fa-bench 2.1.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (146) hide show
  1. fa_bench-2.1.0.dist-info/METADATA +430 -0
  2. fa_bench-2.1.0.dist-info/RECORD +146 -0
  3. fa_bench-2.1.0.dist-info/WHEEL +5 -0
  4. fa_bench-2.1.0.dist-info/entry_points.txt +2 -0
  5. fa_bench-2.1.0.dist-info/licenses/LICENSE +202 -0
  6. fa_bench-2.1.0.dist-info/licenses/NOTICE +26 -0
  7. fa_bench-2.1.0.dist-info/licenses/fabench/dataprep/noisemix/kaldi/COPYING +364 -0
  8. fa_bench-2.1.0.dist-info/top_level.txt +1 -0
  9. fabench/__init__.py +11 -0
  10. fabench/__main__.py +9 -0
  11. fabench/aligners/__init__.py +96 -0
  12. fabench/aligners/base.py +110 -0
  13. fabench/aligners/bfa/__init__.py +7 -0
  14. fabench/aligners/bfa/adapter.py +32 -0
  15. fabench/aligners/charsiu/__init__.py +7 -0
  16. fabench/aligners/charsiu/adapter.py +30 -0
  17. fabench/aligners/crisperwhisper_fa/__init__.py +7 -0
  18. fabench/aligners/crisperwhisper_fa/adapter.py +38 -0
  19. fabench/aligners/elevenlabs_fa/__init__.py +7 -0
  20. fabench/aligners/elevenlabs_fa/adapter.py +29 -0
  21. fabench/aligners/falcon/__init__.py +6 -0
  22. fabench/aligners/falcon/adapter.py +79 -0
  23. fabench/aligners/maps/__init__.py +7 -0
  24. fabench/aligners/maps/adapter.py +171 -0
  25. fabench/aligners/mfa/__init__.py +7 -0
  26. fabench/aligners/mfa/adapter.py +191 -0
  27. fabench/aligners/nemo_fa/__init__.py +6 -0
  28. fabench/aligners/nemo_fa/adapter.py +47 -0
  29. fabench/aligners/neufa/__init__.py +6 -0
  30. fabench/aligners/neufa/adapter.py +37 -0
  31. fabench/aligners/nyra_fa/__init__.py +6 -0
  32. fabench/aligners/nyra_fa/adapter.py +27 -0
  33. fabench/aligners/olign/__init__.py +20 -0
  34. fabench/aligners/olign/adapter.py +236 -0
  35. fabench/aligners/olign_v1/__init__.py +7 -0
  36. fabench/aligners/olign_v1/adapter.py +27 -0
  37. fabench/aligners/pocketsphinx/__init__.py +8 -0
  38. fabench/aligners/pocketsphinx/adapter.py +24 -0
  39. fabench/aligners/qwen3_fa/__init__.py +7 -0
  40. fabench/aligners/qwen3_fa/adapter.py +160 -0
  41. fabench/aligners/relabel.py +122 -0
  42. fabench/aligners/runner.py +327 -0
  43. fabench/aligners/stable_ts/__init__.py +7 -0
  44. fabench/aligners/stable_ts/adapter.py +41 -0
  45. fabench/aligners/subprocess_aligner.py +245 -0
  46. fabench/aligners/torchaudio_fa/__init__.py +7 -0
  47. fabench/aligners/torchaudio_fa/adapter.py +64 -0
  48. fabench/aligners/unity2/__init__.py +6 -0
  49. fabench/aligners/unity2/adapter.py +32 -0
  50. fabench/aligners/whisperx/__init__.py +7 -0
  51. fabench/aligners/whisperx/adapter.py +31 -0
  52. fabench/audio.py +148 -0
  53. fabench/cli.py +275 -0
  54. fabench/config.py +618 -0
  55. fabench/config.yaml +58 -0
  56. fabench/dataprep/__init__.py +18 -0
  57. fabench/dataprep/datasets/__init__.py +171 -0
  58. fabench/dataprep/datasets/acquire.py +131 -0
  59. fabench/dataprep/datasets/en/__init__.py +11 -0
  60. fabench/dataprep/datasets/en/buckeye/__init__.py +17 -0
  61. fabench/dataprep/datasets/en/buckeye/download.py +55 -0
  62. fabench/dataprep/datasets/en/buckeye/processor.py +588 -0
  63. fabench/dataprep/datasets/en/l2arctic/__init__.py +11 -0
  64. fabench/dataprep/datasets/en/l2arctic/download.py +44 -0
  65. fabench/dataprep/datasets/en/l2arctic/processor.py +84 -0
  66. fabench/dataprep/datasets/en/timit/__init__.py +17 -0
  67. fabench/dataprep/datasets/en/timit/download.py +51 -0
  68. fabench/dataprep/datasets/en/timit/processor.py +241 -0
  69. fabench/dataprep/datasets/export_refs.py +138 -0
  70. fabench/dataprep/datasets/plausibility.py +84 -0
  71. fabench/dataprep/datasets/runner.py +88 -0
  72. fabench/dataprep/datasets/textgrid.py +44 -0
  73. fabench/dataprep/noisemix/__init__.py +10 -0
  74. fabench/dataprep/noisemix/datadirs.py +114 -0
  75. fabench/dataprep/noisemix/kaldi/COPYING +364 -0
  76. fabench/dataprep/noisemix/kaldi/PROVENANCE.md +76 -0
  77. fabench/dataprep/noisemix/kaldi/steps/data/augment_data_dir.py +303 -0
  78. fabench/dataprep/noisemix/kaldi/steps/data/data_dir_manipulation_lib.py +17 -0
  79. fabench/dataprep/noisemix/kaldi/steps/data/make_musan.py +178 -0
  80. fabench/dataprep/noisemix/kaldi/steps/data/make_musan.sh +71 -0
  81. fabench/dataprep/noisemix/kaldi/steps/data/reverberate_data_dir.py +687 -0
  82. fabench/dataprep/noisemix/kaldi/steps/libs/__init__.py +11 -0
  83. fabench/dataprep/noisemix/kaldi/steps/libs/common.py +498 -0
  84. fabench/dataprep/noisemix/make_noisy.py +346 -0
  85. fabench/dataprep/noisemix/make_noisy.sh +167 -0
  86. fabench/dataprep/noisemix/manifest.py +102 -0
  87. fabench/dataprep/noisemix/materialise.py +73 -0
  88. fabench/dataprep/noisemix/prepare.py +237 -0
  89. fabench/dataprep/noisemix/shadow_root.py +251 -0
  90. fabench/dataprep/noisemix/utt2dur.py +60 -0
  91. fabench/envfile.py +102 -0
  92. fabench/gates.py +261 -0
  93. fabench/init.py +318 -0
  94. fabench/noise/__init__.py +2 -0
  95. fabench/noise/config.yaml +34 -0
  96. fabench/noise/manifest.py +112 -0
  97. fabench/noise/mixer.py +149 -0
  98. fabench/noise/musan.py +165 -0
  99. fabench/noise/provider.py +63 -0
  100. fabench/noise/runner.py +90 -0
  101. fabench/noise/synth.py +44 -0
  102. fabench/normalize/__init__.py +93 -0
  103. fabench/normalize/check.py +52 -0
  104. fabench/normalize/en/__init__.py +91 -0
  105. fabench/normalize/en/canonical.py +89 -0
  106. fabench/normalize/en/config.yaml +12 -0
  107. fabench/normalize/en/maps.py +208 -0
  108. fabench/paths.py +211 -0
  109. fabench/pipeline.py +121 -0
  110. fabench/report/__init__.py +2 -0
  111. fabench/report/runner.py +76 -0
  112. fabench/schema.py +314 -0
  113. fabench/score/__init__.py +2 -0
  114. fabench/score/aggregate.py +159 -0
  115. fabench/score/boundary.py +125 -0
  116. fabench/score/config.yaml +29 -0
  117. fabench/score/core.py +363 -0
  118. fabench/score/matched.py +206 -0
