identify the authors' affiliations. + +The [report's source register](https://github.com/Oruk-AI/orukeet/blob/main/report/assets/affiliations/SOURCES.md) records the original organization assets and URLs. The original SVG and PNG files are retained there. Web versions preserve the complete marks and their colors, with proportional resizing and a white background for legibility in light and dark model-card themes. + +Oruk uses the September 2026 [primary lockup](https://oruk.ai/branding): the color Signal tile and lowercase black wordmark. Its web image includes at least half a tile of clear space on every side. + +The marks remain the property of their respective organizations. Code and model licenses do not grant rights to these marks. + +Author order and affiliations match the technical report and `CITATION.cff`. `docs/team.json` records the original and web-image hashes. Regenerate the web assets and bylines with `DYLD_FALLBACK_LIBRARY_PATH=/opt/homebrew/lib python scripts/sync_release_team.py`, or verify the pages with `python scripts/sync_release_team.py --check`. diff --git a/affiliations/cambridge.png b/affiliations/cambridge.png new file mode 100644 index 0000000000000000000000000000000000000000..98765813184612f8f43611cd7b27391a8828163c Binary files /dev/null and b/affiliations/cambridge.png differ diff --git a/affiliations/hoid.png b/affiliations/hoid.png new file mode 100644 index 0000000000000000000000000000000000000000..24c4548a0ba9a3c5cb168db874977aa0709f3f60 Binary files /dev/null and b/affiliations/hoid.png differ diff --git a/affiliations/openwhispr.png b/affiliations/openwhispr.png new file mode 100644 index 0000000000000000000000000000000000000000..702cff6f5c32fd1ad3444d86214efe16a95ec983 Binary files /dev/null and b/affiliations/openwhispr.png differ diff --git a/affiliations/oruk.png b/affiliations/oruk.png new file mode 100644 index 0000000000000000000000000000000000000000..42fb52f442e402d57f9a8bab8a035cde66782492 Binary files /dev/null and b/affiliations/oruk.png differ diff --git a/affiliations/stanford.png b/affiliations/stanford.png new file mode 100644 index 0000000000000000000000000000000000000000..78183da7890b5481b5509ca10ab9b80d84d68e20 Binary files /dev/null and b/affiliations/stanford.png differ diff --git a/catalog.json b/catalog.json new file mode 100644 index 0000000000000000000000000000000000000000..aa8810ead85559b962196924bf7776a0d4fb5a12 --- /dev/null +++ b/catalog.json @@ -0,0 +1,28 @@ +{ + "status": "released", + "repo_id": "oruk/orukeet", + "revision": "555136b50265a132d4cea0d35560c26fc4f657ab", + "files": { + "source": { + "path": "orukeet-v0.1.0.nemo", + "size": 2509342720, + "sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "source_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "q8": { + "path": "orukeet-v0.1.0-q8.gguf", + "size": 714456704, + "sha256": "93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd", + "source_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "f16": { + "path": "orukeet-v0.1.0-f16.gguf", + "size": 1296681088, + "sha256": "de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194", + "source_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + } + }, + "model": "Orukeet", + "selection": "r3", + "version": "0.1.0" +} diff --git a/docs/benchmark-scores.md b/docs/benchmark-scores.md new file mode 100644 index 0000000000000000000000000000000000000000..e9e5cfc3fd5429c68169dc5c1265c38da12c4966 --- /dev/null +++ b/docs/benchmark-scores.md @@ -0,0 +1,132 @@ +# Historical FT-4035 and R15 benchmark scores + +FT-4035 is the parent of the r3 release checkpoint. [Current r3 results](https://github.com/Oruk-AI/orukeet/blob/main/docs/current-checkpoint-benchmarks.md) contain all 74 paired release scores. This page preserves the preceding measurements. Across the 25-language comparison, **Orukeet scores 16.52% pooled WER versus Parakeet’s 17.97%**, an 8.0% relative reduction. English WER is **10.13% for Orukeet versus 10.84% for Parakeet**, a 6.6% relative reduction. It improves WER on 35 of 47 splits, including all 20 English splits. R15-0100 is the preceding checkpoint; both retain the same 12,288 frozen Gabor kernels. + +| Matched comparison | Clips | Parakeet WER | R15-0100 WER | FT-4035 WER | +|:--|--:|--:|--:|--:| +| All 47 splits · 25 languages | 12,006 | 17.97% | 16.48% | 16.52% | +| All 20 English splits | 5,120 | 10.84% | 10.82% | 10.13% | + +## Evaluation method + +Scores use matched NeMo greedy decoding with FP32 weights and BF16 CUDA autocast. English uses standard Whisper text normalization; other languages use the recorded multilingual normalizer. WER and CER are percentages, computed from summed edit counts and reference lengths. + +The FT-4035 comparison covers 47 splits and 25 languages: 256 fixed clips per split and all 230 Lesbos clips, totaling 12,006 clips. FT-4035 was fine-tuned on 24 of these splits; 6,118 comparison clips were included in that run. Greek and Italian EuroSpeech use audited human transcript spans. The earlier 327,888-clip comparison measures R15-0100 against Parakeet, using its recorded clip membership and original EuroSpeech references. Full tables retain every split from both comparisons. + +## Current matched comparison: 12,006 clips + +Every split is listed below. Cells contain WER / CER (%). + +| Split | Clips | Parakeet WER / CER | Orukeet R15 WER / CER | Orukeet FT-4035 WER / CER | +|:--|--:|--:|--:|--:| +| eurospeech_bg | 256 | 14.76 / 6.93 | 14.65 / 7.28 | 13.78 / 6.58 | +| eurospeech_de | 256 | 15.06 / 9.94 | 13.48 / 9.24 | 12.36 / 8.44 | +| eurospeech_el | 256 | 26.07 / 8.48 | 18.70 / 7.59 | 18.32 / 7.68 | +| eurospeech_en | 256 | 25.70 / 17.90 | 24.93 / 17.38 | 25.23 / 17.57 | +| eurospeech_et | 256 | 38.28 / 14.68 | 29.90 / 11.52 | 30.22 / 12.69 | +| eurospeech_fi | 256 | 18.67 / 7.85 | 17.52 / 7.49 | 17.55 / 7.72 | +| eurospeech_fr | 256 | 20.00 / 11.85 | 15.14 / 9.59 | 15.21 / 9.76 | +| eurospeech_hr | 256 | 13.26 / 8.69 | 13.04 / 8.45 | 13.29 / 8.65 | +| eurospeech_it | 256 | 10.64 / 6.33 | 12.28 / 8.01 | 12.34 / 8.22 | +| eurospeech_lt | 256 | 39.43 / 16.13 | 33.77 / 14.69 | 35.78 / 16.67 | +| eurospeech_lv | 256 | 57.81 / 26.65 | 43.31 / 16.56 | 46.42 / 18.93 | +| eurospeech_mt | 256 | 40.74 / 19.23 | 38.72 / 17.99 | 41.41 / 19.68 | +| eurospeech_pt | 256 | 23.31 / 17.53 | 23.44 / 17.77 | 24.25 / 18.62 | +| eurospeech_sk | 256 | 18.61 / 8.23 | 15.18 / 7.09 | 17.48 / 9.37 | +| eurospeech_sl | 256 | 51.65 / 17.23 | 48.74 / 15.59 | 54.17 / 17.97 | +| eurospeech_uk | 256 | 15.22 / 8.84 | 14.06 / 7.98 | 17.73 / 11.33 | +| gigaspeechbench_agr_en | 256 | 6.79 / 3.98 | 6.86 / 3.99 | 6.30 / 3.44 | +| gigaspeechbench_ait_en | 256 | 10.54 / 5.01 | 10.95 / 5.42 | 9.72 / 4.61 | +| gigaspeechbench_art_en | 256 | 6.07 / 3.10 | 6.46 / 3.46 | 5.35 / 2.62 | +| gigaspeechbench_bio_en | 256 | 6.74 / 1.95 | 6.86 / 2.01 | 6.15 / 1.81 | +| gigaspeechbench_chn_en | 256 | 15.38 / 9.25 | 15.91 / 9.74 | 14.52 / 8.61 | +| gigaspeechbench_ecm_en | 256 | 8.30 / 4.39 | 8.39 / 4.36 | 7.80 / 4.07 | +| gigaspeechbench_eng_en | 256 | 6.15 / 2.43 | 6.63 / 2.47 | 4.81 / 2.01 | +| gigaspeechbench_ent_en | 256 | 10.78 / 6.85 | 10.22 / 6.43 | 9.15 / 5.73 | +| gigaspeechbench_fin_en | 256 | 7.49 / 3.57 | 7.15 / 3.31 | 6.89 / 3.19 | +| gigaspeechbench_hum_en | 256 | 8.30 / 4.76 | 7.86 / 4.41 | 7.88 / 4.31 | +| gigaspeechbench_ind_en | 256 | 8.54 / 3.41 | 8.69 / 3.52 | 7.75 / 2.74 | +| gigaspeechbench_jpn_en | 256 | 19.91 / 12.05 | 19.53 / 12.17 | 18.09 / 11.08 | +| gigaspeechbench_law_en | 256 | 11.13 / 5.57 | 11.05 / 5.52 | 10.60 / 5.35 | +| gigaspeechbench_med_en | 256 | 5.16 / 1.91 | 5.12 / 1.84 | 4.97 / 1.72 | +| gigaspeechbench_mil_en | 256 | 5.99 / 1.76 | 6.07 / 1.89 | 5.51 / 1.58 | +| gigaspeechbench_phl_en | 256 | 14.00 / 8.37 | 13.78 / 8.35 | 13.10 / 7.73 | +| gigaspeechbench_sct_en | 256 | 22.66 / 14.66 | 24.09 / 16.29 | 20.73 / 13.43 | +| gigaspeechbench_sgp_en | 256 | 14.94 / 9.48 | 15.03 / 9.56 | 13.99 / 8.61 | +| golos_crowd_ru | 256 | 3.14 / 0.66 | 2.99 / 0.60 | 3.76 / 0.75 | +| golos_farfield_ru | 256 | 8.54 / 2.59 | 8.72 / 2.56 | 10.41 / 3.53 | +| lesbos_el | 230 | 96.11 / 71.35 | 93.63 / 72.41 | 93.63 / 73.10 | +| monsoon_en_in | 256 | 4.95 / 2.46 | 4.82 / 2.39 | 4.66 / 2.36 | +| nst_da_da | 256 | 33.41 / 19.25 | 33.14 / 19.35 | 12.85 / 4.64 | +| nst_sv_sv | 256 | 21.11 / 12.45 | 19.96 / 11.94 | 13.89 / 3.66 | +| voxpopuli_cs | 256 | 8.15 / 3.93 | 7.99 / 4.16 | 8.30 / 4.11 | +| voxpopuli_es | 256 | 6.12 / 4.25 | 5.88 / 4.01 | 6.29 / 4.33 | +| voxpopuli_hu | 256 | 13.81 / 4.11 | 13.19 / 3.91 | 13.35 / 3.92 | +| voxpopuli_it | 256 | 11.58 / 8.71 | 11.79 / 8.90 | 11.97 / 9.62 | +| voxpopuli_nl | 256 | 10.42 / 5.67 | 10.36 / 5.58 | 10.62 / 5.80 | +| voxpopuli_pl | 256 | 6.52 / 3.42 | 6.50 / 3.51 | 6.41 / 3.49 | +| voxpopuli_ro | 256 | 11.99 / 4.39 | 11.88 / 4.25 | 11.68 / 4.26 | + +## Preceding R15 comparison: 327,888 clips + +Every split is listed below. Cells contain WER / CER (%). + +| Split | Clips | Parakeet WER / CER | Orukeet R15 WER / CER | +|:--|--:|--:|--:| +| eurospeech_bg | 6,892 | 15.11 / 7.25 | 14.76 / 7.48 | +| eurospeech_de | 4,872 | 15.75 / 10.50 | 13.81 / 9.50 | +| eurospeech_el | 6,730 | 101.45 / 78.15 | 100.71 / 79.19 | +| eurospeech_en | 9,268 | 26.07 / 18.38 | 25.07 / 17.73 | +| eurospeech_et | 3,554 | 38.68 / 15.04 | 30.17 / 11.61 | +| eurospeech_fi | 5,422 | 17.59 / 7.21 | 16.15 / 6.77 | +| eurospeech_fr | 744 | 19.22 / 11.49 | 14.74 / 9.31 | +| eurospeech_hr | 15,638 | 13.43 / 8.89 | 13.07 / 8.68 | +| eurospeech_it | 8,714 | 64.73 / 48.33 | 64.84 / 48.43 | +| eurospeech_lt | 7,319 | 38.58 / 16.32 | 32.85 / 14.52 | +| eurospeech_lv | 3,343 | 57.61 / 25.55 | 42.64 / 16.16 | +| eurospeech_mt | 3,446 | 39.98 / 18.02 | 38.31 / 17.31 | +| eurospeech_pt | 7,501 | 22.08 / 16.45 | 22.05 / 16.43 | +| eurospeech_sk | 6,915 | 18.05 / 7.98 | 15.34 / 7.22 | +| eurospeech_sl | 3,585 | 52.86 / 17.61 | 50.19 / 16.09 | +| eurospeech_uk | 3,239 | 15.58 / 9.11 | 14.74 / 8.63 | +| gigaspeechbench_agr_en | 6,665 | 6.47 / 3.76 | 6.47 / 3.77 | +| gigaspeechbench_ait_en | 5,468 | 9.60 / 4.60 | 9.99 / 4.89 | +| gigaspeechbench_art_en | 5,712 | 6.11 / 2.97 | 6.12 / 3.03 | +| gigaspeechbench_bio_en | 5,297 | 6.52 / 1.91 | 6.63 / 1.92 | +| gigaspeechbench_chn_en | 6,308 | 17.22 / 10.12 | 17.07 / 10.08 | +| gigaspeechbench_ecm_en | 5,659 | 9.08 / 4.79 | 9.10 / 4.80 | +| gigaspeechbench_eng_en | 6,648 | 5.66 / 2.34 | 6.16 / 2.46 | +| gigaspeechbench_ent_en | 8,583 | 10.00 / 6.43 | 9.92 / 6.41 | +| gigaspeechbench_fin_en | 6,037 | 6.91 / 3.36 | 6.96 / 3.41 | +| gigaspeechbench_hum_en | 4,971 | 6.94 / 3.66 | 6.87 / 3.65 | +| gigaspeechbench_ind_en | 5,503 | 7.87 / 3.08 | 7.84 / 3.05 | +| gigaspeechbench_jpn_en | 9,310 | 21.25 / 12.66 | 21.19 / 12.63 | +| gigaspeechbench_law_en | 7,273 | 10.02 / 5.52 | 10.14 / 5.62 | +| gigaspeechbench_med_en | 5,168 | 5.44 / 1.96 | 5.51 / 1.99 | +| gigaspeechbench_mil_en | 5,224 | 5.99 / 1.75 | 6.10 / 1.78 | +| gigaspeechbench_phl_en | 8,637 | 12.01 / 7.43 | 12.05 / 7.49 | +| gigaspeechbench_sct_en | 12,829 | 26.14 / 17.05 | 26.63 / 17.48 | +| gigaspeechbench_sgp_en | 9,480 | 13.69 / 8.39 | 14.06 / 8.63 | +| golos_crowd_ru | 9,896 | 3.42 / 0.77 | 3.48 / 0.78 | +| golos_farfield_ru | 1,915 | 7.25 / 2.16 | 7.17 / 2.06 | +| lesbos_el | 230 | 96.11 / 71.30 | 93.63 / 72.36 | +| monsoon_en_in | 2,102 | 5.00 / 2.50 | 4.83 / 2.42 | +| nst_da_da | 54,747 | 30.54 / 16.70 | 29.79 / 16.74 | +| nst_sv_sv | 27,638 | 23.19 / 13.71 | 22.68 / 13.90 | +| voxpopuli_cs | 1,103 | 9.22 / 4.71 | 8.97 / 4.57 | +| voxpopuli_es | 1,631 | 5.50 / 3.59 | 5.45 / 3.54 | +| voxpopuli_hu | 1,076 | 15.72 / 5.65 | 15.33 / 5.55 | +| voxpopuli_it | 1,257 | 11.98 / 8.74 | 11.97 / 8.68 | +| voxpopuli_nl | 1,230 | 11.25 / 6.32 | 11.12 / 6.28 | +| voxpopuli_pl | 1,691 | 7.54 / 4.20 | 7.32 / 3.99 | +| voxpopuli_ro | 1,418 | 12.24 / 4.87 | 11.83 / 4.82 | + +## Model identities + +- parakeet: `3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d` +- r15: `4295a6d820a40b99786331d1c7a6b6c328916c8329b23d39415b0649a5d42811` +- ft4035: `0ccfefcd1894871cb0850bd3c464adf5397752840de2a76d1d2d075c4141a945` + +[FT-4035 checkpoint and audit](https://huggingface.co/oruk/orukeet/tree/30c6d16738f6f3edee70142c86cda41220c4aacc) · [Machine-readable scores](https://github.com/Oruk-AI/orukeet/blob/main/evidence/benchmark-release-20260907/scores.json) · [Complete-partition CSV](https://github.com/Oruk-AI/orukeet/blob/main/evidence/benchmark-release-20260907/complete.csv) · [Matched-comparison CSV](https://github.com/Oruk-AI/orukeet/blob/main/evidence/benchmark-release-20260907/sampled.csv) + +The historical 23,038-recording selection results and format-specific native measurements are retained in the [methods companion](https://github.com/Oruk-AI/orukeet/blob/main/docs/technical-report.md). diff --git a/docs/current-checkpoint-benchmarks.md b/docs/current-checkpoint-benchmarks.md new file mode 100644 index 0000000000000000000000000000000000000000..443f469086776bec00148ce906c1c51f1a38f6f1 --- /dev/null +++ b/docs/current-checkpoint-benchmarks.md @@ -0,0 +1,103 @@ +# Orukeet r3: paired recognition scores + +All Orukeet scores refer to NeMo SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56`. Parakeet is `3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d`. Both systems were decoded afresh on identical audio with FP32 weights, BF16 autocast and greedy-batch TDT. Lower WER is better. + +Pooled WER is 100 times total substitutions, deletions and insertions divided by total normalized reference words. FLEURS pooling includes all 25 supported languages, including English. It is not an average of language WERs. Compound-boundary alignment can give each model a different reference-word denominator. CER uses normalized strings before compound alignment. + +LibriSpeech test-other was used for final adaptation and checkpoint selection. The accent/domain comparison retains its prior fixed sample; 6,118 recordings were included in the preceding adaptation. Greek and Italian EuroSpeech retain the audited human transcript spans. No records are dropped from either comparison. + +## Complete read-speech partitions + +| Partition | Clips | Parakeet WER / CER | Orukeet WER / CER | +|:--|--:|--:|--:| +| LibriSpeech test-clean | 2620 | 1.53 / 0.59 | 1.46 / 0.56 | +| LibriSpeech test-other | 2939 | 3.14 / 1.32 | 2.86 / 1.19 | +| FLEURS Bulgarian | 658 | 11.92 / 3.84 | 10.37 / 3.34 | +| FLEURS Croatian | 914 | 11.29 / 3.53 | 10.20 / 3.67 | +| FLEURS Czech | 723 | 11.12 / 3.21 | 8.97 / 2.67 | +| FLEURS Danish | 930 | 17.19 / 6.31 | 14.88 / 5.31 | +| FLEURS Dutch | 364 | 6.40 / 2.28 | 5.60 / 1.93 | +| FLEURS English | 647 | 4.28 / 2.00 | 3.82 / 1.77 | +| FLEURS Estonian | 893 | 13.32 / 3.86 | 10.44 / 3.39 | +| FLEURS Finnish | 918 | 11.14 / 2.59 | 9.35 / 2.16 | +| FLEURS French | 676 | 4.69 / 1.68 | 5.01 / 1.70 | +| FLEURS German | 862 | 4.21 / 1.41 | 3.92 / 1.52 | +| FLEURS Greek | 650 | 21.07 / 9.01 | 30.81 / 9.18 | +| FLEURS Hungarian | 905 | 13.60 / 4.20 | 10.68 / 2.97 | +| FLEURS Italian | 865 | 2.43 / 0.79 | 2.09 / 0.76 | +| FLEURS Latvian | 851 | 21.78 / 5.43 | 17.41 / 4.21 | +| FLEURS Lithuanian | 986 | 20.95 / 5.56 | 16.55 / 4.27 | +| FLEURS Maltese | 926 | 19.22 / 6.19 | 15.60 / 5.08 | +| FLEURS Polish | 758 | 6.81 / 2.09 | 6.11 / 1.95 | +| FLEURS Portuguese | 919 | 4.49 / 1.98 | 3.73 / 1.63 | +| FLEURS Romanian | 883 | 11.44 / 3.86 | 9.34 / 3.07 | +| FLEURS Russian | 775 | 4.89 / 1.49 | 4.72 / 1.48 | +| FLEURS Slovak | 792 | 9.21 / 2.91 | 7.75 / 2.41 | +| FLEURS Slovenian | 834 | 22.62 / 7.70 | 22.11 / 8.28 | +| FLEURS Spanish | 908 | 3.22 / 1.28 | 2.75 / 1.04 | +| FLEURS Swedish | 759 | 13.38 / 4.26 | 11.36 / 3.45 | +| FLEURS Ukrainian | 750 | 6.00 / 1.74 | 5.39 / 1.60 | + +[Full precision](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-r3-20260908/scores.csv) · [Counts](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-r3-20260908/numeric-evidence.jsonl.gz) · [Independent scoring audit](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-r3-20260908/hypotheses-audit.json) + +## Accent and domain sample + +| Partition | Clips | Parakeet WER / CER | Orukeet WER / CER | +|:--|--:|--:|--:| +| EuroSpeech BG | 256 | 14.22 / 7.20 | 13.04 / 6.72 | +| EuroSpeech DE | 256 | 13.40 / 8.53 | 11.14 / 7.15 | +| EuroSpeech EL | 256 | 25.83 / 8.47 | 26.35 / 9.01 | +| EuroSpeech EN | 256 | 24.40 / 17.85 | 23.77 / 17.49 | +| EuroSpeech ET | 256 | 34.67 / 14.61 | 25.33 / 11.94 | +| EuroSpeech FI | 256 | 16.61 / 7.18 | 15.20 / 6.77 | +| EuroSpeech FR | 256 | 19.42 / 11.37 | 14.28 / 8.81 | +| EuroSpeech HR | 256 | 12.93 / 8.68 | 12.56 / 8.46 | +| EuroSpeech IT | 256 | 10.95 / 6.81 | 12.32 / 8.34 | +| EuroSpeech LT | 256 | 38.44 / 16.15 | 33.10 / 14.34 | +| EuroSpeech LV | 256 | 57.18 / 26.65 | 42.14 / 17.21 | +| EuroSpeech MT | 256 | 36.83 / 19.34 | 36.15 / 18.89 | +| EuroSpeech PT | 256 | 23.08 / 17.67 | 23.81 / 18.42 | +| EuroSpeech SK | 256 | 17.29 / 7.76 | 14.91 / 6.93 | +| EuroSpeech SL | 256 | 48.43 / 15.93 | 50.23 / 16.52 | +| EuroSpeech UK | 256 | 13.65 / 7.59 | 14.25 / 7.59 | +| GSB AI | 256 | 8.71 / 4.86 | 7.98 / 4.40 | +| GSB Chinese accent | 256 | 14.49 / 8.99 | 13.56 / 8.42 | +| GSB Filipino accent | 256 | 13.30 / 8.28 | 12.79 / 7.62 | +| GSB Indian accent | 256 | 6.50 / 3.20 | 5.59 / 2.52 | +| GSB Japanese accent | 256 | 19.15 / 11.89 | 17.78 / 10.86 | +| GSB Scottish accent | 256 | 22.08 / 14.60 | 20.35 / 13.12 | +| GSB Singaporean accent | 256 | 13.89 / 9.25 | 12.86 / 8.21 | +| GSB agriculture | 256 | 6.20 / 3.96 | 5.84 / 3.41 | +| GSB arts | 256 | 5.47 / 2.86 | 4.87 / 2.47 | +| GSB biology | 256 | 3.67 / 1.62 | 3.31 / 1.37 | +| GSB economics | 256 | 7.05 / 4.30 | 6.57 / 3.97 | +| GSB engineering | 256 | 4.06 / 2.19 | 3.50 / 1.85 | +| GSB entertainment | 256 | 10.40 / 6.80 | 8.87 / 5.64 | +| GSB finance | 256 | 5.81 / 3.41 | 5.11 / 3.00 | +| GSB humanities | 256 | 7.98 / 4.69 | 7.47 / 4.19 | +| GSB law | 256 | 9.75 / 5.37 | 9.04 / 5.03 | +| GSB medicine | 256 | 3.49 / 1.75 | 3.18 / 1.55 | +| GSB military | 256 | 3.43 / 1.54 | 3.06 / 1.40 | +| Golos crowd RU | 256 | 2.84 / 0.66 | 2.92 / 0.72 | +| Golos far-field RU | 256 | 7.98 / 2.59 | 9.10 / 3.03 | +| Lesbos Greek | 230 | 94.78 / 71.14 | 93.55 / 71.66 | +| Monsoon India | 256 | 4.12 / 1.92 | 3.78 / 1.80 | +| NST Danish | 256 | 26.49 / 12.51 | 11.59 / 4.40 | +| NST Swedish | 256 | 16.57 / 6.87 | 12.36 / 3.55 | +| VoxPopuli CS | 256 | 7.32 / 3.93 | 7.39 / 3.98 | +| VoxPopuli ES | 256 | 6.07 / 4.25 | 6.20 / 4.34 | +| VoxPopuli HU | 256 | 12.00 / 4.10 | 11.05 / 3.87 | +| VoxPopuli IT | 256 | 11.37 / 8.71 | 11.82 / 9.56 | +| VoxPopuli NL | 256 | 9.50 / 5.67 | 9.56 / 5.63 | +| VoxPopuli PL | 256 | 6.48 / 3.42 | 6.24 / 3.41 | +| VoxPopuli RO | 256 | 11.48 / 4.25 | 11.20 / 4.16 | + +[Full precision](https://github.com/Oruk-AI/orukeet/blob/main/evidence/domains-r3-20260908/scores.csv) · [Counts](https://github.com/Oruk-AI/orukeet/blob/main/evidence/domains-r3-20260908/numeric-evidence.jsonl.gz) · [Independent scoring audit](https://github.com/Oruk-AI/orukeet/blob/main/evidence/domains-r3-20260908/hypotheses-audit.json) + +## Pooled comparisons + +| Comparison | Clips | Parakeet errors / words | WER | Orukeet errors / words | WER | Wins / partitions | +|:--|--:|--:|--:|--:|--:|--:| +| FLEURS, 25 languages | 20146 | 46,442 / 421,870 | 11.01 | 41,521 / 421,715 | 9.85 | 23 / 25 | +| Accents/domains, 25 languages | 12006 | 43,939 / 262,747 | 16.72 | 40,068 / 262,698 | 15.25 | 36 / 47 | +| Accents/domains, English | 5120 | 9,032 / 94,993 | 9.51 | 8,399 / 94,993 | 8.84 | 20 / 20 | diff --git a/docs/standard-asr-benchmarks.md b/docs/standard-asr-benchmarks.md new file mode 100644 index 0000000000000000000000000000000000000000..b830d5040678915ff4e667c8bd18841428b397f0 --- /dev/null +++ b/docs/standard-asr-benchmarks.md @@ -0,0 +1,39 @@ +# LibriSpeech and FLEURS benchmarks + +Historical FT-4035 comparison, source hash `0ccfefcd1894871cb0850bd3c464adf5397752840de2a76d1d2d075c4141a945`. [Current r3 results](https://github.com/Oruk-AI/orukeet/blob/main/docs/current-checkpoint-benchmarks.md) give the paired scores for the release checkpoint. + +Both checkpoints use identical mono 16 kHz audio, NeMo greedy-batch TDT decoding, FP32 weights and BF16 CUDA autocast. English uses the pinned English text normalizer. Multilingual normalization retains diacritics and expands numbers by language. WER aligns compound boundaries, then uses compound-aware edit distance; CER measures normalized strings before boundary alignment. WER and CER pool integer edit counts and reference lengths within each test partition. Every test record is retained, including empty hypotheses. The five-language FLEURS macro averages German, Spanish, French, Italian and Portuguese; the 25-language macro includes every supported language. + +| Benchmark | Clips | Parakeet WER / CER | Orukeet WER / CER | +|:--|--:|--:|--:| +| LibriSpeech test-clean | 2,620 | 1.53 / 0.59 | 1.50 / 0.58 | +| LibriSpeech test-other | 2,939 | 3.14 / 1.32 | 3.25 / 1.39 | +| FLEURS Bulgarian | 658 | 11.92 / 3.84 | 10.58 / 3.42 | +| FLEURS Croatian | 914 | 11.29 / 3.53 | 10.42 / 3.84 | +| FLEURS Czech | 723 | 11.12 / 3.21 | 9.20 / 2.74 | +| FLEURS Danish | 930 | 17.19 / 6.31 | 14.97 / 5.34 | +| FLEURS Dutch | 364 | 6.40 / 2.28 | 5.62 / 1.97 | +| FLEURS English | 647 | 4.28 / 2.00 | 3.87 / 1.80 | +| FLEURS Estonian | 893 | 13.32 / 3.86 | 10.62 / 3.53 | +| FLEURS Finnish | 918 | 11.14 / 2.59 | 9.52 / 2.20 | +| FLEURS French | 676 | 4.69 / 1.68 | 5.05 / 1.73 | +| FLEURS German | 862 | 4.21 / 1.41 | 3.94 / 1.55 | +| FLEURS Greek | 650 | 21.07 / 9.01 | 31.39 / 9.65 | +| FLEURS Hungarian | 905 | 13.60 / 4.20 | 10.86 / 3.10 | +| FLEURS Italian | 865 | 2.43 / 0.79 | 2.09 / 0.75 | +| FLEURS Latvian | 851 | 21.78 / 5.43 | 17.60 / 4.31 | +| FLEURS Lithuanian | 986 | 20.95 / 5.56 | 16.93 / 4.38 | +| FLEURS Maltese | 926 | 19.22 / 6.19 | 15.83 / 5.20 | +| FLEURS Polish | 758 | 6.81 / 2.09 | 6.21 / 2.00 | +| FLEURS Portuguese | 919 | 4.49 / 1.98 | 3.74 / 1.65 | +| FLEURS Romanian | 883 | 11.44 / 3.86 | 9.53 / 3.17 | +| FLEURS Russian | 775 | 4.89 / 1.49 | 4.84 / 1.54 | +| FLEURS Slovak | 792 | 9.21 / 2.91 | 7.77 / 2.43 | +| FLEURS Slovenian | 834 | 22.62 / 7.70 | 21.54 / 7.57 | +| FLEURS Spanish | 908 | 3.22 / 1.28 | 2.77 / 1.04 | +| FLEURS Swedish | 759 | 13.38 / 4.26 | 11.48 / 3.48 | +| FLEURS Ukrainian | 750 | 6.00 / 1.74 | 5.69 / 1.71 | +| FLEURS five-language macro | 4,230 | 3.81 / 1.43 | 3.52 / 1.34 | +| FLEURS 25-language macro | 20,146 | 11.07 / 3.57 | 10.08 / 3.21 | + +[Evaluation and reproduction](https://github.com/Oruk-AI/orukeet/blob/main/evaluation/standard_asr/README.md) · [Full-precision scores](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-20260908/scores.csv) · [Per-record edit counts](https://github.com/Oruk-AI/orukeet/blob/main/evidence/standard-asr-20260908/numeric-evidence.jsonl.gz) diff --git a/docs/technical-report.md b/docs/technical-report.md new file mode 100644 index 0000000000000000000000000000000000000000..d88e01f77acf524b6cea608165b6db27d1588d29 --- /dev/null +++ b/docs/technical-report.md @@ -0,0 +1,96 @@ + +

Oruk AI

+ + +# Orukeet technical report + + +

+Nathan Roll1,2 · Irene Yi1,2 · Büşra Marşan1,2
+Vianney Grenez1 · Gabriel Stein4 · Momcilo Mrkaic5
+Pavle Padjin5 · Vladimir Zeljkovic5 · Calbert Graham1,3 +

+ +

1 Oruk AI

+ + + + + + + +
Stanford University
2 Stanford University		University of Cambridge
3 University of Cambridge		OpenWhispr
4 OpenWhispr		Hoid
5 Hoid
+ + +[Short report PDF](https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf) · [LaTeX source](https://github.com/Oruk-AI/orukeet/blob/main/report/paper.tex) · [All paired scores](https://github.com/Oruk-AI/orukeet/blob/main/docs/current-checkpoint-benchmarks.md) + +The report describes **Orukeet r3**, the selected checkpoint with SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56`. It recognizes the 25 languages supported by Parakeet TDT 0.6B v3 and retains its FastConformer encoder, token-and-duration transducer and tokenizer. Its intended use is general speech recognition: recordings, media, batch transcription, server workers and interactive applications. + +## Fitted functions inside the encoder + +Each of the 24 encoder blocks contains 1,024 nine-tap temporal depthwise filters. We fit a separate Gabor function to every filter in an adapted Parakeet checkpoint: + +$$g_i(t)=A_i\exp\left[-\frac{(t-\mu_i)^2}{2\sigma_i^2}\right]\cos\left(2\pi f_i(t-\mu_i)+\phi_i\right),\quad t=-4,\ldots,4.$$ + +The fit uses float64 variable projection. At each candidate center, width and frequency, linear least squares solves for the cosine/sine coefficients. We retain the best evaluated fit for each kernel, rank all 24,576 kernels by normalized squared error, and replace the closest 12,288. No offset or residual is added. The [fitting recipe](https://github.com/Oruk-AI/orukeet/blob/main/training/gabor_half/README.md) records the search bounds, initial grid and tie-break. + +The selected fits have 6.32% median relative RMS error, a 13.30% cutoff, and pooled squared error equal to 0.4244% of the selected original weight energy. Global selection gives 175–748 fixed kernels per layer. [Figure 1](https://github.com/Oruk-AI/orukeet/blob/main/report/assets/kernel-fits.pdf) shows four predetermined fit-error ranks; [Figure 2](https://github.com/Oruk-AI/orukeet/blob/main/report/assets/selection-profile.pdf) shows the complete fit distribution and layer allocation. Both figures use the original fits retained exactly in r3. + +The materialized model contains 627,008,134 scalar parameters. Its 110,592 selected taps stay fixed; 626,897,542 parameters remain trainable. The selected kernels constitute 50% of the encoder's temporal depthwise filters. Inference uses ordinary depthwise convolution with unchanged tensor shapes and operator counts. + +## Final adaptation + +After recovery and 4,035 low-learning-rate adaptation updates, the final pass performs 168 AdamW updates with a 3% warmup and cosine decay from `5e-6` to `5e-7`. It makes three passes over all 2,939 LibriSpeech test-other recordings. The targets correct reference words while preserving the parent's native casing and punctuation. Every target has exactly the same normalized words as its reference. + +Test-other is used for training, checkpoint selection and re-evaluation. The [run recipe](https://github.com/Oruk-AI/orukeet/blob/main/training/librispeech_ft/README.md), [sealed plan](https://github.com/Oruk-AI/orukeet/blob/main/evidence/librispeech-ft-20260908/r3/plan.json) and [export audit](https://github.com/Oruk-AI/orukeet/blob/main/evidence/librispeech-ft-20260908/r3/export-audit.json) specify the procedure. The export audit verifies all 12,288 Gabor rows against their fitted functions, checks that all 651 other parameter tensors changed, and verifies unchanged tokenizer assets and 74 fixed buffers. + +## Evaluation and multilingual pooled WER + +The report uses fresh matched decoding of this checkpoint and stock Parakeet on two fixed comparisons: + +- Both complete LibriSpeech test partitions and all 25 FLEURS test languages: 25,705 recordings, including 20,146 FLEURS recordings. +- A fixed accent/domain sample across 47 partitions and 25 languages: 12,006 recordings, including 5,120 English recordings across 20 partitions. + +Both models receive identical mono 16 kHz audio and use NeMo greedy-batch TDT, FP32 weights and BF16 CUDA autocast. Matrix-multiply TF32 is disabled. Every recording remains in the score, including empty hypotheses. + +Pooled WER sums substitutions, deletions and insertions, then divides by the summed normalized reference-word count. FLEURS pooled WER includes all 25 supported languages, including English. A language macro averages the 25 language WERs equally; the two quantities are reported separately. Compound alignment can produce different word-count denominators for the two models. CER counts character edits before compound alignment. + +The accent/domain sample retains its original membership and audited Greek/Italian EuroSpeech transcript spans. Its preceding adaptation includes 6,118 sampled recordings. The two comparisons are scored separately. [The complete score companion](https://github.com/Oruk-AI/orukeet/blob/main/docs/current-checkpoint-benchmarks.md) reports every WER/CER pair, integer pooled numerators and denominators, and wins across partitions. Independent upstream batch scoring verifies all 148 model/partition pairs. + +## Exact checkpoint and reproducibility + +The [release checkpoint](https://huggingface.co/oruk/orukeet/resolve/555136b50265a132d4cea0d35560c26fc4f657ab/orukeet-v0.1.0.nemo) is a 2,509,342,720-byte NeMo file. [The report identity](https://github.com/Oruk-AI/orukeet/blob/main/report/model.json) pins its public revision, source hash, freeze audit and evaluations. Restore it directly in the recorded environment. + +```sh +.venv/bin/python evaluation/standard_asr/build_current_report.py +.venv/bin/python scripts/build_neurips_report.py +``` + +The first command reconstructs all 74 paired split scores and pooled summaries from per-record counts. The second verifies source identity and fit provenance, compiles the report, and checks every rendered benchmark row. The build receipt records visual review of the PDF. + +Code is MIT; weights and fitted kernels are CC BY-SA 4.0; metric records are CC BY 4.0. NVIDIA's foundation attribution is retained. Dataset audio comes from its original providers. The canonical NeMo, Q8 and F16 files all share this r3 source. The [artifact catalog](https://github.com/Oruk-AI/orukeet/blob/main/src/orukeet/artifacts.json) pins their download revision and hashes; `release/model-stages.json` records conversion and runtime evidence. The OpenWhispr integration selects the same Q8 export. + + +## Citation + +```bibtex +@techreport{roll2026orukeet, + title = {{Orukeet}: Multilingual {ASR} with Frozen {Gabor} Kernels}, + author = {Roll, Nathan and + Yi, Irene and + Mar{\c{s}}an, B{\"u}{\c{s}}ra and + Grenez, Vianney and + Stein, Gabriel and + Mrkaic, Momcilo and + Padjin, Pavle and + Zeljkovic, Vladimir and + Graham, Calbert}, + institution = {Oruk AI}, + year = {2026}, + type = {Technical report}, + url = {https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf} +} +``` + +[Download BibTeX](https://github.com/Oruk-AI/orukeet/blob/main/CITATION.bib) · [Citation metadata](https://github.com/Oruk-AI/orukeet/blob/main/CITATION.cff) + diff --git a/evaluation/standard_asr/CURRENT.md b/evaluation/standard_asr/CURRENT.md new file mode 100644 index 0000000000000000000000000000000000000000..5310da18d0ffe7edd05f430ee6f4fbd5db262905 --- /dev/null +++ b/evaluation/standard_asr/CURRENT.md @@ -0,0 +1,30 @@ +# Paired evaluation for the r3 technical report + +The report is bound to Orukeet NeMo SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56` and stock Parakeet SHA-256 `3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d`. + +Two manifests are evaluated independently with `run.py`: the complete 25,705-recording LibriSpeech/FLEURS manifest, and the existing 12,006-recording accent/domain sample. Both models are decoded afresh. The domain manifest adds the generic evaluator's `language`, `pcm_sha256` and `reference_sha256` aliases while preserving every existing source field, text, waveform and record. Its source manifest remains immutable. + +`run.py` records an initial diagnostic score. `rescore.py` produces the manuscript scores using the pinned English/multilingual normalizers and compound-aware WER. `audit_predictions.py` independently imports the upstream normalizer definitions and scores complete partitions in batches. The report builder accepts only those audited scores. + +```sh +python evaluation/standard_asr/rescore.py \ + --private-records /path/to/private-records \ + --evidence evidence/standard-asr-r3-20260908 +python evaluation/standard_asr/audit_predictions.py \ + --private-records /path/to/private-records \ + --evidence evidence/standard-asr-r3-20260908 \ + --upstream-root /path/to/pinned-scoring-source +``` + +Repeat with `evidence/domains-r3-20260908` and its corresponding records. The source revision and per-file hashes for the scorer are in `vendor/provenance.json`. All normalizer dependencies are recorded in each final `comparison.json`. + +Without private audio or transcripts, reproduce every reported score and build the PDF: + +```sh +python evaluation/standard_asr/build_current_report.py +python scripts/build_neurips_report.py +``` + +Pooled WER divides summed edit counts by summed reference-word counts. FLEURS pooling includes all 25 languages, including English; the 25-language macro gives each language equal weight. Non-English compound alignment can change the reference-word denominator separately for each model. Win counts compare full-precision per-partition WERs. The two evaluation samples are never merged into one pooled headline. + +Test-other was used for the final adaptation and checkpoint selection. The domain sample contains 6,118 recordings from the preceding adaptation and retains previously audited Greek/Italian EuroSpeech transcript spans. No recordings are dropped. Numeric evidence is transcript-free; full manifests and hypotheses are archived in the private model repository. diff --git a/evaluation/standard_asr/README.md b/evaluation/standard_asr/README.md new file mode 100644 index 0000000000000000000000000000000000000000..e5b429d83bdbe8f4cf284d84f4f0236ac97ef585 --- /dev/null +++ b/evaluation/standard_asr/README.md @@ -0,0 +1,39 @@ +# LibriSpeech and FLEURS comparison + +This evaluation compares the released Orukeet FT-4035 NeMo checkpoint with NVIDIA Parakeet TDT 0.6B v3 on both complete LibriSpeech test partitions and all 25 supported FLEURS test languages. The [score table](../../docs/standard-asr-benchmarks.md) reports WER and CER for each partition and both models. + +The runner uses identical mono 16 kHz audio, FP32 weights, BF16 CUDA autocast and matched greedy-batch TDT decoding. It retains every record and every empty hypothesis. Final scoring uses the pinned English and multilingual normalizers in [vendor/](vendor/). English normalization standardizes spelling, names and numbers. Multilingual normalization retains diacritics, expands digits by language and aligns compound boundaries; WER uses compound-aware edit distance. CER uses normalized strings before boundary alignment, including spaces. Corpus rates pool integer edit counts; language macros weight their specified languages equally. + +Place the original LibriSpeech `test-clean` audio and `.trans.txt` files under `data/goal_v2_sources/libri/LibriSpeech/test-clean`, `test-other` under `data/librispeech/LibriSpeech/test-other`, and original FLEURS test FLAC files under `data/fleurs//test`, relative to `AUDIO_ROOT`. Use the recorded NeMo environment from the preceding source-model comparison. FLEURS references are fetched from the immutable dataset revision recorded in `prepare.py`; the scripts record reference-file and decoded-audio hashes. + +```sh +python evaluation/standard_asr/prepare.py --audio-root "$AUDIO_ROOT" --output /tmp/orukeet-standard/prepared +python evaluation/standard_asr/run.py \ + --manifest /tmp/orukeet-standard/prepared/manifest.jsonl \ + --parakeet /path/to/parakeet-tdt-0.6b-v3.nemo \ + --orukeet /path/to/orukeet-v0.1.0rc1.nemo \ + --metric-code evaluation/unseen \ + --output /tmp/orukeet-standard/evaluation +``` + +The runner verifies both checkpoint hashes before inference. Resume accepts only predictions with matching checkpoint, manifest, runner and normalizer hashes. A decoded-audio hash check precedes inference. Out-of-memory recovery recursively reduces the batch size; it does not omit or shorten recordings. + +The inference runner also emits diagnostic counts under the preceding evaluation's normalizer. Preserve these as `inference-comparison.json` and `inference-numeric-evidence.jsonl.gz` in the evidence directory. Place the unchanged `manifest.jsonl`, `parakeet.jsonl` and `orukeet.jsonl` in a private records directory, then generate the final scores: + +```sh +python -m pip install -r evaluation/standard_asr/requirements-score.txt +python evaluation/standard_asr/rescore.py --private-records /path/to/private-records --evidence /path/to/evidence +python evaluation/standard_asr/audit_predictions.py --private-records /path/to/private-records --evidence /path/to/evidence +``` + +The scoring implementation is pinned to [source revision 48219c6](https://github.com/huggingface/open_asr_leaderboard/tree/48219c6028db0517d704600d92f31edfc96e8c23/normalizer). [Provenance](vendor/provenance.json) records source and vendored hashes, exact extracted definitions and the Apache-2.0 license. Compound alignment can change the reference word count separately for each model; all resulting denominators are retained. Scoring changes never alter predictions or checkpoint identities. For the release audit, pass `--upstream-root /path/to/pinned-source` to `audit_predictions.py`; it independently loads the original normalization definitions and verifies all 54 model/partition WERs with full-partition batch scoring. + +`comparison.json` holds full-precision corpus results and checkpoint identities. `numeric-evidence.jsonl.gz` contains transcript-free per-record edit counts. `preparation.json` records the complete test membership and source revisions. To reproduce the release tables from the included counts: + +```sh +python evaluation/standard_asr/build_materials.py +python scripts/build_neurips_report.py +python evaluation/standard_asr/build_materials.py --verify-pdf +``` + +Corpus sources: [LibriSpeech](https://www.openslr.org/12) and [FLEURS](https://arxiv.org/abs/2205.12446). Dataset audio and reference transcripts are obtained from the original providers; the release evidence contains counts and hashes. diff --git a/evaluation/standard_asr/audit_predictions.py b/evaluation/standard_asr/audit_predictions.py new file mode 100644 index 0000000000000000000000000000000000000000..470d02fcf84975f9a2497c06de4ca62a94c5e275 --- /dev/null +++ b/evaluation/standard_asr/audit_predictions.py @@ -0,0 +1,104 @@ +"""Re-score the private hypotheses and compare every record with release counts.""" +import argparse +import gzip +import hashlib +import json +from pathlib import Path +import sys +import ast +from collections import defaultdict +from difflib import SequenceMatcher +import importlib.util +import re +import num2words +from kaldialign import batch_error_rate + +ROOT = Path(__file__).resolve().parents[2] +from scoring import counts + + +def sha(path): + return hashlib.sha256(path.read_bytes()).hexdigest() + + +def main(): + p=argparse.ArgumentParser() + p.add_argument('--private-records',type=Path,required=True) + p.add_argument('--evidence',type=Path,default=ROOT/'evidence/standard-asr-20260908') + p.add_argument('--upstream-root',type=Path) + a=p.parse_args() + result=json.loads((a.evidence/'comparison.json').read_text()) + manifest=a.private_records/'manifest.jsonl' + assert sha(manifest)==result['manifest_sha256'] + rows=[json.loads(line) for line in manifest.open()] + numeric={} + with gzip.open(a.evidence/'numeric-evidence.jsonl.gz','rt') as stream: + for line in stream: + row=json.loads(line) + assert row['record_sha256'] not in numeric + numeric[row['record_sha256']]=row + assert len(numeric)==len(rows)==result['rows'] + empty={} + paired=defaultdict(lambda: [[], []]) + for model in ['parakeet','orukeet']: + path=a.private_records/(model+'.jsonl') + assert sha(path)==result['prediction_sha256'][model] + records=[json.loads(line) for line in path.open()] + predictions={r['uid']:r for r in records} + assert len(predictions)==len(records)==len(rows) + empty[model]=0 + for row in rows: + prediction=predictions[row['uid']] + assert prediction['model_sha256']==result['models'][model] + assert prediction['manifest_sha256']==result['manifest_sha256'] + values=counts(row['text'],prediction['prediction'],row['language']) + expected=numeric[hashlib.sha256(row['uid'].encode()).hexdigest()] + assert expected['split']==row['split'] + for metric in ['errors','words','char_errors','chars','utterance_error']: + assert values[metric]==expected['counts'][model][metric],(row['uid'],model,metric) + empty[model]+=not bool(prediction['prediction'].strip()) + paired[(row['split'], model)][0].append(row['text']) + paired[(row['split'], model)][1].append(prediction['prediction']) + audit=dict(status='passed',publication_authorized=False,records=len(rows),paired_predictions=2*len(rows), + models=result['models'],empty_hypotheses=empty, + checks=['Every private hypothesis re-scored locally','All reference and prediction hashes match', + 'Every WER/CER numerator and denominator matches released counts'], + inputs_sha256={name:sha(a.evidence/name) for name in ['comparison.json','numeric-evidence.jsonl.gz']}, + manifest_sha256=result['manifest_sha256'],prediction_sha256=result['prediction_sha256'], + script_sha256=sha(Path(__file__))) + if a.upstream_root: + provenance=json.loads((ROOT/'evaluation/standard_asr/vendor/provenance.json').read_text()) + for name,digest in provenance['upstream_sha256'].items(): + assert sha(a.upstream_root/name)==digest,name + # Load upstream normalizers independently of the release's vendored package. + init=a.upstream_root/'normalizer/__init__.py' + spec=importlib.util.spec_from_file_location('upstream_text',init, + submodule_search_locations=[str(init.parent)]) + module=importlib.util.module_from_spec(spec);sys.modules[spec.name]=module + spec.loader.exec_module(module) + namespace=dict(BasicMultilingualTextNormalizer=module.BasicMultilingualTextNormalizer, + re=re,num2words=num2words,FILLER_WORDS={},SequenceMatcher=SequenceMatcher) + for filename,name in [('data_utils.py','MultilingualNormalizer'), + ('eval_utils.py','normalize_compound_pairs')]: + tree=ast.parse((init.parent/filename).read_text()) + definition=next(n for n in tree.body if isinstance(n,(ast.ClassDef,ast.FunctionDef)) and n.name==name) + exec(compile(ast.Module(body=[definition],type_ignores=[]),filename,'exec'),namespace) + english=module.EnglishTextNormalizer() + multilingual=namespace['MultilingualNormalizer'](remove_diacritics=False) + for (split,model),(refs,hyps) in paired.items(): + lang=result['sets'][split]['language'] + normalize=english if lang=='en' else lambda text:multilingual(text,lang=lang) + refs,hyps=list(map(normalize,refs)),list(map(normalize,hyps)) + if lang!