  119. fabench/score/runner.py +250 -0
  120. fabench/score/segmentation.py +205 -0
  121. fabench/score/word.py +129 -0
  122. fabench/selftest.py +231 -0
  123. fabench/timestamp_asrs/__init__.py +41 -0
  124. fabench/timestamp_asrs/base.py +82 -0
  125. fabench/timestamp_asrs/cloud/__init__.py +37 -0
  126. fabench/timestamp_asrs/cloud/adapter.py +746 -0
  127. fabench/timestamp_asrs/cloud/awssig.py +159 -0
  128. fabench/timestamp_asrs/cloud/http.py +232 -0
  129. fabench/timestamp_asrs/cloud/providers.py +1264 -0
  130. fabench/timestamp_asrs/crisperwhisper/__init__.py +7 -0
  131. fabench/timestamp_asrs/crisperwhisper/adapter.py +27 -0
  132. fabench/timestamp_asrs/granite_speech_plus/__init__.py +7 -0
  133. fabench/timestamp_asrs/granite_speech_plus/adapter.py +28 -0
  134. fabench/timestamp_asrs/parakeet_tdt/__init__.py +7 -0
  135. fabench/timestamp_asrs/parakeet_tdt/adapter.py +25 -0
  136. fabench/timestamp_asrs/qwen3_asr/__init__.py +7 -0
  137. fabench/timestamp_asrs/qwen3_asr/adapter.py +30 -0
  138. fabench/timestamp_asrs/subprocess_asr.py +45 -0
  139. fabench/timestamp_asrs/torchaudio_asr/__init__.py +7 -0
  140. fabench/timestamp_asrs/torchaudio_asr/adapter.py +25 -0
  141. fabench/timestamp_asrs/whisper3/__init__.py +7 -0
  142. fabench/timestamp_asrs/whisper3/adapter.py +26 -0
  143. fabench/timestamp_asrs/whisper_ts/__init__.py +6 -0
  144. fabench/timestamp_asrs/whisper_ts/adapter.py +27 -0
  145. fabench/timestamp_asrs/whisperx_asr/__init__.py +6 -0
  146. fabench/timestamp_asrs/whisperx_asr/adapter.py +28 -0
@@ -0,0 +1,430 @@
1
+ Metadata-Version: 2.4
2
+ Name: fa-bench
3
+ Version: 2.1.0
4
+ Summary: Word and phone timestamp accuracy of forced aligners and timestamped ASR on clean and noisy speech, scored against hand-marked TIMIT and Buckeye boundaries.
5
+ Author: Olewave, LLC
6
+ License-Expression: Apache-2.0
7
+ Project-URL: Homepage, https://github.com/olewave/fa-bench
8
+ Project-URL: Leaderboard, https://olewave.com/en/leaderboard/fa-bench
9
+ Project-URL: Paper, https://arxiv.org/abs/2609.32396
10
+ Project-URL: Changelog, https://github.com/olewave/fa-bench/blob/main/CHANGELOG.md
11
+ Classifier: Programming Language :: Python :: 3
12
+ Classifier: Intended Audience :: Science/Research
13
+ Classifier: Topic :: Scientific/Engineering :: Information Analysis
14
+ Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
15
+ Requires-Python: >=3.11
16
+ Description-Content-Type: text/markdown
17
+ License-File: LICENSE
18
+ License-File: NOTICE
19
+ License-File: fabench/dataprep/noisemix/kaldi/COPYING
20
+ Requires-Dist: numpy
21
+ Requires-Dist: scipy
22
+ Requires-Dist: soundfile
23
+ Requires-Dist: pandas
24
+ Requires-Dist: pyarrow
25
+ Requires-Dist: pyyaml
26
+ Requires-Dist: praatio
27
+ Requires-Dist: tqdm
28
+ Requires-Dist: requests
29
+ Provides-Extra: test
30
+ Requires-Dist: pytest; extra == "test"
31
+ Provides-Extra: maps
32
+ Requires-Dist: g2p-en; extra == "maps"
33
+ Dynamic: license-file
34
+
35
+ <p align="center">
36
+ <img src="https://raw.githubusercontent.com/olewave/fa-bench/v2.1.0/docs/FA-Bench.jpeg" alt="FA-Bench" width="200">
37
+ </p>
38
+
39
+ <p align="center">
40
+ <a href="https://olewave.com/en/leaderboard/fa-bench"><img src="https://img.shields.io/badge/leaderboard-latest-orange" alt="Leaderboard (latest)"></a>
41
+ <a href="https://arxiv.org/abs/2609.32396"><img src="https://img.shields.io/badge/arXiv-2609.32396-b31b1b" alt="Paper on arXiv"></a>
42
+ <a href="https://github.com/olewave/fa-bench/actions/workflows/ci.yml"><img src="https://github.com/olewave/fa-bench/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
43
+ <a href="https://github.com/olewave/fa-bench/blob/v2.1.0/LICENSE"><img src="https://img.shields.io/badge/license-Apache%202.0-blue" alt="Apache 2.0 license"></a>
44
+ <img src="https://img.shields.io/badge/python-3.11%2B-blue" alt="Python 3.11+">
45
+ </p>
46
+
47
+ # <a href="https://arxiv.org/abs/2609.32396" title="The FA-Bench paper on arXiv"><img src="https://raw.githubusercontent.com/olewave/fa-bench/v2.1.0/docs/pdf-icon.svg" alt="Paper (PDF)" height="32"></a> FA-Bench: A Benchmark for Evaluating Phone- and Word-Level Timestamp Accuracy in Forced Aligners
48
+
49
+ <p align="right">
50
+ <b> Click ⭐ at the top right to save FA-Bench into your speech toolbox. Save it now! ↗️</b><br>
51
+ <b> 点击右上角⭐,将FA-Bench收藏进您的语音工具箱,现在就收藏! ↗️</b><br>
52
+ <b> 오른쪽 상단의 ⭐를 클릭해 FA-Bench를 음성 툴박스에 저장하세요. 지금 바로! ↗️</b><br>
53
+ <b> 右上の⭐をクリックして、FA-Bench を音声ツールボックスに保存しましょう。今すぐ!↗️</b><br>
54
+ <b dir="rtl"> انقر على ⭐ في أعلى اليمين لحفظ FA-Bench في صندوق أدوات الكلام لديك. احفظه الآن! ↗️</b>
55
+ </p>
56
+
57
+ FA-Bench measures how well a system places word and phone boundaries in
58
+ speech, and how much of that accuracy survives noise. The reference is the
59
+ boundaries that human linguists marked by hand in two English corpora, read
60
+ speech in [TIMIT](https://github.com/olewave/fa-bench/blob/v2.1.0/records/README.md#about-timit)
61
+ and conversation in
62
+ [Buckeye](https://github.com/olewave/fa-bench/blob/v2.1.0/records/README.md#about-buckeye). We add
63
+ **no new annotation**. Times are used as annotated and labels are folded into
64
+ the shared TIMIT-39 phone set. The [FA-Bench paper](https://arxiv.org/abs/2609.32396) describes
65
+ the protocol and what the results show.
66
+
67
+ Every system runs on **clean** audio and on
68
+ [four degradations](https://github.com/olewave/fa-bench/blob/v2.1.0/records/README.md#the-four-conditions)
69
+ (`reverb`, `noise`, `music`, `babble`). Systems are scored in **two tracks
70
+ that never share a leaderboard**. A forced aligner is given the reference
71
+ transcript, so only its timing is tested. A timestamped ASR decodes its own
72
+ words, so its timing error carries its recognition error too. Every run is
73
+ seeded and reproducible from a single command.
74
+
75
+ **33 systems scored**, 24 open models and 9 commercial APIs.
76
+
77
+ - **Forced aligners**, given the reference transcript.
78
+ [BFA](https://github.com/tabahi/bournemouth-forced-aligner) ·
79
+ [Charsiu](https://github.com/lingjzhu/charsiu) ·
80
+ [CrisperWhisper](https://github.com/nyrahealth/CrisperWhisper) ·
81
+ [FALCON](https://github.com/MLSpeech/FALCON) ·
82
+ [MAPS](https://github.com/MasonPhonLab/MAPS) ·
83
+ [Montreal Forced Aligner (MFA)](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) 2.0 and 3.4 ·
84
+ [MMS-FA](https://docs.pytorch.org/audio/stable/generated/torchaudio.pipelines.MMS_FA.html) ·
85
+ [NeMo Forced Aligner](https://github.com/NVIDIA-NeMo/Speech/tree/main/tools/nemo_forced_aligner), Conformer and FastConformer checkpoints ·
86
+ [NeuFA](https://github.com/thuhcsi/NeuFA) ·
87
+ [nyra forced aligner](https://huggingface.co/nyralabs/nyra_forced_aligner_en) ·
88
+ [PocketSphinx](https://github.com/cmusphinx/pocketsphinx) ·
89
+ [Qwen3-ForcedAligner](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B) ·
90
+ [stable-ts](https://github.com/jianfch/stable-ts) ·
91
+ [TorchAudio-FA](https://github.com/pytorch/audio) ·
92
+ [UnitY2](https://github.com/facebookresearch/seamless_communication) ·
93
+ [WhisperX](https://github.com/m-bain/whisperx)
94
+ - **Open ASR with timestamps**, decoding their own words.
95
+ [Granite Speech 4.1 Plus](https://huggingface.co/ibm-granite/granite-speech-4.1-2b-plus) ·
96
+ [Parakeet-TDT](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) ·
97
+ [Qwen3-ASR](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) ·
98
+ [TorchAudio wav2vec2 ASR](https://docs.pytorch.org/audio/stable/generated/torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H.html) ·
99
+ [Whisper large-v3](https://huggingface.co/openai/whisper-large-v3) ·
100
+ [Whisper-timestamped](https://github.com/linto-ai/whisper-timestamped).
101
+ CrisperWhisper runs here too.
102
+ - **Commercial APIs.**
103
+ [Olign](https://olewave.com/en/olign-olewaves-lancet-accurate-speech-to-text-forced-alignment-service/) (forced alignment) ·
104
+ [Amazon Transcribe](https://aws.amazon.com/transcribe/) ·
105
+ [AssemblyAI Universal 3.5](https://www.assemblyai.com/) ·
106
+ [Azure AI Speech](https://azure.microsoft.com/en-us/products/ai-foundry/tools/speech) ·
107
+ [Deepgram Nova-3](https://deepgram.com/) ·
108
+ [ElevenLabs Scribe v2](https://elevenlabs.io/speech-to-text) ·
109
+ [Google Chirp 2](https://docs.cloud.google.com/speech-to-text/docs/models/chirp-2) ·
110
+ [IBM Watson Speech to Text](https://www.ibm.com/products/speech-to-text) ·
111
+ [Speechmatics](https://www.speechmatics.com/)
112
+
113
+ We also score 17 cascades, in which one of the aligners re-times the words an
114
+ ASR produced. A cascade holds the recognizer constant, which an ASR cannot, so
115
+ it is not ranked among the ASRs. It is reported beside the aligner on Track 1
116
+ instead, as what recognition errors cost that aligner.
117
+
118
+ **Latest results** are on the
119
+ [leaderboard](https://olewave.com/en/leaderboard/fa-bench).
120
+ [`records/`](https://github.com/olewave/fa-bench/blob/v2.1.0/records/README.md) lists the snapshots and describes the corpora
121
+ and the four conditions.
122
+
123
+ > **Licensed under the [Apache License 2.0](https://github.com/olewave/fa-bench/blob/v2.1.0/LICENSE)** from release 2.0.0.
124
+ > Copyright 2026 Olewave, LLC. The name FA-Bench stays with Olewave, as
125
+ > [the name policy](#the-name) explains. Contributions are welcome under
126
+ > [`CONTRIBUTING.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/CONTRIBUTING.md), and vendored third-party code keeps its
127
+ > own licence ([`NOTICE`](https://github.com/olewave/fa-bench/blob/v2.1.0/NOTICE)).
128
+
129
+ ---
130
+
131
+ ## Install
132
+
133
+ ```bash
134
+ uv venv --python 3.12 .venv && . .venv/bin/activate
135
+ uv pip install -e ".[test]" # the core library and pytest
136
+ ```
137
+
138
+ The scorer is also published on PyPI as `fa-bench`, and `pip install
139
+ fa-bench` gives the library and the `fabench` command. The commands that build
140
+ and run the benchmark read the dataset definitions and system recipes from a
141
+ checkout of this repository, so clone it, or unpack the sdist, to reproduce
142
+ the published numbers.
143
+
144
+ That is enough to run the self-test and the sanity gates. Each system
145
+ installs itself from its own recipe,
146
+ `evals/<kind>/<tool>/download_and_install.sh`, into its own environment. See
147
+ [`evals/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/evals/README.md).
148
+
149
+ ---
150
+
151
+ ## Quickstart
152
+
153
+ ```bash
154
+ fabench selftest # prove the whole chain with NO licensed data
155
+ fabench gates # sanity gates that need no data
156
+ fabench init # say where your corpora are (interactive)
157
+ fabench config # inspect the fully resolved run
158
+ ```
159
+
160
+ If you see `fabench: command not found`, the venv is not active. Run
161
+ `. .venv/bin/activate`, or call `./bin/fabench`, which always uses this
162
+ checkout.
163
+
164
+ `gates` is the known-answer check on the measurement chain. Run it on a fresh
165
+ clone. Every gate has one answer fixed in advance, so a broken chain fails
166
+ loudly instead of reporting plausible numbers. Gate #2 needs licensed audio
167
+ and runs inside `fabench ingest`. `selftest` builds the synthetic corpus the
168
+ other gates use.
169
+
170
+ ---
171
+
172
+ ## Running it
173
+
174
+ ```bash
175
+ fabench ingest # staged corpora -> canonical gold, plus the plausibility gate
176
+ fabench noise fetch # MUSAN (about 11 GB, once), only when needed
177
+ fabench run # mix -> align -> score -> report
178
+ ```
179
+
180
+ `fabench run` mixes its own noise on the fly, white, pink, ambient and babble
181
+ noise at set SNRs (`fabench/noise/config.yaml`). It is a quick check that the
182
+ chain works. **The published noisy numbers use a different set.** They come
183
+ from the four Kaldi conditions, reverb, noise, music and babble.
184
+ `fabench prepare` builds them from your staged corpora, together with the
185
+ shadow roots every noisy cell reads and a sha256 manifest of both
186
+ ([`datasets/prep/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/datasets/prep/README.md)).
187
+
188
+ ```bash
189
+ fabench prepare --corpus timit --subset core_test --out /data/fabench/noisy
190
+ ```
191
+
192
+ Reproducing a published number takes three commands. `prepare` builds the
193
+ degraded audio, `run` runs one system on one cell through its adapter, and
194
+ `score` turns the saved outputs into the headline metrics. A cell config names
195
+ one system and one cell, `<corpus>/<subset>/<condition>`, and
196
+ `evals/gen_config.py` writes one for any cell.
197
+
198
+ ```bash
199
+ fabench prepare --corpus timit,buckeye --out /data/fabench/noisy # the four degraded conditions
200
+ fabench run --config <cell>.yaml # one system on one cell
201
+ fabench score --config <cell>.yaml # saved outputs -> the published metrics
202
+ ```
203
+
204
+ For sweeps across systems and cells, use the staged driver. It takes the tools
205
+ to run by name. With no names it runs `mfa charsiu maps bfa`.