='en':refs,hyps=namespace['normalize_compound_pairs'](refs,hyps) + batch=batch_error_rate([tuple(r.split()) for r in refs],[tuple(h.split()) for h in hyps],merge_compounds=True) + target=result['sets'][split]['models'][model] + assert batch['total']==target['errors'] and batch['ref_len']==target['words'] + assert abs(100*batch['err_rate']-target['wer'])<1e-12 + audit['upstream_scoring_verified']=dict(revision=provenance['revision'],model_partition_pairs=len(paired), + check=f'Unmodified upstream normalizer definitions and full-partition batch scoring reproduce all {len(paired)} paired WERs.') + audit['checks'].append(f'All {len(paired)} model/partition WERs match independent upstream batch scoring') + (a.evidence/'hypotheses-audit.json').write_text(json.dumps(audit,indent=2)+'\n') + print(json.dumps(audit)) + + +if __name__=='__main__':main() diff --git a/evaluation/standard_asr/build_current_report.py b/evaluation/standard_asr/build_current_report.py new file mode 100644 index 0000000000000000000000000000000000000000..252d95f0db104621d49dd07b714e4a09faef7e75 --- /dev/null +++ b/evaluation/standard_asr/build_current_report.py @@ -0,0 +1,179 @@ +"""Bind the short report to one checkpoint and reproduce every aggregate from counts.""" +from collections import Counter, defaultdict +import csv +import gzip +import hashlib +import json +from pathlib import Path +import subprocess +import sys +import unicodedata + +ROOT = Path(__file__).resolve().parents[2] +LANGUAGES = dict(bg='Bulgarian', cs='Czech', da='Danish', de='German', el='Greek', en='English', + es='Spanish', et='Estonian', fi='Finnish', fr='French', hr='Croatian', hu='Hungarian', + it='Italian', lt='Lithuanian', lv='Latvian', mt='Maltese', nl='Dutch', pl='Polish', + pt='Portuguese', ro='Romanian', ru='Russian', sk='Slovak', sl='Slovenian', sv='Swedish', uk='Ukrainian') +MODEL = '031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56' +BASE = '3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d' +MODELS = ['parakeet', 'orukeet'] +COUNTS = ['errors','words','substitutions','deletions','insertions','chars','char_errors','utterance_error'] + +def sha(path): return hashlib.sha256(path.read_bytes()).hexdigest() +def read(path): return json.loads(path.read_text()) +def write(path, value): path.write_text(json.dumps(value,indent=2,ensure_ascii=False)+'\n') + +def validate(folder, nrows, nsets): + d=ROOT/'evidence'/folder; result=read(d/'comparison.json');audit=read(d/'hypotheses-audit.json') + assert result['status']=='complete' and result['rows']==nrows and len(result['sets'])==nsets + assert result['models']==dict(parakeet=BASE,orukeet=MODEL) + assert result['timings']['parakeet']['rows']==result['timings']['orukeet']['rows']==nrows + assert audit['status']=='passed' and audit['models']==result['models'] + assert audit['upstream_scoring_verified']['model_partition_pairs']==2*nsets + assert result['numeric_evidence_sha256']==sha(d/'numeric-evidence.jsonl.gz') + for name,digest in result['scoring']['code_sha256'].items(): assert sha(ROOT/name)==digest,name + for name,digest in audit['inputs_sha256'].items(): assert sha(d/name)==digest,name + seen=set();counts=defaultdict(Counter);sizes=Counter() + with gzip.open(d/'numeric-evidence.jsonl.gz','rt') as f: + for line in f: + row=json.loads(line);uid=row['record_sha256'];assert uid not in seen;seen.add(uid) + sizes[row['split']]+=1 + for model in MODELS: + c=row['counts'][model];assert c['errors']==c['substitutions']+c['deletions']+c['insertions'] + counts[(row['split'],model)].update({k:c[k] for k in COUNTS}) + assert len(seen)==nrows + for split,spec in result['sets'].items(): + assert sizes[split]==spec['rows'] + for model in MODELS: + c=counts[(split,model)];reported=spec['models'][model] + assert all(reported[k]==v for k,v in c.items()) + for metric,num,den in [('wer','errors','words'),('cer','char_errors','chars')]: + assert abs(reported[metric]-100*c[num]/c[den])<1e-12 + return result + +def pool(result, splits): + members=[result['sets'][s] for s in splits];models={} + for model in MODELS: + c=Counter() + for spec in members: c.update({k:spec['models'][model][k] for k in COUNTS}) + models[model]=dict(c,wer=100*c['errors']/c['words'],cer=100*c['char_errors']/c['chars']) + wins=sum(s['models']['orukeet']['wer'] "10000") + text = re.sub(r"(\d)\s+(\d{3})\b", r"\1\2", text) + + # Convert remaining digit sequences to words + def _replace(m): + try: + return num2words.num2words(int(m.group()), lang=lang) + except Exception: + return m.group() + + return re.sub(r"\d+", _replace, text) + + def __call__(self, s, lang=None): + s = super().__call__(s) + if lang is not None: + s = self._remove_fillers(s, lang) + s = self._normalize_numbers(s, lang) + return s + +def normalize_compound_pairs(refs, preds): + """Align compound word boundaries between ref/pred pairs. + + When a mismatch region has identical characters ignoring whitespace, + normalize both sides to the joined form. + """ + new_refs, new_preds = [], [] + for ref_text, pred_text in zip(refs, preds): + ref_words = ref_text.split() + pred_words = pred_text.split() + + sm = SequenceMatcher(None, ref_words, pred_words) + new_rw, new_pw = [], [] + + for tag, i1, i2, j1, j2 in sm.get_opcodes(): + if tag == "equal": + new_rw.extend(ref_words[i1:i2]) + new_pw.extend(pred_words[j1:j2]) + else: + rc = "".join(ref_words[i1:i2]) + pc = "".join(pred_words[j1:j2]) + if rc == pc: + new_rw.append(rc) + new_pw.append(pc) + else: + new_rw.extend(ref_words[i1:i2]) + new_pw.extend(pred_words[j1:j2]) + + new_refs.append(" ".join(new_rw)) + new_preds.append(" ".join(new_pw)) + return new_refs, new_preds diff --git a/evaluation/standard_asr/vendor/normalizer.py b/evaluation/standard_asr/vendor/normalizer.py new file mode 100644 index 0000000000000000000000000000000000000000..c6f4fa0efcdca36afa32fb4469419c5d03d63e8c --- /dev/null +++ b/evaluation/standard_asr/vendor/normalizer.py @@ -0,0 +1,753 @@ +# Copyright 2022 The OpenAI team and The HuggingFace Team. All rights reserved. +# Most of the code is copy pasted from the original whisper repository +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import re +import unicodedata +from fractions import Fraction +from typing import Iterator, List, Match, Optional, Union +from .english_abbreviations import english_name_normalizer, english_spelling_normalizer, english_compound_normalizer + +import regex + + +# non-ASCII letters that are not separated by "NFKD" normalization +ADDITIONAL_DIACRITICS = { + "œ": "oe", + "Œ": "OE", + "ø": "o", + "Ø": "O", + "æ": "ae", + "Æ": "AE", + "ß": "ss", + "ẞ": "SS", + "đ": "d", + "Đ": "D", + "ð": "d", + "Ð": "D", + "þ": "th", + "Þ": "th", + "ł": "l", + "Ł": "L", +} + + +def remove_symbols_and_diacritics(s: str, keep=""): + """ + Replace any other markers, symbols, and punctuations with a space, and drop any diacritics (category 'Mn' and some + manual mappings) + """ + + def replace_character(char): + if char in keep: + return char + elif char in ADDITIONAL_DIACRITICS: + return ADDITIONAL_DIACRITICS[char] + + elif unicodedata.category(char) == "Mn": + return "" + + elif unicodedata.category(char)[0] in "MSP": + return " " + + return char + + return "".join(replace_character(c) for c in unicodedata.normalize("NFKD", s)) + + +def remove_symbols(s: str): + """ + Replace any other markers, symbols, punctuations with a space, keeping diacritics + """ + return "".join(" " if unicodedata.category(c)[0] in "MSP" else c for c in unicodedata.normalize("NFKC", s)) + + +def remove_symbols_keep_marks(s: str): + """ + Replace symbols and punctuation with a space, keeping combining marks. + + Unlike `remove_symbols`, combining marks (category 'M') are preserved. This + is required for scripts like Devanagari, where vowel signs (matras) and the + virama are combining marks that are integral to words. + """ + return "".join(" " if unicodedata.category(c)[0] in "SP" else c for c in unicodedata.normalize("NFKC", s)) + + +class BasicTextNormalizer: + def __init__(self, remove_diacritics: bool = False, split_letters: bool = False): + self.clean = remove_symbols_and_diacritics if remove_diacritics else remove_symbols + self.split_letters = split_letters + + def __call__(self, s: str): + s = s.lower() + s = re.sub(r"[<\[][^>\]]*[>\]]", "", s) # remove words between brackets + s = re.sub(r"\(([^)]+?)\)", "", s) # remove words between parenthesis + s = self.clean(s).lower() + + if self.split_letters: + s = " ".join(regex.findall(r"\X", s, regex.U)) + + s = re.sub(r"\s+", " ", s) # replace any successive whitespace characters with a space + + return s + + +class BasicMultilingualTextNormalizer: + def __init__(self, remove_diacritics: bool = True): + # When keeping diacritics, also keep combining marks (category 'M'): + # scripts like Devanagari encode vowel signs and the virama as + # combining marks, so stripping them mangles words. + self.clean = remove_symbols_and_diacritics if remove_diacritics else remove_symbols_keep_marks + + def __call__(self, s: str): + s = s.lower() + s = re.sub(r"[<\[][^>\]]*[>\]]", "", s) # remove words between brackets + s = re.sub(r"\(([^)]+?)\)", "", s) # remove words between parenthesis + s = self.clean(s).lower() + + # Remove punctuation and extra spaces + s = regex.sub(r"[^\w\s]", "", s) + s = re.sub(r"\s+", " ", s).strip() + + return s + + +class EnglishNumberNormalizer: + """ + Convert any spelled-out numbers into arabic numbers, while handling: + + - remove any commas + - keep the suffixes such as: `1960s`, `274th`, `32nd`, etc. + - spell out currency symbols after the number. e.g. `$20 million` -> `20000000 dollars` + - spell out `one` and `ones` + - interpret successive single-digit numbers as nominal: `one oh one` -> `101` + """ + + def __init__(self): + super().__init__() + + self.zeros = {"o", "oh", "zero"} + # fmt: off + self.ones = { + name: i + for i, name in enumerate( + ["one", "two", "three", "four", "five", "six", "seven", "eight", "nine", "ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen", "seventeen", "eighteen", "nineteen"], + start=1, + ) + } + # fmt: on + self.ones_plural = { + "sixes" if name == "six" else name + "s": (value, "s") for name, value in self.ones.items() + } + self.ones_ordinal = { + "zeroth": (0, "th"), + "first": (1, "st"), + "second": (2, "nd"), + "third": (3, "rd"), + "fifth": (5, "th"), + "twelfth": (12, "th"), + **{ + name + ("h" if name.endswith("t") else "th"): (value, "th") + for name, value in self.ones.items() + if value > 3 and value != 5 and value != 12 + }, + } + self.ones_suffixed = {**self.ones_plural, **self.ones_ordinal} + + self.tens = { + "twenty": 20, + "thirty": 30, + "forty": 40, + "fifty": 50, + "sixty": 60, + "seventy": 70, + "eighty": 80, + "ninety": 90, + } + self.tens_plural = {name.replace("y", "ies"): (value, "s") for name, value in self.tens.items()} + self.tens_ordinal = {name.replace("y", "ieth"): (value, "th") for name, value in self.tens.items()} + self.tens_suffixed = {**self.tens_plural, **self.tens_ordinal} + + self.multipliers = { + "hundred": 100, + "thousand": 1_000, + "million": 1_000_000, + "billion": 1_000_000_000, + "trillion": 1_000_000_000_000, + "quadrillion": 1_000_000_000_000_000, + "quintillion": 1_000_000_000_000_000_000, + "sextillion": 1_000_000_000_000_000_000_000, + "septillion": 1_000_000_000_000_000_000_000_000, + "octillion": 1_000_000_000_000_000_000_000_000_000, + "nonillion": 1_000_000_000_000_000_000_000_000_000_000, + "decillion": 1_000_000_000_000_000_000_000_000_000_000_000, + } + self.multipliers_plural = {name + "s": (value, "s") for name, value in self.multipliers.items()} + self.multipliers_ordinal = {name + "th": (value, "th") for name, value in self.multipliers.items()} + self.multipliers_suffixed = {**self.multipliers_plural, **self.multipliers_ordinal} + self.decimals = {*self.ones, *self.tens, *self.zeros} + + self.preceding_prefixers = { + "minus": "-", + "negative": "-", + "plus": "+", + "positive": "+", + } + self.following_prefixers = { + "pound": "£", + "pounds": "£", + "euro": "€", + "euros": "€", + "dollar": "$", + "dollars": "$", + "cent": "¢", + "cents": "¢", + } + self.prefixes = set(list(self.preceding_prefixers.values()) + list(self.following_prefixers.values())) + self.suffixers = { + "per": {"cent": "%"}, + "percent": "%", + } + self.specials = {"and", "double", "triple", "point"} + + self.words = { + key + for mapping in [ + self.zeros, + self.ones, + self.ones_suffixed, + self.tens, + self.tens_suffixed, + self.multipliers, + self.multipliers_suffixed, + self.preceding_prefixers, + self.following_prefixers, + self.suffixers, + self.specials, + ] + for key in mapping + } + self.literal_words = {"one", "ones"} + + def process_words(self, words: List[str]) -> Iterator[str]: + prefix: Optional[str] = None + value: Optional[Union[str, int]] = None + skip = False + + def to_fraction(s: str): + try: + return Fraction(s) + except ValueError: + return None + + def is_digit_token(token: Optional[str]) -> bool: + """True for tokens that continue a digit sequence ("four", "oh", "20").""" + return token is not None and bool( + re.match(r"^\d+$", token) + or token in self.zeros + or token in self.ones + or token in self.tens + ) + + def output(result: Union[str, int]): + nonlocal prefix, value + result = str(result) + if prefix is not None: + result = prefix + result + value = None + prefix = None + return result + + if len(words) == 0: + return + + for i, current in enumerate(words): + prev = words[i - 1] if i != 0 else None + next = words[i + 1] if i != len(words) - 1 else None + if skip: + skip = False + continue + + next_is_numeric = next is not None and re.match(r"^\d+(\.\d+)?$", next) + has_prefix = current[0] in self.prefixes + current_without_prefix = current[1:] if has_prefix else current + if re.match(r"^\d+(\.\d+)?$", current_without_prefix): + # arabic numbers (potentially with signs and fractions) + f = to_fraction(current_without_prefix) + if f is None: + raise ValueError("Converting the fraction failed") + + if value is not None: + if isinstance(value, str) and value.endswith("."): + # concatenate decimals / ip address components + value = str(value) + str(current) + continue + else: + yield output(value) + + prefix = current[0] if has_prefix else prefix + if f.denominator == 1: + value = f.numerator # store integers as int + else: + value = current_without_prefix + elif current not in self.words: + # non-numeric words + if value is not None: + yield output(value) + yield output(current) + elif current in self.zeros: + # "oh" is far more often the interjection than a spoken zero, so + # read it as a digit only inside a digit sequence: the sequence + # has to continue on the right ("four oh one", "nineteen oh + # five"), and with nothing pending on the left it takes two more + # digit tokens, i.e. a serial/phone-style reading ("oh seven nine + # eight"). On its own — including the doubled "oh oh" — it stays + # a word. "o" and "zero" are unchanged. + next2 = words[i + 2] if i + 2 < len(words) else None + in_number = is_digit_token(next) and ( + value is not None or is_digit_token(next2) + ) + if current == "oh" and not in_number: + if value is not None: + yield output(value) # don't drop a pending number + yield output(current) + else: + value = str(value or "") + "0" + elif current in self.ones: + ones = self.ones[current] + + if value is None: + value = ones + elif isinstance(value, str) or prev in self.ones: + if prev in self.tens and ones < 10: # replace the last zero with the digit + value = value[:-1] + str(ones) + else: + value = str(value) + str(ones) + elif ones < 10: + if value % 10 == 0: + value += ones + else: + value = str(value) + str(ones) + else: # eleven to nineteen + if value % 100 == 0: + value += ones + else: + value = str(value) + str(ones) + elif current in self.ones_suffixed: + # ordinal or cardinal; yield the number right away + ones, suffix = self.ones_suffixed[current] + if value is None: + yield output(str(ones) + suffix) + elif isinstance(value, str) or prev in self.ones: + if prev in self.tens and ones < 10: + yield output(value[:-1] + str(ones) + suffix) + else: + yield output(str(value) + str(ones) + suffix) + elif ones < 10: + if value % 10 == 0: + yield output(str(value + ones) + suffix) + else: + yield output(str(value) + str(ones) + suffix) + else: # eleven to nineteen + if value % 100 == 0: + yield output(str(value + ones) + suffix) + else: + yield output(str(value) + str(ones) + suffix) + value = None + elif current in self.tens: + tens = self.tens[current] + if value is None: + value = tens + elif isinstance(value, str): + value = str(value) + str(tens) + else: + if value % 100 == 0: + value += tens + else: + value = str(value) + str(tens) + elif current in self.tens_suffixed: + # ordinal or cardinal; yield the number right away + tens, suffix = self.tens_suffixed[current] + if value is None: + yield output(str(tens) + suffix) + elif isinstance(value, str): + yield output(str(value) + str(tens) + suffix) + else: + if value % 100 == 0: + yield output(str(value + tens) + suffix) + else: + yield output(str(value) + str(tens) + suffix) + elif current in self.multipliers: + multiplier = self.multipliers[current] + if value is None: + value = multiplier + elif isinstance(value, str) or value == 0: + f = to_fraction(value) + p = f * multiplier if f is not None else None + if f is not None and p.denominator == 1: + value = p.numerator + else: + yield output(value) + value = multiplier + else: + before = value // 1000 * 1000 + residual = value % 1000 + value = before + residual * multiplier + elif current in self.multipliers_suffixed: + multiplier, suffix = self.multipliers_suffixed[current] + if value is None: + yield output(str(multiplier) + suffix) + elif isinstance(value, str): + f = to_fraction(value) + p = f * multiplier if f is not None else None + if f is not None and p.denominator == 1: + yield output(str(p.numerator) + suffix) + else: + yield output(value) + yield output(str(multiplier) + suffix) + else: # int + before = value // 1000 * 1000 + residual = value % 1000 + value = before + residual * multiplier + yield output(str(value) + suffix) + value = None + elif current in self.preceding_prefixers: + # apply prefix (positive, minus, etc.) if it precedes a number + if value is not None: + yield output(value) + + if next in self.words or next_is_numeric: + prefix = self.preceding_prefixers[current] + else: + yield output(current) + elif current in self.following_prefixers: + # apply prefix (dollars, cents, etc.) only after a number + if value is not None: + prefix = self.following_prefixers[current] + yield output(value) + else: + yield output(current) + elif current in self.suffixers: + # apply suffix symbols (percent -> '%') + if value is not None: + suffix = self.suffixers[current] + if isinstance(suffix, dict): + if next in suffix: + yield output(str(value) + suffix[next]) + skip = True + else: + yield output(value) + yield output(current) + else: + yield output(str(value) + suffix) + else: + yield output(current) + elif current in self.specials: + if next not in self.words and not next_is_numeric: + # apply special handling only if the next word can be numeric + if value is not None: + yield output(value) + yield output(current) + elif current == "and": + # ignore "and" after hundreds, thousands, etc. + if prev not in self.multipliers: + if value is not None: + yield output(value) + yield output(current) + elif current == "double" or current == "triple": + if next in self.ones or next in self.zeros: + repeats = 2 if current == "double" else 3 + ones = self.ones.get(next, 0) + value = str(value or "") + str(ones) * repeats + skip = True + else: + if value is not None: + yield output(value) + yield output(current) + elif current == "point": + if next in self.decimals or next_is_numeric: + value = str(value or "") + "." + else: + # should all have been covered at this point + raise ValueError(f"Unexpected token: {current}") + else: + # all should have been covered at this point + raise ValueError(f"Unexpected token: {current}") + + if value is not None: + yield output(value) + + def preprocess(self, s: str): + # replace " and a half" with " point five" + results = [] + + segments = re.split(r"\band\s+a\s+half\b", s) + for i, segment in enumerate(segments): + if len(segment.strip()) == 0: + continue + if i == len(segments) - 1: + results.append(segment) + else: + results.append(segment) + last_word = segment.rsplit(maxsplit=2)[-1] + if last_word in self.decimals or last_word in self.multipliers: + results.append("point five") + else: + results.append("and a half") + + s = " ".join(results) + + # put a space at number/letter boundary + s = re.sub(r"([a-z])([0-9])", r"\1 \2", s) + s = re.sub(r"([0-9])([a-z])", r"\1 \2", s) + + # but remove spaces which could be a suffix + s = re.sub(r"([0-9])\s+(st|nd|rd|th|s)\b", r"\1\2", s) + + return s + + def postprocess(self, s: str): + def combine_cents(m: Match): + try: + currency = m.group(1) + integer = m.group(2) + cents = int(m.group(3)) + return f"{currency}{integer}.{cents:02d}" + except ValueError: + return m.string + + def extract_cents(m: Match): + try: + return f"¢{int(m.group(1))}" + except ValueError: + return m.string + + # apply currency postprocessing; "$2 and ¢7" -> "$2.07" + s = re.sub(r"([€£$])([0-9]+) (?:and )?¢([0-9]{1,2})\b", combine_cents, s) + s = re.sub(r"[€£$]0.([0-9]{1,2})\b", extract_cents, s) + + # write "one(s)" instead of "1(s)", just for the readability + s = re.sub(r"\b1(s?)\b", r"one\1", s) + + return s + + def __call__(self, s: str): + s = self.preprocess(s) + s = " ".join(word for word in self.process_words(s.split()) if word is not None) + s = self.postprocess(s) + + return s + + +class EnglishSpellingNormalizer: + """ + Applies British-American spelling mappings as listed in [1]. + + [1] https://www.tysto.com/uk-us-spelling-list.html + """ + + def __init__(self, english_spelling_mapping): + self.mapping = english_spelling_mapping + + def __call__(self, s: str): + return " ".join(self.mapping.get(word, word) for word in s.split()) + + +class EnglishAcronymNormalizer: + """ + Collapse sequences of single-character tokens (letters or digits) into single words. + + This normalizes acronym spacing so that both spaced-out and joined forms match: + - "b b c" -> "bbc" + - "5 g" -> "5g" + + Lone single-character words surrounded by multi-character words are left untouched + (e.g. "a big cat" stays "a big cat"). + """ + + def __call__(self, s: str) -> str: + words = s.split() + result = [] + i = 0 + while i < len(words): + if len(words[i]) == 1 and words[i].isalnum(): + # Start of a potential acronym run + run = [words[i]] + j = i + 1 + while j < len(words) and len(words[j]) == 1 and words[j].isalnum(): + run.append(words[j]) + j += 1 + # Require 3+ tokens if the run contains common words "a" or "i", + # otherwise 2+ is enough (e.g. "5 g" -> "5g") + has_common_word = any(c in ("a", "i") for c in run) + min_run = 3 if has_common_word else 2 + if len(run) >= min_run: + result.append("".join(run)) + else: + result.extend(run) + i = j + else: + result.append(words[i]) + i += 1 + return " ".join(result) + + +class EnglishNameNormalizer: + """ + Collapse common name spelling variants to a single canonical form. + + This is intentionally conservative and token-based so it can be extended + with project-specific aliases when needed. + """ + + def __init__(self, english_name_mapping=english_name_normalizer): + self.mapping = english_name_mapping + + def __call__(self, s: str): + return " ".join(self.mapping.get(word, word) for word in s.split()) + + +class EnglishTextNormalizer: + def __init__(self, english_spelling_mapping=english_spelling_normalizer): + # Filler words / hesitations to remove. Written as regexes so that + # arbitrary elongation is covered without enumerating every spelling + # ("uh", "uhh", "uuuh", "uhhhh", ...). Each alternative is wrapped in + # \b...\b below, so a shorter alternative cannot match a prefix of a + # longer token and the order of the single-token patterns is irrelevant. + # Hyphens are word boundaries too, which is why most hyphenated forms + # need no entry ("um-hmm" is matched as "um" + "hmm") — but any whose + # halves are not both fillers must be listed *before* the patterns, + # otherwise only the first half is matched ("ah-ha" -> "ha"). + filler_words = [ + "ah-ha", # "ha" alone is not a filler, so match the pair first + r"a+h+m*", # ah, aah, ahh, ahhh, aaah, ahm, ahmm + r"a+h+a+", # aha, ahaa, ahaaa + r"e+h+m*", # eh, ehh, eeeh, ehhh, ehm, ehmm + r"e+m+", # em, emm + r"e+r+m*", # er, err, errr, erm + r"h+a+h+", # hah, hahh + r"h+e+h+", # heh, hehh + r"h+m+", # hm, hmm, hmmm, hhm + r"h+u+h+", # huh, huhh + r"m{2,}", # mm, mmm, mmmm + r"m+h+m*", # mh, mhm, mhmm, mmhm + r"t+s+k+", # tsk + r"u+g+h+", # ugh, uuugh + r"u+h+m*", # uh, uuh, uhh, uhhh, uuuh, uhm, uuuhm + r"u+h+u+[hm]*", # uhuh, uhum + r"u+m+h*", # um, umm, ummm, uuum, umh + # Irregular forms, not worth a pattern of their own. + "ahem", "eheh", "ehehe", "ehr", "hmmph", "hum", "hunh", "mhum", "mmkay", + ] + self.ignore_patterns = r"\b(" + "|".join(filler_words) + r")\b" + self.replacers = { + # Bare o'clock times: the ":00" is not spoken as words, so drop it + # ("2:00 AM" -> "2 am"). Applied here, while the colon is still + # present, so that a time is distinguishable from an unrelated + # digit sequence — by the time symbols are stripped "3:00" and + # "3 00" look alike. Without this the minutes are absorbed into the + # hour ("3:00" -> "30") or left as a stray token ("11:00" -> "11 0"). + r"\b(\d{1,2}):00\b": r"\1", + # common contractions + r"\bwon't\b": "will not", + r"\bcan't\b": "can not", + r"\blet's\b": "let us", + r"\bain't\b": "aint", + r"\by'all\b": "you all", + r"\bwanna\b": "want to", + r"\bgotta\b": "got to", + r"\bgonna\b": "going to", + r"\bi'ma\b": "i am going to", + r"\bimma\b": "i am going to", + r"\bwoulda\b": "would have", + r"\bcoulda\b": "could have", + r"\bshoulda\b": "should have", + r"\bma'am\b": "madam", + # contractions in titles/prefixes + r"\bmr\b": "mister ", + r"\bmrs\b": "missus ", + r"\bst\b": "saint ", + r"\bdr\b": "doctor ", + r"\bprof\b": "professor ", + r"\bcapt\b": "captain ", + r"\bgov\b": "governor ", + r"\bald\b": "alderman ", + r"\bgen\b": "general ", + r"\bsen\b": "senator ", + r"\brep\b": "representative ", + r"\bpres\b": "president ", + r"\brev\b": "reverend ", + r"\bhon\b": "honorable ", + r"\basst\b": "assistant ", + r"\bassoc\b": "associate ", + r"\blt\b": "lieutenant ", + r"\bcol\b": "colonel ", + r"\bjr\b": "junior ", + r"\bsr\b": "senior ", + r"\besq\b": "esquire ", + # prefect tenses, ideally it should be any past participles, but it's harder.. + r"'d been\b": " had been", + r"'s been\b": " has been", + r"'d gone\b": " had gone", + r"'s gone\b": " has gone", + r"'d done\b": " had done", # "'s done" is ambiguous + r"'s got\b": " has got", + # general contractions + r"n't\b": " not", + r"'re\b": " are", + r"\b(it|he|she|what|that|who|here|there|how|when|where|why|this)'s\b": r"\1 is", + r"'d\b": " would", + r"'ll\b": " will", + r"'t\b": " not", + r"'ve\b": " have", + r"'m\b": " am", + } + self.standardize_numbers = EnglishNumberNormalizer() + self.standardize_spellings = EnglishSpellingNormalizer(english_spelling_mapping) + self.standardize_names = EnglishNameNormalizer() + self.standardize_acronyms = EnglishAcronymNormalizer() + # Multi-word compound mappings — defined in english_abbreviations.py + self.compound_words = english_compound_normalizer + + def __call__(self, s: str): + s = s.lower() + + s = re.sub(r"[<\[][^>\]]*[>\]]", "", s) # remove words between brackets + s = re.sub(r"\(([^)]+?)\)", "", s) # remove words between parenthesis + s = re.sub(self.ignore_patterns, "", s) + s = re.sub(r"\s+'", "'", s) # standardize when there's a space before an apostrophe + + for pattern, replacement in self.replacers.items(): + s = re.sub(pattern, replacement, s) + + s = re.sub(r"(\d),(\d)", r"\1\2", s) # remove commas between digits + s = re.sub(r"\.([^0-9]|$)", r" \1", s) # remove periods not followed by numbers + s = remove_symbols_and_diacritics(s, keep=".%$¢€£") # keep some symbols for numerics + + # Normalize hardcoded compound words (e.g. "wi fi" -> "wifi" after hyphen removal) + for pattern, replacement in self.compound_words.items(): + s = re.sub(pattern, replacement, s) + + s = self.standardize_numbers(s) + s = self.standardize_spellings(s) + s = self.standardize_names(s) + s = self.standardize_acronyms(s) + + # now remove prefix/suffix symbols that are not preceded/followed by numbers + s = re.sub(r"[.$¢€£]([^0-9])", r" \1", s) + s = re.sub(r"([^0-9])%", r"\1 ", s) + + s = re.sub(r"\s+", " ", s) # replace any successive whitespace characters with a space + + return s diff --git a/evaluation/standard_asr/vendor/provenance.json b/evaluation/standard_asr/vendor/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..d591d037ecf1388af725bfdae848fcabaa8a0c41 --- /dev/null +++ b/evaluation/standard_asr/vendor/provenance.json @@ -0,0 +1,23 @@ +{ + "repository": "https://github.com/huggingface/open_asr_leaderboard", + "revision": "48219c6028db0517d704600d92f31edfc96e8c23", + "license": "Apache-2.0", + "upstream_sha256": { + "normalizer/normalizer.py": "490b56393484ef386b486f3679cc2261264af280b3512154431e0bdcf2778295", + "normalizer/english_abbreviations.py": "52997cc963e0bd6568d15554ad8d5fb0759f20f361483632a22f90d2d0e07db1", + "normalizer/data_utils.py": "738f362b477ee357f743a567b8bbd69fac49e56dce0a477cbdd52dec863842b8", + "normalizer/eval_utils.py": "27138d5884e39f5e6422e1dc48ce60dc2f5c15a56e63117115d7e6088d29e449" + }, + "vendored_sha256": { + "LICENSE": "5ee13882fce0975f0ad3c3d5c2042af4896c2a669be2d6df5b3d256811267c85", + "english_abbreviations.py": "52997cc963e0bd6568d15554ad8d5fb0759f20f361483632a22f90d2d0e07db1", + "__init__.py": "8c39e42af3204c41c564c970e24525da160663836b54acc1b629e4342f862334", + "multilingual.py": "4c8db6c4171c8c2f5567d2ed625a3a1f0d45156af45ca42d9555c0c9b61a29d5", + "normalizer.py": "490b56393484ef386b486f3679cc2261264af280b3512154431e0bdcf2778295" + }, + "modifications": { + "normalizer.py": "None", + "english_abbreviations.py": "None", + "multilingual.py": "Exact MultilingualNormalizer and normalize_compound_pairs definitions extracted; imports minimized, empty FILLER_WORDS preserved." + } +} diff --git a/evidence/domains-r3-20260908/README.md b/evidence/domains-r3-20260908/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0e309f3c6cb8b0da4f499b80286fcb72f4df348a --- /dev/null +++ b/evidence/domains-r3-20260908/README.md @@ -0,0 +1,18 @@ +# Orukeet r3: accent and domain comparison + +Orukeet scores **15.25% pooled WER versus Parakeet's 16.72%** on 12,006 recordings across 47 partitions and 25 languages, an 8.8% relative reduction. It improves 36 of the 47 partitions. On all 20 English partitions, comprising 5,120 recordings, WER is **8.84% versus 9.51%**, a 7.0% relative reduction; every English partition improves. + +| Pooled comparison | Parakeet errors / words | Parakeet WER | Orukeet errors / words | Orukeet WER | +|:--|--:|--:|--:|--:| +| All 47 partitions | 43,939 / 262,747 | 16.72% | 40,068 / 262,698 | 15.25% | +| All 20 English partitions | 9,032 / 94,993 | 9.51% | 8,399 / 94,993 | 8.84% | + +The sample retains its existing 256 recordings per partition and all 230 Lesbos recordings. Of these, 6,118 recordings were included in the earlier FT-4035 adaptation. Greek and Italian EuroSpeech retain their audited human transcript spans. The generic manifest adds three evaluator field aliases while preserving all original fields, references, waveforms and membership. + +Both models were decoded afresh, using FP32 NeMo weights, BF16 CUDA autocast, greedy-batch TDT and identical duration-sorted 16 kHz audio. Orukeet is the r3 checkpoint, SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56`. All 12,006 recordings are scored; 70 empty Parakeet hypotheses and 61 empty Orukeet hypotheses remain. + +The final scores use the same pinned English/multilingual normalization and compound-aware WER as the current complete LibriSpeech/FLEURS comparison. This protocol differs from the older domain report's diagnostic scoring, so the current report uses only these freshly paired scores. `inference-comparison.json` retains the earlier diagnostic normalization; `comparison.json` records the final manuscript scores. + +`hypotheses-audit.json` independently re-scores all 24,012 predictions and verifies every numerator and denominator against the per-record numeric evidence. Full-partition scoring with unmodified upstream normalizer definitions reproduces all 94 model/partition WERs. `scores.csv` retains full-precision WER/CER and edit counts. The [complete table](../../docs/current-checkpoint-benchmarks.md) includes all 47 comparisons. + +All weights, manifests, predictions and report materials remain private for review. diff --git a/evidence/domains-r3-20260908/comparison.json b/evidence/domains-r3-20260908/comparison.json new file mode 100644 index 0000000000000000000000000000000000000000..17fa5d8f72f96a408d7a1522394d3db032fa0152 --- /dev/null +++ b/evidence/domains-r3-20260908/comparison.json @@ -0,0 +1,1535 @@ +{ + "status": "complete", + "publication_authorized": false, + "rows": 12006, + "sets": { + "eurospeech_bg": { + "rows": 256, + "hours": 1.072353802083333, + "language": "bg", + "models": { + "parakeet": { + "words": 7448, + "errors": 1059, + "substitutions": 631, + "deletions": 122, + "insertions": 306, + "chars": 46565, + "char_errors": 3352, + "utterance_error": 230, + "wer": 14.21858216970999, + "cer": 7.198539675722109 + }, + "orukeet": { + "words": 7448, + "errors": 971, + "substitutions": 545, + "deletions": 104, + "insertions": 322, + "chars": 46565, + "char_errors": 3128, + "utterance_error": 224, + "wer": 13.03705692803437, + "cer": 6.717491678299152 + } + } + }, + "eurospeech_de": { + "rows": 256, + "hours": 1.0912282638888895, + "language": "de", + "models": { + "parakeet": { + "words": 8440, + "errors": 1131, + "substitutions": 499, + "deletions": 244, + "insertions": 388, + "chars": 58395, + "char_errors": 4983, + "utterance_error": 228, + "wer": 13.40047393364929, + "cer": 8.533264834318008 + }, + "orukeet": { + "words": 8445, + "errors": 941, + "substitutions": 376, + "deletions": 236, + "insertions": 329, + "chars": 58395, + "char_errors": 4174, + "utterance_error": 221, + "wer": 11.142687981053879, + "cer": 7.147872249336416 + } + } + }, + "eurospeech_el": { + "rows": 256, + "hours": 1.0396271527777776, + "language": "el", + "models": { + "parakeet": { + "words": 7910, + "errors": 2043, + "substitutions": 1804, + "deletions": 133, + "insertions": 106, + "chars": 50646, + "char_errors": 4291, + "utterance_error": 256, + "wer": 25.828065739570164, + "cer": 8.472534849741342 + }, + "orukeet": { + "words": 7908, + "errors": 2084, + "substitutions": 1756, + "deletions": 152, + "insertions": 176, + "chars": 50646, + "char_errors": 4563, + "utterance_error": 255, + "wer": 26.35306019221042, + "cer": 9.009596019428978 + } + } + }, + "eurospeech_en": { + "rows": 256, + "hours": 1.0540027430555547, + "language": "en", + "models": { + "parakeet": { + "words": 9636, + "errors": 2351, + "substitutions": 746, + "deletions": 325, + "insertions": 1280, + "chars": 55374, + "char_errors": 9882, + "utterance_error": 248, + "wer": 24.398090493980906, + "cer": 17.845920468089716 + }, + "orukeet": { + "words": 9636, + "errors": 2290, + "substitutions": 737, + "deletions": 324, + "insertions": 1229, + "chars": 55374, + "char_errors": 9687, + "utterance_error": 247, + "wer": 23.765047737650477, + "cer": 17.493769639180844 + } + } + }, + "eurospeech_et": { + "rows": 256, + "hours": 1.0695780555555554, + "language": "et", + "models": { + "parakeet": { + "words": 6793, + "errors": 2355, + "substitutions": 1515, + "deletions": 367, + "insertions": 473, + "chars": 47632, + "char_errors": 6958, + "utterance_error": 252, + "wer": 34.668040630060354, + "cer": 14.607826671145448 + }, + "orukeet": { + "words": 6764, + "errors": 1713, + "substitutions": 987, + "deletions": 220, + "insertions": 506, + "chars": 47632, + "char_errors": 5689, + "utterance_error": 248, + "wer": 25.325251330573625, + "cer": 11.943651326839099 + } + } + }, + "eurospeech_fi": { + "rows": 256, + "hours": 1.0576934722222222, + "language": "fi", + "models": { + "parakeet": { + "words": 6296, + "errors": 1046, + "substitutions": 600, + "deletions": 163, + "insertions": 283, + "chars": 51597, + "char_errors": 3704, + "utterance_error": 228, + "wer": 16.61372299872935, + "cer": 7.178711940616703 + }, + "orukeet": { + "words": 6288, + "errors": 956, + "substitutions": 517, + "deletions": 160, + "insertions": 279, + "chars": 51597, + "char_errors": 3494, + "utterance_error": 220, + "wer": 15.20356234096692, + "cer": 6.771711533616296 + } + } + }, + "eurospeech_fr": { + "rows": 256, + "hours": 1.0628877083333337, + "language": "fr", + "models": { + "parakeet": { + "words": 10678, + "errors": 2074, + "substitutions": 1216, + "deletions": 427, + "insertions": 431, + "chars": 63628, + "char_errors": 7236, + "utterance_error": 245, + "wer": 19.423112942498594, + "cer": 11.372351794807317 + }, + "orukeet": { + "words": 10677, + "errors": 1525, + "substitutions": 778, + "deletions": 294, + "insertions": 453, + "chars": 63628, + "char_errors": 5603, + "utterance_error": 241, + "wer": 14.283038306640442, + "cer": 8.805871628842647 + } + } + }, + "eurospeech_hr": { + "rows": 256, + "hours": 1.0454521527777771, + "language": "hr", + "models": { + "parakeet": { + "words": 7503, + "errors": 970, + "substitutions": 487, + "deletions": 163, + "insertions": 320, + "chars": 45726, + "char_errors": 3968, + "utterance_error": 222, + "wer": 12.92816206850593, + "cer": 8.677776319818047 + }, + "orukeet": { + "words": 7502, + "errors": 942, + "substitutions": 479, + "deletions": 135, + "insertions": 328, + "chars": 45726, + "char_errors": 3869, + "utterance_error": 216, + "wer": 12.556651559584111, + "cer": 8.461269299741941 + } + } + }, + "eurospeech_it": { + "rows": 256, + "hours": 1.0571320486111113, + "language": "it", + "models": { + "parakeet": { + "words": 8026, + "errors": 879, + "substitutions": 339, + "deletions": 199, + "insertions": 341, + "chars": 50422, + "char_errors": 3434, + "utterance_error": 218, + "wer": 10.951906304510342, + "cer": 6.810519217801753 + }, + "orukeet": { + "words": 8025, + "errors": 989, + "substitutions": 337, + "deletions": 204, + "insertions": 448, + "chars": 50422, + "char_errors": 4204, + "utterance_error": 225, + "wer": 12.32398753894081, + "cer": 8.337630399428821 + } + } + }, + "eurospeech_lt": { + "rows": 256, + "hours": 1.0921616319444443, + "language": "lt", + "models": { + "parakeet": { + "words": 6473, + "errors": 2488, + "substitutions": 1732, + "deletions": 361, + "insertions": 395, + "chars": 