206
+
207
+ ```bash
208
+ evals/run_evals.sh mfa # align clean, then score
209
+ evals/run_evals.sh --use-noisy-dataset mfa # the same cells on the Kaldi conditions
210
+ evals/run_evals.sh --stage 3 mfa # rescore from saved output, no GPU
211
+ ```
212
+
213
+ Alignments are kept on disk, so a change to a metric or a grouping is a
214
+ **rescore** (`evals/run_evals.sh --stage 3`), which takes seconds per cell
215
+ instead of GPU hours.
216
+
217
+ TIMIT and Buckeye are **licensed, and you stage them yourself**. FA-Bench never
218
+ downloads them. `fabench init` asks where they are and writes `.fabench.env`.
219
+ A corpus it cannot find is left unset rather than guessed, and `fabench run`
220
+ skips it and prints how to obtain it.
221
+
222
+ ---
223
+
224
+ ## Where things live
225
+
226
+ ```
227
+ fabench/ the library, one folder per subsystem, each with its own README
228
+ evals/ how each system was run, with recipes, environments and raw output
229
+ datasets/ split lists and per-corpus config, with prep/ recording how gold was built
230
+ summary/ leaderboards and reports, SCRIPT OUTPUT, gitignored
231
+ records/ what each published snapshot covers, numbers on the leaderboard
232
+ ```
233
+
234
+ Every path below `evals/` and `summary/` shares one cell key,
235
+ `<lang>/<corpus>/<subset>/<condition>`, with `origin` for un-augmented audio.
236
+
237
+ | Path | Holds | Tracked? |
238
+ |---|---|---|
239
+ | `evals/<kind>/<tool>/<cell>/` | one tool's run, with its config, `hyp.jsonl` and scores together | no |
240
+ | `summary/<kind>/<cell>/` | the cross-tool leaderboard for that track | no, regenerated |
241
+ | `summary/local/<cell>/` | your own `fabench run` | no |
242
+
243
+ Nothing you run is tracked, so reproducing the benchmark leaves `git status`
244
+ clean. **Publishing is a separate, deliberate act.** Olewave posts the
245
+ numbers of a snapshot on the
246
+ [leaderboard](https://olewave.com/en/leaderboard/fa-bench) and names the
247
+ public release that produced them.
248
+
249
+ Further reading. [`evals/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/evals/README.md) explains how the systems
250
+ were run, [`datasets/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/datasets/README.md) covers staging, splits and
251
+ licensing, and [`datasets/prep/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/datasets/prep/README.md) has the
252
+ exact data-prep and augmentation commands.
253
+
254
+ ---
255
+
256
+ ## Configuration
257
+
258
+ **There is no run config to write.** Every default composes at load time from
259
+ the part of the tree that owns it, so `fabench run` with no `--config` is a
260
+ complete, valid run. The defaults live in
261
+
262
+ - `fabench/config.yaml` and `fabench/noise/config.yaml`,
263
+ - `datasets/languages/<lang>/config.yaml`,
264
+ - `evals/config.yaml` and `evals/<kind>/<tool>/config.yaml`,
265
+ - `fabench/normalize/<lang>/config.yaml` and `fabench/score/config.yaml`.
266
+
267
+ Two files carry what is specific to you.
268
+
269
+ | File | Holds | Tracked? |
270
+ |---|---|---|
271
+ | `.fabench.env` | what the **machine** is, with staged corpus roots, GPUs, threads and tool paths | no (template in `.fabench.env.example`) |
272
+ | a YAML you name | a deliberate **variant**, passed as `--config my.yaml` or `$FABENCH_CONFIG` | your choice |
273
+
274
+ Precedence is **run config > environment > composed defaults**. There is no
275
+ magic filename. A config is only ever loaded because you named it. Check what
276
+ anything resolved to with `fabench config`.
277
+
278
+ ---
279
+
280
+ ## Metrics
281
+
282
+ FA-Bench scores both tiers, words and phones, at 20 ms. From release 2.0.0 the
283
+ word tier leads with the **word timestamp F1**, and the phone tier keeps the
284
+ label-checked boundary F1. The October 2026 snapshot was published before
285
+ 2.0.0 and reports the label-checked F1 on both tiers.
286
+
287
+ - **Word timestamp F1 at 20 ms** (`wts_f1`, with `wts_precision` and
288
+ `wts_recall`, new in 2.0.0). Every reference word owns two timestamps, its
289
+ start and its end. A timestamp is a hit when the system returned the same
290
+ word, matched by label on the scorer's alignment, and placed that word's
291
+ start or end within 20 ms. Recall divides the hits by two per reference word
292
+ and precision by two per system word, so a misrecognized, dropped or
293
+ invented word costs its own two timestamps and nothing else. `wts_mae_ms`
294
+ averages the start and end errors of the matched words, and
295
+ `wts_start_mae_ms` and `wts_end_mae_ms` report them apart.
296
+ - **Boundary MAE** (`wbe_ms` for words, `mae_ms` for phones). Units are
297
+ matched to the reference by label first. A boundary is then scored only
298
+ when every unit beside it matched, so a dropped or misrecognized unit
299
+ leaves the average instead of worsening it. `n_word_boundaries` and
300
+ `n_boundaries` count the boundaries each average covers.
301
+ - **F1 at 20 ms, labels checked** (`wbnd_f1_all_20ms`, `bnd_f1_all_20ms`). A
302
+ boundary is a hit only when the units on **both** sides of it match the
303
+ reference and it lies within 20 ms. The two utterance edges count, against
304
+ silence. Every reference boundary is in the denominator, including those of
305
+ an utterance a system returned nothing for, so skipping hard material costs
306
+ recall. The `n_*_gold_all` and `n_*_hyp_all` columns hold the two
307
+ denominators.
308
+ - **WER and PER** add up the units a system relabelled, dropped or invented,
309
+ over the number of reference units. A system that skips hard units to
310
+ flatter its MAE pays for them here.
311
+
312
+ Each row also says how many utterances the system returned (`n_utts`), how
313
+ many it returned nothing for (`n_absent`) and how many came back empty
314
+ (`n_empty`). Further analysis metrics, among them the other tolerances and
315
+ the boundary classes of the paper, are part of Olewave's analysis work, and
316
+ this repository does not compute them.
317
+
318
+ Every table spans **clean and the four degradations on one row**, because the
319
+ question is how much of a system's accuracy survives degradation.
320
+
321
+ ---
322
+
323
+ ## Status
324
+
325
+ The roster is listed at the top. Track 1 is given the transcript. Track 2
326
+ decodes its own words. Every system is scored on both corpora, on all four
327
+ scored splits and on clean and degraded audio. There is one exception.
328
+ ElevenLabs has degraded cells on the two test splits only, because its credit
329
+ ran out.
330
+
331
+ **Olign gives accurate and cost-effective forced alignment and ASR with
332
+ timestamps.** It is a hosted service at a launch price of $0.18 per audio hour
333
+ (list price $0.21), and every new account comes with free credits. Register at
334
+ <https://dashboard.olewave.com/#signup> and create an API key under Profile → API
335
+ key. Every other system installs from its own recipe and needs nothing from us.
336
+
337
+ **Cloud-service outputs and scores never go into training, tuning or target
338
+ setting for Olign.**
339
+
340
+ Not every system reaches every tier. On Track 1, only **BFA, Charsiu, FALCON,
341
+ MAPS, MFA, NeuFA, Olign, PocketSphinx and TorchAudio emit phones**. The others,
342
+ and every Track 2 system, emit words only and appear in the word tables alone.