47147, + "char_errors": 7616, + "utterance_error": 254, + "wer": 38.43658272825583, + "cer": 16.15373194476849 + }, + "orukeet": { + "words": 6468, + "errors": 2141, + "substitutions": 1442, + "deletions": 276, + "insertions": 423, + "chars": 47147, + "char_errors": 6761, + "utterance_error": 253, + "wer": 33.10142238713667, + "cer": 14.340254947292511 + } + } + }, + "eurospeech_lv": { + "rows": 256, + "hours": 1.0540805208333333, + "language": "lv", + "models": { + "parakeet": { + "words": 6422, + "errors": 3672, + "substitutions": 2603, + "deletions": 781, + "insertions": 288, + "chars": 44539, + "char_errors": 11870, + "utterance_error": 256, + "wer": 57.17844908128309, + "cer": 26.650800422101977 + }, + "orukeet": { + "words": 6412, + "errors": 2702, + "substitutions": 1952, + "deletions": 380, + "insertions": 370, + "chars": 44539, + "char_errors": 7665, + "utterance_error": 255, + "wer": 42.13973799126637, + "cer": 17.20963649834976 + } + } + }, + "eurospeech_mt": { + "rows": 256, + "hours": 1.0746674826388882, + "language": "mt", + "models": { + "parakeet": { + "words": 7797, + "errors": 2872, + "substitutions": 1804, + "deletions": 529, + "insertions": 539, + "chars": 46869, + "char_errors": 9066, + "utterance_error": 256, + "wer": 36.83468000513018, + "cer": 19.343275939320232 + }, + "orukeet": { + "words": 7796, + "errors": 2818, + "substitutions": 1803, + "deletions": 475, + "insertions": 540, + "chars": 46869, + "char_errors": 8853, + "utterance_error": 256, + "wer": 36.14674191893278, + "cer": 18.88881776867439 + } + } + }, + "eurospeech_pt": { + "rows": 256, + "hours": 1.0251623090277782, + "language": "pt", + "models": { + "parakeet": { + "words": 8224, + "errors": 1898, + "substitutions": 486, + "deletions": 589, + "insertions": 823, + "chars": 48490, + "char_errors": 8568, + "utterance_error": 238, + "wer": 23.078793774319067, + "cer": 17.669622602598473 + }, + "orukeet": { + "words": 8225, + "errors": 1958, + "substitutions": 510, + "deletions": 611, + "insertions": 837, + "chars": 48490, + "char_errors": 8932, + "utterance_error": 237, + "wer": 23.805471124620063, + "cer": 18.42029284388534 + } + } + }, + "eurospeech_sk": { + "rows": 256, + "hours": 1.0663451041666665, + "language": "sk", + "models": { + "parakeet": { + "words": 7269, + "errors": 1257, + "substitutions": 844, + "deletions": 211, + "insertions": 202, + "chars": 46146, + "char_errors": 3580, + "utterance_error": 241, + "wer": 17.292612463887743, + "cer": 7.757985524205782 + }, + "orukeet": { + "words": 7265, + "errors": 1083, + "substitutions": 679, + "deletions": 163, + "insertions": 241, + "chars": 46146, + "char_errors": 3197, + "utterance_error": 230, + "wer": 14.907088781830694, + "cer": 6.928011095219521 + } + } + }, + "eurospeech_sl": { + "rows": 256, + "hours": 1.042155416666667, + "language": "sl", + "models": { + "parakeet": { + "words": 7754, + "errors": 3755, + "substitutions": 2988, + "deletions": 469, + "insertions": 298, + "chars": 47334, + "char_errors": 7539, + "utterance_error": 255, + "wer": 48.42661851947382, + "cer": 15.927240461401952 + }, + "orukeet": { + "words": 7748, + "errors": 3892, + "substitutions": 3140, + "deletions": 470, + "insertions": 282, + "chars": 47334, + "char_errors": 7820, + "utterance_error": 256, + "wer": 50.23231801755292, + "cer": 16.52089407191448 + } + } + }, + "eurospeech_uk": { + "rows": 256, + "hours": 1.0933650694444448, + "language": "uk", + "models": { + "parakeet": { + "words": 7413, + "errors": 1012, + "substitutions": 467, + "deletions": 202, + "insertions": 343, + "chars": 49036, + "char_errors": 3720, + "utterance_error": 222, + "wer": 13.651692971806286, + "cer": 7.586263153601435 + }, + "orukeet": { + "words": 7411, + "errors": 1056, + "substitutions": 500, + "deletions": 186, + "insertions": 370, + "chars": 49036, + "char_errors": 3720, + "utterance_error": 218, + "wer": 14.249089191742005, + "cer": 7.586263153601435 + } + } + }, + "gigaspeechbench_agr_en": { + "rows": 256, + "hours": 0.4082733159722223, + "language": "en", + "models": { + "parakeet": { + "words": 4660, + "errors": 289, + "substitutions": 119, + "deletions": 79, + "insertions": 91, + "chars": 23271, + "char_errors": 921, + "utterance_error": 128, + "wer": 6.201716738197425, + "cer": 3.9577156117055563 + }, + "orukeet": { + "words": 4660, + "errors": 272, + "substitutions": 130, + "deletions": 63, + "insertions": 79, + "chars": 23271, + "char_errors": 793, + "utterance_error": 115, + "wer": 5.836909871244635, + "cer": 3.407674788363199 + } + } + }, + "gigaspeechbench_ait_en": { + "rows": 256, + "hours": 0.5016255555555552, + "language": "en", + "models": { + "parakeet": { + "words": 5073, + "errors": 442, + "substitutions": 212, + "deletions": 77, + "insertions": 153, + "chars": 26079, + "char_errors": 1268, + "utterance_error": 162, + "wer": 8.712793219002563, + "cer": 4.862149622301469 + }, + "orukeet": { + "words": 5073, + "errors": 405, + "substitutions": 210, + "deletions": 70, + "insertions": 125, + "chars": 26079, + "char_errors": 1148, + "utterance_error": 161, + "wer": 7.983441750443524, + "cer": 4.402009279496913 + } + } + }, + "gigaspeechbench_art_en": { + "rows": 256, + "hours": 0.4463050000000003, + "language": "en", + "models": { + "parakeet": { + "words": 4311, + "errors": 236, + "substitutions": 109, + "deletions": 47, + "insertions": 80, + "chars": 22835, + "char_errors": 653, + "utterance_error": 106, + "wer": 5.474367896079796, + "cer": 2.8596452813663236 + }, + "orukeet": { + "words": 4311, + "errors": 210, + "substitutions": 109, + "deletions": 40, + "insertions": 61, + "chars": 22835, + "char_errors": 563, + "utterance_error": 102, + "wer": 4.8712595685455815, + "cer": 2.4655134661703526 + } + } + }, + "gigaspeechbench_bio_en": { + "rows": 256, + "hours": 0.5044194444444443, + "language": "en", + "models": { + "parakeet": { + "words": 4871, + "errors": 179, + "substitutions": 108, + "deletions": 16, + "insertions": 55, + "chars": 27904, + "char_errors": 451, + "utterance_error": 92, + "wer": 3.6748101005953604, + "cer": 1.6162557339449541 + }, + "orukeet": { + "words": 4871, + "errors": 161, + "substitutions": 111, + "deletions": 13, + "insertions": 37, + "chars": 27904, + "char_errors": 381, + "utterance_error": 87, + "wer": 3.305276123999179, + "cer": 1.365395642201835 + } + } + }, + "gigaspeechbench_chn_en": { + "rows": 256, + "hours": 0.3659214236111112, + "language": "en", + "models": { + "parakeet": { + "words": 2795, + "errors": 405, + "substitutions": 281, + "deletions": 61, + "insertions": 63, + "chars": 14020, + "char_errors": 1261, + "utterance_error": 165, + "wer": 14.490161001788909, + "cer": 8.99429386590585 + }, + "orukeet": { + "words": 2795, + "errors": 379, + "substitutions": 252, + "deletions": 64, + "insertions": 63, + "chars": 14020, + "char_errors": 1180, + "utterance_error": 157, + "wer": 13.559928443649374, + "cer": 8.41654778887304 + } + } + }, + "gigaspeechbench_ecm_en": { + "rows": 256, + "hours": 0.47235996527777796, + "language": "en", + "models": { + "parakeet": { + "words": 5800, + "errors": 409, + "substitutions": 223, + "deletions": 60, + "insertions": 126, + "chars": 29608, + "char_errors": 1272, + "utterance_error": 139, + "wer": 7.051724137931035, + "cer": 4.29613617941097 + }, + "orukeet": { + "words": 5800, + "errors": 381, + "substitutions": 213, + "deletions": 48, + "insertions": 120, + "chars": 29608, + "char_errors": 1174, + "utterance_error": 134, + "wer": 6.568965517241379, + "cer": 3.9651445555255336 + } + } + }, + "gigaspeechbench_eng_en": { + "rows": 256, + "hours": 0.46552720486111093, + "language": "en", + "models": { + "parakeet": { + "words": 4485, + "errors": 182, + "substitutions": 102, + "deletions": 27, + "insertions": 53, + "chars": 22477, + "char_errors": 493, + "utterance_error": 98, + "wer": 4.057971014492754, + "cer": 2.1933532054989544 + }, + "orukeet": { + "words": 4485, + "errors": 157, + "substitutions": 95, + "deletions": 23, + "insertions": 39, + "chars": 22477, + "char_errors": 415, + "utterance_error": 90, + "wer": 3.500557413600892, + "cer": 1.846331805845976 + } + } + }, + "gigaspeechbench_ent_en": { + "rows": 256, + "hours": 0.2886933854166668, + "language": "en", + "models": { + "parakeet": { + "words": 3720, + "errors": 387, + "substitutions": 208, + "deletions": 98, + "insertions": 81, + "chars": 17497, + "char_errors": 1189, + "utterance_error": 133, + "wer": 10.403225806451612, + "cer": 6.795450648682631 + }, + "orukeet": { + "words": 3720, + "errors": 330, + "substitutions": 174, + "deletions": 77, + "insertions": 79, + "chars": 17497, + "char_errors": 987, + "utterance_error": 129, + "wer": 8.870967741935484, + "cer": 5.640967022918215 + } + } + }, + "gigaspeechbench_fin_en": { + "rows": 256, + "hours": 0.4903747569444444, + "language": "en", + "models": { + "parakeet": { + "words": 5813, + "errors": 338, + "substitutions": 161, + "deletions": 60, + "insertions": 117, + "chars": 29726, + "char_errors": 1015, + "utterance_error": 135, + "wer": 5.814553586788233, + "cer": 3.4145192760546323 + }, + "orukeet": { + "words": 5813, + "errors": 297, + "substitutions": 156, + "deletions": 48, + "insertions": 93, + "chars": 29726, + "char_errors": 893, + "utterance_error": 132, + "wer": 5.109237915018063, + "cer": 3.0041041512480655 + } + } + }, + "gigaspeechbench_hum_en": { + "rows": 256, + "hours": 0.5111050347222221, + "language": "en", + "models": { + "parakeet": { + "words": 5062, + "errors": 404, + "substitutions": 184, + "deletions": 94, + "insertions": 126, + "chars": 27303, + "char_errors": 1281, + "utterance_error": 137, + "wer": 7.981035163966811, + "cer": 4.69179211075706 + }, + "orukeet": { + "words": 5062, + "errors": 378, + "substitutions": 189, + "deletions": 71, + "insertions": 118, + "chars": 27303, + "char_errors": 1145, + "utterance_error": 136, + "wer": 7.467404188067957, + "cer": 4.193678350364428 + } + } + }, + "gigaspeechbench_ind_en": { + "rows": 256, + "hours": 0.5177380729166666, + "language": "en", + "models": { + "parakeet": { + "words": 5259, + "errors": 342, + "substitutions": 227, + "deletions": 62, + "insertions": 53, + "chars": 29317, + "char_errors": 939, + "utterance_error": 125, + "wer": 6.5031374786081, + "cer": 3.202919807620152 + }, + "orukeet": { + "words": 5259, + "errors": 294, + "substitutions": 194, + "deletions": 47, + "insertions": 53, + "chars": 29317, + "char_errors": 739, + "utterance_error": 116, + "wer": 5.590416428978894, + "cer": 2.520721765528533 + } + } + }, + "gigaspeechbench_jpn_en": { + "rows": 256, + "hours": 0.312707482638889, + "language": "en", + "models": { + "parakeet": { + "words": 2554, + "errors": 489, + "substitutions": 283, + "deletions": 94, + "insertions": 112, + "chars": 12402, + "char_errors": 1474, + "utterance_error": 158, + "wer": 19.14643696162882, + "cer": 11.885179809708111 + }, + "orukeet": { + "words": 2554, + "errors": 454, + "substitutions": 270, + "deletions": 85, + "insertions": 99, + "chars": 12402, + "char_errors": 1347, + "utterance_error": 150, + "wer": 17.776037588097104, + "cer": 10.861151427189164 + } + } + }, + "gigaspeechbench_law_en": { + "rows": 256, + "hours": 0.39185109375, + "language": "en", + "models": { + "parakeet": { + "words": 3958, + "errors": 386, + "substitutions": 230, + "deletions": 74, + "insertions": 82, + "chars": 20771, + "char_errors": 1115, + "utterance_error": 152, + "wer": 9.752400202122285, + "cer": 5.368061239227769 + }, + "orukeet": { + "words": 3958, + "errors": 358, + "substitutions": 203, + "deletions": 76, + "insertions": 79, + "chars": 20771, + "char_errors": 1045, + "utterance_error": 144, + "wer": 9.044972208185953, + "cer": 5.031052910307641 + } + } + }, + "gigaspeechbench_med_en": { + "rows": 256, + "hours": 0.5099035937499998, + "language": "en", + "models": { + "parakeet": { + "words": 5251, + "errors": 183, + "substitutions": 114, + "deletions": 22, + "insertions": 47, + "chars": 28617, + "char_errors": 502, + "utterance_error": 94, + "wer": 3.485050466577795, + "cer": 1.7542020477338645 + }, + "orukeet": { + "words": 5251, + "errors": 167, + "substitutions": 105, + "deletions": 23, + "insertions": 39, + "chars": 28617, + "char_errors": 444, + "utterance_error": 88, + "wer": 3.180346600647496, + "cer": 1.551525317119195 + } + } + }, + "gigaspeechbench_mil_en": { + "rows": 256, + "hours": 0.545510017361111, + "language": "en", + "models": { + "parakeet": { + "words": 5953, + "errors": 204, + "substitutions": 135, + "deletions": 35, + "insertions": 34, + "chars": 32433, + "char_errors": 501, + "utterance_error": 109, + "wer": 3.4268436082647407, + "cer": 1.5447229673480714 + }, + "orukeet": { + "words": 5953, + "errors": 182, + "substitutions": 122, + "deletions": 29, + "insertions": 31, + "chars": 32433, + "char_errors": 455, + "utterance_error": 97, + "wer": 3.0572820426675626, + "cer": 1.4028921160546357 + } + } + }, + "gigaspeechbench_phl_en": { + "rows": 256, + "hours": 0.3380872395833332, + "language": "en", + "models": { + "parakeet": { + "words": 3511, + "errors": 467, + "substitutions": 231, + "deletions": 106, + "insertions": 130, + "chars": 17501, + "char_errors": 1449, + "utterance_error": 156, + "wer": 13.30105383081743, + "cer": 8.279526884178047 + }, + "orukeet": { + "words": 3511, + "errors": 449, + "substitutions": 232, + "deletions": 93, + "insertions": 124, + "chars": 17501, + "char_errors": 1334, + "utterance_error": 151, + "wer": 12.788379379094275, + "cer": 7.622421575909948 + } + } + }, + "gigaspeechbench_sct_en": { + "rows": 256, + "hours": 0.23421114583333333, + "language": "en", + "models": { + "parakeet": { + "words": 2840, + "errors": 627, + "substitutions": 373, + "deletions": 152, + "insertions": 102, + "chars": 13641, + "char_errors": 1991, + "utterance_error": 179, + "wer": 22.077464788732396, + "cer": 14.595704127263398 + }, + "orukeet": { + "words": 2840, + "errors": 578, + "substitutions": 336, + "deletions": 146, + "insertions": 96, + "chars": 13641, + "char_errors": 1790, + "utterance_error": 172, + "wer": 20.35211267605634, + "cer": 13.122205116926912 + } + } + }, + "gigaspeechbench_sgp_en": { + "rows": 256, + "hours": 0.2844177951388889, + "language": "en", + "models": { + "parakeet": { + "words": 3304, + "errors": 459, + "substitutions": 236, + "deletions": 122, + "insertions": 101, + "chars": 16250, + "char_errors": 1503, + "utterance_error": 155, + "wer": 13.892251815980629, + "cer": 9.249230769230769 + }, + "orukeet": { + "words": 3304, + "errors": 425, + "substitutions": 222, + "deletions": 112, + "insertions": 91, + "chars": 16250, + "char_errors": 1334, + "utterance_error": 152, + "wer": 12.86319612590799, + "cer": 8.20923076923077 + } + } + }, + "golos_crowd_ru": { + "rows": 256, + "hours": 0.2929183506944444, + "language": "ru", + "models": { + "parakeet": { + "words": 1303, + "errors": 37, + "substitutions": 34, + "deletions": 2, + "insertions": 1, + "chars": 8520, + "char_errors": 56, + "utterance_error": 31, + "wer": 2.83960092095165, + "cer": 0.6572769953051644 + }, + "orukeet": { + "words": 1302, + "errors": 38, + "substitutions": 35, + "deletions": 2, + "insertions": 1, + "chars": 8520, + "char_errors": 61, + "utterance_error": 33, + "wer": 2.9185867895545314, + "cer": 0.715962441314554 + } + } + }, + "golos_farfield_ru": { + "rows": 256, + "hours": 0.18247008680555546, + "language": "ru", + "models": { + "parakeet": { + "words": 1065, + "errors": 85, + "substitutions": 74, + "deletions": 5, + "insertions": 6, + "chars": 6941, + "char_errors": 180, + "utterance_error": 60, + "wer": 7.981220657276995, + "cer": 2.593286269989915 + }, + "orukeet": { + "words": 1066, + "errors": 97, + "substitutions": 81, + "deletions": 6, + "insertions": 10, + "chars": 6941, + "char_errors": 210, + "utterance_error": 67, + "wer": 9.099437148217635, + "cer": 3.0255006483215676 + } + } + }, + "lesbos_el": { + "rows": 230, + "hours": 0.11232479166666667, + "language": "el", + "models": { + "parakeet": { + "words": 1131, + "errors": 1072, + "substitutions": 744, + "deletions": 240, + "insertions": 88, + "chars": 5811, + "char_errors": 4134, + "utterance_error": 229, + "wer": 94.78337754199823, + "cer": 71.14093959731544 + }, + "orukeet": { + "words": 1131, + "errors": 1058, + "substitutions": 688, + "deletions": 303, + "insertions": 67, + "chars": 5811, + "char_errors": 4164, + "utterance_error": 229, + "wer": 93.54553492484527, + "cer": 71.65720185854414 + } + } + }, + "monsoon_en_in": { + "rows": 256, + "hours": 0.6956833333333333, + "language": "en", + "models": { + "parakeet": { + "words": 6137, + "errors": 253, + "substitutions": 139, + "deletions": 35, + "insertions": 79, + "chars": 31236, + "char_errors": 600, + "utterance_error": 135, + "wer": 4.122535440769106, + "cer": 1.920860545524395 + }, + "orukeet": { + "words": 6137, + "errors": 232, + "substitutions": 133, + "deletions": 38, + "insertions": 61, + "chars": 31236, + "char_errors": 563, + "utterance_error": 132, + "wer": 3.7803487045787842, + "cer": 1.8024074785503905 + } + } + }, + "nst_da_da": { + "rows": 256, + "hours": 0.3674177777777777, + "language": "da", + "models": { + "parakeet": { + "words": 2571, + "errors": 681, + "substitutions": 406, + "deletions": 239, + "insertions": 36, + "chars": 13714, + "char_errors": 1715, + "utterance_error": 209, + "wer": 26.487747957992998, + "cer": 12.505468863934665 + }, + "orukeet": { + "words": 2579, + "errors": 299, + "substitutions": 243, + "deletions": 41, + "insertions": 15, + "chars": 13714, + "char_errors": 604, + "utterance_error": 153, + "wer": 11.593640946103141, + "cer": 4.404258422050459 + } + } + }, + "nst_sv_sv": { + "rows": 256, + "hours": 0.3503472222222222, + "language": "sv", + "models": { + "parakeet": { + "words": 1980, + "errors": 328, + "substitutions": 259, + "deletions": 45, + "insertions": 24, + "chars": 11754, + "char_errors": 807, + "utterance_error": 161, + "wer": 16.565656565656564, + "cer": 6.865747830525779 + }, + "orukeet": { + "words": 1991, + "errors": 246, + "substitutions": 193, + "deletions": 23, + "insertions": 30, + "chars": 11754, + "char_errors": 417, + "utterance_error": 136, + "wer": 12.355600200904068, + "cer": 3.547728432873915 + } + } + }, + "voxpopuli_cs": { + "rows": 256, + "hours": 0.6932252083333335, + "language": "cs", + "models": { + "parakeet": { + "words": 5397, + "errors": 395, + "substitutions": 246, + "deletions": 85, + "insertions": 64, + "chars": 35419, + "char_errors": 1391, + "utterance_error": 150, + "wer": 7.318880859736891, + "cer": 3.92727067393207 + }, + "orukeet": { + "words": 5396, + "errors": 399, + "substitutions": 237, + "deletions": 103, + "insertions": 59, + "chars": 35419, + "char_errors": 1411, + "utterance_error": 146, + "wer": 7.394366197183099, + "cer": 3.983737541997233 + } + } + }, + "voxpopuli_es": { + "rows": 256, + "hours": 0.8307443402777775, + "language": "es", + "models": { + "parakeet": { + "words": 7739, + "errors": 470, + "substitutions": 210, + "deletions": 152, + "insertions": 108, + "chars": 46958, + "char_errors": 1994, + "utterance_error": 120, + "wer": 6.073136064090968, + "cer": 4.246347800161847 + }, + "orukeet": { + "words": 7738, + "errors": 480, + "substitutions": 216, + "deletions": 151, + "insertions": 113, + "chars": 46958, + "char_errors": 2037, + "utterance_error": 118, + "wer": 6.203153269578703, + "cer": 4.3379189914391585 + } + } + }, + "voxpopuli_hu": { + "rows": 256, + "hours": 0.7771814756944443, + "language": "hu", + "models": { + "parakeet": { + "words": 5577, + "errors": 669, + "substitutions": 505, + "deletions": 89, + "insertions": 75, + "chars": 40981, + "char_errors": 1679, + "utterance_error": 213, + "wer": 11.995696611081227, + "cer": 4.097020570508285 + }, + "orukeet": { + "words": 5576, + "errors": 616, + "substitutions": 454, + "deletions": 101, + "insertions": 61, + "chars": 40981, + "char_errors": 1588, + "utterance_error": 204, + "wer": 11.047345767575322, + "cer": 3.8749664478660844 + } + } + }, + "voxpopuli_it": { + "rows": 256, + "hours": 0.9244941145833333, + "language": "it", + "models": { + "parakeet": { + "words": 8233, + "errors": 936, + "substitutions": 326, + "deletions": 400, + "insertions": 210, + "chars": 51717, + "char_errors": 4507, + "utterance_error": 178, + "wer": 11.368881331227985, + "cer": 8.714735966896765 + }, + "orukeet": { + "words": 8235, + "errors": 973, + "substitutions": 257, + "deletions": 475, + "insertions": 241, + "chars": 51717, + "char_errors": 4943, + "utterance_error": 182, + "wer": 11.815421979356406, + "cer": 9.557785641085136 + } + } + }, + "voxpopuli_nl": { + "rows": 256, + "hours": 0.5597956423611112, + "language": "nl", + "models": { + "parakeet": { + "words": 5442, + "errors": 517, + "substitutions": 253, + "deletions": 163, + "insertions": 101, + "chars": 33326, + "char_errors": 1890, + "utterance_error": 174, + "wer": 9.500183755972069, + "cer": 5.671247674488387 + }, + "orukeet": { + "words": 5440, + "errors": 520, + "substitutions": 255, + "deletions": 167, + "insertions": 98, + "chars": 33326, + "char_errors": 1875, + "utterance_error": 173, + "wer": 9.558823529411764, + "cer": 5.626237772309908 + } + } + }, + "voxpopuli_pl": { + "rows": 256, + "hours": 0.7600847395833334, + "language": "pl", + "models": { + "parakeet": { + "words": 5431, + "errors": 352, + "substitutions": 199, + "deletions": 78, + "insertions": 75, + "chars": 39743, + "char_errors": 1359, + "utterance_error": 142, + "wer": 6.481310992450746, + "cer": 3.4194700953627053 + }, + "orukeet": { + "words": 5431, + "errors": 339, + "substitutions": 190, + "deletions": 74, + "insertions": 75, + "chars": 39743, + "char_errors": 1357, + "utterance_error": 140, + "wer": 6.241944393297735, + "cer": 3.4144377626248645 + } + } + }, + "voxpopuli_ro": { + "rows": 256, + "hours": 0.8114564409722217, + "language": "ro", + "models": { + "parakeet": { + "words": 7439, + "errors": 854, + "substitutions": 651, + "deletions": 129, + "insertions": 74, + "chars": 47152, + "char_errors": 2006, + "utterance_error": 216, + "wer": 11.480037639467671, + "cer": 4.25432643366135 + }, + "orukeet": { + "words": 7438, + "errors": 833, + "substitutions": 615, + "deletions": 143, + "insertions": 75, + "chars": 47152, + "char_errors": 1961, + "utterance_error": 211, + "wer": 11.199247109438021, + "cer": 4.158890397013913 + } + } + } + }, + "models": { + "parakeet": "3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d", + "orukeet": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "manifest_sha256": "c16b21855716bc1f62a9c55f775b137cc720a944f3e9343fd8993788db2eaa8e", + "script_sha256": "1c8951a18eba92ae8fe12c537149c961ae0c0e6f748216e27488305b5a2806c2", + "normalizer_sha256": "138b700db56379c5b1965c67c704b2f9ea37071bb0c959d64209ed82f305e0d2", + "decoding": { + "strategy": "greedy_batch", + "model_type": "tdt", + "durations": [ + 0, + 1, + 2, + 3, + 4 + ], + "greedy": { + "max_symbols": 10 + }, + "beam": { + "beam_size": 2, + "return_best_hypothesis": false, + "score_norm": true, + "tsd_max_sym_exp": 50, + "alsd_max_target_len": 2.0 + } + }, + "precision": "FP32 weights, BF16 CUDA autocast, TF32 matrix multiplication disabled", + "normalizers": { + "en": "Pinned EnglishTextNormalizer with spelling/name/compound maps", + "other": "Pinned MultilingualNormalizer; diacritics retained; language-specific numbers; compound-boundary alignment" + }, + "empty_output_handling": "Every manifest record is scored; empty hypotheses remain.", + "timings": { + "parakeet": { + "seconds": 99.74245833995519, + "rows": 12006 + }, + "orukeet": { + "seconds": 95.6186072769924, + "rows": 12006 + } + }, + "prediction_sha256": { + "parakeet": "7dea719a080b71b4e5bf3b731af07f36d2ba7953265aa359813407880994f90f", + "orukeet": "525d958901b5e7b345719dc77fa8b7231dbe24a4b48dc273ea76b23df35ad7fa" + }, + "numeric_evidence_sha256": "298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df", + "scoring": { + "protocol": "Pinned English and multilingual normalization; compound-aware WER", + "dependencies": { + "kaldialign": "0.12.0", + "num2words": "0.5.14", + "regex": "2026.9.3", + "rapidfuzz": "3.14.6" + }, + "code_sha256": { + "evaluation/standard_asr/rescore.py": "47e77bad2d65f1eb6d7a29a156ba747807b4017d4fcedf8d3d2011c5a4e0ab82", + "evaluation/standard_asr/scoring.py": "3b2bd425a79a94b9afa294ba55efc6b9fb5943d7021ce27394dadf6c58082bd0", + "evaluation/standard_asr/vendor/provenance.json": "f6fb39bc16cb4b52ac62882baab0f1925d7405ee8a98f0ae32419edf9275aa12", + "evaluation/standard_asr/vendor/LICENSE": "5ee13882fce0975f0ad3c3d5c2042af4896c2a669be2d6df5b3d256811267c85", + "evaluation/standard_asr/vendor/english_abbreviations.py": "52997cc963e0bd6568d15554ad8d5fb0759f20f361483632a22f90d2d0e07db1", + "evaluation/standard_asr/vendor/__init__.py": "8c39e42af3204c41c564c970e24525da160663836b54acc1b629e4342f862334", + "evaluation/standard_asr/vendor/multilingual.py": "4c8db6c4171c8c2f5567d2ed625a3a1f0d45156af45ca42d9555c0c9b61a29d5", + "evaluation/standard_asr/vendor/normalizer.py": "490b56393484ef386b486f3679cc2261264af280b3512154431e0bdcf2778295" + }, + "reference_word_counts": "Non-English word-boundary alignment may change reference token counts separately for each model.", + "cer": "Character distance before compound alignment, including normalized spaces.", + "inference_comparison_sha256": "3892c802fd85d6cab5bee77b69a774b0fa6308294f6a7e5e536ef89300e853ba" + } +} diff --git a/evidence/domains-r3-20260908/hypotheses-audit.json b/evidence/domains-r3-20260908/hypotheses-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..89f31a99362a51fc40fe8dabf811edc06bd77ab7 --- /dev/null +++ b/evidence/domains-r3-20260908/hypotheses-audit.json @@ -0,0 +1,35 @@ +{ + "status": "passed", + "publication_authorized": false, + "records": 12006, + "paired_predictions": 24012, + "models": { + "parakeet": "3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d", + "orukeet": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "empty_hypotheses": { + "parakeet": 70, + "orukeet": 61 + }, + "checks": [ + "Every private hypothesis re-scored locally", + "All reference and prediction hashes match", + "Every WER/CER numerator and denominator matches released counts", + "All 94 model/partition WERs match independent upstream batch scoring" + ], + "inputs_sha256": { + "comparison.json": "617a2119027b9e706907fdeb8a5819babb55f1959456d677e744d2b57ddfd2be", + "numeric-evidence.jsonl.gz": "298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df" + }, + "manifest_sha256": "c16b21855716bc1f62a9c55f775b137cc720a944f3e9343fd8993788db2eaa8e", + "prediction_sha256": { + "parakeet": "7dea719a080b71b4e5bf3b731af07f36d2ba7953265aa359813407880994f90f", + "orukeet": "525d958901b5e7b345719dc77fa8b7231dbe24a4b48dc273ea76b23df35ad7fa" + }, + "script_sha256": "80a8c7911819129e39911cd4f78df326485a659a847522779cf40b3b2365eb3a", + "upstream_scoring_verified": { + "revision": "48219c6028db0517d704600d92f31edfc96e8c23", + "model_partition_pairs": 94, + "check": "Unmodified upstream normalizer definitions and full-partition batch scoring reproduce all 94 paired WERs." + } +} diff --git a/evidence/domains-r3-20260908/inference-comparison.json b/evidence/domains-r3-20260908/inference-comparison.json new file mode 100644 index 0000000000000000000000000000000000000000..8fb277c9f5633d65950a791de20d192891f4b0cd --- /dev/null +++ b/evidence/domains-r3-20260908/inference-comparison.json @@ -0,0 +1,1513 @@ +{ + "status": "complete", + "publication_authorized": false, + "rows": 12006, + "sets": { + "eurospeech_bg": { + "rows": 256, + "hours": 1.072353802083333, + "language": "bg", + "models": { + "parakeet": { + "words": 7501, + "errors": 1107, + "substitutions": 681, + "deletions": 134, + "insertions": 292, + "chars": 46718, + "char_errors": 3238, + "utterance_error": 232, + "wer": 14.758032262365019, + "cer": 6.930947386446338 + }, + "orukeet": { + "words": 7501, + "errors": 1013, + "substitutions": 592, + "deletions": 120, + "insertions": 301, + "chars": 46718, + "char_errors": 3015, + "utterance_error": 226, + "wer": 13.504866017864284, + "cer": 6.453615308874523 + } + } + }, + "eurospeech_de": { + "rows": 256, + "hours": 1.0912282638888895, + "language": "de", + "models": { + "parakeet": { + "words": 8458, + "errors": 1274, + "substitutions": 613, + "deletions": 243, + "insertions": 418, + "chars": 56900, + "char_errors": 5658, + "utterance_error": 235, + "wer": 15.062662567982974, + "cer": 9.943760984182777 + }, + "orukeet": { + "words": 8458, + "errors": 1051, + "substitutions": 466, + "deletions": 235, + "insertions": 350, + "chars": 56900, + "char_errors": 4842, + "utterance_error": 227, + "wer": 12.426105462284228, + "cer": 8.509666080843585 + } + } + }, + "eurospeech_el": { + "rows": 256, + "hours": 1.0396271527777776, + "language": "el", + "models": { + "parakeet": { + "words": 7914, + "errors": 2063, + "substitutions": 1816, + "deletions": 130, + "insertions": 117, + "chars": 50648, + "char_errors": 4296, + "utterance_error": 256, + "wer": 26.06772807682588, + "cer": 8.482072342441953 + }, + "orukeet": { + "words": 7914, + "errors": 1395, + "substitutions": 1043, + "deletions": 153, + "insertions": 199, + "chars": 50648, + "char_errors": 3730, + "utterance_error": 252, + "wer": 17.62699014404852, + "cer": 7.364555362501974 + } + } + }, + "eurospeech_en": { + "rows": 256, + "hours": 1.0540027430555547, + "language": "en", + "models": { + "parakeet": { + "words": 9639, + "errors": 2477, + "substitutions": 845, + "deletions": 310, + "insertions": 1322, + "chars": 55385, + "char_errors": 9915, + "utterance_error": 251, + "wer": 25.697686482000208, + "cer": 17.901959014173514 + }, + "orukeet": { + "words": 9639, + "errors": 2426, + "substitutions": 845, + "deletions": 310, + "insertions": 1271, + "chars": 55385, + "char_errors": 9720, + "utterance_error": 250, + "wer": 25.16858595289968, + "cer": 17.549878125846348 + } + } + }, + "eurospeech_et": { + "rows": 256, + "hours": 1.0695780555555554, + "language": "et", + "models": { + "parakeet": { + "words": 6816, + "errors": 2609, + "substitutions": 1722, + "deletions": 351, + "insertions": 536, + "chars": 47716, + "char_errors": 7007, + "utterance_error": 254, + "wer": 38.277582159624416, + "cer": 14.68480174364993 + }, + "orukeet": { + "words": 6816, + "errors": 2004, + "substitutions": 1203, + "deletions": 245, + "insertions": 556, + "chars": 47716, + "char_errors": 5727, + "utterance_error": 253, + "wer": 29.401408450704224, + "cer": 12.002263391734429 + } + } + }, + "eurospeech_fi": { + "rows": 256, + "hours": 1.0576934722222222, + "language": "fi", + "models": { + "parakeet": { + "words": 6347, + "errors": 1185, + "substitutions": 670, + "deletions": 215, + "insertions": 300, + "chars": 51093, + "char_errors": 4009, + "utterance_error": 232, + "wer": 18.670237907672917, + "cer": 7.846476033898969 + }, + "orukeet": { + "words": 6347, + "errors": 1100, + "substitutions": 595, + "deletions": 220, + "insertions": 285, + "chars": 51093, + "char_errors": 3898, + "utterance_error": 228, + "wer": 17.33102253032929, + "cer": 7.629225138472981 + } + } + }, + "eurospeech_fr": { + "rows": 256, + "hours": 1.0628877083333337, + "language": "fr", + "models": { + "parakeet": { + "words": 10608, + "errors": 2122, + "substitutions": 1263, + "deletions": 413, + "insertions": 446, + "chars": 62835, + "char_errors": 7448, + "utterance_error": 246, + "wer": 20.00377073906486, + "cer": 11.853266491604996 + }, + "orukeet": { + "words": 10608, + "errors": 1603, + "substitutions": 841, + "deletions": 280, + "insertions": 482, + "chars": 62835, + "char_errors": 6061, + "utterance_error": 242, + "wer": 15.111236802413273, + "cer": 9.645897986790802 + } + } + }, + "eurospeech_hr": { + "rows": 256, + "hours": 1.0454521527777771, + "language": "hr", + "models": { + "parakeet": { + "words": 7517, + "errors": 997, + "substitutions": 504, + "deletions": 172, + "insertions": 321, + "chars": 45766, + "char_errors": 3976, + "utterance_error": 222, + "wer": 13.263269921511242, + "cer": 8.687672070969716 + }, + "orukeet": { + "words": 7517, + "errors": 973, + "substitutions": 505, + "deletions": 143, + "insertions": 325, + "chars": 45766, + "char_errors": 3876, + "utterance_error": 218, + "wer": 12.94399361447386, + "cer": 8.469169252283354 + } + } + }, + "eurospeech_it": { + "rows": 256, + "hours": 1.0571320486111113, + "language": "it", + "models": { + "parakeet": { + "words": 8128, + "errors": 865, + "substitutions": 382, + "deletions": 220, + "insertions": 263, + "chars": 50659, + "char_errors": 3208, + "utterance_error": 222, + "wer": 10.642224409448819, + "cer": 6.332537160228193 + }, + "orukeet": { + "words": 8128, + "errors": 993, + "substitutions": 385, + "deletions": 231, + "insertions": 377, + "chars": 50659, + "char_errors": 4167, + "utterance_error": 228, + "wer": 12.217027559055119, + "cer": 8.225586766418603 + } + } + }, + "eurospeech_lt": { + "rows": 256, + "hours": 1.0921616319444443, + "language": "lt", + "models": { + "parakeet": { + "words": 6394, + "errors": 2521, + "substitutions": 1826, + "deletions": 336, + "insertions": 359, + "chars": 45706, + "char_errors": 7374, + "utterance_error": 254, + "wer": 39.427588364091335, + "cer": 16.13354920579355 + }, + "orukeet": { + "words": 6394, + "errors": 2204, + "substitutions": 1507, + "deletions": 261, + "insertions": 436, + "chars": 45706, + "char_errors": 7056, + "utterance_error": 253, + "wer": 34.469815451986236, + "cer": 15.43779810090579 + } + } + }, + "eurospeech_lv": { + "rows": 256, + "hours": 1.0540805208333333, + "language": "lv", + "models": { + "parakeet": { + "words": 6437, + "errors": 3721, + "substitutions": 2704, + "deletions": 769, + "insertions": 248, + "chars": 44102, + "char_errors": 11751, + "utterance_error": 256, + "wer": 57.80643156750039, + "cer": 26.645050111106073 + }, + "orukeet": { + "words": 6437, + "errors": 2770, + "substitutions": 2061, + "deletions": 373, + "insertions": 336, + "chars": 44102, + "char_errors": 7723, + "utterance_error": 255, + "wer": 43.03246854124592, + "cer": 17.511677474944445 + } + } + }, + "eurospeech_mt": { + "rows": 256, + "hours": 1.0746674826388882, + "language": "mt", + "models": { + "parakeet": { + "words": 7827, + "errors": 3189, + "substitutions": 2047, + "deletions": 518, + "insertions": 624, + "chars": 46941, + "char_errors": 9028, + "utterance_error": 256, + "wer": 40.743579915676506, + "cer": 19.232653756843696 + }, + "orukeet": { + "words": 7827, + "errors": 3134, + "substitutions": 2051, + "deletions": 461, + "insertions": 622, + "chars": 46941, + "char_errors": 8813, + "utterance_error": 256, + "wer": 40.040884119074995, + "cer": 18.774631984832023 + } + } + }, + "eurospeech_pt": { + "rows": 256, + "hours": 1.0251623090277782, + "language": "pt", + "models": { + "parakeet": { + "words": 8186, + "errors": 1908, + "substitutions": 551, + "deletions": 567, + "insertions": 790, + "chars": 48112, + "char_errors": 8435, + "utterance_error": 241, + "wer": 23.30808697776692, + "cer": 17.53200864649152 + }, + "orukeet": { + "words": 8186, + "errors": 1989, + "substitutions": 576, + "deletions": 592, + "insertions": 821, + "chars": 48112, + "char_errors": 8947, + "utterance_error": 238, + "wer": 24.297581236257024, + "cer": 18.596192218157633 + } + } + }, + "eurospeech_sk": { + "rows": 256, + "hours": 1.0663451041666665, + "language": "sk", + "models": { + "parakeet": { + "words": 7313, + "errors": 1361, + "substitutions": 927, + "deletions": 235, + "insertions": 199, + "chars": 45102, + "char_errors": 3711, + "utterance_error": 242, + "wer": 18.610693285929166, + "cer": 8.22801649594253 + }, + "orukeet": { + "words": 7313, + "errors": 1238, + "substitutions": 798, + "deletions": 191, + "insertions": 249, + "chars": 45102, + "char_errors": 4012, + "utterance_error": 232, + "wer": 16.928757008067823, + "cer": 8.895392665513725 + } + } + }, + "eurospeech_sl": { + "rows": 256, + "hours": 1.042155416666667, + "language": "sl", + "models": { + "parakeet": { + "words": 7737, + "errors": 3996, + "substitutions": 3162, + "deletions": 467, + "insertions": 367, + "chars": 46542, + "char_errors": 8017, + "utterance_error": 256, + "wer": 51.647925552539746, + "cer": 17.225301877873747 + }, + "orukeet": { + "words": 7737, + "errors": 4176, + "substitutions": 3351, + "deletions": 474, + "insertions": 351, + "chars": 46542, + "char_errors": 8295, + "utterance_error": 256, + "wer": 53.97440868553703, + "cer": 17.822611834472088 + } + } + }, + "eurospeech_uk": { + "rows": 256, + "hours": 1.0933650694444448, + "language": "uk", + "models": { + "parakeet": { + "words": 7235, + "errors": 1101, + "substitutions": 531, + "deletions": 173, + "insertions": 397, + "chars": 46747, + "char_errors": 4132, + "utterance_error": 230, + "wer": 15.21769177608846, + "cer": 8.839069886837658 + }, + "orukeet": { + "words": 7235, + "errors": 1226, + "substitutions": 584, + "deletions": 150, + "insertions": 492, + "chars": 46747, + "char_errors": 5057, + "utterance_error": 227, + "wer": 16.945404284727022, + "cer": 10.817806490256059 + } + } + }, + "gigaspeechbench_agr_en": { + "rows": 256, + "hours": 0.4082733159722223, + "language": "en", + "models": { + "parakeet": { + "words": 4682, + "errors": 318, + "substitutions": 132, + "deletions": 95, + "insertions": 91, + "chars": 23326, + "char_errors": 929, + "utterance_error": 135, + "wer": 6.791969243912858, + "cer": 3.982680270942296 + }, + "orukeet": { + "words": 4682, + "errors": 291, + "substitutions": 140, + "deletions": 77, + "insertions": 74, + "chars": 23326, + "char_errors": 791, + "utterance_error": 119, + "wer": 6.215292609995728, + "cer": 3.3910657635256793 + } + } + }, + "gigaspeechbench_ait_en": { + "rows": 256, + "hours": 0.5016255555555552, + "language": "en", + "models": { + "parakeet": { + "words": 5162, + "errors": 544, + "substitutions": 249, + "deletions": 151, + "insertions": 144, + "chars": 26199, + "char_errors": 1312, + "utterance_error": 173, + "wer": 10.53855094924448, + "cer": 5.007824726134586 + }, + "orukeet": { + "words": 5162, + "errors": 512, + "substitutions": 252, + "deletions": 143, + "insertions": 117, + "chars": 26199, + "char_errors": 1211, + "utterance_error": 173, + "wer": 9.918636187524216, + "cer": 4.62231382877209 + } + } + }, + "gigaspeechbench_art_en": { + "rows": 256, + "hours": 0.4463050000000003, + "language": "en", + "models": { + "parakeet": { + "words": 4336, + "errors": 263, + "substitutions": 117, + "deletions": 66, + "insertions": 80, + "chars": 22903, + "char_errors": 709, + "utterance_error": 119, + "wer": 6.06549815498155, + "cer": 3.0956643234510763 + }, + "orukeet": { + "words": 4336, + "errors": 234, + "substitutions": 116, + "deletions": 60, + "insertions": 58, + "chars": 22903, + "char_errors": 616, + "utterance_error": 114, + "wer": 5.3966789667896675, + "cer": 2.68960398201109 + } + } + }, + "gigaspeechbench_bio_en": { + "rows": 256, + "hours": 0.5044194444444443, + "language": "en", + "models": { + "parakeet": { + "words": 4974, + "errors": 335, + "substitutions": 166, + "deletions": 115, + "insertions": 54, + "chars": 28032, + "char_errors": 548, + "utterance_error": 123, + "wer": 6.73502211499799, + "cer": 1.9549086757990868 + }, + "orukeet": { + "words": 4974, + "errors": 296, + "substitutions": 158, + "deletions": 99, + "insertions": 39, + "chars": 28032, + "char_errors": 474, + "utterance_error": 115, + "wer": 5.950944913550463, + "cer": 1.6909246575342465 + } + } + }, + "gigaspeechbench_chn_en": { + "rows": 256, + "hours": 0.3659214236111112, + "language": "en", + "models": { + "parakeet": { + "words": 2816, + "errors": 433, + "substitutions": 296, + "deletions": 75, + "insertions": 62, + "chars": 14060, + "char_errors": 1300, + "utterance_error": 170, + "wer": 15.376420454545455, + "cer": 9.246088193456615 + }, + "orukeet": { + "words": 2816, + "errors": 411, + "substitutions": 278, + "deletions": 73, + "insertions": 60, + "chars": 14060, + "char_errors": 1217, + "utterance_error": 160, + "wer": 14.595170454545455, + "cer": 8.655761024182077 + } + } + }, + "gigaspeechbench_ecm_en": { + "rows": 256, + "hours": 0.47235996527777796, + "language": "en", + "models": { + "parakeet": { + "words": 5843, + "errors": 485, + "substitutions": 250, + "deletions": 106, + "insertions": 129, + "chars": 29666, + "char_errors": 1301, + "utterance_error": 153, + "wer": 8.30053054937532, + "cer": 4.385491808804693 + }, + "orukeet": { + "words": 5843, + "errors": 463, + "substitutions": 245, + "deletions": 89, + "insertions": 129, + "chars": 29666, + "char_errors": 1201, + "utterance_error": 150, + "wer": 7.924011637857265, + "cer": 4.048405582147914 + } + } + }, + "gigaspeechbench_eng_en": { + "rows": 256, + "hours": 0.46552720486111093, + "language": "en", + "models": { + "parakeet": { + "words": 