343
+ The tables show a dash rather than a number wherever that is so. Per-system
344
+ notes are in each tool's README under `fabench/aligners/`. The numbers and
345
+ their caveats are on the
346
+ [leaderboard](https://olewave.com/en/leaderboard/fa-bench).
347
+
348
+ Reproduce with `.venv/bin/python -m pytest -q` and `fabench gates`. The only
349
+ gate that needs restricted data is gold plausibility, which runs inside
350
+ `fabench ingest`.
351
+
352
+ ---
353
+
354
+ ## License
355
+
356
+ FA-Bench is licensed under the [Apache License 2.0](https://github.com/olewave/fa-bench/blob/v2.1.0/LICENSE) from release
357
+ 2.0.0. Copyright 2026 Olewave, LLC. [`NOTICE`](https://github.com/olewave/fa-bench/blob/v2.1.0/NOTICE) names the third-party
358
+ material in the tree, and each vendored tree keeps its licence beside it.
359
+ Releases up to 1.1.3 were published under PolyForm Noncommercial and are no
360
+ longer distributed. If you publish results produced with this software,
361
+ please cite the [paper](https://arxiv.org/abs/2609.32396), whose entry is in
362
+ [`CITATION.cff`](https://github.com/olewave/fa-bench/blob/v2.1.0/CITATION.cff).
363
+
364
+ ### The name
365
+
366
+ FA-Bench is a name of Olewave, LLC. Results that Olewave runs under the
367
+ published protocol are FA-Bench results. Results you produce with this code
368
+ are welcome, and you can describe them as scored with the FA-Bench scorer,
369
+ version 2.0. The Apache 2.0 license covers the code. The name stays with
370
+ Olewave.
371
+
372
+ ### What this repository leaves out, and why
373
+
374
+ - **The corpora.** TIMIT and Buckeye are licensed by their publishers, and
375
+ their audio, transcripts and hand-marked boundaries may not be passed on.
376
+ FA-Bench ships the split lists and the recipes that turn your own licensed
377
+ copy into the benchmark data, and it never downloads either corpus.
378
+ - **System outputs.** A forced aligner's output carries the reference
379
+ transcript word for word, and a recogniser's output comes close to it, so
380
+ every hypothesis file is as restricted as the corpus. They stay on the
381
+ machine that produced them, and the published tables hold numbers only.
382
+ - **Olewave's analysis tools.** The metrics beyond the headline set and the
383
+ per-utterance error analysis are part of Olewave's paid service. So are the
384
+ tools that compare vendors utterance by utterance, the result tables of
385
+ each snapshot and the code that builds them. The scorer that produces every
386
+ headline number is here in full.
387
+ - **Machine settings.** `.fabench.env` holds your corpus paths and API keys.
388
+ The tracked template is `.fabench.env.example`.
389
+
390
+ ### Getting the corpora
391
+
392
+ - **TIMIT** is catalog item
393
+ [LDC93S1](https://catalog.ldc.upenn.edu/LDC93S1) of the Linguistic Data
394
+ Consortium. Obtain it under an LDC licence. Members of the consortium
395
+ receive it with their membership, and others pay a fee.
396
+ - **Buckeye** is distributed by The Ohio State University through
397
+ [buckeyecorpus.osu.edu](https://buckeyecorpus.osu.edu/), after registration
398
+ on that site.
399
+
400
+ Stage each copy in any folder and tell FA-Bench where it is with
401
+ `fabench init`, which writes `FABENCH_TIMIT_ROOT` and `FABENCH_BUCKEYE_ROOT`
402
+ into `.fabench.env`.
403
+
404
+ ### Systems under test
405
+
406
+ Each system is installed by its own recipe under its own terms, and FA-Bench
407
+ redistributes none of their code or weights. Some carry restrictions worth
408
+ knowing before you run them. MMS-FA, CrisperWhisper and the UnitY2 tokenizer
409
+ are released for noncommercial use only, BFA and whisper-timestamped are under
410
+ the GPL family, NeuFA and its G2P carry no licence at all, and WebMAUS is
411
+ offered for academic research. [`evals/README.md`](https://github.com/olewave/fa-bench/blob/v2.1.0/evals/README.md) lists each
412
+ recipe with its licence.
413
+
414
+ ---
415
+
416
+ ## Paper
417
+
418
+ The FA-Bench paper is on arXiv as
419
+ [arXiv 2609.32396](https://arxiv.org/abs/2609.32396), and every link to the
420
+ paper on this page opens it there.
421
+
422
+ ---
423
+
424
+ ## Related benchmarks
425
+
426
+ The [nyra speech timing benchmark](https://github.com/nyrahealth/nyra_speech_timing_benchmark)
427
+ from nyra labs also scores forced aligners and speech APIs on word timing. It
428
+ adds FluencyBank for stuttered speech and it reports speed. It cuts Buckeye into
429
+ chunks of up to 20 s and averages the error over each word's start and end, so
430
+ its numbers differ from the ones on this page.
@@ -0,0 +1,146 @@
1
+ fa_bench-2.1.0.dist-info/licenses/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
2
+ fa_bench-2.1.0.dist-info/licenses/NOTICE,sha256=CzMqMklJjtpvVUkaQBo4oEU2xFGVV59MAdAm0ag5glc,1166
3
+ fa_bench-2.1.0.dist-info/licenses/fabench/dataprep/noisemix/kaldi/COPYING,sha256=5o8G8ySFU7R6oAFNF2aXw8q6uIOV3Qc_KIcxJ0L6gLk,17264
4
+ fabench/__init__.py,sha256=_hGewSfQJ9_SE6b329Aq9pE0ISnTY5G_Wna3P3ZyMZs,374
5
+ fabench/__main__.py,sha256=woA0lwHb1kteGmfjp6qYX7ds6G1_3YEcyC9PrrxZ05E,159
6
+ fabench/audio.py,sha256=4c8it3DKBzeJ8AzKUPCxmX7WlZDXnvwKvZKLtVLwutc,5309
7
+ fabench/cli.py,sha256=-IPC7nSqeYlHXqQTCCpXqPDR2h3iVBENIn0YSDKG0rM,10888
8
+ fabench/config.py,sha256=r35uEh6W4X3yoN-shdmbsGtGUCvURBV78pLeK8SLUag,25025
9
+ fabench/config.yaml,sha256=84xWXXFpgSmlP2ByFIjSmywmr5e_N7m6uXj8Ao9po64,3075
10
+ fabench/envfile.py,sha256=RPW_wQYVPkx6GVtiIsRNSKRq-gFCHzWDmfD7PVN_lC0,3786
11
+ fabench/gates.py,sha256=Z-U6ilEN0CmMbjTc99hc_fLo7XxGunvS9v1Uqc4chLU,11931
12
+ fabench/init.py,sha256=6JCJ50EXi3tD1AnJXVyAyMNeUuVlLlY-nRAe6mj2lwc,13710
13
+ fabench/paths.py,sha256=-wT49_FD778UjXL8x2d0Lj5t2ydL3v1_gvpnJcwvRHc,8481
14
+ fabench/pipeline.py,sha256=fX8IfaFEftoqw_nZXpr93WNbCRIf-iHGY1uGZpAWWso,5051
15
+ fabench/schema.py,sha256=WxAdjxridxXICXorImy1170xs9Lz9YinTsD7f_r0oDM,10847
16
+ fabench/selftest.py,sha256=LBzptSHqfNt3nbbpqJjd0quASF_v4lpPAeAV3jizanw,9376
17
+ fabench/aligners/__init__.py,sha256=kt7gJUJTeocJUWEb_g8lGnhFpQQNLMt_xSksH4OkQ68,5335
18
+ fabench/aligners/base.py,sha256=-Z01TtfTNNxay5iSirRDTKTLaEY-dta27VTVzULHrco,3802
19
+ fabench/aligners/relabel.py,sha256=6-Xmny7R6l7HiRNBTG-KxIKOaLTTv_NVjB_p39_ipvM,5754
20
+ fabench/aligners/runner.py,sha256=mKbttUWYbyKq2qqfeWLblVK_5dndY6AHGn6UBpdHfwA,14762
21
+ fabench/aligners/subprocess_aligner.py,sha256=-XGcCtZRD7LVgzz72mFzlsILm9Zd284-t46aae3X8Vs,11095
22
+ fabench/aligners/bfa/__init__.py,sha256=Q9KFpv_LebGaVwwnyiQilnc9QAvx0vzcqZdW40FyM1w,208
23
+ fabench/aligners/bfa/adapter.py,sha256=jiRgj0j-dXRiy5Y78d81e3B8mG8mlLI4wECcsqRGYHw,1273
24
+ fabench/aligners/charsiu/__init__.py,sha256=RKpyP72Q5UVZZk_Mp2r9VrXEyzTyQo7onlhywOREn9s,224
25
+ fabench/aligners/charsiu/adapter.py,sha256=WpD8TkbJk-vqqZ7XFT1z8Q106WeeK7BW9Ep5KGX_J3s,1246
26
+ fabench/aligners/crisperwhisper_fa/__init__.py,sha256=qoiyfcrh26Wl7PjcKEfC5ulK0cAnhBVcZ_p3Lcw4Ml0,249
27
+ fabench/aligners/crisperwhisper_fa/adapter.py,sha256=CDaUzK2xyPc7xKmjDztqLgYQPxkWtBe8h_8cgHlqQp4,1567
28
+ fabench/aligners/elevenlabs_fa/__init__.py,sha256=lk2lIEYGniiM8hvx_Y9jbt_-O2byPIAaf6ZIv_tAQ6A,235
29
+ fabench/aligners/elevenlabs_fa/adapter.py,sha256=yqc0mi7h-vY6XN4gnGEKVaVYc7FmtvGYgSUmL2A1Jqs,1255
30
+ fabench/aligners/falcon/__init__.py,sha256=mjguH1zH-WsRrFCsJqtPQgVYm86UGlAuq8IeBWsmx54,142
31
+ fabench/aligners/falcon/adapter.py,sha256=J_CO1W4iri-W7X7jjDvdGQB2g6gmQr1G9gxl-8MRC-Y,4003
32
+ fabench/aligners/maps/__init__.py,sha256=ZeJr_ogU5VR99amSYQzPTZDnQf8AdvK_d3W5NxLNX8o,212
33
+ fabench/aligners/maps/adapter.py,sha256=1bSFcE9aU_1LlJqltH_mYOgg18KudKsweRj4KDLWErw,7313
34
+ fabench/aligners/mfa/__init__.py,sha256=mC5DJE1DoXdAtQn10gXPGsRT5xGNnvrMj0u2elPlJZg,208
35
+ fabench/aligners/mfa/adapter.py,sha256=p1am-dLr3GBLBNNcVxefEUZf0Y_Q-y-YQ8-ttsAfFJg,8470
36
+ fabench/aligners/nemo_fa/__init__.py,sha256=KO7kudYklRNm23ayYrBQEsu4O1y5MIgScHNDtRb7yFA,143
37
+ fabench/aligners/nemo_fa/adapter.py,sha256=4okOmU3i3QZAtbjtogOouqog7ZwM5G6aEbSo2CVz2yk,2128
38
+ fabench/aligners/neufa/__init__.py,sha256=88Dm6_rdBOgM4nspE8puU1SK4szNnxhOtMEq3nLIC_o,139
39
+ fabench/aligners/neufa/adapter.py,sha256=ZOe8Yz3lfwU5vJRTBD-eM08lHmjS_v4WJnfN2IfoHPg,1532
40
+ fabench/aligners/nyra_fa/__init__.py,sha256=acotGtL8X4P_4xcfAIm6T6TrIRSBBInvG63YMFhpEsM,143
41
+ fabench/aligners/nyra_fa/adapter.py,sha256=MeWWCpOA_kEIwOfGOheqIIj9fF3JqhuIo2nnjcX1L2M,1046
42
+ fabench/aligners/olign/__init__.py,sha256=LaeZFICjIGE4chE1w6mWT4mZtTeaIYCe9qG0msS703I,634
43
+ fabench/aligners/olign/adapter.py,sha256=gGJqy1yj2DdJa9WqzDw6D1DHzikY3Z3afWwVb9aIU7g,9633
44
+ fabench/aligners/olign_v1/__init__.py,sha256=hmP4YNxGYx9_yThG1i3dkVyhnZ7PKrePgtIxGGFdfZs,217
45
+ fabench/aligners/olign_v1/adapter.py,sha256=yWE21bw5DkZYpjprHSN1tzAF4kmQJiQd2Zu14loPVDQ,1003
46
+ fabench/aligners/pocketsphinx/__init__.py,sha256=SedzLvNu-pwc59J2IfIT79CvCZoVvlvU3MlrvJ6s-bQ,245
47
+ fabench/aligners/pocketsphinx/adapter.py,sha256=jEk3vECGxAxIwlEnVjQH0BAn_CA5Pwx54tSymfMk23g,933
48
+ fabench/aligners/qwen3_fa/__init__.py,sha256=eVWbCztuuuYF_r3HZeYICfeKz6VwO5gk27vuizCb-zU,210
49
+ fabench/aligners/qwen3_fa/adapter.py,sha256=vvzdjOrXhSckG8xUhmtWJIB4UiU3D9uxvDQVIQ-23X8,6869
50
+ fabench/aligners/stable_ts/__init__.py,sha256=TS6QohsRXU0M2RAu7TyfeQW50C8DykDQoulWHep2pmY,182
51
+ fabench/aligners/stable_ts/adapter.py,sha256=Pb4LyoO1EomNq_-hiBgDqhRoOg5fIkSpEnUGGxWQ7do,1746
52
+ fabench/aligners/torchaudio_fa/__init__.py,sha256=GlxNr0ZsW_rS1dPmXnlRGxaBYYK2TfYCuKYF45i86os,245
53