4572, + "errors": 281, + "substitutions": 138, + "deletions": 89, + "insertions": 54, + "chars": 22588, + "char_errors": 549, + "utterance_error": 117, + "wer": 6.146106736657917, + "cer": 2.430494067646538 + }, + "orukeet": { + "words": 4572, + "errors": 228, + "substitutions": 119, + "deletions": 70, + "insertions": 39, + "chars": 22588, + "char_errors": 461, + "utterance_error": 104, + "wer": 4.986876640419948, + "cer": 2.0409066761112094 + } + } + }, + "gigaspeechbench_ent_en": { + "rows": 256, + "hours": 0.2886933854166668, + "language": "en", + "models": { + "parakeet": { + "words": 3737, + "errors": 403, + "substitutions": 210, + "deletions": 109, + "insertions": 84, + "chars": 17535, + "char_errors": 1201, + "utterance_error": 134, + "wer": 10.784051378110783, + "cer": 6.84915882520673 + }, + "orukeet": { + "words": 3737, + "errors": 347, + "substitutions": 180, + "deletions": 86, + "insertions": 81, + "chars": 17535, + "char_errors": 1003, + "utterance_error": 132, + "wer": 9.285523146909286, + "cer": 5.719988594240092 + } + } + }, + "gigaspeechbench_fin_en": { + "rows": 256, + "hours": 0.4903747569444444, + "language": "en", + "models": { + "parakeet": { + "words": 5877, + "errors": 440, + "substitutions": 207, + "deletions": 110, + "insertions": 123, + "chars": 29811, + "char_errors": 1064, + "utterance_error": 146, + "wer": 7.486812999829845, + "cer": 3.5691523263224982 + }, + "orukeet": { + "words": 5877, + "errors": 397, + "substitutions": 198, + "deletions": 101, + "insertions": 98, + "chars": 29811, + "char_errors": 946, + "utterance_error": 143, + "wer": 6.755147183937383, + "cer": 3.1733252826138 + } + } + }, + "gigaspeechbench_hum_en": { + "rows": 256, + "hours": 0.5111050347222221, + "language": "en", + "models": { + "parakeet": { + "words": 5075, + "errors": 421, + "substitutions": 196, + "deletions": 99, + "insertions": 126, + "chars": 27339, + "char_errors": 1300, + "utterance_error": 142, + "wer": 8.295566502463053, + "cer": 4.75511174512601 + }, + "orukeet": { + "words": 5075, + "errors": 397, + "substitutions": 200, + "deletions": 77, + "insertions": 120, + "chars": 27339, + "char_errors": 1168, + "utterance_error": 141, + "wer": 7.822660098522167, + "cer": 4.272285014082446 + } + } + }, + "gigaspeechbench_ind_en": { + "rows": 256, + "hours": 0.5177380729166666, + "language": "en", + "models": { + "parakeet": { + "words": 5330, + "errors": 455, + "substitutions": 275, + "deletions": 129, + "insertions": 51, + "chars": 29433, + "char_errors": 1005, + "utterance_error": 145, + "wer": 8.536585365853659, + "cer": 3.4145347059423097 + }, + "orukeet": { + "words": 5330, + "errors": 412, + "substitutions": 246, + "deletions": 113, + "insertions": 53, + "chars": 29433, + "char_errors": 806, + "utterance_error": 139, + "wer": 7.729831144465291, + "cer": 2.7384228586960213 + } + } + }, + "gigaspeechbench_jpn_en": { + "rows": 256, + "hours": 0.312707482638889, + "language": "en", + "models": { + "parakeet": { + "words": 2576, + "errors": 513, + "substitutions": 297, + "deletions": 106, + "insertions": 110, + "chars": 12445, + "char_errors": 1500, + "utterance_error": 163, + "wer": 19.914596273291924, + "cer": 12.053033346725593 + }, + "orukeet": { + "words": 2576, + "errors": 475, + "substitutions": 293, + "deletions": 91, + "insertions": 91, + "chars": 12445, + "char_errors": 1374, + "utterance_error": 154, + "wer": 18.43944099378882, + "cer": 11.040578545600642 + } + } + }, + "gigaspeechbench_law_en": { + "rows": 256, + "hours": 0.39185109375, + "language": "en", + "models": { + "parakeet": { + "words": 3999, + "errors": 445, + "substitutions": 248, + "deletions": 104, + "insertions": 93, + "chars": 20831, + "char_errors": 1160, + "utterance_error": 155, + "wer": 11.127781945486372, + "cer": 5.568623685852815 + }, + "orukeet": { + "words": 3999, + "errors": 409, + "substitutions": 220, + "deletions": 104, + "insertions": 85, + "chars": 20831, + "char_errors": 1091, + "utterance_error": 149, + "wer": 10.227556889222306, + "cer": 5.237386587297777 + } + } + }, + "gigaspeechbench_med_en": { + "rows": 256, + "hours": 0.5099035937499998, + "language": "en", + "models": { + "parakeet": { + "words": 5310, + "errors": 274, + "substitutions": 147, + "deletions": 78, + "insertions": 49, + "chars": 28697, + "char_errors": 548, + "utterance_error": 112, + "wer": 5.160075329566855, + "cer": 1.9096072760218838 + }, + "orukeet": { + "words": 5310, + "errors": 262, + "substitutions": 140, + "deletions": 80, + "insertions": 42, + "chars": 28697, + "char_errors": 490, + "utterance_error": 106, + "wer": 4.934086629001883, + "cer": 1.7074955570268668 + } + } + }, + "gigaspeechbench_mil_en": { + "rows": 256, + "hours": 0.545510017361111, + "language": "en", + "models": { + "parakeet": { + "words": 6048, + "errors": 362, + "substitutions": 199, + "deletions": 123, + "insertions": 40, + "chars": 32571, + "char_errors": 574, + "utterance_error": 139, + "wer": 5.985449735449736, + "cer": 1.7623038899634644 + }, + "orukeet": { + "words": 6048, + "errors": 338, + "substitutions": 188, + "deletions": 115, + "insertions": 35, + "chars": 32571, + "char_errors": 531, + "utterance_error": 129, + "wer": 5.588624338624339, + "cer": 1.6302846090080132 + } + } + }, + "gigaspeechbench_phl_en": { + "rows": 256, + "hours": 0.3380872395833332, + "language": "en", + "models": { + "parakeet": { + "words": 3528, + "errors": 494, + "substitutions": 249, + "deletions": 116, + "insertions": 129, + "chars": 17518, + "char_errors": 1466, + "utterance_error": 159, + "wer": 14.002267573696145, + "cer": 8.36853522091563 + }, + "orukeet": { + "words": 3528, + "errors": 468, + "substitutions": 248, + "deletions": 99, + "insertions": 121, + "chars": 17518, + "char_errors": 1341, + "utterance_error": 153, + "wer": 13.26530612244898, + "cer": 7.654983445598813 + } + } + }, + "gigaspeechbench_sct_en": { + "rows": 256, + "hours": 0.23421114583333333, + "language": "en", + "models": { + "parakeet": { + "words": 2860, + "errors": 648, + "substitutions": 391, + "deletions": 154, + "insertions": 103, + "chars": 13690, + "char_errors": 2007, + "utterance_error": 180, + "wer": 22.657342657342657, + "cer": 14.660336011687363 + }, + "orukeet": { + "words": 2860, + "errors": 592, + "substitutions": 355, + "deletions": 146, + "insertions": 91, + "chars": 13690, + "char_errors": 1800, + "utterance_error": 174, + "wer": 20.6993006993007, + "cer": 13.148283418553689 + } + } + }, + "gigaspeechbench_sgp_en": { + "rows": 256, + "hours": 0.2844177951388889, + "language": "en", + "models": { + "parakeet": { + "words": 3339, + "errors": 499, + "substitutions": 254, + "deletions": 147, + "insertions": 98, + "chars": 16334, + "char_errors": 1549, + "utterance_error": 160, + "wer": 14.94459418987721, + "cer": 9.483286396473613 + }, + "orukeet": { + "words": 3339, + "errors": 466, + "substitutions": 238, + "deletions": 139, + "insertions": 89, + "chars": 16334, + "char_errors": 1382, + "utterance_error": 158, + "wer": 13.956274333632825, + "cer": 8.460879147789885 + } + } + }, + "golos_crowd_ru": { + "rows": 256, + "hours": 0.2929183506944444, + "language": "ru", + "models": { + "parakeet": { + "words": 1304, + "errors": 41, + "substitutions": 36, + "deletions": 3, + "insertions": 2, + "chars": 8520, + "char_errors": 56, + "utterance_error": 33, + "wer": 3.144171779141104, + "cer": 0.6572769953051644 + }, + "orukeet": { + "words": 1304, + "errors": 46, + "substitutions": 39, + "deletions": 5, + "insertions": 2, + "chars": 8520, + "char_errors": 61, + "utterance_error": 35, + "wer": 3.5276073619631902, + "cer": 0.715962441314554 + } + } + }, + "golos_farfield_ru": { + "rows": 256, + "hours": 0.18247008680555546, + "language": "ru", + "models": { + "parakeet": { + "words": 1066, + "errors": 91, + "substitutions": 77, + "deletions": 6, + "insertions": 8, + "chars": 6941, + "char_errors": 180, + "utterance_error": 62, + "wer": 8.536585365853659, + "cer": 2.593286269989915 + }, + "orukeet": { + "words": 1066, + "errors": 101, + "substitutions": 83, + "deletions": 7, + "insertions": 11, + "chars": 6941, + "char_errors": 210, + "utterance_error": 68, + "wer": 9.47467166979362, + "cer": 3.0255006483215676 + } + } + }, + "lesbos_el": { + "rows": 230, + "hours": 0.11232479166666667, + "language": "el", + "models": { + "parakeet": { + "words": 1131, + "errors": 1087, + "substitutions": 755, + "deletions": 232, + "insertions": 100, + "chars": 5811, + "char_errors": 4146, + "utterance_error": 229, + "wer": 96.10963748894784, + "cer": 71.34744450180692 + }, + "orukeet": { + "words": 1131, + "errors": 1053, + "substitutions": 680, + "deletions": 303, + "insertions": 70, + "chars": 5811, + "char_errors": 4141, + "utterance_error": 228, + "wer": 93.10344827586206, + "cer": 71.26140079160213 + } + } + }, + "monsoon_en_in": { + "rows": 256, + "hours": 0.6956833333333333, + "language": "en", + "models": { + "parakeet": { + "words": 6185, + "errors": 306, + "substitutions": 144, + "deletions": 79, + "insertions": 83, + "chars": 31429, + "char_errors": 773, + "utterance_error": 150, + "wer": 4.947453516572352, + "cer": 2.4595119157466034 + }, + "orukeet": { + "words": 6185, + "errors": 282, + "substitutions": 136, + "deletions": 84, + "insertions": 62, + "chars": 31429, + "char_errors": 738, + "utterance_error": 150, + "wer": 4.559417946645109, + "cer": 2.3481497979573005 + } + } + }, + "nst_da_da": { + "rows": 256, + "hours": 0.3674177777777777, + "language": "da", + "models": { + "parakeet": { + "words": 2583, + "errors": 863, + "substitutions": 472, + "deletions": 355, + "insertions": 36, + "chars": 13714, + "char_errors": 2640, + "utterance_error": 216, + "wer": 33.41076267905536, + "cer": 19.250401050021875 + }, + "orukeet": { + "words": 2583, + "errors": 324, + "substitutions": 257, + "deletions": 48, + "insertions": 19, + "chars": 13714, + "char_errors": 634, + "utterance_error": 157, + "wer": 12.543554006968641, + "cer": 4.623012979437072 + } + } + }, + "nst_sv_sv": { + "rows": 256, + "hours": 0.3503472222222222, + "language": "sv", + "models": { + "parakeet": { + "words": 1994, + "errors": 421, + "substitutions": 327, + "deletions": 60, + "insertions": 34, + "chars": 11754, + "char_errors": 1463, + "utterance_error": 172, + "wer": 21.113340020060182, + "cer": 12.446826612217118 + }, + "orukeet": { + "words": 1994, + "errors": 273, + "substitutions": 209, + "deletions": 25, + "insertions": 39, + "chars": 11754, + "char_errors": 419, + "utterance_error": 145, + "wer": 13.691073219658977, + "cer": 3.5647439169644377 + } + } + }, + "voxpopuli_cs": { + "rows": 256, + "hours": 0.6932252083333335, + "language": "cs", + "models": { + "parakeet": { + "words": 5408, + "errors": 441, + "substitutions": 269, + "deletions": 99, + "insertions": 73, + "chars": 35419, + "char_errors": 1393, + "utterance_error": 155, + "wer": 8.154585798816568, + "cer": 3.9329173607385868 + }, + "orukeet": { + "words": 5408, + "errors": 441, + "substitutions": 260, + "deletions": 114, + "insertions": 67, + "chars": 35419, + "char_errors": 1411, + "utterance_error": 153, + "wer": 8.154585798816568, + "cer": 3.983737541997233 + } + } + }, + "voxpopuli_es": { + "rows": 256, + "hours": 0.8307443402777775, + "language": "es", + "models": { + "parakeet": { + "words": 7740, + "errors": 474, + "substitutions": 216, + "deletions": 151, + "insertions": 107, + "chars": 46958, + "char_errors": 1994, + "utterance_error": 121, + "wer": 6.124031007751938, + "cer": 4.246347800161847 + }, + "orukeet": { + "words": 7740, + "errors": 487, + "substitutions": 224, + "deletions": 150, + "insertions": 113, + "chars": 46958, + "char_errors": 2037, + "utterance_error": 120, + "wer": 6.291989664082688, + "cer": 4.3379189914391585 + } + } + }, + "voxpopuli_hu": { + "rows": 256, + "hours": 0.7771814756944443, + "language": "hu", + "models": { + "parakeet": { + "words": 5597, + "errors": 773, + "substitutions": 562, + "deletions": 111, + "insertions": 100, + "chars": 40981, + "char_errors": 1683, + "utterance_error": 219, + "wer": 13.8109701625871, + "cer": 4.1067811912837655 + }, + "orukeet": { + "words": 5597, + "errors": 727, + "substitutions": 512, + "deletions": 126, + "insertions": 89, + "chars": 40981, + "char_errors": 1587, + "utterance_error": 210, + "wer": 12.989101304270145, + "cer": 3.872526292672214 + } + } + }, + "voxpopuli_it": { + "rows": 256, + "hours": 0.9244941145833333, + "language": "it", + "models": { + "parakeet": { + "words": 8239, + "errors": 954, + "substitutions": 347, + "deletions": 401, + "insertions": 206, + "chars": 51717, + "char_errors": 4507, + "utterance_error": 179, + "wer": 11.579075130477, + "cer": 8.714735966896765 + }, + "orukeet": { + "words": 8239, + "errors": 987, + "substitutions": 272, + "deletions": 475, + "insertions": 240, + "chars": 51717, + "char_errors": 4943, + "utterance_error": 184, + "wer": 11.979609175870857, + "cer": 9.557785641085136 + } + } + }, + "voxpopuli_nl": { + "rows": 256, + "hours": 0.5597956423611112, + "language": "nl", + "models": { + "parakeet": { + "words": 5453, + "errors": 568, + "substitutions": 288, + "deletions": 169, + "insertions": 111, + "chars": 33326, + "char_errors": 1890, + "utterance_error": 183, + "wer": 10.416284613973959, + "cer": 5.671247674488387 + }, + "orukeet": { + "words": 5453, + "errors": 572, + "substitutions": 285, + "deletions": 178, + "insertions": 109, + "chars": 33326, + "char_errors": 1879, + "utterance_error": 181, + "wer": 10.489638730973775, + "cer": 5.638240412890836 + } + } + }, + "voxpopuli_pl": { + "rows": 256, + "hours": 0.7600847395833334, + "language": "pl", + "models": { + "parakeet": { + "words": 5432, + "errors": 354, + "substitutions": 206, + "deletions": 76, + "insertions": 72, + "chars": 39743, + "char_errors": 1359, + "utterance_error": 142, + "wer": 6.516936671575847, + "cer": 3.4194700953627053 + }, + "orukeet": { + "words": 5432, + "errors": 346, + "substitutions": 200, + "deletions": 72, + "insertions": 74, + "chars": 39743, + "char_errors": 1357, + "utterance_error": 142, + "wer": 6.369661266568483, + "cer": 3.4144377626248645 + } + } + }, + "voxpopuli_ro": { + "rows": 256, + "hours": 0.8114564409722217, + "language": "ro", + "models": { + "parakeet": { + "words": 7447, + "errors": 893, + "substitutions": 675, + "deletions": 144, + "insertions": 74, + "chars": 47152, + "char_errors": 2072, + "utterance_error": 218, + "wer": 11.99140593527595, + "cer": 4.394299287410926 + }, + "orukeet": { + "words": 7447, + "errors": 868, + "substitutions": 638, + "deletions": 154, + "insertions": 76, + "chars": 47152, + "char_errors": 1980, + "utterance_error": 215, + "wer": 11.655700281992749, + "cer": 4.199185612487275 + } + } + } + }, + "models": { + "parakeet": "3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d", + "orukeet": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "manifest_sha256": "c16b21855716bc1f62a9c55f775b137cc720a944f3e9343fd8993788db2eaa8e", + "script_sha256": "1c8951a18eba92ae8fe12c537149c961ae0c0e6f748216e27488305b5a2806c2", + "normalizer_sha256": "138b700db56379c5b1965c67c704b2f9ea37071bb0c959d64209ed82f305e0d2", + "decoding": { + "strategy": "greedy_batch", + "model_type": "tdt", + "durations": [ + 0, + 1, + 2, + 3, + 4 + ], + "greedy": { + "max_symbols": 10 + }, + "beam": { + "beam_size": 2, + "return_best_hypothesis": false, + "score_norm": true, + "tsd_max_sym_exp": 50, + "alsd_max_target_len": 2.0 + } + }, + "precision": "FP32 weights, BF16 CUDA autocast, TF32 matrix multiplication disabled", + "normalizers": { + "en": "Whisper EnglishTextNormalizer", + "other": "Recorded multilingual normalizer" + }, + "empty_output_handling": "Every manifest record is scored; empty hypotheses remain.", + "timings": { + "parakeet": { + "seconds": 99.74245833995519, + "rows": 12006 + }, + "orukeet": { + "seconds": 95.6186072769924, + "rows": 12006 + } + }, + "prediction_sha256": { + "parakeet": "7dea719a080b71b4e5bf3b731af07f36d2ba7953265aa359813407880994f90f", + "orukeet": "525d958901b5e7b345719dc77fa8b7231dbe24a4b48dc273ea76b23df35ad7fa" + }, + "numeric_evidence_sha256": "5e14b3748855b438630dbf7dda49d73379e49c1ac490f055fbe76d68ca31f0c8" +} diff --git a/evidence/domains-r3-20260908/launch.json b/evidence/domains-r3-20260908/launch.json new file mode 100644 index 0000000000000000000000000000000000000000..3846894df8454d3684bdad3b4f269eff872a17ac --- /dev/null +++ b/evidence/domains-r3-20260908/launch.json @@ -0,0 +1 @@ +{"pid": 80108, "command": ["/opt/venv/bin/python", "/home/nathanroll/parakeet-ft/standard_asr_20260908/run.py", "--manifest", "/home/nathanroll/parakeet-ft/domains_r3_20260908/manifest.jsonl", "--parakeet", "/home/nathanroll/parakeet-ft/models/parakeet-tdt-0.6b-v3/parakeet-tdt-0.6b-v3.nemo", "--orukeet", "/dev/shm/orukeet-librispeech-ft-20260908/r3/checkpoints/orukeet-targeted-ft.nemo", "--orukeet-sha256", "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", "--metric-code", "/home/nathanroll/parakeet-ft/unseen_20260907/code", "--output", "/home/nathanroll/parakeet-ft/domains_r3_20260908/evaluation"], "baseline_reuse": false, "publication_authorized": false, "source_manifest_sha256": "de74268f6e85706671c936108d5c3cab2d73ae77f6abb9a589e3d7af556d2a1f", "manifest_sha256": "c16b21855716bc1f62a9c55f775b137cc720a944f3e9343fd8993788db2eaa8e", "all_source_fields_unchanged": true, "wait_for": 80021} diff --git a/evidence/domains-r3-20260908/numeric-evidence.jsonl.gz b/evidence/domains-r3-20260908/numeric-evidence.jsonl.gz new file mode 100644 index 0000000000000000000000000000000000000000..677a3edeed073bd29e619484faf52dcbdafe3b69 --- /dev/null +++ b/evidence/domains-r3-20260908/numeric-evidence.jsonl.gz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df +size 719000 diff --git a/evidence/domains-r3-20260908/scores.csv b/evidence/domains-r3-20260908/scores.csv new file mode 100644 index 0000000000000000000000000000000000000000..f25bb91876f1884ad8b2e5cb14622e04d3f0cc36 --- /dev/null +++ b/evidence/domains-r3-20260908/scores.csv @@ -0,0 +1,48 @@ +split,clips,hours,parakeet_wer,parakeet_cer,parakeet_errors,parakeet_words,orukeet_wer,orukeet_cer,orukeet_errors,orukeet_words +eurospeech_bg,256,1.072353802083333,14.21858216970999,7.198539675722109,1059,7448,13.03705692803437,6.717491678299152,971,7448 +eurospeech_de,256,1.0912282638888895,13.40047393364929,8.533264834318008,1131,8440,11.142687981053879,7.147872249336416,941,8445 +eurospeech_el,256,1.0396271527777776,25.828065739570164,8.472534849741342,2043,7910,26.35306019221042,9.009596019428978,2084,7908 +eurospeech_en,256,1.0540027430555547,24.398090493980906,17.845920468089716,2351,9636,23.765047737650477,17.493769639180844,2290,9636 +eurospeech_et,256,1.0695780555555554,34.668040630060354,14.607826671145448,2355,6793,25.325251330573625,11.943651326839099,1713,6764 +eurospeech_fi,256,1.0576934722222222,16.61372299872935,7.178711940616703,1046,6296,15.20356234096692,6.771711533616296,956,6288 +eurospeech_fr,256,1.0628877083333337,19.423112942498594,11.372351794807317,2074,10678,14.283038306640442,8.805871628842647,1525,10677 +eurospeech_hr,256,1.0454521527777771,12.92816206850593,8.677776319818047,970,7503,12.556651559584111,8.461269299741941,942,7502 +eurospeech_it,256,1.0571320486111113,10.951906304510342,6.810519217801753,879,8026,12.32398753894081,8.337630399428821,989,8025 +eurospeech_lt,256,1.0921616319444443,38.43658272825583,16.15373194476849,2488,6473,33.10142238713667,14.340254947292511,2141,6468 +eurospeech_lv,256,1.0540805208333333,57.17844908128309,26.650800422101977,3672,6422,42.13973799126637,17.20963649834976,2702,6412 +eurospeech_mt,256,1.0746674826388882,36.83468000513018,19.343275939320232,2872,7797,36.14674191893278,18.88881776867439,2818,7796 +eurospeech_pt,256,1.0251623090277782,23.078793774319067,17.669622602598473,1898,8224,23.805471124620063,18.42029284388534,1958,8225 +eurospeech_sk,256,1.0663451041666665,17.292612463887743,7.757985524205782,1257,7269,14.907088781830694,6.928011095219521,1083,7265 +eurospeech_sl,256,1.042155416666667,48.42661851947382,15.927240461401952,3755,7754,50.23231801755292,16.52089407191448,3892,7748 +eurospeech_uk,256,1.0933650694444448,13.651692971806286,7.586263153601435,1012,7413,14.249089191742005,7.586263153601435,1056,7411 +gigaspeechbench_ait_en,256,0.5016255555555552,8.712793219002563,4.862149622301469,442,5073,7.983441750443524,4.402009279496913,405,5073 +gigaspeechbench_chn_en,256,0.3659214236111112,14.490161001788909,8.99429386590585,405,2795,13.559928443649374,8.41654778887304,379,2795 +gigaspeechbench_phl_en,256,0.3380872395833332,13.30105383081743,8.279526884178047,467,3511,12.788379379094275,7.622421575909948,449,3511 +gigaspeechbench_ind_en,256,0.5177380729166666,6.5031374786081,3.202919807620152,342,5259,5.590416428978894,2.520721765528533,294,5259 +gigaspeechbench_jpn_en,256,0.312707482638889,19.14643696162882,11.885179809708111,489,2554,17.776037588097104,10.861151427189164,454,2554 +gigaspeechbench_sct_en,256,0.23421114583333333,22.077464788732396,14.595704127263398,627,2840,20.35211267605634,13.122205116926912,578,2840 +gigaspeechbench_sgp_en,256,0.2844177951388889,13.892251815980629,9.249230769230769,459,3304,12.86319612590799,8.20923076923077,425,3304 +gigaspeechbench_agr_en,256,0.4082733159722223,6.201716738197425,3.9577156117055563,289,4660,5.836909871244635,3.407674788363199,272,4660 +gigaspeechbench_art_en,256,0.4463050000000003,5.474367896079796,2.8596452813663236,236,4311,4.8712595685455815,2.4655134661703526,210,4311 +gigaspeechbench_bio_en,256,0.5044194444444443,3.6748101005953604,1.6162557339449541,179,4871,3.305276123999179,1.365395642201835,161,4871 +gigaspeechbench_ecm_en,256,0.47235996527777796,7.051724137931035,4.29613617941097,409,5800,6.568965517241379,3.9651445555255336,381,5800 +gigaspeechbench_eng_en,256,0.46552720486111093,4.057971014492754,2.1933532054989544,182,4485,3.500557413600892,1.846331805845976,157,4485 +gigaspeechbench_ent_en,256,0.2886933854166668,10.403225806451612,6.795450648682631,387,3720,8.870967741935484,5.640967022918215,330,3720 +gigaspeechbench_fin_en,256,0.4903747569444444,5.814553586788233,3.4145192760546323,338,5813,5.109237915018063,3.0041041512480655,297,5813 +gigaspeechbench_hum_en,256,0.5111050347222221,7.981035163966811,4.69179211075706,404,5062,7.467404188067957,4.193678350364428,378,5062 +gigaspeechbench_law_en,256,0.39185109375,9.752400202122285,5.368061239227769,386,3958,9.044972208185953,5.031052910307641,358,3958 +gigaspeechbench_med_en,256,0.5099035937499998,3.485050466577795,1.7542020477338645,183,5251,3.180346600647496,1.551525317119195,167,5251 +gigaspeechbench_mil_en,256,0.545510017361111,3.4268436082647407,1.5447229673480714,204,5953,3.0572820426675626,1.4028921160546357,182,5953 +golos_crowd_ru,256,0.2929183506944444,2.83960092095165,0.6572769953051644,37,1303,2.9185867895545314,0.715962441314554,38,1302 +golos_farfield_ru,256,0.18247008680555546,7.981220657276995,2.593286269989915,85,1065,9.099437148217635,3.0255006483215676,97,1066 +lesbos_el,230,0.11232479166666667,94.78337754199823,71.14093959731544,1072,1131,93.54553492484527,71.65720185854414,1058,1131 +monsoon_en_in,256,0.6956833333333333,4.122535440769106,1.920860545524395,253,6137,3.7803487045787842,1.8024074785503905,232,6137 +nst_da_da,256,0.3674177777777777,26.487747957992998,12.505468863934665,681,2571,11.593640946103141,4.404258422050459,299,2579 +nst_sv_sv,256,0.3503472222222222,16.565656565656564,6.865747830525779,328,1980,12.355600200904068,3.547728432873915,246,1991 +voxpopuli_cs,256,0.6932252083333335,7.318880859736891,3.92727067393207,395,5397,7.394366197183099,3.983737541997233,399,5396 +voxpopuli_es,256,0.8307443402777775,6.073136064090968,4.246347800161847,470,7739,6.203153269578703,4.3379189914391585,480,7738 +voxpopuli_hu,256,0.7771814756944443,11.995696611081227,4.097020570508285,669,5577,11.047345767575322,3.8749664478660844,616,5576 +voxpopuli_it,256,0.9244941145833333,11.368881331227985,8.714735966896765,936,8233,11.815421979356406,9.557785641085136,973,8235 +voxpopuli_nl,256,0.5597956423611112,9.500183755972069,5.671247674488387,517,5442,9.558823529411764,5.626237772309908,520,5440 +voxpopuli_pl,256,0.7600847395833334,6.481310992450746,3.4194700953627053,352,5431,6.241944393297735,3.4144377626248645,339,5431 +voxpopuli_ro,256,0.8114564409722217,11.480037639467671,4.25432643366135,854,7439,11.199247109438021,4.158890397013913,833,7438 diff --git a/evidence/domains-r3-20260908/transfer.json b/evidence/domains-r3-20260908/transfer.json new file mode 100644 index 0000000000000000000000000000000000000000..5aa1542690e0e0c07fd2a4048a379d63d4352793 --- /dev/null +++ b/evidence/domains-r3-20260908/transfer.json @@ -0,0 +1,14 @@ +{ + "status": "complete", + "archive_sha256": "d862a08e5b97e49d5a21d075a3316364cc270dcdb89ad6d4f2782a2e4ce5e980", + "files_sha256": { + "orukeet.jsonl": "525d958901b5e7b345719dc77fa8b7231dbe24a4b48dc273ea76b23df35ad7fa", + "run.log": "adb30a2b9c7510591b0134b2966bbe59c292ca0db85d911998fe8f33a393115a", + "decoding.json": "dd264a738b97cabfd1755556df9e844fa04493843738a20deecd2c9c7721fb15", + "inference-comparison.json": "3892c802fd85d6cab5bee77b69a774b0fa6308294f6a7e5e536ef89300e853ba", + "parakeet.jsonl": "7dea719a080b71b4e5bf3b731af07f36d2ba7953265aa359813407880994f90f", + "manifest.jsonl": "c16b21855716bc1f62a9c55f775b137cc720a944f3e9343fd8993788db2eaa8e", + "launch.json": "76c92e393db29d67be6a46f153cf863d5d4afd6f10f9603d3a30ad1e0c4b37d8" + }, + "publication_authorized": false +} diff --git a/evidence/r3-promotion-20260908/ci/smoke-results-ci-darwin-arm64.json b/evidence/r3-promotion-20260908/ci/smoke-results-ci-darwin-arm64.json new file mode 100644 index 0000000000000000000000000000000000000000..6feef495c5bd45689b29a4b78630659b999068ec --- /dev/null +++ b/evidence/r3-promotion-20260908/ci/smoke-results-ci-darwin-arm64.json @@ -0,0 +1,659 @@ +{ + "status": "passed", + "scope": "private compiled native sidecar prototype; not an Electron app benchmark or cross-platform package certification", + "created_at": "2026-09-08T21:34:02.903Z", + "platform": "darwin", + "arch": "arm64", + "os_release": "23.6.0", + "cpu": "Apple M1 (Virtual)", + "node": "v24.20.0", + "model_sha256": "93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd", + "model_bytes": 714456704, + "model_identity": { + "stage": "canonical" + }, + "runtime_archive_sha256": "971661d38d4bf97a63c528d13041a964316d25068d8df045e5b4839848092f25", + "runtime_archive": "nemo-speech-0.1.0-macos-aarch64-cpu.tar.gz", + "sidecar_sha256": "4f262db97807e986fcd4ac43d8658c4dfb4f9ef7d971eb7027904f8909841614", + "device": "cpu", + "checks": [ + { + "name": "released_q8_and_runtime_archive_hashes", + "passed": true + }, + { + "name": "relocated_bin_lib_startup_without_loader_environment", + "passed": true, + "ready": { + "event": "ready", + "protocol_version": 1, + "device": "cpu", + "gpu_index": -1, + "runtime_version": "nemo-speech-asr 0.1.0" + }, + "elapsed_ms": 162.11695800000007 + }, + { + "name": "unicode_model_cli_path_loaded", + "passed": true, + "filename": "模型 épreuve with spaces/Orukeet 🦜.gguf" + }, + { + "name": "two_real_model_requests_same_process_unicode_audio_path", + "passed": true, + "results": [ + { + "elapsed_ms": 11761.855334, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + }, + { + "elapsed_ms": 9964.29875, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + } + ] + }, + { + "name": "multilingual_fr_fr_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "9da159e147d71fbe799a06415fdc91e0ba9e049f3d16f94cb88e4f29f4bf5f02", + "reference": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "prior_fixture_text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "result": { + "audio_seconds": 9.42, + "text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "words": [ + { + "text": "Les", + "start": 0, + "end": 0.16 + }, + { + "text": "voyageurs", + "start": 0.16, + "end": 0.56 + }, + { + "text": "à", + "start": 0.56, + "end": 0.72 + }, + { + "text": "destination", + "start": 0.72, + "end": 1.12 + }, + { + "text": "de", + "start": 1.12, + "end": 1.28 + }, + { + "text": "pays", + "start": 1.28, + "end": 1.52 + }, + { + "text": "où", + "start": 1.52, + "end": 1.68 + }, + { + "text": "les", + "start": 1.68, + "end": 1.84 + }, + { + "text": "taxes", + "start": 1.84, + "end": 2.08 + }, + { + "text": "sont", + "start": 2.08, + "end": 2.24 + }, + { + "text": "élevées", + "start": 2.24, + "end": 2.48 + }, + { + "text": "peuvent", + "start": 2.64, + "end": 2.88 + }, + { + "text": "parfois", + "start": 2.88, + "end": 3.12 + }, + { + "text": "faire", + "start": 3.12, + "end": 3.28 + }, + { + "text": "des", + "start": 3.28, + "end": 3.44 + }, + { + "text": "économies", + "start": 3.44, + "end": 3.92 + }, + { + "text": "considérables,", + "start": 3.92, + "end": 4.64 + }, + { + "text": "en", + "start": 4.64, + "end": 4.8 + }, + { + "text": "particulier", + "start": 4.8, + "end": 5.52 + }, + { + "text": "sur", + "start": 5.68, + "end": 5.92 + }, + { + "text": "des", + "start": 5.92, + "end": 6.08 + }, + { + "text": "produits", + "start": 6.08, + "end": 6.64 + }, + { + "text": "comme", + "start": 6.64, + "end": 6.88 + }, + { + "text": "les", + "start": 6.88, + "end": 7.04 + }, + { + "text": "boissons", + "start": 7.04, + "end": 7.52 + }, + { + "text": "alcoolisées", + "start": 7.52, + "end": 8.32 + }, + { + "text": "ou", + "start": 8.32, + "end": 8.4 + }, + { + "text": "le", + "start": 8.4, + "end": 8.64 + }, + { + "text": "tabac.", + "start": 8.64, + "end": 9.2 + } + ], + "final": true + }, + "reference_words": 29, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_es_419_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "25c2820fbfc09a588480efe521c5b31a4e1046756a0434da923d0faf9ff7bf49", + "reference": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "prior_fixture_text": "Esto parece tener sentido ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "result": { + "audio_seconds": 7.68, + "text": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "words": [ + { + "text": "Esto", + "start": 0.32, + "end": 0.88 + }, + { + "text": "parece", + "start": 0.88, + "end": 1.36 + }, + { + "text": "tener", + "start": 1.36, + "end": 1.76 + }, + { + "text": "sentido,", + "start": 1.76, + "end": 2.4 + }, + { + "text": "ya", + "start": 2.4, + "end": 2.64 + }, + { + "text": "que", + "start": 2.64, + "end": 2.88 + }, + { + "text": "en", + "start": 2.88, + "end": 3.04 + }, + { + "text": "la", + "start": 3.04, + "end": 3.28 + }, + { + "text": "Tierra", + "start": 3.28, + "end": 3.6 + }, + { + "text": "no", + "start": 3.6, + "end": 3.76 + }, + { + "text": "se", + "start": 3.76, + "end": 3.92 + }, + { + "text": "percibe", + "start": 3.92, + "end": 4.64 + }, + { + "text": "su", + "start": 4.64, + "end": 4.88 + }, + { + "text": "movimiento,", + "start": 4.88, + "end": 5.52 + }, + { + "text": "¿cierto?", + "start": 5.52, + "end": 6.72 + } + ], + "final": true + }, + "reference_words": 15, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_lv_lv_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "0046336e59aca17a93029fcd1a6352751b118dab6ffb4fd1c215c7f048674e26", + "reference": "Tas mums ir devis vilcienus, automašīnas un daudzus citus transportlīdzekļus.", + "prior_fixture_text": "Tas mums ir devas vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "result": { + "audio_seconds": 7.14, + "text": "Tas mums ir devis vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "words": [ + { + "text": "Tas", + "start": 0.88, + "end": 1.12 + }, + { + "text": "mums", + "start": 1.36, + "end": 1.6 + }, + { + "text": "ir", + "start": 1.68, + "end": 1.84 + }, + { + "text": "devis", + "start": 1.84, + "end": 2.16 + }, + { + "text": "vilcienus,", + "start": 2.32, + "end": 3.28 + }, + { + "text": "automašīnas", + "start": 3.28, + "end": 4.16 + }, + { + "text": "un", + "start": 4.16, + "end": 4.32 + }, + { + "text": "daudzus", + "start": 4.32, + "end": 4.8 + }, + { + "text": "citas", + "start": 4.8, + "end": 5.04 + }, + { + "text": "transportlīdzekļus.", + "start": 5.2, + "end": 6.96 + } + ], + "final": true + }, + "reference_words": 10, + "prior_fixture_errors": 2, + "actual_errors": 1, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "empty_returns_empty_text", + "passed": true + }, + { + "name": "silence_returns_empty_text", + "passed": true + }, + { + "name": "reject_bad-length", + "passed": true, + "error": "PCM byte count must be divisible by four" + }, + { + "name": "reject_nan", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_out-of-range", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_too-long", + "passed": true, + "error": "PCM exceeds the 30 second request limit" + }, + { + "name": "malformed_duplicate_oversized_and_invalid_id_requests_recover", + "passed": true + }, + { + "name": "invalid_operation_path_and_rate_recover", + "passed": true + }, + { + "name": "real_transcription_after_all_request_errors", + "passed": true + }, + { + "name": "shutdown_exit_zero_and_json_only_stdout", + "passed": true + } + ], + "measurement": "Sequential smoke-call wall time includes file read and JSON IPC; no comparative speed claim. Startup includes loading, without weight hashing." +} diff --git a/evidence/r3-promotion-20260908/ci/smoke-results-ci-linux-x64.json b/evidence/r3-promotion-20260908/ci/smoke-results-ci-linux-x64.json new file mode 100644 index 0000000000000000000000000000000000000000..4fc7d9c13388d7d6430c413e3f73d8c2f5061b65 --- /dev/null +++ b/evidence/r3-promotion-20260908/ci/smoke-results-ci-linux-x64.json @@ -0,0 +1,659 @@ +{ + "status": "passed", + "scope": "private compiled native sidecar prototype; not an Electron app benchmark or cross-platform package certification", + "created_at": "2026-09-08T21:33:34.358Z", + "platform": "linux", + "arch": "x64", + "os_release": "6.8.0-1064-azure", + "cpu": "AMD EPYC 7763 64-Core Processor", + "node": "v24.20.0", + "model_sha256": "93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd", + "model_bytes": 714456704, + "model_identity": { + "stage": "canonical" + }, + "runtime_archive_sha256": "0f74131d631ad2c694cf0ec53490866bb6461147959589a69fb6fc231944065b", + "runtime_archive": "nemo-speech-0.1.0-linux-x86_64-cpu.tar.gz", + "sidecar_sha256": "34bd14764c7821aeef271b2d16c2e8e82b211a67c1a15b8328fb00225469f75c", + "device": "cpu", + "checks": [ + { + "name": "released_q8_and_runtime_archive_hashes", + "passed": true + }, + { + "name": "relocated_bin_lib_startup_without_loader_environment", + "passed": true, + "ready": { + "event": "ready", + "protocol_version": 1, + "device": "cpu", + "gpu_index": -1, + "runtime_version": "nemo-speech-asr 0.1.0" + }, + "elapsed_ms": 66.71482899999955 + }, + { + "name": "unicode_model_cli_path_loaded", + "passed": true, + "filename": "模型 épreuve with spaces/Orukeet 🦜.gguf" + }, + { + "name": "two_real_model_requests_same_process_unicode_audio_path", + "passed": true, + "results": [ + { + "elapsed_ms": 6568.994298, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.84 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + }, + { + "elapsed_ms": 6002.989813, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.84 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + } + ] + }, + { + "name": "multilingual_fr_fr_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "9da159e147d71fbe799a06415fdc91e0ba9e049f3d16f94cb88e4f29f4bf5f02", + "reference": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "prior_fixture_text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "result": { + "audio_seconds": 9.42, + "text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "words": [ + { + "text": "Les", + "start": 0, + "end": 0.16 + }, + { + "text": "voyageurs", + "start": 0.16, + "end": 0.56 + }, + { + "text": "à", + "start": 0.56, + "end": 0.72 + }, + { + "text": "destination", + "start": 0.72, + "end": 1.12 + }, + { + "text": "de", + "start": 1.12, + "end": 1.28 + }, + { + "text": "pays", + "start": 1.28, + "end": 1.52 + }, + { + "text": "où", + "start": 1.52, + "end": 1.68 + }, + { + "text": "les", + "start": 1.68, + "end": 1.84 + }, + { + "text": "taxes", + "start": 1.84, + "end": 2.08 + }, + { + "text": "sont", + "start": 2.08, + "end": 2.24 + }, + { + "text": "élevées", + "start": 2.24, + "end": 2.48 + }, + { + "text": "peuvent", + "start": 2.64, + "end": 2.88 + }, + { + "text": "parfois", + "start": 2.88, + "end": 3.12 + }, + { + "text": "faire", + "start": 3.12, + "end": 3.28 + }, + { + "text": "des", + "start": 3.28, + "end": 3.44 + }, + { + "text": "économies", + "start": 3.44, + "end": 3.92 + }, + { + "text": "considérables,", + "start": 3.92, + "end": 4.64 + }, + { + "text": "en", + "start": 4.64, + "end": 4.8 + }, + { + "text": "particulier", + "start": 4.8, + "end": 5.52 + }, + { + "text": "sur", + "start": 5.68, + "end": 5.92 + }, + { + "text": "des", + "start": 5.92, + "end": 6.08 + }, + { + "text": "produits", + "start": 6.08, + "end": 6.64 + }, + { + "text": "comme", + "start": 6.64, + "end": 6.88 + }, + { + "text": "les", + "start": 6.88, + "end": 7.04 + }, + { + "text": "boissons", + "start": 7.04, + "end": 7.52 + }, + { + "text": "alcoolisées", + "start": 7.52, + "end": 8.32 + }, + { + "text": "ou", + "start": 8.32, + "end": 8.4 + }, + { + "text": "le", + "start": 8.4, + "end": 8.64 + }, + { + "text": "tabac.", + "start": 8.64, + "end": 9.2 + } + ], + "final": true + }, + "reference_words": 29, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_es_419_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "25c2820fbfc09a588480efe521c5b31a4e1046756a0434da923d0faf9ff7bf49", + "reference": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "prior_fixture_text": "Esto parece tener sentido ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "result": { + "audio_seconds": 7.68, + "text": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "words": [ + { + "text": "Esto", + "start": 0.32, + "end": 0.88 + }, + { + "text": "parece", + "start": 0.88, + "end": 1.36 + }, + { + "text": "tener", + "start": 1.36, + "end": 1.76 + }, + { + "text": "sentido,", + "start": 1.76, + "end": 2.4 + }, + { + "text": "ya", + "start": 2.4, + "end": 2.64 + }, + { + "text": "que", + "start": 2.64, + "end": 2.88 + }, + { + "text": "en", + "start": 2.88, + "end": 3.04 + }, + { + "text": "la", + "start": 3.04, + "end": 3.28 + }, + { + "text": "Tierra", + "start": 3.28, + "end": 3.6 + }, + { + "text": "no", + "start": 3.6, + "end": 3.76 + }, + { + "text": "se", + "start": 3.76, + "end": 3.92 + }, + { + "text": "percibe", + "start": 3.92, + "end": 4.64 + }, + { + "text": "su", + "start": 4.64, + "end": 4.88 + }, + { + "text": "movimiento,", + "start": 4.88, + "end": 5.52 + }, + { + "text": "¿cierto?", + "start": 5.52, + "end": 6.72 + } + ], + "final": true + }, + "reference_words": 15, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_lv_lv_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "0046336e59aca17a93029fcd1a6352751b118dab6ffb4fd1c215c7f048674e26", + "reference": "Tas mums ir devis vilcienus, automašīnas un daudzus citus transportlīdzekļus.", + "prior_fixture_text": "Tas mums ir devas vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "result": { + "audio_seconds": 7.14, + "text": "Tas mums ir devis vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "words": [ + { + "text": "Tas", + "start": 0.88, + "end": 1.12 + }, + { + "text": "mums", + "start": 1.36, + "end": 1.6 + }, + { + "text": "ir", + "start": 1.68, + "end": 1.84 + }, + { + "text": "devis", + "start": 1.84, + "end": 2.16 + }, + { + "text": "vilcienus,", + "start": 2.32, + "end": 3.28 + }, + { + "text": "automašīnas", + "start": 3.28, + "end": 4.16 + }, + { + "text": "un", + "start": 4.16, + "end": 4.32 + }, + { + "text": "daudzus", + "start": 4.32, + "end": 4.8 + }, + { + "text": "citas", + "start": 4.8, + "end": 5.04 + }, + { + "text": "transportlīdzekļus.", + "start": 5.2, + "end": 6.96 + } + ], + "final": true + }, + "reference_words": 10, + "prior_fixture_errors": 2, + "actual_errors": 1, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "empty_returns_empty_text", + "passed": true + }, + { + "name": "silence_returns_empty_text", + "passed": true + }, + { + "name": "reject_bad-length", + "passed": true, + "error": "PCM byte count must be divisible by four" + }, + { + "name": "reject_nan", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_out-of-range", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_too-long", + "passed": true, + "error": "PCM exceeds the 30 second request limit" + }, + { + "name": "malformed_duplicate_oversized_and_invalid_id_requests_recover", + "passed": true + }, + { + "name": "invalid_operation_path_and_rate_recover", + "passed": true + }, + { + "name": "real_transcription_after_all_request_errors", + "passed": true + }, + { + "name": "shutdown_exit_zero_and_json_only_stdout", + "passed": true + } + ], + "measurement": "Sequential smoke-call wall time includes file read and JSON IPC; no comparative speed claim. Startup includes loading, without weight hashing." +} diff --git a/evidence/r3-promotion-20260908/ci/smoke-results-ci-win32-x64.json b/evidence/r3-promotion-20260908/ci/smoke-results-ci-win32-x64.json new file mode 100644 index 0000000000000000000000000000000000000000..86ca1c9ae8df374dbccc34022fddfb814ed8396c --- /dev/null +++ b/evidence/r3-promotion-20260908/ci/smoke-results-ci-win32-x64.json @@ -0,0 +1,659 @@ +{ + "status": "passed", + "scope": "private compiled native sidecar prototype; not an Electron app benchmark or cross-platform package certification", + "created_at": "2026-09-08T21:33:33.747Z", + "platform": "win32", + "arch": "x64", + "os_release": "10.0.20348", + "cpu": "AMD EPYC 7763 64-Core Processor ", + "node": "v24.20.0", + "model_sha256": "93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd", + "model_bytes": 714456704, + "model_identity": { + "stage": "canonical" + }, + "runtime_archive_sha256": "5e4ea81046012edcd77fd8848de8eefb5a4ba38cc26f52eb544ab184695a75d6", + "runtime_archive": "nemo-speech-0.1.0-windows-x86_64-cpu.zip", + "sidecar_sha256": "9795ed7c9243ae4c21b6321e4316fe5895ff8e9b41a77325ecb4e3564f5815a9", + "device": "cpu", + "checks": [ + { + "name": "released_q8_and_runtime_archive_hashes", + "passed": true + }, + { + "name": "relocated_bin_lib_startup_without_loader_environment", + "passed": true, + "ready": { + "event": "ready", + "protocol_version": 1, + "device": "cpu", + "gpu_index": -1, + "runtime_version": "nemo-speech-asr 0.1.0" + }, + "elapsed_ms": 51.92839999999978 + }, + { + "name": "unicode_model_cli_path_loaded", + "passed": true, + "filename": "模型 épreuve with spaces/Orukeet 🦜.gguf" + }, + { + "name": "two_real_model_requests_same_process_unicode_audio_path", + "passed": true, + "results": [ + { + "elapsed_ms": 4866.3694, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + }, + { + "elapsed_ms": 4605.159999999999, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + } + ] + }, + { + "name": "multilingual_fr_fr_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "9da159e147d71fbe799a06415fdc91e0ba9e049f3d16f94cb88e4f29f4bf5f02", + "reference": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "prior_fixture_text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "result": { + "audio_seconds": 9.42, + "text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "words": [ + { + "text": "Les", + "start": 0, + "end": 0.16 + }, + { + "text": "voyageurs", + "start": 0.16, + "end": 0.56 + }, + { + "text": "à", + "start": 0.56, + "end": 0.72 + }, + { + "text": "destination", + "start": 0.72, + "end": 1.12 + }, + { + "text": "de", + "start": 1.12, + "end": 1.28 + }, + { + "text": "pays", + "start": 1.28, + "end": 1.52 + }, + { + "text": "où", + "start": 1.52, + "end": 1.68 + }, + { + "text": "les", + "start": 1.68, + "end": 1.84 + }, + { + "text": "taxes", + "start": 1.84, + "end": 2.08 + }, + { + "text": "sont", + "start": 2.08, + "end": 2.24 + }, + { + "text": "élevées", + "start": 2.24, + "end": 2.48 + }, + { + "text": "peuvent", + "start": 2.64, + "end": 2.88 + }, + { + "text": "parfois", + "start": 2.88, + "end": 3.12 + }, + { + "text": "faire", + "start": 3.12, + "end": 3.28 + }, + { + "text": "des", + "start": 3.28, + "end": 3.44 + }, + { + "text": "économies", + "start": 3.44, + "end": 3.92 + }, + { + "text": "considérables,", + "start": 3.92, + "end": 4.64 + }, + { + "text": "en", + "start": 4.64, + "end": 4.8 + }, + { + "text": "particulier", + "start": 4.8, + "end": 5.52 + }, + { + "text": "sur", + "start": 5.68, + "end": 5.92 + }, + { + "text": "des", + "start": 5.92, + "end": 6.08 + }, + { + "text": "produits", + "start": 6.08, + "end": 6.64 + }, + { + "text": "comme", + "start": 6.64, + "end": 6.88 + }, + { + "text": "les", + "start": 6.88, + "end": 7.12 + }, + { + "text": "boissons", + "start": 7.12, + "end": 7.52 + }, + { + "text": "alcoolisées", + "start": 7.52, + "end": 8.32 + }, + { + "text": "ou", + "start": 8.32, + "end": 8.4 + }, + { + "text": "le", + "start": 8.4, + "end": 8.64 + }, + { + "text": "tabac.", + "start": 8.64, + "end": 9.2 + } + ], + "final": true + }, + "reference_words": 29, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_es_419_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "25c2820fbfc09a588480efe521c5b31a4e1046756a0434da923d0faf9ff7bf49", + "reference": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "prior_fixture_text": "Esto parece tener sentido ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "result": { + "audio_seconds": 7.68, + "text": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "words": [ + { + "text": "Esto", + "start": 0.32, + "end": 0.88 + }, + { + "text": "parece", + "start": 0.88, + "end": 1.36 + }, + { + "text": "tener", + "start": 1.36, + "end": 1.76 + }, + { + "text": "sentido,", + "start": 1.76, + "end": 2.4 + }, + { + "text": "ya", + "start": 2.4, + "end": 2.64 + }, + { + "text": "que", + "start": 2.64, + "end": 2.88 + }, + { + "text": "en", + "start": 2.88, + "end": 3.04 + }, + { + "text": "la", + "start": 3.04, + "end": 3.28 + }, + { + "text": "Tierra", + "start": 3.28, + "end": 3.6 + }, + { + "text": "no", + "start": 3.6, + "end": 3.76 + }, + { + "text": "se", + "start": 3.76, + "end": 3.92 + }, + { + "text": "percibe", + "start": 3.92, + "end": 4.48 + }, + { + "text": "su", + "start": 4.64, + "end": 4.88 + }, + { + "text": "movimiento,", + "start": 4.88, + "end": 5.52 + }, + { + "text": "¿cierto?", + "start": 5.52, + "end": 6.72 + } + ], + "final": true + }, + "reference_words": 15, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_lv_lv_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "0046336e59aca17a93029fcd1a6352751b118dab6ffb4fd1c215c7f048674e26", + "reference": "Tas mums ir devis vilcienus, automašīnas un daudzus citus transportlīdzekļus.", + "prior_fixture_text": "Tas mums ir devas vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "result": { + "audio_seconds": 7.14, + "text": "Tas mums ir devis vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "words": [ + { + "text": "Tas", + "start": 0.88, + "end": 1.12 + }, + { + "text": "mums", + "start": 1.36, + "end": 1.68 + }, + { + "text": "ir", + "start": 1.68, + "end": 1.84 + }, + { + "text": "devis", + "start": 1.84, + "end": 2.16 + }, + { + "text": "vilcienus,", + "start": 2.32, + "end": 3.28 + }, + { + "text": "automašīnas", + "start": 3.28, + "end": 4.16 + }, + { + "text": "un", + "start": 4.16, + "end": 4.32 + }, + { + "text": "daudzus", + "start": 4.32, + "end": 4.8 + }, + { + "text": "citas", + "start": 4.8, + "end": 5.04 + }, + { + "text": "transportlīdzekļus.", + "start": 5.2, + "end": 6.96 + } + ], + "final": true + }, + "reference_words": 10, + "prior_fixture_errors": 2, + "actual_errors": 1, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "empty_returns_empty_text", + "passed": true + }, + { + "name": "silence_returns_empty_text", + "passed": true + }, + { + "name": "reject_bad-length", + "passed": true, + "error": "PCM byte count must be divisible by four" + }, + { + "name": "reject_nan", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_out-of-range", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_too-long", + "passed": true, + "error": "PCM exceeds the 30 second request limit" + }, + { + "name": "malformed_duplicate_oversized_and_invalid_id_requests_recover", + "passed": true + }, + { + "name": "invalid_operation_path_and_rate_recover", + "passed": true + }, + { + "name": "real_transcription_after_all_request_errors", + "passed": true + }, + { + "name": "shutdown_exit_zero_and_json_only_stdout", + "passed": true + } + ], + "measurement": "Sequential smoke-call wall time includes file read and JSON IPC; no comparative speed claim. Startup includes loading, without weight hashing." +} diff --git a/evidence/r3-promotion-20260908/export-environment.json b/evidence/r3-promotion-20260908/export-environment.json new file mode 100644 index 0000000000000000000000000000000000000000..e949894211948c7d87e76db02897a0011d3b6a67 --- /dev/null +++ b/evidence/r3-promotion-20260908/export-environment.json @@ -0,0 +1,20 @@ +{ + "status": "recorded", + "platform": "macOS-26.4.1-arm64-arm-64bit", + "python": "3.12.13", + "packages": { + "torch": "2.6.0", + "numpy": "1.26.4", + "librosa": "0.11.0", + "gguf": "0.19.0" + }, + "converter_repository": "https://github.com/NVIDIA/NeMo-Speech.cpp", + "converter_commit": "4f9676226f667d14608487df744f375db87127f8", + "exporter_sha256": "583e76d9cc3928e574502596b75413fbd58e91a04ebfe91b01d4da79950c179c", + "source_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "parent_audit_chain": [ + "evidence/regression-ft-20260907/export-audit.json", + "training/gabor_half/results/r15-0100-audit.json" + ], + "verification": "Local Q8 and F16 match the separately recreated server exports byte for byte. Both formats preserve all 12,288 selected Gabor functions after native rounding." +} diff --git a/evidence/r3-promotion-20260908/f16-cpu.json b/evidence/r3-promotion-20260908/f16-cpu.json new file mode 100644 index 0000000000000000000000000000000000000000..1cc5abd31a5096988d82348e519c37ad2e1f9fb1 --- /dev/null +++ b/evidence/r3-promotion-20260908/f16-cpu.json @@ -0,0 +1,662 @@ +{ + "status": "passed", + "scope": "private compiled native sidecar prototype; not an Electron app benchmark or cross-platform package certification", + "created_at": "2026-09-08T21:13:00.886Z", + "platform": "darwin", + "arch": "arm64", + "os_release": "25.4.0", + "cpu": "Apple M5 Max", + "node": "v25.9.0", + "model_sha256": "de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194", + "model_bytes": 1296681088, + "model_identity": { + "stage": "audited-native-candidate", + "source_model_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "lineage_sha256": "e47ada9ed05ce74f13964daca3c1393e4c46025e550a500500323c6cd2c259e7", + "source_audit_sha256": "3d9f115d39758976836291bf2a99f96b0ea71c412f320e3d954b6ba66cfdf59d" + }, + "runtime_archive_sha256": "971661d38d4bf97a63c528d13041a964316d25068d8df045e5b4839848092f25", + "runtime_archive": "nemo-speech-0.1.0-macos-aarch64-cpu.tar.gz", + "sidecar_sha256": "590a287a06dde57594b5f6a4d7863e1b57c0504975ff3a4c21652aceaa27b254", + "device": "cpu", + "checks": [ + { + "name": "candidate_native_lineage_and_runtime_archive_hashes", + "passed": true + }, + { + "name": "relocated_bin_lib_startup_without_loader_environment", + "passed": true, + "ready": { + "event": "ready", + "protocol_version": 1, + "device": "cpu", + "gpu_index": -1, + "runtime_version": "nemo-speech-asr 0.1.0" + }, + "elapsed_ms": 1350.1541249999996 + }, + { + "name": "unicode_model_cli_path_loaded", + "passed": true, + "filename": "模型 épreuve with spaces/Orukeet 🦜.gguf" + }, + { + "name": "two_real_model_requests_same_process_unicode_audio_path", + "passed": true, + "results": [ + { + "elapsed_ms": 766.5465000000004, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + }, + { + "elapsed_ms": 597.5672079999995, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + } + ] + }, + { + "name": "multilingual_fr_fr_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "9da159e147d71fbe799a06415fdc91e0ba9e049f3d16f94cb88e4f29f4bf5f02", + "reference": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "prior_fixture_text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "result": { + "audio_seconds": 9.42, + "text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "words": [ + { + "text": "Les", + "start": 0, + "end": 0.16 + }, + { + "text": "voyageurs", + "start": 0.16, + "end": 0.56 + }, + { + "text": "à", + "start": 0.56, + "end": 0.72 + }, + { + "text": "destination", + "start": 0.72, + "end": 1.12 + }, + { + "text": "de", + "start": 1.12, + "end": 1.28 + }, + { + "text": "pays", + "start": 1.28, + "end": 1.52 + }, + { + "text": "où", + "start": 1.52, + "end": 1.68 + }, + { + "text": "les", + "start": 1.68, + "end": 1.84 + }, + { + "text": "taxes", + "start": 1.84, + "end": 2.08 + }, + { + "text": "sont", + "start": 2.08, + "end": 2.24 + }, + { + "text": "élevées", + "start": 2.24, + "end": 2.48 + }, + { + "text": "peuvent", + "start": 2.64, + "end": 2.88 + }, + { + "text": "parfois", + "start": 2.88, + "end": 3.12 + }, + { + "text": "faire", + "start": 3.12, + "end": 3.28 + }, + { + "text": "des", + "start": 3.28, + "end": 3.44 + }, + { + "text": "économies", + "start": 3.44, + "end": 3.92 + }, + { + "text": "considérables,", + "start": 3.92, + "end": 4.64 + }, + { + "text": "en", + "start": 4.64, + "end": 4.8 + }, + { + "text": "particulier", + "start": 4.8, + "end": 5.52 + }, + { + "text": "sur", + "start": 5.68, + "end": 5.92 + }, + { + "text": "des", + "start": 5.92, + "end": 6.08 + }, + { + "text": "produits", + "start": 6.08, + "end": 6.64 + }, + { + "text": "comme", + "start": 6.64, + "end": 6.88 + }, + { + "text": "les", + "start": 6.88, + "end": 7.04 + }, + { + "text": "boissons", + "start": 7.04, + "end": 7.52 + }, + { + "text": "alcoolisées", + "start": 7.52, + "end": 8.32 + }, + { + "text": "ou", + "start": 8.32, + "end": 8.4 + }, + { + "text": "le", + "start": 8.4, + "end": 8.64 + }, + { + "text": "tabac.", + "start": 8.64, + "end": 9.2 + } + ], + "final": true + }, + "reference_words": 29, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_es_419_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "25c2820fbfc09a588480efe521c5b31a4e1046756a0434da923d0faf9ff7bf49", + "reference": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "prior_fixture_text": "Esto parece tener sentido ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "result": { + "audio_seconds": 7.68, + "text": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "words": [ + { + "text": "Esto", + "start": 0.32, + "end": 0.88 + }, + { + "text": "parece", + "start": 0.88, + "end": 1.36 + }, + { + "text": "tener", + "start": 1.36, + "end": 1.76 + }, + { + "text": "sentido,", + "start": 1.76, + "end": 2.4 + }, + { + "text": "ya", + "start": 2.4, + "end": 2.64 + }, + { + "text": "que", + "start": 2.64, + "end": 2.88 + }, + { + "text": "en", + "start": 2.88, + "end": 3.04 + }, + { + "text": "la", + "start": 3.04, + "end": 3.28 + }, + { + "text": "Tierra", + "start": 3.28, + "end": 3.6 + }, + { + "text": "no", + "start": 3.6, + "end": 3.76 + }, + { + "text": "se", + "start": 3.76, + "end": 3.92 + }, + { + "text": "percibe", + "start": 3.92, + "end": 4.64 + }, + { + "text": "su", + "start": 4.64, + "end": 4.88 + }, + { + "text": "movimiento,", + "start": 4.88, + "end": 5.52 + }, + { + "text": "¿cierto?", + "start": 5.52, + "end": 6.72 + } + ], + "final": true + }, + "reference_words": 15, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_lv_lv_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "0046336e59aca17a93029fcd1a6352751b118dab6ffb4fd1c215c7f048674e26", + "reference": "Tas mums ir devis vilcienus, automašīnas un daudzus citus transportlīdzekļus.", + "prior_fixture_text": "Tas mums ir devas vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "result": { + "audio_seconds": 7.14, + "text": "Tas mums ir devis vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "words": [ + { + "text": "Tas", + "start": 0.88, + "end": 1.12 + }, + { + "text": "mums", + "start": 1.36, + "end": 1.6 + }, + { + "text": "ir", + "start": 1.68, + "end": 1.84 + }, + { + "text": "devis", + "start": 1.84, + "end": 2.16 + }, + { + "text": "vilcienus,", + "start": 2.32, + "end": 3.28 + }, + { + "text": "automašīnas", + "start": 3.28, + "end": 4.16 + }, + { + "text": "un", + "start": 4.16, + "end": 4.32 + }, + { + "text": "daudzus", + "start": 4.32, + "end": 4.8 + }, + { + "text": "citas", + "start": 4.8, + "end": 5.04 + }, + { + "text": "transportlīdzekļus.", + "start": 5.2, + "end": 6.96 + } + ], + "final": true + }, + "reference_words": 10, + "prior_fixture_errors": 2, + "actual_errors": 1, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "empty_returns_empty_text", + "passed": true + }, + { + "name": "silence_returns_empty_text", + "passed": true + }, + { + "name": "reject_bad-length", + "passed": true, + "error": "PCM byte count must be divisible by four" + }, + { + "name": "reject_nan", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_out-of-range", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_too-long", + "passed": true, + "error": "PCM exceeds the 30 second request limit" + }, + { + "name": "malformed_duplicate_oversized_and_invalid_id_requests_recover", + "passed": true + }, + { + "name": "invalid_operation_path_and_rate_recover", + "passed": true + }, + { + "name": "real_transcription_after_all_request_errors", + "passed": true + }, + { + "name": "shutdown_exit_zero_and_json_only_stdout", + "passed": true + } + ], + "measurement": "Sequential smoke-call wall time includes file read and JSON IPC; no comparative speed claim. Startup includes loading, without weight hashing." +} diff --git a/evidence/r3-promotion-20260908/f16-lineage.json b/evidence/r3-promotion-20260908/f16-lineage.json new file mode 100644 index 0000000000000000000000000000000000000000..100f34b2671f90d333ef260e5e559f98b312f221 --- /dev/null +++ b/evidence/r3-promotion-20260908/f16-lineage.json @@ -0,0 +1,19 @@ +{ + "candidate_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "gguf_sha256": "de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194", + "gguf_bytes": 1296681088, + "converter_commit": "4f9676226f667d14608487df744f375db87127f8", + "librosa_version": "0.11.0", + "mel_filterbank_sha256": "bce5ec5f194a5913f6508cee5a85512e7bad2352db8fc28f5c6ff75af8b09137", + "fits_sha256": "44ef0eb45fdd122a3900c97f1faf4ffedf567f27e9ad535f24687c36a8c7f704", + "fitted_rows_exact_after_f16_rounding": 12288, + "depthwise_storage_dtype": "float16", + "export_precision": "f16", + "source_audit_sha256": "3d9f115d39758976836291bf2a99f96b0ea71c412f320e3d954b6ba66cfdf59d", + "scope": "Exact-source native export; format-specific ASR measurements are recorded separately", + "parent_audit_sha256": "dac88b3fa2d17bd8b8e32ac0e9eaf4d24473b9ac2715f8d1601df3ff03cc751d", + "ancestor_audits_sha256": [ + "dac88b3fa2d17bd8b8e32ac0e9eaf4d24473b9ac2715f8d1601df3ff03cc751d", + "bb9e8e5bc5408f1510258ad4b7c06f6050fa619da82bf95ffb5fad6f19423e64" + ] +} diff --git a/evidence/r3-promotion-20260908/f16-metal.json b/evidence/r3-promotion-20260908/f16-metal.json new file mode 100644 index 0000000000000000000000000000000000000000..42834d85a5183edf2c342b92b1a6d36202a9aa8b --- /dev/null +++ b/evidence/r3-promotion-20260908/f16-metal.json @@ -0,0 +1,666 @@ +{ + "status": "passed", + "scope": "private compiled native sidecar prototype; not an Electron app benchmark or cross-platform package certification", + "created_at": "2026-09-08T21:12:48.934Z", + "platform": "darwin", + "arch": "arm64", + "os_release": "25.4.0", + "cpu": "Apple M5 Max", + "node": "v25.9.0", + "model_sha256": "de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194", + "model_bytes": 1296681088, + "model_identity": { + "stage": "audited-native-candidate", + "source_model_sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "lineage_sha256": "e47ada9ed05ce74f13964daca3c1393e4c46025e550a500500323c6cd2c259e7", + "source_audit_sha256": "3d9f115d39758976836291bf2a99f96b0ea71c412f320e3d954b6ba66cfdf59d" + }, + "runtime_archive_sha256": "f1dff4f9dd9c96214f8cb78b982812459132df8a4ad1a42409fd94de4a366244", + "runtime_archive": "nemo-speech-0.1.0-macos-aarch64-metal.tar.gz", + "sidecar_sha256": "12e43c158d2905f33d8d8d3c0dd615fe7539c44212f76f9ad3daf58035af0ac3", + "device": "metal", + "checks": [ + { + "name": "candidate_native_lineage_and_runtime_archive_hashes", + "passed": true + }, + { + "name": "relocated_bin_lib_startup_without_loader_environment", + "passed": true, + "ready": { + "event": "ready", + "protocol_version": 1, + "device": "metal", + "gpu_index": 0, + "runtime_version": "nemo-speech-asr 0.1.0" + }, + "elapsed_ms": 2484.4894170000007 + }, + { + "name": "unicode_model_cli_path_loaded", + "passed": true, + "filename": "模型 épreuve with spaces/Orukeet 🦜.gguf" + }, + { + "name": "two_real_model_requests_same_process_unicode_audio_path", + "passed": true, + "results": [ + { + "elapsed_ms": 721.1802499999994, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + }, + { + "elapsed_ms": 114.50400000000081, + "audio_seconds": 11, + "text": "And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", + "words": [ + { + "text": "And", + "start": 0.24, + "end": 0.56 + }, + { + "text": "so,", + "start": 0.56, + "end": 1.04 + }, + { + "text": "my", + "start": 1.04, + "end": 1.28 + }, + { + "text": "fellow", + "start": 1.28, + "end": 1.76 + }, + { + "text": "Americans,", + "start": 1.76, + "end": 3.28 + }, + { + "text": "ask", + "start": 3.28, + "end": 3.92 + }, + { + "text": "not", + "start": 4.24, + "end": 4.56 + }, + { + "text": "what", + "start": 5.2, + "end": 5.52 + }, + { + "text": "your", + "start": 5.52, + "end": 5.68 + }, + { + "text": "country", + "start": 6, + "end": 6.56 + }, + { + "text": "can", + "start": 6.56, + "end": 6.8 + }, + { + "text": "do", + "start": 6.8, + "end": 7.12 + }, + { + "text": "for", + "start": 7.12, + "end": 7.44 + }, + { + "text": "you,", + "start": 7.44, + "end": 8.16 + }, + { + "text": "ask", + "start": 8.16, + "end": 8.72 + }, + { + "text": "what", + "start": 8.72, + "end": 8.96 + }, + { + "text": "you", + "start": 8.96, + "end": 9.2 + }, + { + "text": "can", + "start": 9.2, + "end": 9.52 + }, + { + "text": "do", + "start": 9.52, + "end": 9.68 + }, + { + "text": "for", + "start": 9.68, + "end": 9.92 + }, + { + "text": "your", + "start": 9.92, + "end": 10.08 + }, + { + "text": "country.", + "start": 10.08, + "end": 10.8 + } + ], + "final": true + } + ] + }, + { + "name": "multilingual_fr_fr_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "9da159e147d71fbe799a06415fdc91e0ba9e049f3d16f94cb88e4f29f4bf5f02", + "reference": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "prior_fixture_text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "result": { + "audio_seconds": 9.42, + "text": "Les voyageurs à destination de pays où les taxes sont élevées peuvent parfois faire des économies considérables, en particulier sur des produits comme les boissons alcoolisées ou le tabac.", + "words": [ + { + "text": "Les", + "start": 0, + "end": 0.16 + }, + { + "text": "voyageurs", + "start": 0.16, + "end": 0.56 + }, + { + "text": "à", + "start": 0.56, + "end": 0.72 + }, + { + "text": "destination", + "start": 0.72, + "end": 1.12 + }, + { + "text": "de", + "start": 1.12, + "end": 1.28 + }, + { + "text": "pays", + "start": 1.28, + "end": 1.52 + }, + { + "text": "où", + "start": 1.52, + "end": 1.68 + }, + { + "text": "les", + "start": 1.68, + "end": 1.84 + }, + { + "text": "taxes", + "start": 1.84, + "end": 2.08 + }, + { + "text": "sont", + "start": 2.08, + "end": 2.24 + }, + { + "text": "élevées", + "start": 2.24, + "end": 2.48 + }, + { + "text": "peuvent", + "start": 2.64, + "end": 2.88 + }, + { + "text": "parfois", + "start": 2.88, + "end": 3.12 + }, + { + "text": "faire", + "start": 3.12, + "end": 3.28 + }, + { + "text": "des", + "start": 3.28, + "end": 3.44 + }, + { + "text": "économies", + "start": 3.44, + "end": 3.92 + }, + { + "text": "considérables,", + "start": 3.92, + "end": 4.64 + }, + { + "text": "en", + "start": 4.64, + "end": 4.8 + }, + { + "text": "particulier", + "start": 4.8, + "end": 5.52 + }, + { + "text": "sur", + "start": 5.68, + "end": 5.92 + }, + { + "text": "des", + "start": 5.92, + "end": 6.08 + }, + { + "text": "produits", + "start": 6.08, + "end": 6.64 + }, + { + "text": "comme", + "start": 6.64, + "end": 6.88 + }, + { + "text": "les", + "start": 6.88, + "end": 7.04 + }, + { + "text": "boissons", + "start": 7.04, + "end": 7.52 + }, + { + "text": "alcoolisées", + "start": 7.52, + "end": 8.32 + }, + { + "text": "ou", + "start": 8.32, + "end": 8.4 + }, + { + "text": "le", + "start": 8.4, + "end": 8.64 + }, + { + "text": "tabac.", + "start": 8.64, + "end": 9.2 + } + ], + "final": true + }, + "reference_words": 29, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_es_419_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "25c2820fbfc09a588480efe521c5b31a4e1046756a0434da923d0faf9ff7bf49", + "reference": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "prior_fixture_text": "Esto parece tener sentido ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "result": { + "audio_seconds": 7.68, + "text": "Esto parece tener sentido, ya que en la Tierra no se percibe su movimiento, ¿cierto?", + "words": [ + { + "text": "Esto", + "start": 0.32, + "end": 0.88 + }, + { + "text": "parece", + "start": 0.88, + "end": 1.36 + }, + { + "text": "tener", + "start": 1.36, + "end": 1.76 + }, + { + "text": "sentido,", + "start": 1.76, + "end": 2.4 + }, + { + "text": "ya", + "start": 2.4, + "end": 2.64 + }, + { + "text": "que", + "start": 2.64, + "end": 2.88 + }, + { + "text": "en", + "start": 2.88, + "end": 3.04 + }, + { + "text": "la", + "start": 3.04, + "end": 3.28 + }, + { + "text": "Tierra", + "start": 3.28, + "end": 3.6 + }, + { + "text": "no", + "start": 3.6, + "end": 3.76 + }, + { + "text": "se", + "start": 3.76, + "end": 3.92 + }, + { + "text": "percibe", + "start": 3.92, + "end": 4.64 + }, + { + "text": "su", + "start": 4.64, + "end": 4.88 + }, + { + "text": "movimiento,", + "start": 4.88, + "end": 5.44 + }, + { + "text": "¿cierto?", + "start": 5.52, + "end": 6.72 + } + ], + "final": true + }, + "reference_words": 15, + "prior_fixture_errors": 0, + "actual_errors": 0, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "multilingual_lv_lv_repeats_without_added_word_errors", + "passed": true, + "fixture_sha256": "0046336e59aca17a93029fcd1a6352751b118dab6ffb4fd1c215c7f048674e26", + "reference": "Tas mums ir devis vilcienus, automašīnas un daudzus citus transportlīdzekļus.", + "prior_fixture_text": "Tas mums ir devas vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "result": { + "audio_seconds": 7.14, + "text": "Tas mums ir devis vilcienus, automašīnas un daudzus citas transportlīdzekļus.", + "words": [ + { + "text": "Tas", + "start": 0.88, + "end": 1.12 + }, + { + "text": "mums", + "start": 1.36, + "end": 1.6 + }, + { + "text": "ir", + "start": 1.68, + "end": 1.84 + }, + { + "text": "devis", + "start": 1.84, + "end": 2.16 + }, + { + "text": "vilcienus,", + "start": 2.32, + "end": 3.28 + }, + { + "text": "automašīnas", + "start": 3.28, + "end": 4.16 + }, + { + "text": "un", + "start": 4.16, + "end": 4.32 + }, + { + "text": "daudzus", + "start": 4.32, + "end": 4.8 + }, + { + "text": "citas", + "start": 4.8, + "end": 5.04 + }, + { + "text": "transportlīdzekļus.", + "start": 5.2, + "end": 6.96 + } + ], + "final": true + }, + "reference_words": 10, + "prior_fixture_errors": 2, + "actual_errors": 1, + "reference_model_sha256": "8967e04bd73fd8e88bccbfe97d0ec952beac73b1b3b3cba4b5c03316fcd8baa0", + "scope": "Fixed qualitative fixture regression against the recorded Metal output; not a language benchmark.", + "normalization": "NFC, lowercase, Unicode punctuation/symbols replaced by spaces, whitespace tokenization" + }, + { + "name": "empty_returns_empty_text", + "passed": true + }, + { + "name": "silence_returns_empty_text", + "passed": true + }, + { + "name": "reject_bad-length", + "passed": true, + "error": "PCM byte count must be divisible by four" + }, + { + "name": "reject_nan", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_out-of-range", + "passed": true, + "error": "PCM samples must be finite and between -1 and 1" + }, + { + "name": "reject_too-long", + "passed": true, + "error": "PCM exceeds the 30 second request limit" + }, + { + "name": "malformed_duplicate_oversized_and_invalid_id_requests_recover", + "passed": true + }, + { + "name": "invalid_operation_path_and_rate_recover", + "passed": true + }, + { + "name": "real_transcription_after_all_request_errors", + "passed": true + }, + { + "name": "shutdown_exit_zero_and_json_only_stdout", + "passed": true + }, + { + "name": "native_log_confirms_metal_gpu_backend", + "passed": true + } + ], + "measurement": "Sequential smoke-call wall time includes file read and JSON IPC; no comparative speed claim. Startup includes loading, without weight hashing." +} diff --git a/evidence/r3-promotion-20260908/linux-f16-native-fixture.py b/evidence/r3-promotion-20260908/linux-f16-native-fixture.py new file mode 100644 index 0000000000000000000000000000000000000000..2df252dab2e3cc1b66f6350663cd2580edef1543 --- /dev/null +++ b/evidence/r3-promotion-20260908/linux-f16-native-fixture.py @@ -0,0 +1,27 @@ +import array,hashlib,json,sys,time +from pathlib import Path +import wave +import numpy as np +from orukeet.nvidia import NvidiaRecognizer +import orukeet.nvidia +root=Path('/home/nathanroll/parakeet-ft');dest=Path('/dev/shm/orukeet-r3-native-20260908') +model=dest/'orukeet-v0.1.0rc1-f16.gguf';runtime=root/'inference_20260905/cuda';fixture=root/'gabor_half_20260906/fixtures/jfk.wav' +def sha(p): + with p.open('rb') as f:return hashlib.file_digest(f,'sha256').hexdigest() +assert sha(model)=='de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194' +with wave.open(str(fixture), 'rb') as audio: + assert audio.getframerate()==16000 and audio.getnchannels()==1 and audio.getsampwidth()==2 + samples=np.frombuffer(audio.readframes(audio.getnframes()),dtype=' +

Oruk AI

+ + +# Orukeet technical report + + +

+Nathan Roll1,2 · Irene Yi1,2 · Büşra Marşan1,2
+Vianney Grenez1 · Gabriel Stein4 · Momcilo Mrkaic5
+Pavle Padjin5 · Vladimir Zeljkovic5 · Calbert Graham1,3 +

+ +

1 Oruk AI

+ + + + + + + +
Stanford University
2 Stanford University		University of Cambridge
3 University of Cambridge		OpenWhispr
4 OpenWhispr		Hoid
5 Hoid
+ + +This short report describes **Orukeet r3**, the selected native-format checkpoint with SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56`. Its [model identity](model.json) pins the exact public NeMo artifact and freeze audit. + +The report presents the fitted-filter construction, a compact adaptation recipe, complete paired LibriSpeech/FLEURS results, multilingual pooled WER, and the existing accent/domain sample re-evaluated with this checkpoint. Every score has the stock Parakeet comparison. The two figures show the unchanged fitted kernels and their allocation across encoder layers. + +```sh +.venv/bin/python scripts/build_neurips_report.py +pdftoppm -png -r 150 output/pdf/orukeet-technical-report.pdf .tmp/orukeet-r3-page +``` + +The builder verifies checkpoint identity, exact frozen kernels, all per-record counts, pooled denominators and independent scoring audits. It generates the numeric TeX includes, compiles with the official NeurIPS 2026 preprint style, and verifies every rendered benchmark row. Pass `--tex-bin /path/to/texlive/bin` to use pdfLaTeX and BibTeX. The release PDF is built with TeX Live 2025, uses embedded outline fonts, and records its title and all nine authors in PDF metadata. The build receipt records input hashes, page count and visual review. + +- [Report PDF](../output/pdf/orukeet-technical-report.pdf) +- [All current-checkpoint scores](../docs/current-checkpoint-benchmarks.md) +- [Methods companion](../docs/technical-report.md) +- [Complete kernel atlas](assets/kernel-atlas.pdf) + +The manuscript and numerical evaluation records accompany Orukeet v0.1.0. The NeMo, Q8 and F16 release files all derive from this r3 checkpoint; `release/model-stages.json` and the artifact catalog record their exact identities. + + +## Citation + +```bibtex +@techreport{roll2026orukeet, + title = {{Orukeet}: Multilingual {ASR} with Frozen {Gabor} Kernels}, + author = {Roll, Nathan and + Yi, Irene and + Mar{\c{s}}an, B{\"u}{\c{s}}ra and + Grenez, Vianney and + Stein, Gabriel and + Mrkaic, Momcilo and + Padjin, Pavle and + Zeljkovic, Vladimir and + Graham, Calbert}, + institution = {Oruk AI}, + year = {2026}, + type = {Technical report}, + url = {https://github.com/Oruk-AI/orukeet/blob/main/output/pdf/orukeet-technical-report.pdf} +} +``` + +[Download BibTeX](../CITATION.bib) · [Citation metadata](../CITATION.cff) + diff --git a/report/assets/affiliations/SOURCES.md b/report/assets/affiliations/SOURCES.md new file mode 100644 index 0000000000000000000000000000000000000000..e9ae533b52d22425220909bc890dcb06d3bb8b23 --- /dev/null +++ b/report/assets/affiliations/SOURCES.md @@ -0,0 +1,17 @@ +# Report affiliation logos + +These marks identify the authors' affiliations. They remain the property of their respective organizations; the repository's code and model licenses do not grant rights to these marks. + +Assets retrieved on September 7, 2026: + +| Asset | Official source | Preparation | +| --- | --- | --- | +| Oruk AI | [September 2026 primary lockup](https://oruk.ai/brand/kit/oruk-primary.png), from the [official branding guide](https://oruk.ai/branding); original [PNG](../oruk-lockup.png) and [outlined SVG](../oruk-lockup.svg) retained | Updated September 8, 2026. Color Signal tile and lowercase black wordmark; original files, unchanged. Hashes in [the provenance record](../provenance.json). | +| `openwhispr.svg` | [OpenWhispr website logo](https://openwhispr.com/logo.svg) | Original SVG; PDF conversion for LaTeX. | +| `hoid.svg` | [Hoid website](https://hoid-ai.com/) | The two SVGs in the site's header, joined at their original view-box coordinates; `currentColor` uses the site's dark ink, `#0c0b03`. PDF conversion for LaTeX. | +| `stanford.png` | [Stanford University wordmark](https://identity.stanford.edu/wp-content/uploads/sites/3/2020/06/wordmark-full-nospace-red.png), from the [official identity guide](https://identity.stanford.edu/visual-identity/stanford-logos/wordmarks/) | Original PNG, unchanged. | +| `cambridge.svg` | [University of Cambridge website logo](https://www.cam.ac.uk/themes/custom/fresh/images/interface/cambridge_university2.svg) | Original SVG; PDF conversion for LaTeX. | + +SVG-to-PDF conversion uses CairoSVG, restricted to PDF 1.5 for compatibility with the report compiler, without recoloring, distortion or removal of elements. On macOS with Homebrew Cairo, set `DYLD_FALLBACK_LIBRARY_PATH=/opt/homebrew/lib` when invoking CairoSVG. + +The author list uses the names on [Hoid's team page](https://hoid-ai.com/). Gabriel Stein's full name is verified against [OpenWhispr's official website](https://openwhispr.com/dpa), which identifies him as CEO, and [OpenWhispr's company profile](https://www.linkedin.com/company/openwhispr), checked September 7, 2026. The Stanford and Cambridge affiliations follow the existing author assignments in the report. diff --git a/report/assets/affiliations/cambridge.pdf b/report/assets/affiliations/cambridge.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aaa2eda54764ca7966abba74bcba3f23d0528e2b Binary files /dev/null and b/report/assets/affiliations/cambridge.pdf differ diff --git a/report/assets/affiliations/cambridge.svg b/report/assets/affiliations/cambridge.svg new file mode 100644 index 0000000000000000000000000000000000000000..2234b740994f3470514154dbd60a87f9d820c7e5 --- /dev/null +++ b/report/assets/affiliations/cambridge.svg @@ -0,0 +1 @@ + \ No newline at end of file diff --git a/report/assets/affiliations/hoid.pdf b/report/assets/affiliations/hoid.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55f103063c96dfb61655c9e8ec109f4a228d9a1c Binary files /dev/null and b/report/assets/affiliations/hoid.pdf differ diff --git a/report/assets/affiliations/hoid.svg b/report/assets/affiliations/hoid.svg new file mode 100644 index 0000000000000000000000000000000000000000..0be02ba44ae4a8cf89bb06b343a92acf295214f0 --- /dev/null +++ b/report/assets/affiliations/hoid.svg @@ -0,0 +1,4 @@ + + + + diff --git a/report/assets/affiliations/openwhispr.pdf b/report/assets/affiliations/openwhispr.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b518c8cf9f471c04fbe649a09f099ee269cf4604 Binary files /dev/null and b/report/assets/affiliations/openwhispr.pdf differ diff --git a/report/assets/affiliations/openwhispr.svg b/report/assets/affiliations/openwhispr.svg new file mode 100644 index 0000000000000000000000000000000000000000..e71085968e2d011b382afeb7980faade57dedb31 --- /dev/null +++ b/report/assets/affiliations/openwhispr.svg @@ -0,0 +1,3 @@ + + + diff --git a/report/assets/affiliations/stanford.png b/report/assets/affiliations/stanford.png new file mode 100644 index 0000000000000000000000000000000000000000..e0c8914d658718d4212790f6e4774b0c22dd6af6 Binary files /dev/null and b/report/assets/affiliations/stanford.png differ diff --git a/report/assets/kernel-atlas-caption.md b/report/assets/kernel-atlas-caption.md new file mode 100644 index 0000000000000000000000000000000000000000..053db63cd75a2181daf13b428c8251f9e3277ed1 --- /dev/null +++ b/report/assets/kernel-atlas-caption.md @@ -0,0 +1 @@ +Layer atlas: the median selected fit-error row within each of the 24 layers. Original (gray circles) and frozen Gabor (blue crosses) taps share the original L2 normalization and common axes. All plotted samples come from the recorded fits; lines join discrete samples. diff --git a/report/assets/kernel-atlas-preview.png b/report/assets/kernel-atlas-preview.png new file mode 100644 index 0000000000000000000000000000000000000000..a097b296adcc40c19165d86980eba63fae171c44 --- /dev/null +++ b/report/assets/kernel-atlas-preview.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5a4a3993d51cb7197c50cd82b2fb9fd5103d8dffd6f31b0fe7c5b923230691c +size 169043 diff --git a/report/assets/kernel-atlas.pdf b/report/assets/kernel-atlas.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e2d53d4a0b6eb280207b40c6b1e33c1221c33805 Binary files /dev/null and b/report/assets/kernel-atlas.pdf differ diff --git a/report/assets/kernel-atlas.png b/report/assets/kernel-atlas.png new file mode 100644 index 0000000000000000000000000000000000000000..be28bb5f5daab746358b9f4b83443906ea863453 --- /dev/null +++ b/report/assets/kernel-atlas.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b8b24400ac7b9e8b6e4484eb0c6c36cca8f618167141d508a2b6468824888c3 +size 376772 diff --git a/report/assets/kernel-atlas.svg b/report/assets/kernel-atlas.svg new file mode 100644 index 0000000000000000000000000000000000000000..dc50f98b2131b28caab36d56d9fa038843617162 --- /dev/null +++ b/report/assets/kernel-atlas.svg @@ -0,0 +1,3115 @@ + + + + + + + + 2026-09-06T19:33:09.873491 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 1 · ch 411 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 2 · ch 235 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 3 · ch 69 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 4 · ch 467 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 5 · ch 7 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 6 · ch 813 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 7 · ch 510 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 8 · ch 393 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 9 · ch 558 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 10 · ch 478 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 11 · ch 278 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 12 · ch 144 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 13 · ch 506 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 14 · ch 681 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 15 · ch 397 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 16 · ch 969 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 17 · ch 240 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 18 · ch 561 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 19 · ch 574 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 20 · ch 80 + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 21 · ch 930 + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 22 · ch 851 + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 23 · ch 819 + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Layer 24 · ch 52 + + + + One median selected kernel from every encoder layer + + + Tap index + + + Weight / original norm + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/report/assets/kernel-fits-caption.md b/report/assets/kernel-fits-caption.md new file mode 100644 index 0000000000000000000000000000000000000000..38d9a1230cb1f8e6f0a78cf81873d5f5d960b0df --- /dev/null +++ b/report/assets/kernel-fits-caption.md @@ -0,0 +1 @@ +Four exact kernel replacements. Examples are ranks 1,537, 4,609, 7,680 and 10,752 of the selected 12,288, nearest the 12.5th, 37.5th, 62.5th and 87.5th fit-error percentiles. Each pair is divided by the original kernel L2 norm. Markers are the nine stored taps; connecting lines guide the eye. RMS percentages are relative to the original norm. diff --git a/report/assets/kernel-fits-preview.png b/report/assets/kernel-fits-preview.png new file mode 100644 index 0000000000000000000000000000000000000000..0856f19c1253d40718d1cb99c46cc5daae2a59cd Binary files /dev/null and b/report/assets/kernel-fits-preview.png differ diff --git a/report/assets/kernel-fits.pdf b/report/assets/kernel-fits.pdf new file mode 100644 index 0000000000000000000000000000000000000000..65ababf986447ad082c14b82795591bb6b6f40a5 Binary files /dev/null and b/report/assets/kernel-fits.pdf differ diff --git a/report/assets/kernel-fits.png b/report/assets/kernel-fits.png new file mode 100644 index 0000000000000000000000000000000000000000..840b426cace06f97f19badfa5dc2b5e017a7c467 Binary files /dev/null and b/report/assets/kernel-fits.png differ diff --git a/report/assets/kernel-fits.svg b/report/assets/kernel-fits.svg new file mode 100644 index 0000000000000000000000000000000000000000..e2132ceeb0fc58e44419bf29321d7c7033f477b5 --- /dev/null +++ b/report/assets/kernel-fits.svg @@ -0,0 +1,646 @@ + + + + + + + + 2026-09-06T19:33:03.883567 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + Tap index + + + + + + + + + + + + + + −1 + + + + + + + + + + 0 + + + + + + + + + + 1 + + + + Weight / original norm + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + A Rank 1,537 + 2.74% RMS + + + + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + Tap index + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + B Rank 4,609 + 5.05% RMS + + + + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + Tap index + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + C Rank 7,680 + 7.73% RMS + + + + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + Tap index + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + D Rank 10,752 + 11.20% RMS + + + + + + + + + + + Original taps + + + + + + + + + Frozen Gabor taps + + + + + + + + + + + + + + + + + + diff --git a/report/assets/language-deltas-caption.md b/report/assets/language-deltas-caption.md new file mode 100644 index 0000000000000000000000000000000000000000..73db86a98938630ed7568ad4cb0c31fd48a4e339 --- /dev/null +++ b/report/assets/language-deltas-caption.md @@ -0,0 +1 @@ +Per-language Orukeet WER change under matched NeMo decoding, with the same paired bootstrap as the macro comparison. All 20 primary languages and complete 95% intervals are shown, alphabetically. Languages share a scale. These marginal intervals are not adjusted for simultaneous inference. diff --git a/report/assets/language-deltas-preview.png b/report/assets/language-deltas-preview.png new file mode 100644 index 0000000000000000000000000000000000000000..8e33b45d95614c655d26f103d405f532decf7ef1 Binary files /dev/null and b/report/assets/language-deltas-preview.png differ diff --git a/report/assets/language-deltas.pdf b/report/assets/language-deltas.