+ fabench/aligners/torchaudio_fa/adapter.py,sha256=B6i8XHVd2lF3CDCjda9gOb5VLHArdziXJVQUDQGn7yY,2762
54
+ fabench/aligners/unity2/__init__.py,sha256=1JgihGvBgohlMcLm2385nLvf6VOJm0KFHXpDCy49fTA,142
55
+ fabench/aligners/unity2/adapter.py,sha256=xOQwztYS_IcTLOiHkvEhGlcM41RfCjmm8jqdb-GFwGM,1375
56
+ fabench/aligners/whisperx/__init__.py,sha256=Nm5n7Zb_6enIQXc8xymyV2yxBeusPfLDG6xkf5dO7bA,228
57
+ fabench/aligners/whisperx/adapter.py,sha256=2ohW_DwyKaBYiiUjCvn4iWBeXLCbc_wRjhcV_OhSKyc,1192
58
+ fabench/dataprep/__init__.py,sha256=LCA5vDwxgmS46PUbQvtu58k0wzRVpRac7ELihz_71PM,786
59
+ fabench/dataprep/datasets/__init__.py,sha256=TvOlCYKDMdH_OST3DqX8XilRC1MhD8Xdl6TWvrGE4g4,6869
60
+ fabench/dataprep/datasets/acquire.py,sha256=J6uL1UgwyV-ushoRj1FIpmJeCi249Ae3vq9NbruvAEY,4557
61
+ fabench/dataprep/datasets/export_refs.py,sha256=ppHWrYIbCxXZXAXTtZJlMewLE1F34ju6FEVVUlPq9LI,5840
62
+ fabench/dataprep/datasets/plausibility.py,sha256=3_e_YRSi_siW0o-8f4QIfDIwqfi0kew0zBYvfdCLar4,2972
63
+ fabench/dataprep/datasets/runner.py,sha256=-4bHiPyLyxj9SHRRMCW1IMuXE5BxPrIwh_248vXfSg8,3123
64
+ fabench/dataprep/datasets/textgrid.py,sha256=gYin7-6GrgYz0RyiVqNljTPH9uISJ1VkqHLFFUxzHkk,1414
65
+ fabench/dataprep/datasets/en/__init__.py,sha256=piYWE6setHWcMQ2cU1GQvBFHcm5ZKM4YQ1w-yn-s-ps,488
66
+ fabench/dataprep/datasets/en/buckeye/__init__.py,sha256=eWM3pQf5AJ216ICcyilGi23YitXQqIKJkQJWatyYzU0,459
67
+ fabench/dataprep/datasets/en/buckeye/download.py,sha256=PWe-l2pKSt-UX5aPeJN4hEQs-wUPpc0Ja910aTApDKk,2254
68
+ fabench/dataprep/datasets/en/buckeye/processor.py,sha256=nFboA3ifj6hg6J_TyK7DeGDMuZKdrmuFKSQCzVUD250,24229
69
+ fabench/dataprep/datasets/en/l2arctic/__init__.py,sha256=I2vqA2wkD-cxeFZh2rHTtqleM3catyhDzDCoX3fbkNQ,293
70
+ fabench/dataprep/datasets/en/l2arctic/download.py,sha256=VVI0kA9AlsfpJhqrXYrxtk7MmXT8qLeacyir-wytExI,1805
71
+ fabench/dataprep/datasets/en/l2arctic/processor.py,sha256=n4FrbgVSmgSBld7B949tueUo3UQjvKCH9EaWHUtoLnk,2832
72
+ fabench/dataprep/datasets/en/timit/__init__.py,sha256=7NaHoV0t8HyXRvmCQViBe0kOpbBnJeyrc6Tpjj029mc,552
73
+ fabench/dataprep/datasets/en/timit/download.py,sha256=v50aj0anZ8FA6DDdahuHHnhe3aLz1C6mUxDGrKUOf9E,2211
74
+ fabench/dataprep/datasets/en/timit/processor.py,sha256=JnpUWsAXTItfHLf0P9FkqJZwLPkU3FTqmWn5jyoDFKU,9499
75
+ fabench/dataprep/noisemix/__init__.py,sha256=hCkPXDzqPyOArzgn_jBPEvvug4wPWCQ046_PwYATreQ,403
76
+ fabench/dataprep/noisemix/datadirs.py,sha256=tf6yW0UrjDoEC82QhHyiQcIWxHr6ZpyDLTBGN3HcB8o,5204
77
+ fabench/dataprep/noisemix/make_noisy.py,sha256=5vgYAhmM-nD5qySbSF3Iyfvu_daoJoxlusXaiLd0svg,15788
78
+ fabench/dataprep/noisemix/make_noisy.sh,sha256=WDzLYXecDmvNTvjqCx5tTUhInOclwh8wAZQyD9bAHJc,8276
79
+ fabench/dataprep/noisemix/manifest.py,sha256=u4V7bw3p8Idfmwe_ifnNAbVXWSBjnSAR360xYR8st8M,3576
80
+ fabench/dataprep/noisemix/materialise.py,sha256=g_K5VK_xHFrYAaZ9MI5RiDf9E3SXw1WoqIU3YI0Wjxs,2398
81
+ fabench/dataprep/noisemix/prepare.py,sha256=BAzzzGClz15ogOuixsGvAIM0Ix_1hNhjuz3lGG53DVI,11149
82
+ fabench/dataprep/noisemix/shadow_root.py,sha256=buAnhAA0dgLrVUPkhyaBWvAQR5zK3T_ehhzerik4sGc,10850
83
+ fabench/dataprep/noisemix/utt2dur.py,sha256=82aw-vYCD8PDTZpaghngaUPzA9U5E_o5arZ1Da7uFzM,2067
84
+ fabench/dataprep/noisemix/kaldi/COPYING,sha256=5o8G8ySFU7R6oAFNF2aXw8q6uIOV3Qc_KIcxJ0L6gLk,17264
85
+ fabench/dataprep/noisemix/kaldi/PROVENANCE.md,sha256=mtNGiz4EOFmKyS82ctFkhz1dwAJs0J6DXQuignB1J68,3423
86
+ fabench/dataprep/noisemix/kaldi/steps/data/augment_data_dir.py,sha256=sWijhV0guToapMFdINfB133NcFATIRoYp8nYbKqVhx4,14274
87
+ fabench/dataprep/noisemix/kaldi/steps/data/data_dir_manipulation_lib.py,sha256=1yCviEkfA3E63fuvgPJCyPkLnYEtJN8kqHNgZVQhfj0,694
88
+ fabench/dataprep/noisemix/kaldi/steps/data/make_musan.py,sha256=9yKUKCGlOuyCPXtDoyavJxBzl5SX-UTdvOfeDtnUfWg,6806
89
+ fabench/dataprep/noisemix/kaldi/steps/data/make_musan.sh,sha256=kv0YJ49rbc7VME2d1GIkWzERDCx_XiQaEVILG7n7g7E,2337
90
+ fabench/dataprep/noisemix/kaldi/steps/data/reverberate_data_dir.py,sha256=lcc3eFibPCFOoi15Jou5af5l7c0vg4S3KqQYxLKvWY8,40606
91
+ fabench/dataprep/noisemix/kaldi/steps/libs/__init__.py,sha256=hguZ_AmTqG5HCMLXLANuCv5wbDd9cCZtrhdESnXSypo,172
92
+ fabench/dataprep/noisemix/kaldi/steps/libs/common.py,sha256=iONZEHjmhKC-aShHl8YALQbbfERGgOoMdbwMFUx0pOw,16212
93
+ fabench/noise/__init__.py,sha256=wT-naI3PBtBoFYv4WQA87XUvJFbKjlLYnucepin1BWs,68
94
+ fabench/noise/config.yaml,sha256=DtkAwLvJ_IJRrUjH8AYdwXQIpexECFjCIFqsqjW5znE,1621
95
+ fabench/noise/manifest.py,sha256=kLjx1hzIhcQNt9hTOv9qHLVtg4eOVk42MIQ6Yc6uNwc,4217
96
+ fabench/noise/mixer.py,sha256=_IwfjENCwvJfCfm6Y6tlrAd_NKL-QVMMsZ8V3-y27EU,5063
97
+ fabench/noise/musan.py,sha256=Nw9EGEJeDdLr2N-ar2EUvGJLGVk5x5awgVSmrs39Wh4,6000
98
+ fabench/noise/provider.py,sha256=q3N9qqukHChD0USI6USyX88_OCzwX2Vx5LTntiebYts,2687
99
+ fabench/noise/runner.py,sha256=0O6musfpzQQv7vZ5ksbAVHpsPJz4syBFFT47jtLuXNU,3251