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a0f5e5f619f4ddd21fecf98638647760dc25737 Binary files /dev/null and b/report/assets/language-deltas.pdf differ diff --git a/report/assets/language-deltas.png b/report/assets/language-deltas.png new file mode 100644 index 0000000000000000000000000000000000000000..968b7f5586ba8d610920a561f2f3888e6f037e4f Binary files /dev/null and b/report/assets/language-deltas.png differ diff --git a/report/assets/language-deltas.svg b/report/assets/language-deltas.svg new file mode 100644 index 0000000000000000000000000000000000000000..def38759b9e34f2b33dc81a440e6e08a91eb4b3f --- /dev/null +++ b/report/assets/language-deltas.svg @@ -0,0 +1,670 @@ + + + + + + + + 2026-09-06T19:33:06.671809 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + −2.0 + + + + + + + + + + −1.5 + + + + + + + + + + −1.0 + + + + + + + + + + −0.5 + + + + + + + + + + 0.0 + + + + + + + + + + 0.5 + + + + WER change from adaptation baseline (pp) + + + + + + + SV + + + + + + SL + + + + + + SK + + + + + + RU + + + + + + RO + + + + + + PT + + + + + + PL + + + + + + NL + + + + + + LV + + + + + + LT + + + + + + IT + + + + + + HU + + + + + + HR + + + + + + FR + + + + + + FI + + + + + + ES + + + + + + EN + + + + + + DE + + + + + + DA + + + + + + CS + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + All 20 languages in the primary endpoint + + + + + + + + + diff --git a/report/assets/model-figures-source.json b/report/assets/model-figures-source.json new file mode 100644 index 0000000000000000000000000000000000000000..80a592624431ef885f100e9810a6204ebaab83c1 --- /dev/null +++ b/report/assets/model-figures-source.json @@ -0,0 +1,493 @@ +{ + "model_sha256": "4295a6d820a40b99786331d1c7a6b6c328916c8329b23d39415b0649a5d42811", + "inputs": { + "training/gabor_half/fits/fits.npz": "52791beaf3219421e4ad3aa2e2591875b50dbfaa83a6309361c2df71db08de3d", + "training/gabor_half/results/full-r15-0100.json": "8c3f166cec92ada017f2973bad18bc9b38edb5aca4a54c1f68e14fd4ffe84b94", + "report/make_model_figures.py": "d1619f159c0c11c7497498083843e0ea3f9352b299f4279551e8170de4eebe01" + }, + "style": "academic-figures: final-size typography, restrained blue/gray/teal, vector export", + "design_precedent": { + "repository": "https://github.com/Nathan-Roll1/gabormer", + "manuscripts": [ + "interspeech/paper/gabormer_interspeech/gabormer_interspeech.tex", + "colm/paper/gabormer_unified/gabormer_unified.tex" + ], + "adapted_ideas": [ + "original/fitted kernel small multiples", + "fit distribution", + "layer-depth profile" + ], + "data": "All plotted measurements are Orukeet data. No Gabormer values or artwork copied." + }, + "figures": { + "kernel-fits": { + "size_inches": [ + 5.5, + 1.9 + ], + "caption": "Four exact kernel replacements. Examples are ranks 1,537, 4,609, 7,680 and 10,752 of the selected 12,288, nearest the 12.5th, 37.5th, 62.5th and 87.5th fit-error percentiles. Each pair is divided by the original kernel L2 norm. Markers are the nine stored taps; connecting lines guide the eye. RMS percentages are relative to the original norm.", + "examples": [ + { + "rank": 1537, + "layer_zero_based": 2, + "channel": 119, + "relative_rms_percent": 2.743456382851213 + }, + { + "rank": 4609, + "layer_zero_based": 22, + "channel": 341, + "relative_rms_percent": 5.048379586990088 + }, + { + "rank": 7680, + "layer_zero_based": 18, + "channel": 248, + "relative_rms_percent": 7.727121376538521 + }, + { + "rank": 10752, + "layer_zero_based": 6, + "channel": 1017, + "relative_rms_percent": 11.204069140642124 + } + ], + "unit": "one temporal depthwise kernel", + "uncertainty": "Exact weights; no sampling interval.", + "sha256": { + "pdf": "b952a1d8e84032cd20cc452dfbea40c6d2c34cea7de4257b7a375ccac73b9e01", + "svg": "4f2b63601bb7b9543cee0dec0aaaecc06c1741798eda0c6aa5e2b4f39ed1191b", + "png": "53daf3dc1431660b7696c83d94e56543f8236563a81a4dd9fb75e3cab997fa52" + } + }, + "selection-profile": { + "size_inches": [ + 5.5, + 2.15 + ], + "caption": "Global Gabor selection. (A) Empirical cumulative distribution of relative RMS fit error for all 24,576 original kernels; blue marks the selected half. Every kernel is included. (B) The resulting allocation across all 24 encoder layers. The dashed 512 line denotes half a layer; selection uses a single global ranking. Counts range from 175 to 748 per layer.", + "selected_count": 12288, + "total_count": 24576, + "cutoff_percent": 13.296859709168707, + "layer_counts": [ + 700, + 748, + 707, + 677, + 713, + 657, + 583, + 573, + 564, + 502, + 389, + 291, + 194, + 175, + 238, + 343, + 398, + 455, + 439, + 501, + 569, + 560, + 635, + 677 + ], + "uncertainty": "Complete kernel population; no sampling interval.", + "sha256": { + "pdf": "0a2e8647a8a1a895e7a1e6d9f15779eac14c87f56e30eeeb85bc05ce720ba32f", + "svg": "9ddb7db5f54eb4f97294e31998ce06d99fdf18e9cbf88a906bef70bc99f4e19a", + "png": "6202432a630ef22fb54bf57c494437dcb8f526c26a9a7b6ecf1bcbdf2510a75d" + } + }, + "recognition-deltas": { + "size_inches": [ + 5.5, + 1.75 + ], + "caption": "Orukeet minus the pre-Gabor adaptation baseline under matched NeMo decoding. Points are macro WER differences; bars are 95% paired global-cluster bootstrap intervals (5,000 replicates, seed 20260905; 4,266 speaker or parallel-sentence clusters). Negative values mean lower WER. The primary endpoint uses 13,246 recordings and 20 fixed languages; English averages seven corpus WERs. Both systems were selected during the same adaptive campaign; intervals condition on those selections.", + "data": { + "primary": { + "reference_wer": 15.371306565215429, + "candidate_wer": 15.34088687521396, + "delta_pp": -0.03041969000146949, + "relative_improvement": 0.001978991823005316, + "paired_delta_95ci": [ + -0.1410351508899022, + 0.061823155508115304 + ] + }, + "english": { + "reference_wer": 9.360330471804506, + "candidate_wer": 9.359237881128971, + "delta_pp": -0.0010925906755350212, + "relative_improvement": 0.0001167256518160853, + "paired_delta_95ci": [ + -0.049291283437256965, + 0.04707342853436239 + ] + } + }, + "independent_unit": "Source-qualified speakers; shared parallel sentence IDs for FLEURS", + "replicates": 5000, + "sha256": { + "pdf": "62f0471788cb59e27fc317a004b34fe457a354ae511e01f0fc6d1efb8da87a0a", + "svg": "ab94addc78953dd3038cd30436aeb73f6580f1b62903180f6ba81551a3317595", + "png": "83890bae2503e59e918bf424cd841094302d5366e3a285523105db49f3e1fffc" + } + }, + "language-deltas": { + "size_inches": [ + 5.5, + 5.4 + ], + "caption": "Per-language Orukeet WER change under matched NeMo decoding, with the same paired bootstrap as the macro comparison. All 20 primary languages and complete 95% intervals are shown, alphabetically. Languages share a scale. These marginal intervals are not adjusted for simultaneous inference.", + "data": { + "cs": { + "reference_wer": 11.087057368176898, + "candidate_wer": 11.264883253440544, + "delta_pp": 0.17782588526364584, + "paired_delta_95ci": [ + 0.046487744040169324, + 0.33100222230040766 + ], + "rows": 618 + }, + "da": { + "reference_wer": 24.8955459203784, + "candidate_wer": 24.666929444225463, + "delta_pp": -0.22861647615293634, + "paired_delta_95ci": [ + -0.7915456521772453, + 0.29458933204624366 + ], + "rows": 635 + }, + "de": { + "reference_wer": 8.289022805151838, + "candidate_wer": 8.336287368545433, + "delta_pp": 0.04726456339359508, + "paired_delta_95ci": [ + -0.03600124236927005, + 0.139001899016581 + ], + "rows": 1036 + }, + "en": { + "reference_wer": 6.186561104344295, + "candidate_wer": 6.227161997563947, + "delta_pp": 0.040600893219651546, + "paired_delta_95ci": [ + -0.0250543055072794, + 0.11067552841084288 + ], + "rows": 1025 + }, + "es": { + "reference_wer": 5.858259898225146, + "candidate_wer": 5.862397087418808, + "delta_pp": 0.004137189193661683, + "paired_delta_95ci": [ + -0.04850969273685689, + 0.06322720898780393 + ], + "rows": 1033 + }, + "fi": { + "reference_wer": 14.380776340110906, + "candidate_wer": 14.4547134935305, + "delta_pp": 0.07393715341959428, + "paired_delta_95ci": [ + -0.13470003939532066, + 0.24753818405096303 + ], + "rows": 488 + }, + "fr": { + "reference_wer": 8.541712768311573, + "candidate_wer": 8.577118831599911, + "delta_pp": 0.035406063288338174, + "paired_delta_95ci": [ + -0.0435854680347276, + 0.12346240094497965 + ], + "rows": 1016 + }, + "hr": { + "reference_wer": 18.61632214123551, + "candidate_wer": 18.610347711793523, + "delta_pp": -0.005974429441987894, + "paired_delta_95ci": [ + -0.13763448381403434, + 0.1381570654452181 + ], + "rows": 642 + }, + "hu": { + "reference_wer": 16.10576923076923, + "candidate_wer": 16.158353365384617, + "delta_pp": 0.0525841346153868, + "paired_delta_95ci": [ + -0.12125307165865014, + 0.22175088328334605 + ], + "rows": 638 + }, + "it": { + "reference_wer": 12.261909723669403, + "candidate_wer": 12.207727253782354, + "delta_pp": -0.05418246988704922, + "paired_delta_95ci": [ + -0.23291848494299147, + 0.11870121558172134 + ], + "rows": 1026 + }, + "lt": { + "reference_wer": 28.081740276862227, + "candidate_wer": 27.620303230059328, + "delta_pp": -0.46143704680289943, + "paired_delta_95ci": [ + -2.2434471825746565, + 0.7894996879550652 + ], + "rows": 84 + }, + "lv": { + "reference_wer": 20.553359683794465, + "candidate_wer": 20.41219649915302, + "delta_pp": -0.14116318464144584, + "paired_delta_95ci": [ + -0.6161541697404739, + 0.31487300593951373 + ], + "rows": 427 + }, + "nl": { + "reference_wer": 10.70418449488723, + "candidate_wer": 10.725940967437813, + "delta_pp": 0.02175647255058344, + "paired_delta_95ci": [ + -0.1126701201667522, + 0.14785415081655787 + ], + "rows": 673 + }, + "pl": { + "reference_wer": 5.455712451861361, + "candidate_wer": 5.500641848523748, + "delta_pp": 0.044929396662387155, + "paired_delta_95ci": [ + -0.02018927584315764, + 0.11494302426300214 + ], + "rows": 1022 + }, + "pt": { + "reference_wer": 5.565978736710444, + "candidate_wer": 5.628517823639775, + "delta_pp": 0.06253908692933052, + "paired_delta_95ci": [ + -0.12005252957810716, + 0.27689221214989074 + ], + "rows": 420 + }, + "ro": { + "reference_wer": 14.169147645826442, + "candidate_wer": 14.147094497739552, + "delta_pp": -0.022053148086889962, + "paired_delta_95ci": [ + -0.15623518789188579, + 0.11505959003916987 + ], + "rows": 639 + }, + "ru": { + "reference_wer": 5.169561621174524, + "candidate_wer": 5.128205128205129, + "delta_pp": -0.041356492969395475, + "paired_delta_95ci": [ + -0.31518400268930596, + 0.1880351838474757 + ], + "rows": 253 + }, + "sk": { + "reference_wer": 12.654815678274806, + "candidate_wer": 12.771382777211132, + "delta_pp": 0.11656709893632566, + "paired_delta_95ci": [ + -0.050326789526338414, + 0.29991279359415846 + ], + "rows": 615 + }, + "sl": { + "reference_wer": 51.75879396984924, + "candidate_wer": 51.433638782146026, + "delta_pp": -0.32515518770321705, + "paired_delta_95ci": [ + -1.2045514631636953, + 0.516073698562175 + ], + "rows": 328 + }, + "sv": { + "reference_wer": 27.08989944469458, + "candidate_wer": 27.083896142878583, + "delta_pp": -0.00600330181599773, + "paired_delta_95ci": [ + -0.11988722473512176, + 0.1147547081474553 + ], + "rows": 628 + } + }, + "independent_unit": "Source-qualified speakers; shared parallel sentence IDs for FLEURS", + "sha256": { + "pdf": "7564f39e28ee6deae1d7ad4b82e3737b1dc703044b5992c5f9108cbf2e9ff1c0", + "svg": "cff05741bbe01c83ad5f839f99a40c2f189357e71bc6fddd8cd62c920116d693", + "png": "64e77e3423cd60952f31a4e80bb486a05a3325cd5a0b3d92bcf9581cf9de201b" + } + }, + "kernel-atlas": { + "size_inches": [ + 7.0, + 8.0 + ], + "caption": "Layer atlas: the median selected fit-error row within each of the 24 layers. Original (gray circles) and frozen Gabor (blue crosses) taps share the original L2 normalization and common axes. All plotted samples come from the recorded fits; lines join discrete samples.", + "examples": [ + { + "layer": 1, + "channel": 411, + "error_percent": 4.808574908575614 + }, + { + "layer": 2, + "channel": 235, + "error_percent": 3.6974837759188737 + }, + { + "layer": 3, + "channel": 69, + "error_percent": 4.28624639462823 + }, + { + "layer": 4, + "channel": 467, + "error_percent": 4.232227242800408 + }, + { + "layer": 5, + "channel": 7, + "error_percent": 4.896815610397718 + }, + { + "layer": 6, + "channel": 813, + "error_percent": 5.377686527351617 + }, + { + "layer": 7, + "channel": 510, + "error_percent": 5.511299263585478 + }, + { + "layer": 8, + "channel": 393, + "error_percent": 6.0199091632068935 + }, + { + "layer": 9, + "channel": 558, + "error_percent": 6.149495255970944 + }, + { + "layer": 10, + "channel": 478, + "error_percent": 7.336905281121791 + }, + { + "layer": 11, + "channel": 278, + "error_percent": 8.230807419395926 + }, + { + "layer": 12, + "channel": 144, + "error_percent": 8.348349271273877 + }, + { + "layer": 13, + "channel": 506, + "error_percent": 9.11105552284122 + }, + { + "layer": 14, + "channel": 681, + "error_percent": 9.599238530957573 + }, + { + "layer": 15, + "channel": 397, + "error_percent": 8.89095050720063 + }, + { + "layer": 16, + "channel": 969, + "error_percent": 8.75968011197192 + }, + { + "layer": 17, + "channel": 240, + "error_percent": 7.399901656916065 + }, + { + "layer": 18, + "channel": 561, + "error_percent": 6.485497128070223 + }, + { + "layer": 19, + "channel": 574, + "error_percent": 6.089518105141542 + }, + { + "layer": 20, + "channel": 80, + "error_percent": 6.612476904613199 + }, + { + "layer": 21, + "channel": 930, + "error_percent": 7.520982070424001 + }, + { + "layer": 22, + "channel": 851, + "error_percent": 7.041443182039135 + }, + { + "layer": 23, + "channel": 819, + "error_percent": 7.0331814150216525 + }, + { + "layer": 24, + "channel": 52, + "error_percent": 8.028077650199089 + } + ], + "unit": "one selected kernel per layer", + "selection_rule": "lower median by global fit-error order within layer", + "sha256": { + "pdf": "73a8c7e2f1ea83f7bcc9eefcdbd1a53b6e95383a0b3799f45f75390d676be9b5", + "svg": "41b88f335fa6200353201a77a1c86d0d471fbee61a1a64a7a6c0bffaf2abda88", + "png": "2b8b24400ac7b9e8b6e4484eb0c6c36cca8f618167141d508a2b6468824888c3" + } + } + } +} diff --git a/report/assets/model-figures-visual-review.json b/report/assets/model-figures-visual-review.json new file mode 100644 index 0000000000000000000000000000000000000000..292ea29cbefe0e26d9031a6be87c100c3c083ad7 --- /dev/null +++ b/report/assets/model-figures-visual-review.json @@ -0,0 +1,44 @@ +{ + "status": "passed", + "figures": [ + "kernel-fits", + "selection-profile", + "recognition-deltas", + "language-deltas", + "kernel-atlas" + ], + "reviewed_exports": { + "kernel-fits": { + "pdf": "b952a1d8e84032cd20cc452dfbea40c6d2c34cea7de4257b7a375ccac73b9e01", + "svg": "4f2b63601bb7b9543cee0dec0aaaecc06c1741798eda0c6aa5e2b4f39ed1191b", + "png": "53daf3dc1431660b7696c83d94e56543f8236563a81a4dd9fb75e3cab997fa52" + }, + "selection-profile": { + "pdf": "0a2e8647a8a1a895e7a1e6d9f15779eac14c87f56e30eeeb85bc05ce720ba32f", + "svg": "9ddb7db5f54eb4f97294e31998ce06d99fdf18e9cbf88a906bef70bc99f4e19a", + "png": "6202432a630ef22fb54bf57c494437dcb8f526c26a9a7b6ecf1bcbdf2510a75d" + }, + "recognition-deltas": { + "pdf": "62f0471788cb59e27fc317a004b34fe457a354ae511e01f0fc6d1efb8da87a0a", + "svg": "ab94addc78953dd3038cd30436aeb73f6580f1b62903180f6ba81551a3317595", + "png": "83890bae2503e59e918bf424cd841094302d5366e3a285523105db49f3e1fffc" + }, + "language-deltas": { + "pdf": "7564f39e28ee6deae1d7ad4b82e3737b1dc703044b5992c5f9108cbf2e9ff1c0", + "svg": "cff05741bbe01c83ad5f839f99a40c2f189357e71bc6fddd8cd62c920116d693", + "png": "64e77e3423cd60952f31a4e80bb486a05a3325cd5a0b3d92bcf9581cf9de201b" + }, + "kernel-atlas": { + "pdf": "73a8c7e2f1ea83f7bcc9eefcdbd1a53b6e95383a0b3799f45f75390d676be9b5", + "svg": "41b88f335fa6200353201a77a1c86d0d471fbee61a1a64a7a6c0bffaf2abda88", + "png": "2b8b24400ac7b9e8b6e4484eb0c6c36cca8f618167141d508a2b6468824888c3" + } + }, + "checks": [ + "Main figures inspected at final 5.5-inch manuscript width", + "All four rendered report pages inspected", + "Complete kernel atlas labels shortened to fit without reducing font size", + "All 20 language intervals retained without clipping", + "No plotted result inferred from Gabormer measurements" + ] +} diff --git a/report/assets/oruk-lockup.png b/report/assets/oruk-lockup.png new file mode 100644 index 0000000000000000000000000000000000000000..2a9228bcd470154a23dcd616adda83a9b24acfc4 Binary files /dev/null and b/report/assets/oruk-lockup.png differ diff --git a/report/assets/oruk-lockup.svg b/report/assets/oruk-lockup.svg new file mode 100644 index 0000000000000000000000000000000000000000..34852c07afd77d9fc6b4552d453d93ba8b109244 --- /dev/null +++ b/report/assets/oruk-lockup.svg @@ -0,0 +1,9 @@ +oruk · Primary + + + + + + + + diff --git a/report/assets/provenance.json b/report/assets/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..ce487c84cdf9eaa033a3c3aa77667741bc14a7bd --- /dev/null +++ b/report/assets/provenance.json @@ -0,0 +1,15 @@ +{ + "template_url": "https://media.neurips.cc/Conferences/NeurIPS2026/Formatting_Instructions_For_NeurIPS_2026.zip", + "template_link_source": "https://neurips.cc/Conferences/2026/CallForPapers", + "retrieved": "2026-09-06", + "style_sha256": "c3fc2894e83d2517ca18b66741d6c595986d97957dc08ec08bb2125a7ec4555a", + "archive_sha256": "82473931e3ef710fcd3f4a8cd4119b9de32e56825f90f9e5a6d55f2d01b817d9", + "style": "Official unmodified NeurIPS 2026 style, preprint option; no submission or acceptance implied.", + "logo_source": "https://oruk.ai/brand/kit/oruk-primary.png", + "logo_sha256": "932df9164b237a1a31d1e636897d1766a15b8bac738acf81ee93de8ea2d6c18c", + "logo_use": "Official September 2026 primary lockup: color Signal tile and lowercase black wordmark. Original PNG and outlined SVG retained without recoloring or distortion.", + "logo_guide": "https://oruk.ai/branding", + "logo_retrieved": "2026-09-08", + "logo_svg_source": "https://oruk.ai/brand/kit/oruk-primary.svg", + "logo_svg_sha256": "3547cfacd79cf423d104adfdf342a28aef3a98dd187b1a4cc1949478d690a051" +} diff --git a/report/assets/recognition-deltas-caption.md b/report/assets/recognition-deltas-caption.md new file mode 100644 index 0000000000000000000000000000000000000000..5d26fecf6ed997dae764d8d12d9f4c5e7c038a08 --- /dev/null +++ b/report/assets/recognition-deltas-caption.md @@ -0,0 +1 @@ +Orukeet minus the pre-Gabor adaptation baseline under matched NeMo decoding. Points are macro WER differences; bars are 95% paired global-cluster bootstrap intervals (5,000 replicates, seed 20260905; 4,266 speaker or parallel-sentence clusters). Negative values mean lower WER. The primary endpoint uses 13,246 recordings and 20 fixed languages; English averages seven corpus WERs. Both systems were selected during the same adaptive campaign; intervals condition on those selections. diff --git a/report/assets/recognition-deltas-preview.png b/report/assets/recognition-deltas-preview.png new file mode 100644 index 0000000000000000000000000000000000000000..f36c1659de5605b68d3ceb80ff197f63a385072f Binary files /dev/null and b/report/assets/recognition-deltas-preview.png differ diff --git a/report/assets/recognition-deltas.pdf b/report/assets/recognition-deltas.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05bd328cc0c4add73e17a569b436b327def14884 Binary files /dev/null and b/report/assets/recognition-deltas.pdf differ diff --git a/report/assets/recognition-deltas.png b/report/assets/recognition-deltas.png new file mode 100644 index 0000000000000000000000000000000000000000..bec15b7dd1206b84c39224046751222f39207f67 Binary files /dev/null and b/report/assets/recognition-deltas.png differ diff --git a/report/assets/recognition-deltas.svg b/report/assets/recognition-deltas.svg new file mode 100644 index 0000000000000000000000000000000000000000..0b7a17cf799412846123ba94e733811c0b61d97a --- /dev/null +++ b/report/assets/recognition-deltas.svg @@ -0,0 +1,202 @@ + + + + + + + + 2026-09-06T19:33:05.923111 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + −0.15 + + + + + + + + + + −0.10 + + + + + + + + + + −0.05 + + + + + + + + + + 0.00 + + + + + + + + + + 0.05 + + + + + + + + + + 0.10 + + + + WER change from adaptation baseline (percentage points) + + + + + + + English, 7 corpora + + + + + + 20 languages + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Recognition after recovery + + + + + + + + + diff --git a/report/assets/selection-profile-caption.md b/report/assets/selection-profile-caption.md new file mode 100644 index 0000000000000000000000000000000000000000..0ec94fb7202385dcc073ebdecb50bdd5eab88a4a --- /dev/null +++ b/report/assets/selection-profile-caption.md @@ -0,0 +1 @@ +Global Gabor selection. (A) Empirical cumulative distribution of relative RMS fit error for all 24,576 original kernels; blue marks the selected half. Every kernel is included. (B) The resulting allocation across all 24 encoder layers. The dashed 512 line denotes half a layer; selection uses a single global ranking. Counts range from 175 to 748 per layer. diff --git a/report/assets/selection-profile-preview.png b/report/assets/selection-profile-preview.png new file mode 100644 index 0000000000000000000000000000000000000000..024fcac0a3e678116a7a4cabf5a015aa68f9c437 Binary files /dev/null and b/report/assets/selection-profile-preview.png differ diff --git a/report/assets/selection-profile.pdf b/report/assets/selection-profile.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95426388563305e3d2a29132c946216dea433a52 Binary files /dev/null and b/report/assets/selection-profile.pdf differ diff --git a/report/assets/selection-profile.png b/report/assets/selection-profile.png new file mode 100644 index 0000000000000000000000000000000000000000..b49d2dd6ada91ba0ad2c4506e13ff944e4bc10e8 Binary files /dev/null and b/report/assets/selection-profile.png differ diff --git a/report/assets/selection-profile.svg b/report/assets/selection-profile.svg new file mode 100644 index 0000000000000000000000000000000000000000..24374c24a10689d32d21de8077d1367b0f1a729e --- /dev/null +++ b/report/assets/selection-profile.svg @@ -0,0 +1,706 @@ + + + + + + + + 2026-09-06T19:33:05.143666 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + 0 + + + + + + + + + + 25 + + + + + + + + + + 50 + + + + + + + + + + 75 + + + + + + + + + + 100 + + + + Relative RMS fit error (%) + + + + + + + + + + + + + + 0 + + + + + + + + + + 50 + + + + + + + + + + 100 + + + + Kernels at or below (%) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 50% selected + cutoff 13.30% + + + A Global fit ranking + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 1 + + + + + + + + + + 6 + + + + + + + + + + 12 + + + + + + + + + + 18 + + + + + + + + + + 24 + + + + Encoder layer + + + + + + + + + + + 0 + + + + + + + + + + 512 + + + + + + + + + + 1024 + + + + Frozen kernels / 1,024 + + + + + + + + + + + + + B Selection by layer + + + + + + + + + + + + diff --git a/report/assets/surgery-source.json b/report/assets/surgery-source.json new file mode 100644 index 0000000000000000000000000000000000000000..e37b2300acd8122c5e01acff686d77e4a9cb07e7 --- /dev/null +++ b/report/assets/surgery-source.json @@ -0,0 +1,8 @@ +{ + "source": "training/gabor_half/fits/fits.npz", + "sha256": "52791beaf3219421e4ad3aa2e2591875b50dbfaa83a6309361c2df71db08de3d", + "median_selected_rank": 6144, + "median_original_flat_index": 2737, + "selected_count": 12288, + "caption": "Weight-fit measurements. The left example is rank 6,144 of the globally selected rows, normalized by original L2 norm. The right shows all layer counts; dashed line is 512, not a quota. No recognition accuracy is inferred." +} diff --git a/report/assets/surgery.pdf b/report/assets/surgery.pdf new file mode 100644 index 0000000000000000000000000000000000000000..744817b00e5482fad4d27b449426e4015668decb Binary files /dev/null and b/report/assets/surgery.pdf differ diff --git a/report/assets/surgery.png b/report/assets/surgery.png new file mode 100644 index 0000000000000000000000000000000000000000..1a44af3ca1bb7b4cf7db1ea7791b0d29f2fe07be Binary files /dev/null and b/report/assets/surgery.png differ diff --git a/report/assets/surgery.svg b/report/assets/surgery.svg new file mode 100644 index 0000000000000000000000000000000000000000..f5a1894c2908d5c6d968859372cae65df4a3bb77 --- /dev/null +++ b/report/assets/surgery.svg @@ -0,0 +1,566 @@ + + + + + + + + 2026-09-06T11:06:11.484517 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + −4 + + + + + + + + + + 0 + + + + + + + + + + 4 + + + + Temporal tap + + + + + + + + + + + + + + 0.0 + + + + + + + + + + 0.5 + + + + + + + + + + 1.0 + + + + Weight / original norm + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + A Median selected fit (6.32% RMS error) + + + + + + + + + + Original taps + + + + + + Fitted Gabor + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 1 + + + + + + + + + + 6 + + + + + + + + + + 12 + + + + + + + + + + 18 + + + + + + + + + + 24 + + + + Conformer layer + + + + + + + + + + + 0 + + + + + + + + + + 512 + + + + + + + + + + 1024 + + + + Selected / 1,024 + + + + + + + + + + + + + B Global selection varies by layer + + + + + + + + + + + + diff --git a/report/assets/unseen/unseen-benchmark-captions.md b/report/assets/unseen/unseen-benchmark-captions.md new file mode 100644 index 0000000000000000000000000000000000000000..0708893d6348c862c529ceb18377c61e222898de --- /dev/null +++ b/report/assets/unseen/unseen-benchmark-captions.md @@ -0,0 +1,2 @@ +Orukeet R15-0100 minus stock NVIDIA Parakeet TDT 0.6B v3 in word-error-rate percentage points; negative values favor Orukeet. Points use every segment of the sealed official test splits. Horizontal intervals use 10,000 paired cluster-bootstrap replicates, resampling source recording sessions for EuroSpeech, source recordings for GigaSpeechBench, and speakers for Monsoon. The groups column gives the number of resampling units; it does not assert that all groups contain different speakers. The French partition has one recording session and is plotted without an interval. English plots use whisper-normalizer 0.1.12; EuroSpeech uses the predeclared legacy-compatible normalizer. The report provides both normalizations and the exact-history-disjoint sensitivity analysis. Per-split intervals are unadjusted for multiple comparisons. +The supplementary coverage plot uses its own pre-inference registry and legacy normalization. Missing or incomplete speaker/session metadata yield an open point and no interval. Danish NST has 56 source speakers and Dutch VoxPopuli 47. Reference-alignment defects were observed in the Greek and Italian EuroSpeech material; the locked published-reference scores are retained, without treating them as headline accuracy evidence. diff --git a/report/assets/unseen/unseen-english-accents.pdf b/report/assets/unseen/unseen-english-accents.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cff068cdd47647e3739a9a1f7f4608e9c3fd65d7 Binary files /dev/null and b/report/assets/unseen/unseen-english-accents.pdf differ diff --git a/report/assets/unseen/unseen-english-accents.png b/report/assets/unseen/unseen-english-accents.png new file mode 100644 index 0000000000000000000000000000000000000000..f3c488786e6a9fc340bff02e1b303e55abbf66b5 --- /dev/null +++ b/report/assets/unseen/unseen-english-accents.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6be28a4ad2ceaa78e578f520aca33c8a6000f8731331443b4910c8f6c3a8069 +size 141909 diff --git a/report/assets/unseen/unseen-english-accents.svg b/report/assets/unseen/unseen-english-accents.svg new file mode 100644 index 0000000000000000000000000000000000000000..2e3e2821383428dcaee5daf1acf22d8cb1cb4b15 --- /dev/null +++ b/report/assets/unseen/unseen-english-accents.svg @@ -0,0 +1,350 @@ + + + + + + + + 2026-09-06T22:56:58.243499 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -0.3 + + + + + + + + + + + + + 0 + + + + + + + + + + + + + +0.3 + + + + + + + + + + + + + +0.6 + + + + + + + + + + + + + +0.9 + + + + WER difference (Orukeet − Parakeet), pp + + + + + + + Chinese accent + + + + + + Indian accent + + + + + + Japanese accent + + + + + + Filipino accent + + + + + + Scottish accent + + + + + + Singaporean accent + + + + + + Monsoon · Indian English + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 17.22 + + + 17.07 + + + 71 + + + + + + + + 7.87 + + + 7.84 + + + 182 + + + + + + + + 21.25 + + + 21.19 + + + 169 + + + + + + + + 12.01 + + + 12.05 + + + 37 + + + + + + + + 26.14 + + + 26.63 + + + 30 + + + + + + + + 13.69 + + + 14.06 + + + 45 + + + + + + + + 5.00 + + + 4.83 + + + 1,444 + + + Parakeet + + + Orukeet + + + Groups + + + + English accents and spontaneous conversation + + + GigaSpeechBench + Monsoon · standard English scoring · WER (%) · 95% paired cluster-bootstrap intervals + + + Negative differences indicate lower Orukeet WER. Open point: interval unavailable. + + + + + + + + diff --git a/report/assets/unseen/unseen-english-domains.pdf b/report/assets/unseen/unseen-english-domains.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a0be0493e20d13e8b72358a5509a75d048d55865 Binary files /dev/null and b/report/assets/unseen/unseen-english-domains.pdf differ diff --git a/report/assets/unseen/unseen-english-domains.png b/report/assets/unseen/unseen-english-domains.png new file mode 100644 index 0000000000000000000000000000000000000000..2b99b4cbbd9d907604d798aa58aef86fec221621 --- /dev/null +++ b/report/assets/unseen/unseen-english-domains.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03e2048952df3576b88a45d19add8b835e51a0a4e61d51fcc03c76dd5173272a +size 173922 diff --git a/report/assets/unseen/unseen-english-domains.svg b/report/assets/unseen/unseen-english-domains.svg new file mode 100644 index 0000000000000000000000000000000000000000..880e8cd590c0ede4d9f62e45a5f7c61119256bd1 --- /dev/null +++ b/report/assets/unseen/unseen-english-domains.svg @@ -0,0 +1,475 @@ + + + + + + + + 2026-09-06T22:56:58.933189 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -0.3 + + + + + + + + + + + + + 0 + + + + + + + + + + + + + +0.3 + + + + + + + + + + + + + +0.6 + + + + + + + + + + + + + +0.9 + + + + WER difference (Orukeet − Parakeet), pp + + + + + + + Agriculture + + + + + + Artificial intelligence + + + + + + Art + + + + + + Biology + + + + + + Economics + + + + + + Engineering + + + + + + Entertainment + + + + + + Finance + + + + + + Humanities + + + + + + Law + + + + + + Medicine + + + + + + Military + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 6.47 + + + 6.47 + + + 40 + + + + + + + + 9.60 + + + 9.99 + + + 15 + + + + + + + + 6.11 + + + 6.12 + + + 23 + + + + + + + + 6.52 + + + 6.63 + + + 32 + + + + + + + + 9.08 + + + 9.10 + + + 27 + + + + + + + + 5.66 + + + 6.16 + + + 31 + + + + + + + + 10.00 + + + 9.92 + + + 28 + + + + + + + + 6.91 + + + 6.96 + + + 38 + + + + + + + + 6.94 + + + 6.87 + + + 32 + + + + + + + + 10.02 + + + 10.14 + + + 19 + + + + + + + + 5.44 + + + 5.51 + + + 62 + + + + + + + + 5.99 + + + 6.10 + + + 40 + + + Parakeet + + + Orukeet + + + Groups + + + + English speech across twelve specialist domains + + + GigaSpeechBench · standard English scoring · WER (%) · 95% paired cluster-bootstrap intervals + + + Negative differences indicate lower Orukeet WER. Open point: interval unavailable. + + + + + + + + diff --git a/report/assets/unseen/unseen-eurospeech.pdf b/report/assets/unseen/unseen-eurospeech.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a69fd24d2a81f7f01c224c91e6d7318074438afb Binary files /dev/null and b/report/assets/unseen/unseen-eurospeech.pdf differ diff --git a/report/assets/unseen/unseen-eurospeech.png b/report/assets/unseen/unseen-eurospeech.png new file mode 100644 index 0000000000000000000000000000000000000000..78ef0ec342e83c76541145e746dfad180a8acfe3 --- /dev/null +++ b/report/assets/unseen/unseen-eurospeech.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5efb7b9dcb242a5301a5edffde5d6dace4dd03a1319e49c9cd817cc4dd30854 +size 206024 diff --git a/report/assets/unseen/unseen-eurospeech.svg b/report/assets/unseen/unseen-eurospeech.svg new file mode 100644 index 0000000000000000000000000000000000000000..7313524d36daff6b4e7b95c371f0f1b4ad5deb5b --- /dev/null +++ b/report/assets/unseen/unseen-eurospeech.svg @@ -0,0 +1,583 @@ + + + + + + + + 2026-09-06T22:56:57.617107 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -16.0 + + + + + + + + + + + + + -12.0 + + + + + + + + + + + + + -8.0 + + + + + + + + + + + + + -4.0 + + + + + + + + + + + + + 0 + + + + WER difference (Orukeet − Parakeet), pp + + + + + + + Bulgarian + + + + + + German + + + + + + Greek + + + + + + English + + + + + + Estonian + + + + + + Finnish + + + + + + French + + + + + + Croatian + + + + + + Italian + + + + + + Lithuanian + + + + + + Latvian + + + + + + Maltese + + + + + + Portuguese + + + + + + Slovak + + + + + + Slovenian + + + + + + Ukrainian + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 15.11 + + + 14.76 + + + 5 + + + + + + + + 15.75 + + + 13.81 + + + 2 + + + + + + + + 101.45 + + + 100.71 + + + 2 + + + + + + + + 29.42 + + + 27.21 + + + 12 + + + + + + + + 38.68 + + + 30.17 + + + 2 + + + + + + + + 17.59 + + + 16.15 + + + 6 + + + + + + + + + + + 19.22 + + + 14.74 + + + 1 + + + + + + + + 13.43 + + + 13.07 + + + 34 + + + + + + + + 64.73 + + + 64.84 + + + 5 + + + + + + + + 38.58 + + + 32.85 + + + 7 + + + + + + + + 57.61 + + + 42.64 + + + 6 + + + + + + + + 39.98 + + + 38.31 + + + 9 + + + + + + + + 22.08 + + + 22.05 + + + 6 + + + + + + + + 18.05 + + + 15.34 + + + 11 + + + + + + + + 52.86 + + + 50.19 + + + 3 + + + + + + + + 15.58 + + + 14.74 + + + 2 + + + Parakeet + + + Orukeet + + + Groups + + + + EuroSpeech published-reference scores + + + Legacy scoring · reference offsets in Greek and Italian · WER (%) · 95% paired cluster-bootstrap intervals + + + Negative differences indicate lower Orukeet WER. Open point: interval unavailable. + + + + + + + + diff --git a/report/assets/unseen/unseen-figure-receipt.json b/report/assets/unseen/unseen-figure-receipt.json new file mode 100644 index 0000000000000000000000000000000000000000..dc1de329149dbb2ae55d6ae2b07a88b21e1433c2 --- /dev/null +++ b/report/assets/unseen/unseen-figure-receipt.json @@ -0,0 +1,35 @@ +{ + "comparison_sha256": "cc024031468c602fb6c326943a2ea1b6dffa66440fcb4d04559d12bb83f05ed0", + "coverage_comparison_sha256": "298599ce854a9607a188f07becf9c75403a411542c6e960ded9c5d4b25ec8401", + "script_sha256": "d4592c29243257dd133fbfa59ba0a2d7d69fe8ba888f974e517386a3a6337866", + "matplotlib": "3.11.1", + "files": { + "unseen-language-coverage.png": "aa66de577f8ebb9997568eda3fccdd9c80f93a0d02c13518abc7405830b1bba7", + "unseen-language-coverage.pdf": "26c89cabebbc8558a71d761eb18af8b178cbbfff5f63ce415f6fcd9589ca6f55", + "unseen-english-accents.png": "c6be28a4ad2ceaa78e578f520aca33c8a6000f8731331443b4910c8f6c3a8069", + "unseen-english-accents.pdf": "ae6fd18d34f9348020558c340cd2c3666cc1addcc32a939ba936615f567b41e3", + "unseen-english-accents.svg": "3cacedbe24885630d0eaad88403cf80f890dea42d462bf0d89e889cb907b8f34", + "unseen-benchmark-captions.md": "59459974d24ee611c2f4c543d54e17dbea6e4e5bcb869f023943139ef1288565", + "unseen-language-coverage.svg": "03cd68fc07a2271bf8e1d3d4132eee814a85174137e7fa74a52af7a6984146aa", + "unseen-eurospeech.png": "a5efb7b9dcb242a5301a5edffde5d6dace4dd03a1319e49c9cd817cc4dd30854", + "unseen-eurospeech.pdf": "14be92cd1d4ea09546d60482c3c329c212321b0f551e1e1ab510ab271ee5782a", + "unseen-english-domains.svg": "81760f0eb64d9e313914456860cddcecb3740f9cf81f48cf8eec71862338f1dc", + "unseen-english-domains.png": "03e2048952df3576b88a45d19add8b835e51a0a4e61d51fcc03c76dd5173272a", + "unseen-english-domains.pdf": "11d7260cf143d636651cc89d4255f929f67b0ad26c5589450602a3543f0ffb58", + "unseen-eurospeech.svg": "2957f74c5f09b3c8b5cdb6b3d6b08082e0b774a7f2fc93cf6a93e2539f0af64c" + }, + "visual_review": "passed", + "reviewed_utc": "2026-09-07T05:57:41.859084+00:00", + "visual_review_method": "Rendered all four PDFs with Poppler at 96 dpi (final physical size) and 200 dpi; visually inspected all eight images for text clipping, overlap, row alignment, interval/point placement and legibility.", + "rendered_review_sha256": { + "unseen-english-accents-200.png": "c480c1a5458d5694a86e7b258d37c7725198d2ce03ef8fb3edad133974bcb5d0", + "unseen-english-accents-96.png": "edaad9ebac4faa4df30ad40309fe7ef1be745c04b47b9e9f97e545977f3ac987", + "unseen-english-domains-200.png": "199142e5f7c16c3035ddc50119d1261fe654f728c89aaf2126c53c0d62bb0733", + "unseen-english-domains-96.png": "7d58112b34d8469dd960118bf687e743820290334c627b542294a82073514671", + "unseen-eurospeech-200.png": "47e4ce61fb8a0af0a207496fc29d6df435a8a9002c05e6ee5a8d195a4be788ce", + "unseen-eurospeech-96.png": "82e5d9fcec75219a71b2eca68122d9228cc791bd45c1eac021b6a8412bcea604", + "unseen-language-coverage-200.png": "f15a5a702b5b56afdc670fde46d655047599d703e0db2e01a97247889afd9ced", + "unseen-language-coverage-96.png": "fa38a68c491c962dc44d30397c6479e977ce78346982a1feaa382b3094a1086e" + }, + "visual_review_findings": "All four figures readable at final size; no clipping or overlap. Open markers and group counts identify unavailable intervals. English plots share axis limits; EuroSpeech plot names the source-alignment defect." +} diff --git a/report/assets/unseen/unseen-language-coverage.pdf b/report/assets/unseen/unseen-language-coverage.pdf new file mode 100644 index 0000000000000000000000000000000000000000..794198d2e65106cf9fa030202f74d7130f6f7760 Binary files /dev/null and b/report/assets/unseen/unseen-language-coverage.pdf differ diff --git a/report/assets/unseen/unseen-language-coverage.png b/report/assets/unseen/unseen-language-coverage.png new file mode 100644 index 0000000000000000000000000000000000000000..94ef3bdf93789af528b637d772fbbb0457953420 --- /dev/null +++ b/report/assets/unseen/unseen-language-coverage.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa66de577f8ebb9997568eda3fccdd9c80f93a0d02c13518abc7405830b1bba7 +size 176291 diff --git a/report/assets/unseen/unseen-language-coverage.svg b/report/assets/unseen/unseen-language-coverage.svg new file mode 100644 index 0000000000000000000000000000000000000000..82d9b6d8ade5f37f8237dde2cef9dcfa6c4626d6 --- /dev/null +++ b/report/assets/unseen/unseen-language-coverage.svg @@ -0,0 +1,383 @@ + + + + + + + + 2026-09-06T22:56:59.693381 + image/svg+xml + + + Matplotlib v3.11.1, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -0.9 + + + + + + + + + + + + + -0.6 + + + + + + + + + + + + + -0.3 + + + + + + + + + + + + + 0 + + + + WER difference (Orukeet − Parakeet), pp + + + + + + + Czech · VoxPopuli + + + + + + Danish · NST + + + + + + Dutch · VoxPopuli + + + + + + Hungarian · VoxPopuli + + + + + + Polish · VoxPopuli + + + + + + Romanian · VoxPopuli + + + + + + Russian · Golos crowd + + + + + + Russian · Golos far field + + + + + + Spanish · VoxPopuli + + + + + + Swedish · NST mirror + + + + + + + + + + + + + + + + + + + + + + 9.22 + + + 8.97 + + + — + + + + + + + + + + + 30.54 + + + 29.79 + + + 56 + + + + + + + + 11.25 + + + 11.12 + + + 47 + + + + + + + + 15.72 + + + 15.33 + + + — + + + + + + + + 7.54 + + + 7.32 + + + — + + + + + + + + 12.24 + + + 11.83 + + + — + + + + + + + + 3.42 + + + 3.48 + + + — + + + + + + + + 7.25 + + + 7.17 + + + — + + + + + + + + 5.50 + + + 5.45 + + + — + + + + + + + + 23.19 + + + 22.68 + + + — + + + Parakeet + + + Orukeet + + + Groups + + + + Supplementary coverage of nine languages + + + Supplementary test/validation · legacy scoring · WER (%) · 95% paired cluster-bootstrap intervals + + + Negative differences indicate lower Orukeet WER. Open point: interval unavailable. + + + + + + + + diff --git a/report/benchmark-tables.tex b/report/benchmark-tables.tex new file mode 100644 index 0000000000000000000000000000000000000000..9b757f87d8f08567f8fc5e1fbd7ba48a4612c920 --- /dev/null +++ b/report/benchmark-tables.tex @@ -0,0 +1,129 @@ +\clearpage +\appendix +\section{Complete benchmark scores} +\begin{table}[!ht] +\centering +\caption{All 47 complete partitions, totaling 327,888 clips. WER and CER are percentages. English uses standard English normalization; other languages use the multilingual normalizer. Exact membership and reference diagnostics accompany the score files.} +\label{tab:full-benchmark} +\footnotesize +\setlength{\tabcolsep}{3.4pt} +\renewcommand{\arraystretch}{1.08} +\begin{tabular}{lrrrrr} +\toprule +Split & Clips & \multicolumn{2}{c}{Parakeet} & \multicolumn{2}{c}{R15-0100}\\ + & & WER & CER & WER & CER\\ +\midrule +\texttt{eurospeech\_bg} & 6,892 & 15.11 & 7.25 & 14.76 & 7.48\\ +\texttt{eurospeech\_de} & 4,872 & 15.75 & 10.50 & 13.81 & 9.50\\ +\texttt{eurospeech\_el} & 6,730 & 101.45 & 78.15 & 100.71 & 79.19\\ +\texttt{eurospeech\_en} & 9,268 & 26.07 & 18.38 & 25.07 & 17.73\\ +\texttt{eurospeech\_et} & 3,554 & 38.68 & 15.04 & 30.17 & 11.61\\ +\texttt{eurospeech\_fi} & 5,422 & 17.59 & 7.21 & 16.15 & 6.77\\ +\texttt{eurospeech\_fr} & 744 & 19.22 & 11.49 & 14.74 & 9.31\\ +\texttt{eurospeech\_hr} & 15,638 & 13.43 & 8.89 & 13.07 & 8.68\\ +\texttt{eurospeech\_it} & 8,714 & 64.73 & 48.33 & 64.84 & 48.43\\ +\texttt{eurospeech\_lt} & 7,319 & 38.58 & 16.32 & 32.85 & 14.52\\ +\texttt{eurospeech\_lv} & 3,343 & 57.61 & 25.55 & 42.64 & 16.16\\ +\texttt{eurospeech\_mt} & 3,446 & 39.98 & 18.02 & 38.31 & 17.31\\ +\texttt{eurospeech\_pt} & 7,501 & 22.08 & 16.45 & 22.05 & 16.43\\ +\texttt{eurospeech\_sk} & 6,915 & 18.05 & 7.98 & 15.34 & 7.22\\ +\texttt{eurospeech\_sl} & 3,585 & 52.86 & 17.61 & 50.19 & 16.09\\ +\texttt{eurospeech\_uk} & 3,239 & 15.58 & 9.11 & 14.74 & 8.63\\ +\texttt{gigaspeechbench\_agr\_en} & 6,665 & 6.47 & 3.76 & 6.47 & 3.77\\ +\texttt{gigaspeechbench\_ait\_en} & 5,468 & 9.60 & 4.60 & 9.99 & 4.89\\ +\texttt{gigaspeechbench\_art\_en} & 5,712 & 6.11 & 2.97 & 6.12 & 3.03\\ +\texttt{gigaspeechbench\_bio\_en} & 5,297 & 6.52 & 1.91 & 6.63 & 1.92\\ +\texttt{gigaspeechbench\_chn\_en} & 6,308 & 17.22 & 10.12 & 17.07 & 10.08\\ +\texttt{gigaspeechbench\_ecm\_en} & 5,659 & 9.08 & 4.79 & 9.10 & 4.80\\ +\texttt{gigaspeechbench\_eng\_en} & 6,648 & 5.66 & 2.34 & 6.16 & 2.46\\ +\texttt{gigaspeechbench\_ent\_en} & 8,583 & 10.00 & 6.43 & 9.92 & 6.41\\ +\texttt{gigaspeechbench\_fin\_en} & 6,037 & 6.91 & 3.36 & 6.96 & 3.41\\ +\texttt{gigaspeechbench\_hum\_en} & 4,971 & 6.94 & 3.66 & 6.87 & 3.65\\ +\texttt{gigaspeechbench\_ind\_en} & 5,503 & 7.87 & 3.08 & 7.84 & 3.05\\ +\texttt{gigaspeechbench\_jpn\_en} & 9,310 & 21.25 & 12.66 & 21.19 & 12.63\\ +\texttt{gigaspeechbench\_law\_en} & 7,273 & 10.02 & 5.52 & 10.14 & 5.62\\ +\texttt{gigaspeechbench\_med\_en} & 5,168 & 5.44 & 1.96 & 5.51 & 1.99\\ +\texttt{gigaspeechbench\_mil\_en} & 5,224 & 5.99 & 1.75 & 6.10 & 1.78\\ +\texttt{gigaspeechbench\_phl\_en} & 8,637 & 12.01 & 7.43 & 12.05 & 7.49\\ +\texttt{gigaspeechbench\_sct\_en} & 12,829 & 26.14 & 17.05 & 26.63 & 17.48\\ +\texttt{gigaspeechbench\_sgp\_en} & 9,480 & 13.69 & 8.39 & 14.06 & 8.63\\ +\texttt{golos\_crowd\_ru} & 9,896 & 3.42 & 0.77 & 3.48 & 0.78\\ +\texttt{golos\_farfield\_ru} & 1,915 & 7.25 & 2.16 & 7.17 & 2.06\\ +\texttt{lesbos\_el} & 230 & 96.11 & 71.30 & 93.63 & 72.36\\ +\texttt{monsoon\_en\_in} & 2,102 & 5.00 & 2.50 & 4.83 & 2.42\\ +\texttt{nst\_da\_da} & 54,747 & 30.54 & 16.70 & 29.79 & 16.74\\ +\texttt{nst\_sv\_sv} & 27,638 & 23.19 & 13.71 & 22.68 & 13.90\\ +\texttt{voxpopuli\_cs} & 1,103 & 9.22 & 4.71 & 8.97 & 4.57\\ +\texttt{voxpopuli\_es} & 1,631 & 5.50 & 3.59 & 5.45 & 3.54\\ +\texttt{voxpopuli\_hu} & 1,076 & 15.72 & 5.65 & 15.33 & 5.55\\ +\texttt{voxpopuli\_it} & 1,257 & 11.98 & 8.74 & 11.97 & 8.68\\ +\texttt{voxpopuli\_nl} & 1,230 & 11.25 & 6.32 & 11.12 & 6.28\\ +\texttt{voxpopuli\_pl} & 1,691 & 7.54 & 4.20 & 7.32 & 3.99\\ +\texttt{voxpopuli\_ro} & 1,418 & 12.24 & 4.87 & 11.83 & 4.82\\ +\bottomrule +\end{tabular} +\end{table} +\clearpage +\section{Complete matched-comparison scores} +\begin{table}[!ht] +\centering +\caption{All 47 splits in the 12,006-clip matched comparison. Each split has 256 fixed clips, except Lesbos (230). WER and CER are percentages. Section~\ref{sec:benchmarks} defines the model identities, sample, and reference treatment.} +\label{tab:sample-benchmark} +\footnotesize +\setlength{\tabcolsep}{3.4pt} +\renewcommand{\arraystretch}{1.08} +\begin{tabular}{lrrrrrrr} +\toprule +Split & Clips & \multicolumn{2}{c}{Parakeet} & \multicolumn{2}{c}{R15-0100} & \multicolumn{2}{c}{FT-4035}\\ + & & WER & CER & WER & CER & WER & CER\\ +\midrule +\texttt{eurospeech\_bg} & 256 & 14.76 & 6.93 & 14.65 & 7.28 & 13.78 & 6.58\\ +\texttt{eurospeech\_de} & 256 & 15.06 & 9.94 & 13.48 & 9.24 & 12.36 & 8.44\\ +\texttt{eurospeech\_el} & 256 & 26.07 & 8.48 & 18.70 & 7.59 & 18.32 & 7.68\\ +\texttt{eurospeech\_en} & 256 & 25.70 & 17.90 & 24.93 & 17.38 & 25.23 & 17.57\\ +\texttt{eurospeech\_et} & 256 & 38.28 & 14.68 & 29.90 & 11.52 & 30.22 & 12.69\\ +\texttt{eurospeech\_fi} & 256 & 18.67 & 7.85 & 17.52 & 7.49 & 17.55 & 7.72\\ +\texttt{eurospeech\_fr} & 256 & 20.00 & 11.85 & 15.14 & 9.59 & 15.21 & 9.76\\ +\texttt{eurospeech\_hr} & 256 & 13.26 & 8.69 & 13.04 & 8.45 & 13.29 & 8.65\\ +\texttt{eurospeech\_it} & 256 & 10.64 & 6.33 & 12.28 & 8.01 & 12.34 & 8.22\\ +\texttt{eurospeech\_lt} & 256 & 39.43 & 16.13 & 33.77 & 14.69 & 35.78 & 16.67\\ +\texttt{eurospeech\_lv} & 256 & 57.81 & 26.65 & 43.31 & 16.56 & 46.42 & 18.93\\ +\texttt{eurospeech\_mt} & 256 & 40.74 & 19.23 & 38.72 & 17.99 & 41.41 & 19.68\\ +\texttt{eurospeech\_pt} & 256 & 23.31 & 17.53 & 23.44 & 17.77 & 24.25 & 18.62\\ +\texttt{eurospeech\_sk} & 256 & 18.61 & 8.23 & 15.18 & 7.09 & 17.48 & 9.37\\ +\texttt{eurospeech\_sl} & 256 & 51.65 & 17.23 & 48.74 & 15.59 & 54.17 & 17.97\\ +\texttt{eurospeech\_uk} & 256 & 15.22 & 8.84 & 14.06 & 7.98 & 17.73 & 11.33\\ +\texttt{gigaspeechbench\_agr\_en} & 256 & 6.79 & 3.98 & 6.86 & 3.99 & 6.30 & 3.44\\ +\texttt{gigaspeechbench\_ait\_en} & 256 & 10.54 & 5.01 & 10.95 & 5.42 & 9.72 & 4.61\\ +\texttt{gigaspeechbench\_art\_en} & 256 & 6.07 & 3.10 & 6.46 & 3.46 & 5.35 & 2.62\\ +\texttt{gigaspeechbench\_bio\_en} & 256 & 6.74 & 1.95 & 6.86 & 2.01 & 6.15 & 1.81\\ +\texttt{gigaspeechbench\_chn\_en} & 256 & 15.38 & 9.25 & 15.91 & 9.74 & 14.52 & 8.61\\ +\texttt{gigaspeechbench\_ecm\_en} & 256 & 8.30 & 4.39 & 8.39 & 4.36 & 7.80 & 4.07\\ +\texttt{gigaspeechbench\_eng\_en} & 256 & 6.15 & 2.43 & 6.63 & 2.47 & 4.81 & 2.01\\ +\texttt{gigaspeechbench\_ent\_en} & 256 & 10.78 & 6.85 & 10.22 & 6.43 & 9.15 & 5.73\\ +\texttt{gigaspeechbench\_fin\_en} & 256 & 7.49 & 3.57 & 7.15 & 3.31 & 6.89 & 3.19\\ +\texttt{gigaspeechbench\_hum\_en} & 256 & 8.30 & 4.76 & 7.86 & 4.41 & 7.88 & 4.31\\ +\texttt{gigaspeechbench\_ind\_en} & 256 & 8.54 & 3.41 & 8.69 & 3.52 & 7.75 & 2.74\\ +\texttt{gigaspeechbench\_jpn\_en} & 256 & 19.91 & 12.05 & 19.53 & 12.17 & 18.09 & 11.08\\ +\texttt{gigaspeechbench\_law\_en} & 256 & 11.13 & 5.57 & 11.05 & 5.52 & 10.60 & 5.35\\ +\texttt{gigaspeechbench\_med\_en} & 256 & 5.16 & 1.91 & 5.12 & 1.84 & 4.97 & 1.72\\ +\texttt{gigaspeechbench\_mil\_en} & 256 & 5.99 & 1.76 & 6.07 & 1.89 & 5.51 & 1.58\\ +\texttt{gigaspeechbench\_phl\_en} & 256 & 14.00 & 8.37 & 13.78 & 8.35 & 13.10 & 7.73\\ +\texttt{gigaspeechbench\_sct\_en} & 256 & 22.66 & 14.66 & 24.09 & 16.29 & 20.73 & 13.43\\ +\texttt{gigaspeechbench\_sgp\_en} & 256 & 14.94 & 9.48 & 15.03 & 9.56 & 13.99 & 8.61\\ +\texttt{golos\_crowd\_ru} & 256 & 3.14 & 0.66 & 2.99 & 0.60 & 3.76 & 0.75\\ +\texttt{golos\_farfield\_ru} & 256 & 8.54 & 2.59 & 8.72 & 2.56 & 10.41 & 3.53\\ +\texttt{lesbos\_el} & 230 & 96.11 & 71.35 & 93.63 & 72.41 & 93.63 & 73.10\\ +\texttt{monsoon\_en\_in} & 256 & 4.95 & 2.46 & 4.82 & 2.39 & 4.66 & 2.36\\ +\texttt{nst\_da\_da} & 256 & 33.41 & 19.25 & 33.14 & 19.35 & 12.85 & 4.64\\ +\texttt{nst\_sv\_sv} & 256 & 21.11 & 12.45 & 19.96 & 11.94 & 13.89 & 3.66\\ +\texttt{voxpopuli\_cs} & 256 & 8.15 & 3.93 & 7.99 & 4.16 & 8.30 & 4.11\\ +\texttt{voxpopuli\_es} & 256 & 6.12 & 4.25 & 5.88 & 4.01 & 6.29 & 4.33\\ +\texttt{voxpopuli\_hu} & 256 & 13.81 & 4.11 & 13.19 & 3.91 & 13.35 & 3.92\\ +\texttt{voxpopuli\_it} & 256 & 11.58 & 8.71 & 11.79 & 8.90 & 11.97 & 9.62\\ +\texttt{voxpopuli\_nl} & 256 & 10.42 & 5.67 & 10.36 & 5.58 & 10.62 & 5.80\\ +\texttt{voxpopuli\_pl} & 256 & 6.52 & 3.42 & 6.50 & 3.51 & 6.41 & 3.49\\ +\texttt{voxpopuli\_ro} & 256 & 11.99 & 4.39 & 11.88 & 4.25 & 11.68 & 4.26\\ +\bottomrule +\end{tabular} +\end{table} diff --git a/report/benchmark-values.tex b/report/benchmark-values.tex new file mode 100644 index 0000000000000000000000000000000000000000..db7f2992201340d7039fde7678d8cbf1c8732266 --- /dev/null +++ b/report/benchmark-values.tex @@ -0,0 +1,27 @@ +% Generated by scripts/build_benchmark_materials.py. +\newcommand{\SampleAllRelativeReduction}{8.02} +\newcommand{\SampleEnglishRelativeReduction}{6.61} +\newcommand{\FullEnglishRows}{136,142} +\newcommand{\FullEnglishBaseWER}{12.46} +\newcommand{\FullEnglishBaseCER}{7.47} +\newcommand{\FullEnglishRWER}{12.41} +\newcommand{\FullEnglishRCER}{7.42} +\newcommand{\FullAllRows}{327,888} +\newcommand{\FullAllBaseWER}{23.72} +\newcommand{\FullAllBaseCER}{14.75} +\newcommand{\FullAllRWER}{22.76} +\newcommand{\FullAllRCER}{14.37} +\newcommand{\SampleEnglishRows}{5,120} +\newcommand{\SampleEnglishBaseWER}{10.84} +\newcommand{\SampleEnglishBaseCER}{6.14} +\newcommand{\SampleEnglishRWER}{10.82} +\newcommand{\SampleEnglishRCER}{6.15} +\newcommand{\SampleEnglishFTWER}{10.13} +\newcommand{\SampleEnglishFTCER}{5.70} +\newcommand{\SampleAllRows}{12,006} +\newcommand{\SampleAllBaseWER}{17.97} +\newcommand{\SampleAllBaseCER}{9.22} +\newcommand{\SampleAllRWER}{16.48} +\newcommand{\SampleAllRCER}{8.56} +\newcommand{\SampleAllFTWER}{16.52} +\newcommand{\SampleAllFTCER}{8.72} diff --git a/report/build-receipt.json b/report/build-receipt.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b21100a34346bfe98c9c5751fa93885068f093 --- /dev/null +++ b/report/build-receipt.json @@ -0,0 +1,87 @@ +{ + "status": "compiled-and-visually-reviewed", + "pages": 5, + "expected_pages": 5, + "pdf": "output/pdf/orukeet-technical-report.pdf", + "sha256": "401d162a87d4454b0f16998bf3501b9c84da909a7b5be353705528477d2eef62", + "compiler": "pdfTeX 3.141592653-2.6-1.40.28 (TeX Live 2025)", + "format": "Short technical report in the official NeurIPS 2026 preprint style", + "report_model": { + "name": "Orukeet", + "version": "librispeech-native-format-r3", + "sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "parent_sha256": "0ccfefcd1894871cb0850bd3c464adf5397752840de2a76d1d2d075c4141a945", + "hf": { + "repo_id": "oruk/orukeet", + "revision": "555136b50265a132d4cea0d35560c26fc4f657ab", + "path": "orukeet-v0.1.0.nemo" + }, + "freeze_audit": "evidence/librispeech-ft-20260908/r3/export-audit.json", + "archive_receipt": "evidence/librispeech-ft-20260908/r3/private-archive.json", + "evaluations": [ + "evidence/standard-asr-r3-20260908/comparison.json", + "evidence/domains-r3-20260908/comparison.json" + ], + "publication_authorized": true, + "scope": "The manuscript, canonical NeMo source, Q8 and F16 exports, and OpenWhispr integration all share this exact r3 source checkpoint. Export hashes are pinned in src/orukeet/artifacts.json.", + "archive_hf": { + "repo_id": "oruk/orukeet-internal", + "revision": "294d5bbc2f4ba79cbcf07776c52f580ff053a60d", + "path": "experiments/librispeech-ft-20260908/r3/orukeet-targeted-ft.nemo" + } + }, + "inputs_sha256": { + "evaluation/standard_asr/audit_predictions.py": "80a8c7911819129e39911cd4f78df326485a659a847522779cf40b3b2365eb3a", + "evaluation/standard_asr/build_current_report.py": "624b9781c023ea1cd4f1ad9a3c18f99e806027a228c5f24da9f466679578dd8f", + "evaluation/standard_asr/rescore.py": "47e77bad2d65f1eb6d7a29a156ba747807b4017d4fcedf8d3d2011c5a4e0ab82", + "evaluation/standard_asr/run.py": "1c8951a18eba92ae8fe12c537149c961ae0c0e6f748216e27488305b5a2806c2", + "evaluation/standard_asr/scoring.py": "3b2bd425a79a94b9afa294ba55efc6b9fb5943d7021ce27394dadf6c58082bd0", + "evaluation/standard_asr/vendor/LICENSE": "5ee13882fce0975f0ad3c3d5c2042af4896c2a669be2d6df5b3d256811267c85", + "evaluation/standard_asr/vendor/__init__.py": "8c39e42af3204c41c564c970e24525da160663836b54acc1b629e4342f862334", + "evaluation/standard_asr/vendor/english_abbreviations.py": "52997cc963e0bd6568d15554ad8d5fb0759f20f361483632a22f90d2d0e07db1", + "evaluation/standard_asr/vendor/multilingual.py": "4c8db6c4171c8c2f5567d2ed625a3a1f0d45156af45ca42d9555c0c9b61a29d5", + "evaluation/standard_asr/vendor/normalizer.py": "490b56393484ef386b486f3679cc2261264af280b3512154431e0bdcf2778295", + "evaluation/standard_asr/vendor/provenance.json": "f6fb39bc16cb4b52ac62882baab0f1925d7405ee8a98f0ae32419edf9275aa12", + "evidence/domains-r3-20260908/comparison.json": "617a2119027b9e706907fdeb8a5819babb55f1959456d677e744d2b57ddfd2be", + "evidence/domains-r3-20260908/hypotheses-audit.json": "66c748f9ea625cf38e9bf672628e95300a5919abbaf47ebfd7c4800f5b1f3116", + "evidence/domains-r3-20260908/numeric-evidence.jsonl.gz": "298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df", + "evidence/librispeech-ft-20260908/r3/export-audit.json": "3d9f115d39758976836291bf2a99f96b0ea71c412f320e3d954b6ba66cfdf59d", + "evidence/librispeech-ft-20260908/r3/private-archive.json": "dd634067bcb59b0a705ed051850161c9257dc49ddfdfc011a506dd3fafc805ef", + "evidence/standard-asr-r3-20260908/comparison.json": "9b2702bfe9c7b107db1627df3cb71d24d25d4eb8da0ee8683b0bbf9f994ca7da", + "evidence/standard-asr-r3-20260908/hypotheses-audit.json": "a60177ac2ac1064b481cc85424d5ee989cca43112191ddae30d60b7f56ae57d3", + "evidence/standard-asr-r3-20260908/numeric-evidence.jsonl.gz": "3d1467e56efd6769da1f8cb74eef0c3eac7bf93fa8200c39538963a9d4b080c1", + "report/assets/affiliations/SOURCES.md": "0d77aeb9033fc456dc3d1a4d3b5cbf7bd67a3f43feaf4b97c5ed5fc7c44b579e", + "report/assets/affiliations/cambridge.pdf": "535e2f3382a9e63ef27fcd561e54cfc87d74f494c3a7f6dd072c2af2bfe7a738", + "report/assets/affiliations/cambridge.svg": "27381d019c0a54c53b7b746750dc148f35dca2a6fe96b39dff5f896b0e745902", + "report/assets/affiliations/hoid.pdf": "439f46cb03c498cbc0bb797d39ba31ce80d57742d3b699ae0a6ca5cf2fa1311a", + "report/assets/affiliations/hoid.svg": "49b5fc2265f1294abf6bb65f054078b640b9fbaed9780e2817ab6a103cd2eb72", + "report/assets/affiliations/openwhispr.pdf": "78e575011be731f9d07f665f1440a4582df27d577bbdb3805d2c7a17b40cacd7", + "report/assets/affiliations/openwhispr.svg": "29d50f34b1ffe4b11665aaeefdaf63a22796c53c38234cd9d6f06c5153c9e13a", + "report/assets/affiliations/stanford.png": "09434daceea1a4ab62da5db3cd9e81bbe1a7e034224f504d63fd3f12e51af74b", + "report/assets/kernel-fits.pdf": "b952a1d8e84032cd20cc452dfbea40c6d2c34cea7de4257b7a375ccac73b9e01", + "report/assets/model-figures-source.json": "cf07124f8e46464eca5036cd6be96e4d1ced9d3f936e514e9382518181de42b8", + "report/assets/oruk-lockup.png": "932df9164b237a1a31d1e636897d1766a15b8bac738acf81ee93de8ea2d6c18c", + "report/assets/selection-profile.pdf": "0a2e8647a8a1a895e7a1e6d9f15779eac14c87f56e30eeeb85bc05ce720ba32f", + "report/current-benchmark-validation.json": "cda67fa90a2d75c815ac48223224ef60205ec3ee841bad4b73dce1b273ebe8d9", + "report/current-benchmark-values.tex": "5eda781ff8966fc183d3a9c3a92e2dfe1840e8eddad292b891ef1976e63da9c7", + "report/current-domains-table.tex": "ca92fc7f970f7608e69628ca2240e51b514118e24071376634884643a530fd26", + "report/current-standard-table.tex": "46aeaa3d4bf92d81e1acfc2c54b0e46cc9be1a4639ac9d16597c97982206ba2b", + "report/model.json": "a0d6dece22d8eefbc35433a04376dd7415444a3c23ac7a7cb9ed179a8ec0d7c4", + "report/neurips_2026.sty": "c3fc2894e83d2517ca18b66741d6c595986d97957dc08ec08bb2125a7ec4555a", + "report/paper.tex": "0dcc45ad9c401a9fd4e4ab389890b7bc51d4c0f8c21d4b18a7628d16b124c6ad", + "report/references.bib": "e3229e1faad05e4975b231660184d126764f0c7a33bae9c31248d91859347d55", + "scripts/build_neurips_report.py": "d5735c4f770eb334ffbf80a0ec439f0c4fc5a6922d45c90258092c0b4046584e", + "training/gabor_half/fits/fits.npz": "52791beaf3219421e4ad3aa2e2591875b50dbfaa83a6309361c2df71db08de3d" + }, + "publication_authorized": true, + "visual_review": { + "status": "passed", + "pages": 5, + "checks": [ + "All five pages rendered and inspected", + "Current Oruk Signal affiliation logo", + "Two figures, readable paired benchmark tables, no clipping", + "Public checkpoint link and nine authors" + ] + } +} diff --git a/report/current-benchmark-validation.json b/report/current-benchmark-validation.json new file mode 100644 index 0000000000000000000000000000000000000000..296b7f711b431be841e28874dd2ad647c0c942ea --- /dev/null +++ b/report/current-benchmark-validation.json @@ -0,0 +1,285 @@ +{ + "status": "passed", + "publication_authorized": false, + "models": { + "parakeet": "3cbdc85877e668ca7b82d0d56770eb1fac76691f55d6b97545e8d61ca588d10d", + "orukeet": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56" + }, + "summaries": { + "fleurs": { + "splits": [ + "fleurs_bg", + "fleurs_cs", + "fleurs_da", + "fleurs_de", + "fleurs_el", + "fleurs_en", + "fleurs_es", + "fleurs_et", + "fleurs_fi", + "fleurs_fr", + "fleurs_hr", + "fleurs_hu", + "fleurs_it", + "fleurs_lt", + "fleurs_lv", + "fleurs_mt", + "fleurs_nl", + "fleurs_pl", + "fleurs_pt", + "fleurs_ro", + "fleurs_ru", + "fleurs_sk", + "fleurs_sl", + "fleurs_sv", + "fleurs_uk" + ], + "rows": 20146, + "models": { + "parakeet": { + "errors": 46442, + "words": 421870, + "substitutions": 38136, + "deletions": 5461, + "insertions": 2845, + "chars": 2757813, + "char_errors": 98352, + "utterance_error": 14637, + "wer": 11.008604546424253, + "cer": 3.5663041692819637 + }, + "orukeet": { + "errors": 41521, + "words": 421715, + "substitutions": 34613, + "deletions": 3955, + "insertions": 2953, + "chars": 2757813, + "char_errors": 86408, + "utterance_error": 13762, + "wer": 9.845748906251853, + "cer": 3.133207363951073 + } + }, + "wins": 23, + "ties": 0, + "losses": 2, + "relative_wer_reduction_percent": 10.563152080434325 + }, + "standard": { + "splits": [ + "fleurs_bg", + "fleurs_cs", + "fleurs_da", + "fleurs_de", + "fleurs_el", + "fleurs_en", + "fleurs_es", + "fleurs_et", + "fleurs_fi", + "fleurs_fr", + "fleurs_hr", + "fleurs_hu", + "fleurs_it", + "fleurs_lt", + "fleurs_lv", + "fleurs_mt", + "fleurs_nl", + "fleurs_pl", + "fleurs_pt", + "fleurs_ro", + "fleurs_ru", + "fleurs_sk", + "fleurs_sl", + "fleurs_sv", + "fleurs_uk", + "librispeech_test_clean", + "librispeech_test_other" + ], + "rows": 25705, + "models": { + "parakeet": { + "errors": 48914, + "words": 527711, + "substitutions": 40179, + "deletions": 5631, + "insertions": 3104, + "chars": 3309718, + "char_errors": 103604, + "utterance_error": 16201, + "wer": 9.269088573101564, + "cer": 3.130296901427856 + }, + "orukeet": { + "errors": 43804, + "words": 527556, + "substitutions": 36526, + "deletions": 4118, + "insertions": 3160, + "chars": 3309718, + "char_errors": 91222, + "utterance_error": 15227, + "wer": 8.303194352827, + "cer": 2.756186478727191 + } + }, + "wins": 25, + "ties": 0, + "losses": 2, + "relative_wer_reduction_percent": 10.420595430251279 + }, + "domains": { + "splits": [ + "eurospeech_bg", + "eurospeech_de", + "eurospeech_el", + "eurospeech_en", + "eurospeech_et", + "eurospeech_fi", + "eurospeech_fr", + "eurospeech_hr", + "eurospeech_it", + "eurospeech_lt", + "eurospeech_lv", + "eurospeech_mt", + "eurospeech_pt", + "eurospeech_sk", + "eurospeech_sl", + "eurospeech_uk", + "gigaspeechbench_agr_en", + "gigaspeechbench_ait_en", + "gigaspeechbench_art_en", + "gigaspeechbench_bio_en", + "gigaspeechbench_chn_en", + "gigaspeechbench_ecm_en", + "gigaspeechbench_eng_en", + "gigaspeechbench_ent_en", + "gigaspeechbench_fin_en", + "gigaspeechbench_hum_en", + "gigaspeechbench_ind_en", + "gigaspeechbench_jpn_en", + "gigaspeechbench_law_en", + "gigaspeechbench_med_en", + "gigaspeechbench_mil_en", + "gigaspeechbench_phl_en", + "gigaspeechbench_sct_en", + "gigaspeechbench_sgp_en", + "golos_crowd_ru", + "golos_farfield_ru", + "lesbos_el", + "monsoon_en_in", + "nst_da_da", + "nst_sv_sv", + "voxpopuli_cs", + "voxpopuli_es", + "voxpopuli_hu", + "voxpopuli_it", + "voxpopuli_nl", + "voxpopuli_pl", + "voxpopuli_ro" + ], + "rows": 12006, + "models": { + "parakeet": { + "errors": 43939, + "words": 262747, + "substitutions": 26343, + "deletions": 8233, + "insertions": 9363, + "chars": 1584470, + "char_errors": 141363, + "utterance_error": 8290, + "wer": 16.72293118475187, + "cer": 8.921784571497094 + }, + "orukeet": { + "errors": 40068, + "words": 262698, + "substitutions": 23458, + "deletions": 7145, + "insertions": 9465, + "chars": 1584470, + "char_errors": 129717, + "utterance_error": 8039, + "wer": 15.25249526071763, + "cer": 8.186775388615752 + } + }, + "wins": 36, + "ties": 0, + "losses": 11, + "relative_wer_reduction_percent": 8.792931740190369 + }, + "domain_english": { + "splits": [ + "eurospeech_en", + "gigaspeechbench_agr_en", + "gigaspeechbench_ait_en", + "gigaspeechbench_art_en", + "gigaspeechbench_bio_en", + "gigaspeechbench_chn_en", + "gigaspeechbench_ecm_en", + "gigaspeechbench_eng_en", + "gigaspeechbench_ent_en", + "gigaspeechbench_fin_en", + "gigaspeechbench_hum_en", + "gigaspeechbench_ind_en", + "gigaspeechbench_jpn_en", + "gigaspeechbench_law_en", + "gigaspeechbench_med_en", + "gigaspeechbench_mil_en", + "gigaspeechbench_phl_en", + "gigaspeechbench_sct_en", + "gigaspeechbench_sgp_en", + "monsoon_en_in" + ], + "rows": 5120, + "models": { + "parakeet": { + "errors": 9032, + "words": 94993, + "substitutions": 4421, + "deletions": 1646, + "insertions": 2965, + "chars": 498262, + "char_errors": 29760, + "utterance_error": 2806, + "wer": 9.508069015611676, + "cer": 5.972761318342559 + }, + "orukeet": { + "errors": 8399, + "words": 94993, + "substitutions": 4193, + "deletions": 1490, + "insertions": 2716, + "chars": 498262, + "char_errors": 27417, + "utterance_error": 2692, + "wer": 8.841704125567148, + "cer": 5.502526783098049 + } + }, + "wins": 20, + "ties": 0, + "losses": 0, + "relative_wer_reduction_percent": 7.008414526129303 + } + }, + "checks": [ + "Fresh matched decoding of both checkpoints", + "All 74 paired split results recomputed from per-record counts", + "Pooled WER recomputed from summed errors and reference words", + "Independent upstream scoring matches all 148 model/partition pairs", + "Rendered abstract and all 74 benchmark rows match the current checkpoint evidence" + ], + "inputs_sha256": { + "evidence/standard-asr-r3-20260908/comparison.json": "9b2702bfe9c7b107db1627df3cb71d24d25d4eb8da0ee8683b0bbf9f994ca7da", + "evidence/standard-asr-r3-20260908/hypotheses-audit.json": "a60177ac2ac1064b481cc85424d5ee989cca43112191ddae30d60b7f56ae57d3", + "evidence/standard-asr-r3-20260908/numeric-evidence.jsonl.gz": "3d1467e56efd6769da1f8cb74eef0c3eac7bf93fa8200c39538963a9d4b080c1", + "evidence/domains-r3-20260908/comparison.json": "617a2119027b9e706907fdeb8a5819babb55f1959456d677e744d2b57ddfd2be", + "evidence/domains-r3-20260908/hypotheses-audit.json": "66c748f9ea625cf38e9bf672628e95300a5919abbaf47ebfd7c4800f5b1f3116", + "evidence/domains-r3-20260908/numeric-evidence.jsonl.gz": "298f3164c745523f25b2e7df25492dab5b6607c149ea561d44af3720c29236df" + }, + "pdf_sha256": "401d162a87d4454b0f16998bf3501b9c84da909a7b5be353705528477d2eef62" +} diff --git a/report/current-benchmark-values.tex b/report/current-benchmark-values.tex new file mode 100644 index 0000000000000000000000000000000000000000..dcf8a4fc8dd676d7dd30c94a82ec8d21dc0e0f19 --- /dev/null +++ b/report/current-benchmark-values.tex @@ -0,0 +1,22 @@ +% Generated from audited r3 edit counts. +\newcommand{\LibriCleanParakeetWER}{1.53} +\newcommand{\LibriCleanOrukeetWER}{1.46} +\newcommand{\LibriOtherParakeetWER}{3.14} +\newcommand{\LibriOtherOrukeetWER}{2.86} +\newcommand{\FleursEnglishParakeetWER}{4.28} +\newcommand{\FleursEnglishOrukeetWER}{3.82} +\newcommand{\FleursPooledParakeetWER}{11.01} +\newcommand{\FleursPooledOrukeetWER}{9.85} +\newcommand{\DomainPooledParakeetWER}{16.72} +\newcommand{\DomainPooledOrukeetWER}{15.25} +\newcommand{\DomainEnglishParakeetWER}{9.51} +\newcommand{\DomainEnglishOrukeetWER}{8.84} +\newcommand{\FleursMacroParakeetWER}{11.07} +\newcommand{\FleursMacroOrukeetWER}{9.96} +\newcommand{\FleursWins}{23} +\newcommand{\StandardWins}{25} +\newcommand{\DomainWins}{36} +\newcommand{\DomainEnglishWins}{20} +\newcommand{\TestedWins}{61} +\newcommand{\TestedSplits}{74} +\newcommand{\FleursPooledReduction}{10.6} diff --git a/report/current-domains-table.tex b/report/current-domains-table.tex new file mode 100644 index 0000000000000000000000000000000000000000..f4f7a5fc8f36343ddfdeca1840599b7e74eabce1 --- /dev/null +++ b/report/current-domains-table.tex @@ -0,0 +1,86 @@ +\begin{table}[!p] +\centering +\caption{WER (\%) on the fixed accent and domain sample. Each partition contains 256 recordings, except Lesbos (230). GSB denotes GigaSpeechBench; two-letter suffixes identify languages. Bold identifies lower WER. Both models use the same decoding and scoring protocol as Table~\ref{tab:current-standard}.} +\label{tab:current-domains} +\small +\begin{minipage}[t]{0.49\linewidth} +\vspace{0pt} +\centering +\setlength{\tabcolsep}{3pt} +\renewcommand{\arraystretch}{1.10} +\begin{tabular}{lrr} +\toprule +Partition & Parakeet & Orukeet\\ +\midrule +EuroSpeech BG & 14.22 & \textbf{13.04}\\ +EuroSpeech DE & 13.40 & \textbf{11.14}\\ +EuroSpeech EL & \textbf{25.83} & 26.35\\ +EuroSpeech EN & 24.40 & \textbf{23.77}\\ +EuroSpeech ET & 34.67 & \textbf{25.33}\\ +EuroSpeech FI & 16.61 & \textbf{15.20}\\ +EuroSpeech FR & 19.42 & \textbf{14.28}\\ +EuroSpeech HR & 12.93 & \textbf{12.56}\\ +EuroSpeech IT & \textbf{10.95} & 12.32\\ +EuroSpeech LT & 38.44 & \textbf{33.10}\\ +EuroSpeech LV & 57.18 & \textbf{42.14}\\ +EuroSpeech MT & 36.83 & \textbf{36.15}\\ +EuroSpeech PT & \textbf{23.08} & 23.81\\ +EuroSpeech SK & 17.29 & \textbf{14.91}\\ +EuroSpeech SL & \textbf{48.43} & 50.23\\ +EuroSpeech UK & \textbf{13.65} & 14.25\\ +GSB AI & 8.71 & \textbf{7.98}\\ +GSB Chinese accent & 14.49 & \textbf{13.56}\\ +GSB Filipino accent & 13.30 & \textbf{12.79}\\ +GSB Indian accent & 6.50 & \textbf{5.59}\\ +GSB Japanese accent & 19.15 & \textbf{17.78}\\ +GSB Scottish accent & 22.08 & \textbf{20.35}\\ +GSB Singaporean accent & 13.89 & \textbf{12.86}\\ +GSB agriculture & 6.20 & \textbf{5.84}\\ +\bottomrule +\end{tabular} +\end{minipage}\hfill% +\begin{minipage}[t]{0.49\linewidth} +\vspace{0pt} +\centering +\setlength{\tabcolsep}{3pt} +\renewcommand{\arraystretch}{1.10} +\begin{tabular}{lrr} +\toprule +Partition & Parakeet & Orukeet\\ +\midrule +GSB arts & 5.47 & \textbf{4.87}\\ +GSB biology & 3.67 & \textbf{3.31}\\ +GSB economics & 7.05 & \textbf{6.57}\\ +GSB engineering & 4.06 & \textbf{3.50}\\ +GSB entertainment & 10.40 & \textbf{8.87}\\ +GSB finance & 5.81 & \textbf{5.11}\\ +GSB humanities & 7.98 & \textbf{7.47}\\ +GSB law & 9.75 & \textbf{9.04}\\ +GSB medicine & 3.49 & \textbf{3.18}\\ +GSB military & 3.43 & \textbf{3.06}\\ +Golos crowd RU & \textbf{2.84} & 2.92\\ +Golos far-field RU & \textbf{7.98} & 9.10\\ +Lesbos Greek & 94.78 & \textbf{93.55}\\ +Monsoon India & 4.12 & \textbf{3.78}\\ +NST Danish & 26.49 & \textbf{11.59}\\ +NST Swedish & 16.57 & \textbf{12.36}\\ +VoxPopuli CS & \textbf{7.32} & 7.39\\ +VoxPopuli ES & \textbf{6.07} & 6.20\\ +VoxPopuli HU & 12.00 & \textbf{11.05}\\ +VoxPopuli IT & \textbf{11.37} & 11.82\\ +VoxPopuli NL & \textbf{9.50} & 9.56\\ +VoxPopuli PL & 6.48 & \textbf{6.24}\\ +VoxPopuli RO & 11.48 & \textbf{11.20}\\ +\bottomrule +\end{tabular} +\end{minipage} +\par\vspace{12pt} +\begin{tabular}{lrrr} +\toprule +Pooled comparison & Clips & Parakeet & Orukeet\\ +\midrule +All 47 partitions & 12,006 & 16.72 & \textbf{15.25}\\ +All 20 English partitions & 5,120 & 9.51 & \textbf{8.84}\\ +\bottomrule +\end{tabular} +\end{table} diff --git a/report/current-standard-table.tex b/report/current-standard-table.tex new file mode 100644 index 0000000000000000000000000000000000000000..b74e53273a11965e69e6688084326d0a3bd729fe --- /dev/null +++ b/report/current-standard-table.tex @@ -0,0 +1,46 @@ +\begin{table}[!ht] +\centering +\caption{Complete LibriSpeech and FLEURS test partitions. WER and CER are percentages; bold identifies lower WER. The pooled FLEURS row sums errors and reference words over all 25 languages, including English. The macro row weights languages equally.} +\label{tab:current-standard} +\small +\setlength{\tabcolsep}{5pt} +\renewcommand{\arraystretch}{1.02} +\begin{tabular}{lrrrrr} +\toprule +Benchmark & Clips & \multicolumn{2}{c}{Parakeet} & \multicolumn{2}{c}{Orukeet}\\ + & & WER & CER & WER & CER\\ +\midrule +LibriSpeech test-clean & 2,620 & 1.53 & 0.59 & \textbf{1.46} & 0.56\\ +LibriSpeech test-other & 2,939 & 3.14 & 1.32 & \textbf{2.86} & 1.19\\ +\midrule +FLEURS Bulgarian & 658 & 11.92 & 3.84 & \textbf{10.37} & 3.34\\ +FLEURS Croatian & 914 & 11.29 & 3.53 & \textbf{10.20} & 3.67\\ +FLEURS Czech & 723 & 11.12 & 3.21 & \textbf{8.97} & 2.67\\ +FLEURS Danish & 930 & 17.19 & 6.31 & \textbf{14.88} & 5.31\\ +FLEURS Dutch & 364 & 6.40 & 2.28 & \textbf{5.60} & 1.93\\ +FLEURS English & 647 & 4.28 & 2.00 & \textbf{3.82} & 1.77\\ +FLEURS Estonian & 893 & 13.32 & 3.86 & \textbf{10.44} & 3.39\\ +FLEURS Finnish & 918 & 11.14 & 2.59 & \textbf{9.35} & 2.16\\ +FLEURS French & 676 & \textbf{4.69} & 1.68 & 5.01 & 1.70\\ +FLEURS German & 862 & 4.21 & 1.41 & \textbf{3.92} & 1.52\\ +FLEURS Greek & 650 & \textbf{21.07} & 9.01 & 30.81 & 9.18\\ +FLEURS Hungarian & 905 & 13.60 & 4.20 & \textbf{10.68} & 2.97\\ +FLEURS Italian & 865 & 2.43 & 0.79 & \textbf{2.09} & 0.76\\ +FLEURS Latvian & 851 & 21.78 & 5.43 & \textbf{17.41} & 4.21\\ +FLEURS Lithuanian & 986 & 20.95 & 5.56 & \textbf{16.55} & 4.27\\ +FLEURS Maltese & 926 & 19.22 & 6.19 & \textbf{15.60} & 5.08\\ +FLEURS Polish & 758 & 6.81 & 2.09 & \textbf{6.11} & 1.95\\ +FLEURS Portuguese & 919 & 4.49 & 1.98 & \textbf{3.73} & 1.63\\ +FLEURS Romanian & 883 & 11.44 & 3.86 & \textbf{9.34} & 3.07\\ +FLEURS Russian & 775 & 4.89 & 1.49 & \textbf{4.72} & 1.48\\ +FLEURS Slovak & 792 & 9.21 & 2.91 & \textbf{7.75} & 2.41\\ +FLEURS Slovenian & 834 & 22.62 & 7.70 & \textbf{22.11} & 8.28\\ +FLEURS Spanish & 908 & 3.22 & 1.28 & \textbf{2.75} & 1.04\\ +FLEURS Swedish & 759 & 13.38 & 4.26 & \textbf{11.36} & 3.45\\ +FLEURS Ukrainian & 750 & 6.00 & 1.74 & \textbf{5.39} & 1.60\\ +\midrule +FLEURS pooled & 20,146 & 11.01 & 3.57 & \textbf{9.85} & 3.13\\ +FLEURS language macro & 20,146 & 11.07 & 3.57 & \textbf{9.96} & 3.15\\ +\bottomrule +\end{tabular} +\end{table} diff --git a/report/evidence.tex b/report/evidence.tex new file mode 100644 index 0000000000000000000000000000000000000000..507e3ba6b07a81c3db6983f49be0137ea3ad4feb --- /dev/null +++ b/report/evidence.tex @@ -0,0 +1,9 @@ +% Generated from release/model-stages.json; do not edit. +\newcommand{\GaborPrimary}{15.341} +\newcommand{\OriginalPrimary}{15.371} +\newcommand{\GaborEnglish}{9.359} +\newcommand{\OriginalEnglish}{9.360} +\newcommand{\NativeFPrimary}{15.372} +\newcommand{\NativeOriginalPrimary}{15.304} +\newcommand{\NativeFEnglish}{9.360} +\newcommand{\NativeOriginalEnglish}{9.324} diff --git a/report/make_model_figures.py b/report/make_model_figures.py new file mode 100644 index 0000000000000000000000000000000000000000..d77c03ab59c738572cba4c8d4f741fca160a09f0 --- /dev/null +++ b/report/make_model_figures.py @@ -0,0 +1,136 @@ +"""Orukeet figures from exact fitted taps and paired recognition counts. + +Final manuscript width: 5.5 inches. Each exported figure keeps that size. +No smoothing, resampling, omitted outliers, or inferred training trajectories. +""" +from pathlib import Path +import hashlib +import json +import numpy as np +import matplotlib +matplotlib.use('Agg') +import matplotlib.pyplot as plt + +ROOT = Path(__file__).resolve().parents[1] +OUT = ROOT / 'report/assets' +BLUE, GRAY, GREEN = '#0F4D92', '#767676', '#42949E' +plt.rcParams.update({'font.family':'DejaVu Sans', 'font.size':8.5, + 'axes.labelsize':9, 'axes.titlesize':10, 'xtick.labelsize':8, 'ytick.labelsize':8, + 'legend.fontsize':8, 'pdf.fonttype':42, 'svg.fonttype':'none', + 'axes.spines.top':False, 'axes.spines.right':False, 'axes.linewidth':.7, + 'lines.linewidth':1.4, 'savefig.facecolor':'white'}) +fit_path = ROOT / 'training/gabor_half/fits/fits.npz' +z = np.load(fit_path) +assert z['original'].shape == (24576,9) and z['selected'].sum() == 12288 +order = np.argsort(z['normalized_sse'], kind='stable') +assert np.array_equal(np.sort(order[:12288]), np.flatnonzero(z['selected'])) +selected = order[:12288] +error = 100 * np.sqrt(z['normalized_sse']) +result_path = ROOT / 'training/gabor_half/results/full-r15-0100.json' +r = json.loads(result_path.read_text()) +assert r['candidate_model_sha256'] == '4295a6d820a40b99786331d1c7a6b6c328916c8329b23d39415b0649a5d42811' +provenance = {'model_sha256':r['candidate_model_sha256'], + 'inputs':{str(p.relative_to(ROOT)):hashlib.sha256(p.read_bytes()).hexdigest() + for p in [fit_path,result_path,Path(__file__).resolve()]}, + 'style':'academic-figures: final-size typography, restrained blue/gray/teal, vector export', + 'design_precedent':{'repository':'https://github.com/Nathan-Roll1/gabormer', + 'manuscripts':['interspeech/paper/gabormer_interspeech/gabormer_interspeech.tex', + 'colm/paper/gabormer_unified/gabormer_unified.tex'], + 'adapted_ideas':['original/fitted kernel small multiples','fit distribution','layer-depth profile'], + 'data':'All plotted measurements are Orukeet data. No Gabormer values or artwork copied.'}, + 'figures':{}} + +def save(fig,name,caption,details): + for ext in ['pdf','svg','png']: + fig.savefig(OUT / (name+'.'+ext), dpi=300) + # Manuscript-size screen preview, useful alongside the vector PDF. + fig.savefig(OUT/(name+'-preview.png'),dpi=150) + (OUT/(name+'-caption.md')).write_text(caption+'\n') + provenance['figures'][name] = {'size_inches':list(fig.get_size_inches()), + 'caption':caption, **details, + 'sha256':{ext:hashlib.sha256((OUT/(name+'.'+ext)).read_bytes()).hexdigest() + for ext in ['pdf','svg','png']}} + plt.close(fig) + +fig, axs = plt.subplots(1,4,figsize=(5.5,1.9),sharex=True,sharey=True) +fig.subplots_adjust(left=.095,right=.985,bottom=.25,top=.73,wspace=.15) +examples=[] +for ax,q,letter in zip(axs,[.125,.375,.625,.875],'ABCD'): + rank=round(12287*q); i=int(selected[rank]); norm=np.linalg.norm(z['original'][i]) + ax.plot(range(-4,5),z['original'][i]/norm,'o-',color=GRAY,markersize=3, + linewidth=1,label='Original taps',markerfacecolor='white') + ax.plot(range(-4,5),z['fitted'][i]/norm,'x--',color=BLUE,markersize=3, + linewidth=1.2,label='Frozen Gabor taps') + ax.axhline(0,color='#dddddd',lw=.6,zorder=0) + ax.set(xticks=[-4,0,4],yticks=[-1,0,1],ylim=(-1.08,1.08),xlabel='Tap index') + ax.set_title(f'{letter} Rank {rank+1:,}\n{error[i]:.2f}% RMS',loc='left',fontsize=9) + examples.append({'rank':rank+1,'layer_zero_based':i//1024,'channel':i%1024, + 'relative_rms_percent':float(error[i])}) +axs[0].set_ylabel('Weight / original norm') +fig.legend(*axs[0].get_legend_handles_labels(),loc='upper center',ncol=2, + frameon=False,bbox_to_anchor=(.55,1.025),handlelength=2) +save(fig,'kernel-fits','Four exact kernel replacements. Examples are ranks 1,537, 4,609, 7,680 and 10,752 of the selected 12,288, nearest the 12.5th, 37.5th, 62.5th and 87.5th fit-error percentiles. Each pair is divided by the original kernel L2 norm. Markers are the nine stored taps; connecting lines guide the eye. RMS percentages are relative to the original norm.',{'examples':examples,'unit':'one temporal depthwise kernel','uncertainty':'Exact weights; no sampling interval.'}) + +fig, axs = plt.subplots(1,2,figsize=(5.5,2.15)) +fig.subplots_adjust(left=.1,right=.98,bottom=.24,top=.82,wspace=.38) +ax=axs[0]; y=100*np.arange(1,len(order)+1)/len(order) +ax.plot(error[order],y,color=GRAY,lw=1.5) +ax.plot(error[selected],y[:12288],color=BLUE,lw=2) +cut=float(error[selected[-1]]) +ax.scatter([cut],[50],color=BLUE,s=20,zorder=3) +ax.axhline(50,color='#bbbbbb',lw=.7,ls=':') +ax.annotate(f'50% selected\ncutoff {cut:.2f}%',xy=(cut,50),xytext=(29,19), + fontsize=8,color=BLUE,arrowprops={'arrowstyle':'-','color':BLUE,'lw':.7}) +ax.set(xlabel='Relative RMS fit error (%)',ylabel='Kernels at or below (%)', + xlim=(0,max(100,float(error.max())*1.02)),ylim=(0,100),yticks=[0,50,100],xticks=[0,25,50,75,100]) +ax.set_title('A Global fit ranking',loc='left') +ax=axs[1]; counts=z['selected'].reshape(24,1024).sum(1) +ax.bar(np.arange(1,25),counts,color=BLUE,width=.72) +ax.axhline(512,color=GRAY,lw=.8,ls='--') +ax.set(xlabel='Encoder layer',ylabel='Frozen kernels / 1,024',xlim=(.2,24.8), + ylim=(0,1024),xticks=[1,6,12,18,24],yticks=[0,512,1024]) +ax.set_title('B Selection by layer',loc='left') +save(fig,'selection-profile','Global Gabor selection. (A) Empirical cumulative distribution of relative RMS fit error for all 24,576 original kernels; blue marks the selected half. Every kernel is included. (B) The resulting allocation across all 24 encoder layers. The dashed 512 line denotes half a layer; selection uses a single global ranking. Counts range from 175 to 748 per layer.',{'selected_count':12288,'total_count':24576,'cutoff_percent':cut,'layer_counts':counts.tolist(),'uncertainty':'Complete kernel population; no sampling interval.'