100
+ fabench/noise/synth.py,sha256=fFkm6TSNvT779P4yxonAbWq8OrkG7STVbEcgaz8L-AY,1345
101
+ fabench/normalize/__init__.py,sha256=y01RAjhd_--FWP8CcZYjxtOUFfUE0kE1Z2K5X16cDs4,3097
102
+ fabench/normalize/check.py,sha256=y6VoSLNUttZr-uH6aEj3pgIb7VTA1jGtw7X3qVNtgkY,1920
103
+ fabench/normalize/en/__init__.py,sha256=s1bEPCdSA3F_ajCLCP8Ga8l8QiVj5rap1u2dp1p4rD0,2854
104
+ fabench/normalize/en/canonical.py,sha256=vY0B42fDtEfTXqpm1Z3O29QoAt-CTX5ClPTnWV5Qy0A,3324
105
+ fabench/normalize/en/config.yaml,sha256=zJJ6WjSdR-qT8Xft4lIM4735JwjMagOmocnF-8zk7jw,573
106
+ fabench/normalize/en/maps.py,sha256=ZaN2bRjmf-TkG3J6Kt9Yfm_6YY11SJeXnxiGQNOpkUE,7316
107
+ fabench/report/__init__.py,sha256=wT-naI3PBtBoFYv4WQA87XUvJFbKjlLYnucepin1BWs,68
108
+ fabench/report/runner.py,sha256=i_DNjdheR1aQswpMOaV2SLqq_he8LL7A4YN2Q382VGk,2727
109
+ fabench/score/__init__.py,sha256=wT-naI3PBtBoFYv4WQA87XUvJFbKjlLYnucepin1BWs,68
110
+ fabench/score/aggregate.py,sha256=Gv2620ht7tG9uBcECawSK8bniFarpsmT80gGBkLuVHU,6808
111
+ fabench/score/boundary.py,sha256=eCU1DG1qlEXhesjTGkaeU-3AGgl5Rjq_YS33PY1zibY,5035
112
+ fabench/score/config.yaml,sha256=xunIRXgSMK8TvXNkXTnem1-mB8oUAQHdwwHrxB6Fges,1788
113
+ fabench/score/core.py,sha256=HlklSUcoR3PoUKEbZA0Fg2TOOZiDNKd7z8A4DYC-phc,17043
114
+ fabench/score/matched.py,sha256=Hfn2ORgOo2tPhX9L6ow3ixyl1sZBjhL95SFLsiVuRp4,7535
115
+ fabench/score/runner.py,sha256=YMWgrDZZpZ11aisj0qGlEYC5PGyKBv1qC8CcUoAteKw,10681
116
+ fabench/score/segmentation.py,sha256=XD7Dsn6mW_LopMBDPx7lLHEZlphDpnl8SArVnmtWLLg,8640
117
+ fabench/score/word.py,sha256=If5OkN0NZmV1EJIF5rKpPbrdzHnJUa5kmk_dU0G8QQI,6071
118
+ fabench/timestamp_asrs/__init__.py,sha256=0aZPXq2ojARJ9Izrvqehv3SC7AkvuozT1rehg7OFO3U,1643
119
+ fabench/timestamp_asrs/base.py,sha256=y8h4w7jbbbj8OXiQ_FpYEO62DUsNDnSXt7HmrBmg9Oc,3113
120
+ fabench/timestamp_asrs/subprocess_asr.py,sha256=5KvIKQN0q7bvLWX0Bqc1uGjefy7Z77JfIeVoZM0Uju0,1857
121
+ fabench/timestamp_asrs/cloud/__init__.py,sha256=yAWJiPbyCxVbr3wSyCIfeIa4dL5yZa1EmyvhapDC5GA,977
122
+ fabench/timestamp_asrs/cloud/adapter.py,sha256=a6l1aYR2tpw8ras8pqOKHdF5UX95mgz8ZtrW0XAXQaA,37076
123
+ fabench/timestamp_asrs/cloud/awssig.py,sha256=7wOJn2-Tn7HkXbtbNDl8IlKupWy4Y6SCWCqakAdtwcg,6029
124
+ fabench/timestamp_asrs/cloud/http.py,sha256=Ze-iYZl2KNvjLy2qzSv-kUlQYo6Bs7M3PmVzLOGtC18,9535
125
+ fabench/timestamp_asrs/cloud/providers.py,sha256=J-2Z6WnY7jpGxAFTtbe8BNFkD6M9k-xxnplx4GN_R_A,61533
126
+ fabench/timestamp_asrs/crisperwhisper/__init__.py,sha256=Cqd72cSzVSw3QvbHU7QzyFJNiWF9I6-Yd1OENiGTSZo,231
127
+ fabench/timestamp_asrs/crisperwhisper/adapter.py,sha256=Js03JUk8QMe8awrCIlmmzRznTOnNq9VUggxWqmX3WPI,1098
128
+ fabench/timestamp_asrs/granite_speech_plus/__init__.py,sha256=UC-TjYf7lz4ksWMgQO_ZlEChQBXJBi7dauPOlO3Bie8,244
129
+ fabench/timestamp_asrs/granite_speech_plus/adapter.py,sha256=aCp86w5W8Od2UjsDOZOrHOsJgKAgAeYO24ejUn3LE3k,1084
130
+ fabench/timestamp_asrs/parakeet_tdt/__init__.py,sha256=6DJHAd0FjhdMc55s53Cy1z06GTeVRLaAdgtUOhFDAgs,221
131
+ fabench/timestamp_asrs/parakeet_tdt/adapter.py,sha256=TYuiiIVjyISJJa-AFMTCnhY6XyxJjrWhoNvGtVeT2Qo,930
132
+ fabench/timestamp_asrs/qwen3_asr/__init__.py,sha256=-pghgvV_Mfdqbn2AHUOHVtnTwlu8Jc3isOu_DXDmbdU,212
133
+ fabench/timestamp_asrs/qwen3_asr/adapter.py,sha256=bg_hAJE7hEDyWzuYAqbx_B9mA0SVvvICrp3o9s4T3z0,1342
134
+ fabench/timestamp_asrs/torchaudio_asr/__init__.py,sha256=n2a5lv0VeTYDXfv6rlc--osmn6ZeX2OkPgXn3I5ffEU,227
135
+ fabench/timestamp_asrs/torchaudio_asr/adapter.py,sha256=29ifLKTvzFRqWZlJwAyUpdoxllIZWEt0FgwPVq0OJug,1041
136
+ fabench/timestamp_asrs/whisper3/__init__.py,sha256=EA5_Rq0v4uMw9cNHRhqU_Mq6J2bzXeqn5Tmw3wxMT7I,211
137
+ fabench/timestamp_asrs/whisper3/adapter.py,sha256=ZjKAQnJ0t98qI72VYz0efki_f0t65CnHRTR5jbNJVsA,1019
138
+ fabench/timestamp_asrs/whisper_ts/__init__.py,sha256=zwhzzTzLO5I3eeI4DYWnn5Lw8eHo7mGaZU5FucvGMlo,158
139
+ fabench/timestamp_asrs/whisper_ts/adapter.py,sha256=BQW4w5hF8YrSaKWrDYCjT3-xGlF0nUEVu8Du1FLOWtA,989
140
+ fabench/timestamp_asrs/whisperx_asr/__init__.py,sha256=l66vA8wo71KfoPlJh60o6gzCr-PFORwKUVVb-1EvpFc,164
141
+ fabench/timestamp_asrs/whisperx_asr/adapter.py,sha256=c4puboh-dMwXAKxH2sGzQrUWQrE9q7kDsVTIBEM1qsU,1245
142
+ fa_bench-2.1.0.dist-info/METADATA,sha256=3D4yo5SLerUnac-Va8ytidUFv1byN-qvrJSSZU3t2iA,21833
143
+ fa_bench-2.1.0.dist-info/WHEEL,sha256=mOdXR9C3wMVov7cSY8PY-XCzo3SmmWCHvGQr5csJoao,91
144
+ fa_bench-2.1.0.dist-info/entry_points.txt,sha256=REcJHtj58hsNu-5oZUsGci13hWcXoYQmN3-hGju2WdQ,45
145
+ fa_bench-2.1.0.dist-info/top_level.txt,sha256=UOWS7OWHFWLulrq72dJeQdMcD2qb7iBubKQ02GMnWpI,8
146
+ fa_bench-2.1.0.dist-info/RECORD,,
@@ -0,0 +1,5 @@
1
+ Wheel-Version: 1.0
2
+ Generator: setuptools (84.2.1)
3
+ Root-Is-Purelib: true
4
+ Tag: py3-none-any
5
+
@@ -0,0 +1,2 @@
1
+ [console_scripts]
2
+ fabench = fabench.cli:main