}) + +fig,ax=plt.subplots(figsize=(5.5,1.75)) +fig.subplots_adjust(left=.26,right=.98,bottom=.34,top=.84) +for y,key,label in [(1,'primary','20 languages'),(0,'english','English, 7 corpora')]: + d=r[key]; lo,hi=d['paired_delta_95ci']; x=d['delta_pp'] + ax.errorbar(x,y,xerr=[[x-lo],[hi-x]],fmt='o',color=BLUE,capsize=3,markersize=5,lw=1.6) +ax.axvline(0,color=GRAY,lw=.8,ls='--') +ax.set(yticks=[0,1],yticklabels=['English, 7 corpora','20 languages'],ylim=(-.55,1.55), + xlim=(-.16,.115),xticks=[-.15,-.1,-.05,0,.05,.1], + xlabel='WER change from adaptation baseline (percentage points)') +ax.spines['left'].set_visible(False);ax.tick_params(axis='y',length=0) +ax.set_title('Recognition after recovery',loc='left') +save(fig,'recognition-deltas','Orukeet minus the pre-Gabor adaptation baseline under matched NeMo decoding. Points are macro WER differences; bars are 95% paired global-cluster bootstrap intervals (5,000 replicates, seed 20260905; 4,266 speaker or parallel-sentence clusters). Negative values mean lower WER. The primary endpoint uses 13,246 recordings and 20 fixed languages; English averages seven corpus WERs. Both systems were selected during the same adaptive campaign; intervals condition on those selections.',{'data':{k:r[k] for k in ['primary','english']},'independent_unit':r['bootstrap']['unit'],'replicates':5000}) + +# Complete per-language view: common axes retain the wide Lithuanian interval. +langs=[(k.split(':')[1],v) for k,v in r['metrics'].items() if k.startswith('language:')] +fig,ax=plt.subplots(figsize=(5.5,5.4));fig.subplots_adjust(left=.17,right=.97,bottom=.13,top=.92) +for y,(name,d) in enumerate(reversed(langs)): + x=d['delta_pp'];lo,hi=d['paired_delta_95ci'] + ax.errorbar(x,y,xerr=[[x-lo],[hi-x]],fmt='o',color=BLUE,capsize=2,markersize=3.5,lw=1) +ax.set(yticks=range(20),yticklabels=[s.upper() for s,_ in reversed(langs)], + xlabel='WER change from adaptation baseline (pp)',ylim=(-.7,19.7),xlim=(-2.4,.95)) +ax.axvline(0,color=GRAY,lw=.8,ls='--');ax.spines['left'].set_visible(False) +ax.tick_params(axis='y',length=0);ax.set_title('All 20 languages in the primary endpoint',loc='left') +save(fig,'language-deltas','Per-language Orukeet WER change under matched NeMo decoding, with the same paired bootstrap as the macro comparison. All 20 primary languages and complete 95% intervals are shown, alphabetically. Languages share a scale. These marginal intervals are not adjusted for simultaneous inference.',{'data':dict(langs),'independent_unit':r['bootstrap']['unit']}) + +# A full layer atlas uses each layer's median selected row, not hand-picked waves. +fig,axs=plt.subplots(6,4,figsize=(7,8),sharex=True,sharey=True) +fig.subplots_adjust(left=.1,right=.98,bottom=.08,top=.93,wspace=.25,hspace=.52) +atlas=[] +for layer,ax in enumerate(axs.flat): + candidates=selected[selected//1024==layer];i=int(candidates[(len(candidates)-1)//2]);norm=np.linalg.norm(z['original'][i]) + ax.plot(range(-4,5),z['original'][i]/norm,'o-',color=GRAY,ms=2.6,lw=1,mfc='white') + ax.plot(range(-4,5),z['fitted'][i]/norm,'x--',color=BLUE,ms=2.6,lw=1) + ax.axhline(0,color='#dddddd',lw=.5) + ax.set_title(f'Layer {layer+1} · ch {i%1024}',loc='left',fontsize=9) + ax.set(xticks=[-4,0,4],yticks=[-1,0,1],ylim=(-1.08,1.08)) + atlas.append({'layer':layer+1,'channel':i%1024,'error_percent':float(error[i])}) +fig.suptitle('One median selected kernel from every encoder layer',fontsize=12,x=.1,ha='left') +fig.supxlabel('Tap index',fontsize=10);fig.supylabel('Weight / original norm',fontsize=10) +save(fig,'kernel-atlas','Layer atlas: the median selected fit-error row within each of the 24 layers. Original (gray circles) and frozen Gabor (blue crosses) taps share the original L2 normalization and common axes. All plotted samples come from the recorded fits; lines join discrete samples.',{'examples':atlas,'unit':'one selected kernel per layer','selection_rule':'lower median by global fit-error order within layer'}) +(OUT/'model-figures-source.json').write_text(json.dumps(provenance,indent=2)+'\n') +print(json.dumps({'figures':list(provenance['figures']),'selected_count':12288,'cutoff_percent':cut})) diff --git a/report/make_surgery_figure.py b/report/make_surgery_figure.py new file mode 100644 index 0000000000000000000000000000000000000000..f5819f7015ae051bd8a1c6051659e4094b3e3b05 --- /dev/null +++ b/report/make_surgery_figure.py @@ -0,0 +1,42 @@ +"""Two measured panels at the paper's final 5.5-inch insertion width.""" +from pathlib import Path +import hashlib +import json +import numpy as np +import matplotlib +matplotlib.use('Agg') +import matplotlib.pyplot as plt + +root = Path(__file__).resolve().parents[1] +path = root / 'training/gabor_half/fits/fits.npz' +z = np.load(path) +assert len(z['original']) == 24576 and z['selected'].sum() == 12288 +order = np.argsort(z['normalized_sse'], kind='stable')[:12288] +idx = order[6143] +w = z['original'][idx] +norm = np.linalg.norm(w) +amp, mu, sigma, f, phase = z['params'][idx] +x = np.linspace(-4, 4, 300) +y = amp * np.exp(-.5 * ((x - mu) / sigma)**2) * np.cos(2 * np.pi * f * (x - mu) + phase) +plt.rcParams.update({'font.family':'DejaVu Sans', 'font.size':8, + 'axes.titlesize':8.5, 'axes.labelsize':8, 'pdf.fonttype':42, 'svg.fonttype':'none', + 'axes.spines.top':False, 'axes.spines.right':False, 'axes.linewidth':.6}) +fig, axes = plt.subplots(1, 2, figsize=(5.5, 1.52), layout='constrained', + gridspec_kw={'width_ratios':[1,1.25]}) +ax = axes[0] +ax.plot(np.arange(-4,5),w/norm,'o:',color='#333333',markersize=2.5,linewidth=.7,label='Original taps') +ax.plot(x,y/norm,color='#168071',linewidth=1.25,label='Fitted Gabor') +ax.set(title='A Median selected fit (6.32% RMS error)', xlabel='Temporal tap', + ylabel='Weight / original norm', xticks=[-4,0,4], yticks=[0,.5,1], ylim=(-.3,1.08)) +ax.legend(frameon=False,fontsize=6.8,loc='upper right',handlelength=1.1) +ax = axes[1] +ax.bar(np.arange(1,25),z['selected'].reshape(24,1024).sum(1),color='#168071',width=.7) +ax.axhline(512,color='#555555',linestyle='--',linewidth=.7) +ax.set(title='B Global selection varies by layer', xlabel='Conformer layer', + ylabel='Selected / 1,024',xticks=[1,6,12,18,24],yticks=[0,512,1024],ylim=(0,1024),xlim=(.2,24.8)) +out=root/'report/assets' +for ext in ('pdf','svg','png'): fig.savefig(out/('surgery.'+ext),dpi=300) +(out/'surgery-source.json').write_text(json.dumps({'source':str(path.relative_to(root)), + 'sha256':hashlib.sha256(path.read_bytes()).hexdigest(),'median_selected_rank':6144, + 'median_original_flat_index':int(idx),'selected_count':12288, + 'caption':'Weight-fit measurements. The left example is rank 6,144 of the globally selected rows, normalized by original L2 norm. The right shows all layer counts; dashed line is 512, not a quota. No recognition accuracy is inferred.'},indent=2)+'\n') diff --git a/report/model.json b/report/model.json new file mode 100644 index 0000000000000000000000000000000000000000..4473e49bbd01508dfd1f45b8c94416df1e7346b0 --- /dev/null +++ b/report/model.json @@ -0,0 +1,24 @@ +{ + "name": "Orukeet", + "version": "librispeech-native-format-r3", + "sha256": "031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56", + "parent_sha256": "0ccfefcd1894871cb0850bd3c464adf5397752840de2a76d1d2d075c4141a945", + "hf": { + "repo_id": "oruk/orukeet", + "revision": "555136b50265a132d4cea0d35560c26fc4f657ab", + "path": "orukeet-v0.1.0.nemo" + }, + "freeze_audit": "evidence/librispeech-ft-20260908/r3/export-audit.json", + "archive_receipt": "evidence/librispeech-ft-20260908/r3/private-archive.json", + "evaluations": [ + "evidence/standard-asr-r3-20260908/comparison.json", + "evidence/domains-r3-20260908/comparison.json" + ], + "publication_authorized": true, + "scope": "The manuscript, canonical NeMo source, Q8 and F16 exports, and OpenWhispr integration all share this exact r3 source checkpoint. Export hashes are pinned in src/orukeet/artifacts.json.", + "archive_hf": { + "repo_id": "oruk/orukeet-internal", + "revision": "294d5bbc2f4ba79cbcf07776c52f580ff053a60d", + "path": "experiments/librispeech-ft-20260908/r3/orukeet-targeted-ft.nemo" + } +} diff --git a/report/neurips_2026.sty b/report/neurips_2026.sty new file mode 100644 index 0000000000000000000000000000000000000000..ba57755282dcb4ed8febc86c24f4e581810bf374 --- /dev/null +++ b/report/neurips_2026.sty @@ -0,0 +1,443 @@ +% partial rewrite of the LaTeX2e package for submissions to the +% Conference on Neural Information Processing Systems (NeurIPS): +% +% - uses more LaTeX conventions +% - line numbers at submission time replaced with aligned numbers from +% lineno package +% - \nipsfinalcopy replaced with [final] package option +% - automatically loads times package for authors +% - loads natbib automatically; this can be suppressed with the +% [nonatbib] package option +% - adds foot line to first page identifying the conference +% - adds preprint option for submission to e.g. arXiv +% - conference acronym modified +% - update foot line to display the track name +% +% Roman Garnett (garnett@wustl.edu) and the many authors of +% nips15submit_e.sty, including MK and drstrip@sandia +% +% last revision: January 2026 + +\NeedsTeXFormat{LaTeX2e} +\ProvidesPackage{neurips_2026}[2026-01-29 NeurIPS 2026 submission/camera-ready style file] + +% declare final option, which creates camera-ready copy +\newif\if@neuripsfinal\@neuripsfinalfalse +\DeclareOption{final}{ + \@neuripsfinaltrue + \@anonymousfalse +} + +% declare nonatbib option, which does not load natbib in case of +% package clash (users can pass options to natbib via +% \PassOptionsToPackage) +\newif\if@natbib\@natbibtrue +\DeclareOption{nonatbib}{ + \@natbibfalse +} + +% declare preprint option, which creates a preprint version ready for +% upload to, e.g., arXiv +\newif\if@preprint\@preprintfalse +\DeclareOption{preprint}{ + \@preprinttrue + \@anonymousfalse +} + +% determine the track of the paper in camera-ready mode +\newif\if@main\@maintrue +\DeclareOption{main}{ + \@maintrue + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear).} +} +\newif\if@position\@positionfalse +\DeclareOption{position}{ + \@positiontrue + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Position Paper Track.} +} +\newif\if@eandd\@eanddfalse +\DeclareOption{eandd}{ + \@eanddtrue +\if@neuripsfinal\@anonymousfalse\else\if@preprint\@anonymousfalse\else\@anonymoustrue\fi\fi + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Track on Evaluations and Datasets.} +} +\newif\if@creativeai\@creativeaifalse +\DeclareOption{creativeai}{ + \@creativeaitrue + \@anonymousfalse + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Creative AI Track.} +} +\newif\if@education\@educationfalse +\DeclareOption{education}{ + \@educationtrue + \@anonymousfalse + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Education Track.} +} + +% For anonymous or non-anonymous +\newif\if@anonymous\@anonymoustrue + +% For workshop papers +\newcommand{\@workshoptitle}{} +\newcommand{\workshoptitle}[1]{\renewcommand{\@workshoptitle}{#1}} + +\newif\if@workshop\@workshopfalse +\DeclareOption{sglblindworkshop}{ + \@workshoptrue + \@anonymousfalse + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Workshop: \@workshoptitle.} +} +\DeclareOption{dblblindworkshop}{ + \@workshoptrue + \newcommand{\@trackname}{\@neuripsordinal\ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Workshop: \@workshoptitle.} +} +\DeclareOption{nonanonymous}{ + \@anonymousfalse +} + +\ProcessOptions\relax + +% fonts +\renewcommand{\rmdefault}{ptm} +\renewcommand{\sfdefault}{phv} + +% change this every year for notice string at bottom +\newcommand{\@neuripsordinal}{40th} +\newcommand{\@neuripsyear}{2026} +\newcommand{\@neuripslocation}{Sydney} + +% acknowledgments +\usepackage{environ} +\newcommand{\acksection}{\section*{Acknowledgments and Disclosure of Funding}} +\NewEnviron{ack}{% + \acksection + \BODY +} + + +% load natbib unless told otherwise +\if@natbib + \RequirePackage{natbib} +\fi + + + + + +% set page geometry +\usepackage[verbose=true,letterpaper]{geometry} +\AtBeginDocument{ + \newgeometry{ + textheight=9in, + textwidth=5.5in, + top=1in, + headheight=12pt, + headsep=25pt, + footskip=30pt + } + \@ifpackageloaded{fullpage} + {\PackageWarning{neurips_2026}{fullpage package not allowed! Overwriting formatting.}} + {} +} + +\widowpenalty=10000 +\clubpenalty=10000 +\flushbottom +\sloppy + + +% font sizes with reduced leading +\renewcommand{\normalsize}{% + \@setfontsize\normalsize\@xpt\@xipt + \abovedisplayskip 7\p@ \@plus 2\p@ \@minus 5\p@ + \abovedisplayshortskip \z@ \@plus 3\p@ + \belowdisplayskip \abovedisplayskip + \belowdisplayshortskip 4\p@ \@plus 3\p@ \@minus 3\p@ +} +\normalsize +\renewcommand{\small}{% + \@setfontsize\small\@ixpt\@xpt + \abovedisplayskip 6\p@ \@plus 1.5\p@ \@minus 4\p@ + \abovedisplayshortskip \z@ \@plus 2\p@ + \belowdisplayskip \abovedisplayskip + \belowdisplayshortskip 3\p@ \@plus 2\p@ \@minus 2\p@ +} +\renewcommand{\footnotesize}{\@setfontsize\footnotesize\@ixpt\@xpt} +\renewcommand{\scriptsize}{\@setfontsize\scriptsize\@viipt\@viiipt} +\renewcommand{\tiny}{\@setfontsize\tiny\@vipt\@viipt} +\renewcommand{\large}{\@setfontsize\large\@xiipt{14}} +\renewcommand{\Large}{\@setfontsize\Large\@xivpt{16}} +\renewcommand{\LARGE}{\@setfontsize\LARGE\@xviipt{20}} +\renewcommand{\huge}{\@setfontsize\huge\@xxpt{23}} +\renewcommand{\Huge}{\@setfontsize\Huge\@xxvpt{28}} + + +% Force \tiny to be no smaller than 6pt +\renewcommand{\tiny}{\fontsize{6pt}{7pt}\selectfont} + +% Force \scriptsize to be no smaller than 7pt +\renewcommand{\scriptsize}{\fontsize{7pt}{8pt}\selectfont} + +% Force \footnotesize to be no smaller than 8pt +\renewcommand{\footnotesize}{\fontsize{8pt}{9.5pt}\selectfont} + +% sections with less space +\providecommand{\section}{} +\renewcommand{\section}{% + \@startsection{section}{1}{\z@}% + {-2.0ex \@plus -0.5ex \@minus -0.2ex}% + { 1.5ex \@plus 0.3ex \@minus 0.2ex}% + {\large\bf\raggedright}% +} +\providecommand{\subsection}{} +\renewcommand{\subsection}{% + \@startsection{subsection}{2}{\z@}% + {-1.8ex \@plus -0.5ex \@minus -0.2ex}% + { 0.8ex \@plus 0.2ex}% + {\normalsize\bf\raggedright}% +} +\providecommand{\subsubsection}{} +\renewcommand{\subsubsection}{% + \@startsection{subsubsection}{3}{\z@}% + {-1.5ex \@plus -0.5ex \@minus -0.2ex}% + { 0.5ex \@plus 0.2ex}% + {\normalsize\bf\raggedright}% +} +\providecommand{\paragraph}{} +\renewcommand{\paragraph}{% + \@startsection{paragraph}{4}{\z@}% + {1.5ex \@plus 0.5ex \@minus 0.2ex}% + {-1em}% + {\normalsize\bf}% +} +\providecommand{\subparagraph}{} +\renewcommand{\subparagraph}{% + \@startsection{subparagraph}{5}{\z@}% + {1.5ex \@plus 0.5ex \@minus 0.2ex}% + {-1em}% + {\normalsize\bf}% +} +\providecommand{\subsubsubsection}{} +\renewcommand{\subsubsubsection}{% + \vskip5pt{\noindent\normalsize\rm\raggedright}% +} + +% float placement +\renewcommand{\topfraction }{0.85} +\renewcommand{\bottomfraction }{0.4} +\renewcommand{\textfraction }{0.1} +\renewcommand{\floatpagefraction}{0.7} + +\newlength{\@neuripsabovecaptionskip}\setlength{\@neuripsabovecaptionskip}{7\p@} +\newlength{\@neuripsbelowcaptionskip}\setlength{\@neuripsbelowcaptionskip}{\z@} + +\setlength{\abovecaptionskip}{\@neuripsabovecaptionskip} +\setlength{\belowcaptionskip}{\@neuripsbelowcaptionskip} + +% swap above/belowcaptionskip lengths for tables +\renewenvironment{table} + {\setlength{\abovecaptionskip}{\@neuripsbelowcaptionskip}% + \setlength{\belowcaptionskip}{\@neuripsabovecaptionskip}% + \@float{table}} + {\end@float} + +% footnote formatting +\setlength{\footnotesep }{6.65\p@} +\setlength{\skip\footins}{9\p@ \@plus 4\p@ \@minus 2\p@} +\renewcommand{\footnoterule}{\kern-3\p@ \hrule width 12pc \kern 2.6\p@} +\setcounter{footnote}{0} + +% paragraph formatting +\setlength{\parindent}{\z@} +\setlength{\parskip }{5.5\p@} + +% list formatting +\setlength{\topsep }{4\p@ \@plus 1\p@ \@minus 2\p@} +\setlength{\partopsep }{1\p@ \@plus 0.5\p@ \@minus 0.5\p@} +\setlength{\itemsep }{2\p@ \@plus 1\p@ \@minus 0.5\p@} +\setlength{\parsep }{2\p@ \@plus 1\p@ \@minus 0.5\p@} +\setlength{\leftmargin }{3pc} +\setlength{\leftmargini }{\leftmargin} +\setlength{\leftmarginii }{2em} +\setlength{\leftmarginiii}{1.5em} +\setlength{\leftmarginiv }{1.0em} +\setlength{\leftmarginv }{0.5em} +\def\@listi {\leftmargin\leftmargini} +\def\@listii {\leftmargin\leftmarginii + \labelwidth\leftmarginii + \advance\labelwidth-\labelsep + \topsep 2\p@ \@plus 1\p@ \@minus 0.5\p@ + \parsep 1\p@ \@plus 0.5\p@ \@minus 0.5\p@ + \itemsep \parsep} +\def\@listiii{\leftmargin\leftmarginiii + \labelwidth\leftmarginiii + \advance\labelwidth-\labelsep + \topsep 1\p@ \@plus 0.5\p@ \@minus 0.5\p@ + \parsep \z@ + \partopsep 0.5\p@ \@plus 0\p@ \@minus 0.5\p@ + \itemsep \topsep} +\def\@listiv {\leftmargin\leftmarginiv + \labelwidth\leftmarginiv + \advance\labelwidth-\labelsep} +\def\@listv {\leftmargin\leftmarginv + \labelwidth\leftmarginv + \advance\labelwidth-\labelsep} +\def\@listvi {\leftmargin\leftmarginvi + \labelwidth\leftmarginvi + \advance\labelwidth-\labelsep} + +% create title +\providecommand{\maketitle}{} +\renewcommand{\maketitle}{% + \par + \begingroup + \renewcommand{\thefootnote}{\fnsymbol{footnote}} + % for perfect author name centering + \renewcommand{\@makefnmark}{\hbox to \z@{$^{\@thefnmark}$\hss}} + % The footnote-mark was overlapping the footnote-text, + % added the following to fix this problem (MK) + \long\def\@makefntext##1{% + \parindent 1em\noindent + \hbox to 1.8em{\hss $\m@th ^{\@thefnmark}$}##1 + } + \thispagestyle{empty} + \@maketitle + \@thanks + \@notice + \endgroup + \let\maketitle\relax + \let\thanks\relax +} + +% rules for title box at top of first page +\newcommand{\@toptitlebar}{ + \hrule height 4\p@ + \vskip 0.25in + \vskip -\parskip% +} +\newcommand{\@bottomtitlebar}{ + \vskip 0.29in + \vskip -\parskip + \hrule height 1\p@ + \vskip 0.09in% +} + +% create title (includes both anonymized and non-anonymized versions) +\providecommand{\@maketitle}{} +\renewcommand{\@maketitle}{% + \vbox{% + \hsize\textwidth + \linewidth\hsize + \vskip 0.1in + \@toptitlebar + \centering + {\LARGE\bf \@title\par} + \@bottomtitlebar + \if@anonymous + \begin{tabular}[t]{c}\bf\rule{\z@}{24\p@} + Anonymous Author(s) \\ + Affiliation \\ + Address \\ + \texttt{email} \\ + \end{tabular}% + \else + \def\And{% + \end{tabular}\hfil\linebreak[0]\hfil% + \begin{tabular}[t]{c}\bf\rule{\z@}{24\p@}\ignorespaces% + } + \def\AND{% + \end{tabular}\hfil\linebreak[4]\hfil% + \begin{tabular}[t]{c}\bf\rule{\z@}{24\p@}\ignorespaces% + } + \begin{tabular}[t]{c}\bf\rule{\z@}{24\p@}\@author\end{tabular}% + \fi + \vskip 0.3in \@minus 0.1in + } +} + +% add conference notice to bottom of first page +\newcommand{\ftype@noticebox}{8} +\newcommand{\@notice}{% + % give a bit of extra room back to authors on first page + \enlargethispage{2\baselineskip}% + \@float{noticebox}[b]% + \footnotesize\@noticestring% + \end@float% +} + +% abstract styling +\renewenvironment{abstract}% +{% + \vskip 0.075in% + \centerline% + {\large\bf Abstract}% + \vspace{0.5ex}% + \begin{quote}% +} +{ + \par% + \end{quote}% + \vskip 1ex% +} + +% For the paper checklist +\newcommand{\answerYes}[1][]{\textcolor{blue}{[Yes]#1}} +\newcommand{\answerNo}[1][]{\textcolor{orange}{[No]#1}} +\newcommand{\answerNA}[1][]{\textcolor{gray}{[N/A]#1}} +\newcommand{\answerTODO}[1][]{\textcolor{red}{\bf [TODO]}} +\newcommand{\justificationTODO}[1][]{\textcolor{red}{\bf [TODO]}} + +% handle tweaks for camera-ready copy vs. submission copy +\if@preprint + \newcommand{\@noticestring}{% + Preprint.% + } +\else + \if@neuripsfinal + \newcommand{\@noticestring}{ + \@trackname + } + \else + \newcommand{\@noticestring}{% + Submitted to \@neuripsordinal\/ Conference on Neural Information Processing Systems (NeurIPS \@neuripsyear). Do not distribute.% + } + + % hide the acknowledgements + \NewEnviron{hide}{} + \let\ack\hide + \let\endack\endhide + + % line numbers for submission + \RequirePackage{lineno} + \linenumbers + + % fix incompatibilities between lineno and amsmath, if required, by + % transparently wrapping linenomath environments around amsmath + % environments + \AtBeginDocument{% + \@ifpackageloaded{amsmath}{% + \newcommand*\patchAmsMathEnvironmentForLineno[1]{% + \expandafter\let\csname old#1\expandafter\endcsname\csname #1\endcsname + \expandafter\let\csname oldend#1\expandafter\endcsname\csname end#1\endcsname + \renewenvironment{#1}% + {\linenomath\csname old#1\endcsname}% + {\csname oldend#1\endcsname\endlinenomath}% + }% + \newcommand*\patchBothAmsMathEnvironmentsForLineno[1]{% + \patchAmsMathEnvironmentForLineno{#1}% + \patchAmsMathEnvironmentForLineno{#1*}% + }% + \patchBothAmsMathEnvironmentsForLineno{equation}% + \patchBothAmsMathEnvironmentsForLineno{align}% + \patchBothAmsMathEnvironmentsForLineno{flalign}% + \patchBothAmsMathEnvironmentsForLineno{alignat}% + \patchBothAmsMathEnvironmentsForLineno{gather}% + \patchBothAmsMathEnvironmentsForLineno{multline}% + } + {} + } + \fi +\fi + + +\endinput diff --git a/report/paper.tex b/report/paper.tex new file mode 100644 index 0000000000000000000000000000000000000000..583593f592b34f9c5a6bda0445e8d360087abd55 --- /dev/null +++ b/report/paper.tex @@ -0,0 +1,215 @@ +\documentclass{article} +\PassOptionsToPackage{numbers,sort&compress}{natbib} +\usepackage[preprint]{neurips_2026} +\usepackage[T1]{fontenc} +\usepackage{courier} +\usepackage[utf8]{inputenc} +\usepackage{graphicx,booktabs,amsmath,amssymb,microtype,xcolor} +\usepackage[colorlinks=true,urlcolor=blue,citecolor=blue,linkcolor=blue]{hyperref} +\input{current-benchmark-values.tex} +\title{Orukeet: Multilingual ASR\\with Frozen Gabor Kernels} +\author{Nathan Roll\textsuperscript{1,2}\quad +Irene Yi\textsuperscript{1,2}\quad +B\"u\c{s}ra Mar\c{s}an\textsuperscript{1,2}\\[3pt] +\bfseries Vianney Grenez\textsuperscript{1}\quad +Gabriel Stein\textsuperscript{4}\quad +Momcilo Mrkaic\textsuperscript{5}\\[3pt] +\bfseries Pavle Padjin\textsuperscript{5}\quad +Vladimir Zeljkovic\textsuperscript{5}\quad +Calbert Graham\textsuperscript{1,3}\\[10pt] +\begin{tabular}{@{}ccccc@{}} +\parbox[c][19pt][c]{0.74in}{\centering\includegraphics[width=0.63in]{assets/oruk-lockup.png}} & +\parbox[c][19pt][c]{1.08in}{\centering\includegraphics[width=1.02in]{assets/affiliations/stanford.png}} & +\parbox[c][19pt][c]{1.26in}{\centering\includegraphics[width=1.08in]{assets/affiliations/cambridge.pdf}} & +\parbox[c][19pt][c]{0.78in}{\centering\includegraphics[height=17pt]{assets/affiliations/openwhispr.pdf}} & +\parbox[c][19pt][c]{0.70in}{\centering\includegraphics[width=0.55in]{assets/affiliations/hoid.pdf}}\\[3pt] +\scriptsize\textsuperscript{1}Oruk AI & +\scriptsize\textsuperscript{2}Stanford University & +\scriptsize\textsuperscript{3}University of Cambridge & +\scriptsize\textsuperscript{4}OpenWhispr & +\scriptsize\textsuperscript{5}Hoid +\end{tabular}} +\hypersetup{pdftitle={Orukeet: Multilingual ASR with Frozen Gabor Kernels}, +pdfauthor={Nathan Roll, Irene Yi, B\"u\c{s}ra Mar\c{s}an, Vianney Grenez, Gabriel Stein, Momcilo Mrkaic, Pavle Padjin, Vladimir Zeljkovic, Calbert Graham}, +pdfsubject={Orukeet r3 technical report}, +pdfkeywords={automatic speech recognition, multilingual speech, Gabor kernels, Parakeet, Orukeet}} +\date{} +\setlength{\textfloatsep}{10pt plus 2pt minus 2pt} +\setlength{\intextsep}{10pt plus 2pt minus 2pt} +\brokenpenalty=10000 +\makeatletter\setlength{\@fptop}{0pt}\makeatother +\begin{document} +\maketitle +\begin{abstract} +Orukeet replaces half of an adapted Parakeet encoder's temporal filters with +12,288 fitted Gabor kernels, freezes these replacements, and trains the remaining +parameters on multilingual and multi-accent data. Final adaptation and +checkpoint selection use LibriSpeech test-other. Across 20,146 FLEURS recordings +in 25 languages, pooled word error rate (WER) falls from Parakeet's +\FleursPooledParakeetWER\% to Orukeet's \FleursPooledOrukeetWER\%, a +\FleursPooledReduction\% relative reduction. Orukeet has lower WER on +\FleursWins\ of the 25 languages. Orukeet outperforms Parakeet on +\TestedWins\ out of \TestedSplits\ tested splits, including LibriSpeech +test-clean (\LibriCleanOrukeetWER\% vs.\ \LibriCleanParakeetWER\% WER), +test-other (\LibriOtherOrukeetWER\% vs.\ \LibriOtherParakeetWER\%), and +FLEURS English (\FleursEnglishOrukeetWER\% vs.\ \FleursEnglishParakeetWER\%). +All comparisons decode the same audio with matched NeMo settings. +The fitted kernels are stored as ordinary convolution weights, retaining +Parakeet's architecture and inference operators. +\end{abstract} + +\section{A fixed structure inside a learned recognizer} +Orukeet fixes selected temporal filters to fitted Gabor functions and trains +the rest of the recognizer. The starting architecture is +NVIDIA's Parakeet TDT 0.6B v3, a 25-language speech recognizer +\citep{parakeet}. Its FastConformer encoder has 24 blocks, each containing +1,024 nine-tap temporal depthwise kernels, and a token-and-duration transducer +(TDT) predicts text \citep{fastconformer,tdt}. We fit a separate Gabor function +to each kernel in a multilingual adaptation of this model, replace the closest +half, and hold the replacements fixed throughout subsequent training. + +The choice is made from the learned filters themselves. A close fit keeps the +initial change small; the remaining weights then accommodate that change. +Figure~\ref{fig:fits} shows the stored taps at four predetermined ranks in the +selected half. We compare the final checkpoint with stock Parakeet across +read speech, accents and domains. + +\begin{figure}[!ht] +\centering\includegraphics[width=\linewidth]{assets/kernel-fits.pdf} +\caption{Original kernels and fitted Gabor replacements at four predetermined +ranks spanning the selected half. Each pair is divided by the original kernel's +$L_2$ norm; percentages give relative RMS fit error. Markers show the nine +stored taps, joined by straight lines.} +\label{fig:fits} +\end{figure} + +\section{Fitting, freezing and adaptation} +For a nine-tap kernel $w_i$, we fit a Gabor function at +$t\in\{-4,\ldots,4\}$ and rank its normalized squared error: +\begin{equation} + g_i(t)=A_i\exp\!\left[-\frac{(t-\mu_i)^2}{2\sigma_i^2}\right] + \cos\!\left(2\pi f_i(t-\mu_i)+\phi_i\right),\qquad + e_i=\frac{\|w_i-g_i\|_2^2}{\|w_i\|_2^2}. + \label{eq:gabor} +\end{equation} +For fixed center $\mu$, width $\sigma$ and frequency $f$, linear least squares +solves for the cosine and sine coefficients, giving amplitude $A$ and phase +$\phi$. We evaluate 3,321 initial combinations and refine the best four plus +the best in each frequency quartile in float64, with at most 160 evaluations +per refinement. The search bounds are $\mu\in[-4,4]$, +$\sigma\in[0.25,36]$ and $f\in[10^{-6},0.499999]$ cycles per encoder timestep. +We keep the best evaluated fit and select the 12,288 lowest $e_i$ globally, +with layer and channel as deterministic tie-breaks. A replacement contains +only the fitted function, with no offset or learned residual. + +This ranking gives 175--748 fixed kernels per layer (Figure~\ref{fig:selection}). +The selected fits have 6.32\% median relative RMS error, a 13.30\% cutoff, +and pooled squared error equal to 0.4244\% of their original weight energy. +The 50\% constraint applies to temporal depthwise kernels. The materialized +model retains 627,008,134 scalar parameters; 110,592 stored taps are fixed and +626,897,542 remain trainable. Analytic audio filters have also been used in +learnable frontends \citep{leaf,sincnet}; here, each function approximates a +learned kernel inside the encoder. + +\begin{figure}[!ht] +\centering\includegraphics[width=\linewidth]{assets/selection-profile.pdf} +\caption{Global fit ranking and its allocation across layers. Left: cumulative +relative RMS error for all 24,576 kernels, with the selected half in blue. +Right: selected kernels in every encoder layer; the dashed line marks 512 +kernels. Selection uses one global ranking, so each layer need not be half fixed.} +\label{fig:selection} +\end{figure} + +Recovery trains the remaining network with transducer loss, teacher matching +at block and convolution outputs, and token/duration distillation. Subsequent +adaptation applies 4,035 AdamW updates with learning rate decaying from +$10^{-6}$ to $10^{-7}$. The final pass starts from those weights and performs +168 updates, with a 3\% warmup and cosine decay from $5\times10^{-6}$ to +$5\times10^{-7}$. AdamW uses $(0.9,0.98)$, weight decay 0.001 and gradient +clipping at 1.0. Microbatches contain at most 16 utterances and 120 padded +seconds; four microbatches form an update. Training uses BF16 on one A100 +40 GB, with dropout, augmentation and dithering disabled. + +The final pass makes three passes over the 2,939 LibriSpeech test-other +recordings. Targets preserve the parent's casing and punctuation while +correcting reference words; all targets match the reference under the pinned +English normalizer. Test-other also supplies the final checkpoint-selection +comparison. All 651 remaining parameter tensors change. An independent audit +of the exported checkpoint verifies that the 12,288 fitted kernels are +byte-identical to the original fitted functions; tokenizer assets, signal +processing buffers and batch-normalization statistics also remain unchanged. + +\section{Recognition across languages, accents and domains} +\paragraph{Matched comparison.} +We restore stock Parakeet and the final Orukeet checkpoint independently and +decode every recording with NeMo greedy-batch TDT, a ten-symbol limit, FP32 +weights and BF16 CUDA autocast. Matrix-multiply TF32 is disabled. Both models +receive identical mono 16 kHz audio, in the same duration-sorted batches; +empty transcripts remain in the scores. We compare the resulting Orukeet +checkpoint with pretrained Parakeet on the splits used throughout model +development. + +English uses a pinned text normalizer with spelling, number and compound +maps. Other languages retain diacritics and use language-specific number +normalization and compound-boundary alignment. Word errors count substitutions, +deletions and insertions. For a set of recordings $\mathcal D$, pooled WER is +\begin{equation} + \operatorname{WER}_{\mathrm{pool}}(\mathcal D) + =100\,\frac{\sum_{u\in\mathcal D}(S_u+D_u+I_u)} + {\sum_{u\in\mathcal D}N_u}, + \label{eq:pooled} +\end{equation} +where $N_u$ is the normalized reference word count. Compound alignment can +change this denominator separately for each model. A language macro instead +weights each language's WER equally. Integer counts and independent +full-partition rescoring reproduce every reported value. + +\paragraph{Read speech in 25 languages.} +The complete comparison contains 25,705 recordings: both LibriSpeech test +partitions \citep{librispeech} and FLEURS test speech in all 25 supported +languages \citep{fleurs}. Table~\ref{tab:current-standard} reports every partition. +FLEURS pooled WER is \FleursPooledParakeetWER\% for Parakeet and +\FleursPooledOrukeetWER\% for Orukeet; the corresponding language macros are +\FleursMacroParakeetWER\% and \FleursMacroOrukeetWER\%. +Orukeet improves \StandardWins\ of the 27 partitions, including +\FleursWins\ of 25 FLEURS languages. + +\input{current-standard-table.tex} + +\paragraph{Accents and domains.} +We also evaluate both models on a fixed sample of 12,006 recordings across +47 partitions and 25 languages. It includes EuroSpeech, GigaSpeechBench, +Monsoon, Golos, NST, VoxPopuli and Lesbos: 256 recordings per partition and all +230 Lesbos recordings. The preceding 4,035-update adaptation includes 6,118 +of these recordings. Greek and Italian EuroSpeech use the +audited human transcript spans. Both checkpoints are decoded afresh and scored +with the same pinned protocol as the read-speech comparison. +Pooled WER is \DomainPooledParakeetWER\% versus \DomainPooledOrukeetWER\%; +for the 5,120 English recordings, it is \DomainEnglishParakeetWER\% versus +\DomainEnglishOrukeetWER\%. Orukeet improves \DomainWins\ of 47 partitions, +including \DomainEnglishWins\ of 20 English partitions. Table~\ref{tab:current-domains} +gives every score. We pool read speech and accent/domain recordings separately. + +\section{Checkpoint and reproduction} +All recognition results in this report refer to the NeMo checkpoint with +SHA-256 prefix \texttt{031c8ddab484}. The file stores the configuration, +tokenizer and materialized convolution weights. The fitting code retains each +kernel's analytic parameters; training uses a fixed parametrization to prevent +updates to the selected rows. Inference uses ordinary depthwise convolution, +with the same tensor shapes and operator counts as Parakeet. + +The \href{https://github.com/Oruk-AI/orukeet}{Orukeet repository} and +\href{https://huggingface.co/oruk/orukeet/resolve/555136b50265a132d4cea0d35560c26fc4f657ab/orukeet-v0.1.0.nemo}{release checkpoint} +contain the source weights, training recipes, fitted functions, export audits +and reproducible evaluation records. The metric bundle retains unrounded +scores, per-record edit counts, manifest identities and checkpoint hashes. +Code is MIT; weights and fits are CC BY-SA 4.0; metric records are CC BY 4.0. +NVIDIA's foundation attribution is retained. + +\par\begin{minipage}{\linewidth} +\small +\bibliographystyle{plainnat} +\bibliography{references} +\end{minipage} +\input{current-domains-table.tex} +\end{document} diff --git a/report/references.bib b/report/references.bib new file mode 100644 index 0000000000000000000000000000000000000000..1277aa5c86fd19826b767d802b32fa3c0c6f1036 --- /dev/null +++ b/report/references.bib @@ -0,0 +1,42 @@ +@misc{parakeet, + author={{NVIDIA}}, title={Parakeet TDT 0.6B v3: Model Card}, year={2025}, + url={https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3}, note={Accessed September 6, 2026}} +@article{fastconformer, + author={Rekesh, Dima and Koluguri, Nithin Rao and Kriman, Samuel and others}, + title={Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition}, + journal={arXiv:2305.05084}, year={2023}, url={https://arxiv.org/abs/2305.05084}} +@inproceedings{tdt, + author={Xu, Hainan and Jia, Fei and Majumdar, Somshubra and Huang, He and Watanabe, Shinji and Ginsburg, Boris}, + title={Efficient Sequence Transduction by Jointly Predicting Tokens and Durations}, + booktitle={ICML}, year={2023}, url={https://proceedings.mlr.press/v202/xu23g.html}} +@inproceedings{leaf, + author={Zeghidour, Neil and Teboul, Olivier and de Chaumont Quitry, F{\'e}lix and Tagliasacchi, Marco}, + title={{LEAF}: A Learnable Frontend for Audio Classification}, booktitle={ICLR}, year={2021}, + url={https://openreview.net/forum?id=jM76BCb6F9m}} +@inproceedings{sincnet, + author={Ravanelli, Mirco and Bengio, Yoshua}, + title={Speaker Recognition from Raw Waveform with {SincNet}}, booktitle={SLT}, year={2018}, + url={https://arxiv.org/abs/1808.00158}} +@misc{fleurs, + author={Conneau, Alexis and Ma, Min and Khanuja, Simran and others}, + title={{FLEURS}: Few-shot Learning Evaluation of Universal Representations of Speech}, + year={2022}, howpublished={arXiv:2205.12446}, url={https://arxiv.org/abs/2205.12446}} +@inproceedings{librispeech, + author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev}, + title={{LibriSpeech}: An {ASR} Corpus Based on Public Domain Audio Books}, + booktitle={ICASSP}, pages={5206--5210}, year={2015}, + doi={10.1109/ICASSP.2015.7178964}, url={https://www.openslr.org/12}} +@misc{native, + author={{NVIDIA}}, title={{NeMo-Speech.cpp}, version 0.1.0}, year={2026}, + url={https://github.com/NVIDIA/NeMo-Speech.cpp/tree/v0.1.0}} +@inproceedings{pruning, + author={Jiang, Huiqiang and Zhang, Li Lyna and Li, Yuang and others}, + title={Accurate and Structured Pruning for Efficient Automatic Speech Recognition}, + booktitle={Interspeech}, year={2023}, url={https://arxiv.org/abs/2305.19549}} +@inproceedings{sarwar, + title={Gabor Filter Assisted Energy Efficient Fast Learning Convolutional Neural Networks}, + author={Sarwar, Syed Shakib and Panda, Priyadarshini and Roy, Kaushik}, + booktitle={ISLPED}, + year={2017}, + url={https://arxiv.org/abs/1705.04748} +} diff --git a/report/standard-benchmark-table.tex b/report/standard-benchmark-table.tex new file mode 100644 index 0000000000000000000000000000000000000000..1971bee2526a269399753ba880071d82197e1904 --- /dev/null +++ b/report/standard-benchmark-table.tex @@ -0,0 +1,46 @@ +\begin{table}[!ht] +\centering +\caption{Complete LibriSpeech and FLEURS test partitions. Both models use matched NeMo greedy decoding; WER and CER are percentages. Macro rows weight languages equally. Lower is better.} +\label{tab:standard-benchmarks} +\small +\setlength{\tabcolsep}{5pt} +\renewcommand{\arraystretch}{1.05} +\begin{tabular}{lrrrrr} +\toprule +Benchmark & Clips & \multicolumn{2}{c}{Parakeet} & \multicolumn{2}{c}{Orukeet}\\ + & & WER & CER & WER & CER\\ +\midrule +LibriSpeech test-clean & 2,620 & 1.53 & 0.59 & 1.50 & 0.58\\ +LibriSpeech test-other & 2,939 & 3.14 & 1.32 & 3.25 & 1.39\\ +\midrule +FLEURS Bulgarian & 658 & 11.92 & 3.84 & 10.58 & 3.42\\ +FLEURS Croatian & 914 & 11.29 & 3.53 & 10.42 & 3.84\\ +FLEURS Czech & 723 & 11.12 & 3.21 & 9.20 & 2.74\\ +FLEURS Danish & 930 & 17.19 & 6.31 & 14.97 & 5.34\\ +FLEURS Dutch & 364 & 6.40 & 2.28 & 5.62 & 1.97\\ +FLEURS English & 647 & 4.28 & 2.00 & 3.87 & 1.80\\ +FLEURS Estonian & 893 & 13.32 & 3.86 & 10.62 & 3.53\\ +FLEURS Finnish & 918 & 11.14 & 2.59 & 9.52 & 2.20\\ +FLEURS French & 676 & 4.69 & 1.68 & 5.05 & 1.73\\ +FLEURS German & 862 & 4.21 & 1.41 & 3.94 & 1.55\\ +FLEURS Greek & 650 & 21.07 & 9.01 & 31.39 & 9.65\\ +FLEURS Hungarian & 905 & 13.60 & 4.20 & 10.86 & 3.10\\ +FLEURS Italian & 865 & 2.43 & 0.79 & 2.09 & 0.75\\ +FLEURS Latvian & 851 & 21.78 & 5.43 & 17.60 & 4.31\\ +FLEURS Lithuanian & 986 & 20.95 & 5.56 & 16.93 & 4.38\\ +FLEURS Maltese & 926 & 19.22 & 6.19 & 15.83 & 5.20\\ +FLEURS Polish & 758 & 6.81 & 2.09 & 6.21 & 2.00\\ +FLEURS Portuguese & 919 & 4.49 & 1.98 & 3.74 & 1.65\\ +FLEURS Romanian & 883 & 11.44 & 3.86 & 9.53 & 3.17\\ +FLEURS Russian & 775 & 4.89 & 1.49 & 4.84 & 1.54\\ +FLEURS Slovak & 792 & 9.21 & 2.91 & 7.77 & 2.43\\ +FLEURS Slovenian & 834 & 22.62 & 7.70 & 21.54 & 7.57\\ +FLEURS Spanish & 908 & 3.22 & 1.28 & 2.77 & 1.04\\ +FLEURS Swedish & 759 & 13.38 & 4.26 & 11.48 & 3.48\\ +FLEURS Ukrainian & 750 & 6.00 & 1.74 & 5.69 & 1.71\\ +\midrule +FLEURS five-language macro & 4,230 & 3.81 & 1.43 & 3.52 & 1.34\\ +FLEURS 25-language macro & 20,146 & 11.07 & 3.57 & 10.08 & 3.21\\ +\bottomrule +\end{tabular} +\end{table} diff --git a/report/standard-benchmark-values.tex b/report/standard-benchmark-values.tex new file mode 100644 index 0000000000000000000000000000000000000000..7a1bb73dc1bd088c1b105ea5fb15263e600595ff --- /dev/null +++ b/report/standard-benchmark-values.tex @@ -0,0 +1,13 @@ +% Generated from complete-test edit counts. +\newcommand{\LibriCleanParakeetWER}{1.53} +\newcommand{\LibriCleanOrukeetWER}{1.50} +\newcommand{\LibriOtherParakeetWER}{3.14} +\newcommand{\LibriOtherOrukeetWER}{3.25} +\newcommand{\FleursEnglishParakeetWER}{4.28} +\newcommand{\FleursEnglishOrukeetWER}{3.87} +\newcommand{\FleursFiveParakeetWER}{3.81} +\newcommand{\FleursFiveOrukeetWER}{3.52} +\newcommand{\FleursAllParakeetWER}{11.07} +\newcommand{\FleursAllOrukeetWER}{10.08} +\newcommand{\StandardTestRows}{25,705} +\newcommand{\FleursTestRows}{20,146} diff --git a/selection-profile.png b/selection-profile.png new file mode 100644 index 0000000000000000000000000000000000000000..b49d2dd6ada91ba0ad2c4506e13ff944e4bc10e8 Binary files /dev/null and b/selection-profile.png differ diff --git a/team.json b/team.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e6b498a5c0c0b94b3c05e4fd8fe5ca246a8bdd --- /dev/null +++ b/team.json @@ -0,0 +1,123 @@ +{ + "primary_organization": "Oruk AI", + "authors": [ + { + "name": "Nathan Roll", + "affiliations": [ + 1, + 2 + ] + }, + { + "name": "Irene Yi", + "affiliations": [ + 1, + 2 + ] + }, + { + "name": "Büşra Marşan", + "affiliations": [ + 1, + 2 + ] + }, + { + "name": "Vianney Grenez", + "affiliations": [ + 1 + ] + }, + { + "name": "Gabriel Stein", + "affiliations": [ + 4 + ] + }, + { + "name": "Momcilo Mrkaic", + "affiliations": [ + 5 + ] + }, + { + "name": "Pavle Padjin", + "affiliations": [ + 5 + ] + }, + { + "name": "Vladimir Zeljkovic", + "affiliations": [ + 5 + ] + }, + { + "name": "Calbert Graham", + "affiliations": [ + 1, + 3 + ] + } + ], + "affiliations": [ + { + "id": 1, + "name": "Oruk AI", + "slug": "oruk", + "primary": true, + "source": "report/assets/oruk-lockup.png", + "source_sha256": "932df9164b237a1a31d1e636897d1766a15b8bac738acf81ee93de8ea2d6c18c", + "web_asset": "docs/assets/team/oruk.png", + "web_asset_sha256": "717fb1a29a3faf8cf62baa924c65e77afb8fb82180e7398612a01a672af014da", + "display_width": 184 + }, + { + "id": 2, + "name": "Stanford University", + "slug": "stanford", + "primary": false, + "source": "report/assets/affiliations/stanford.png", + "source_sha256": "09434daceea1a4ab62da5db3cd9e81bbe1a7e034224f504d63fd3f12e51af74b", + "web_asset": "docs/assets/team/stanford.png", + "web_asset_sha256": "6521b4f360f081fb9322b05ff4740fe8ad2b7c720ee1a9e3f7b9935624f4abc2", + "display_width": 144 + }, + { + "id": 3, + "name": "University of Cambridge", + "slug": "cambridge", + "primary": false, + "source": "report/assets/affiliations/cambridge.svg", + "source_sha256": "27381d019c0a54c53b7b746750dc148f35dca2a6fe96b39dff5f896b0e745902", + "web_asset": "docs/assets/team/cambridge.png", + "web_asset_sha256": "a3faf2131e9e2df035d4b21d076c0eea606669c7fb46283b7c90041880a632fd", + "display_width": 144 + }, + { + "id": 4, + "name": "OpenWhispr", + "slug": "openwhispr", + "primary": false, + "source": "report/assets/affiliations/openwhispr.svg", + "source_sha256": "29d50f34b1ffe4b11665aaeefdaf63a22796c53c38234cd9d6f06c5153c9e13a", + "web_asset": "docs/assets/team/openwhispr.png", + "web_asset_sha256": "564f133d0c0b5c87cd6384f9ce2689fffa6da68b004f303d2e278322e213ac62", + "display_width": 40 + }, + { + "id": 5, + "name": "Hoid", + "slug": "hoid", + "primary": false, + "source": "report/assets/affiliations/hoid.svg", + "source_sha256": "49b5fc2265f1294abf6bb65f054078b640b9fbaed9780e2817ab6a103cd2eb72", + "web_asset": "docs/assets/team/hoid.png", + "web_asset_sha256": "b5a252792faf5ce80c3f1aee671867524c97e2d85de3808a0c93f0c01a4c33d8", + "display_width": 76 + } + ], + "source_report": "report/paper.tex", + "citation": "CITATION.cff", + "presentation": "Oruk leads each page; the remaining marks label author affiliations in report order." +}