asr-ja-bench v1.1: sample lists, corrections, hashes and rebuild script

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

LICENSE.md

@@ -0,0 +1,16 @@
+# Licences
+
+asr-ja-bench combines material under different licences:
+
+- `rebuild.py` and `README.md` are licensed under Apache-2.0 (https://www.apache.org/licenses/LICENSE-2.0).
+- Each `data/<subset>/` folder is derived from its source dataset and is licensed under that dataset's licence, as listed below.
+
+| path | source dataset | licence |
+|---|---|---|
+| `data/jsut/` | JSUT basic5000 | CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) |
+| `data/jvnv/` | JVNV | CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) |
+| `data/tedxjp10k/` | TEDxJP-10K | Apache-2.0 (https://www.apache.org/licenses/LICENSE-2.0) |
+| `data/fleurs/` | FLEURS (ja_jp, test) | CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) |
+| `data/speech-bsd/` | SpeechBSD (ja, test) | CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/) |
+
+The audio and transcripts of the source datasets are not included. Obtaining and using them is subject to each dataset's own terms.

README.md

@@ -0,0 +1,135 @@
+---
+language:
+- ja
+task_categories:
+- automatic-speech-recognition
+pretty_name: asr-ja-bench
+license: other
+license_name: per-subset
+license_link: LICENSE.md
+configs:
+- config_name: jsut
+  default: true
+  data_files:
+  - split: test
+    path: data/jsut/samples.jsonl
+- config_name: jvnv
+  data_files:
+  - split: test
+    path: data/jvnv/samples.jsonl
+- config_name: tedxjp10k
+  data_files:
+  - split: test
+    path: data/tedxjp10k/samples.jsonl
+- config_name: fleurs
+  data_files:
+  - split: test
+    path: data/fleurs/samples.jsonl
+- config_name: speech-bsd
+  data_files:
+  - split: test
+    path: data/speech-bsd/samples.jsonl
+---
+
+# asr-ja-bench
+
+A Japanese speech recognition benchmark built from five public test sets: [JSUT](https://sites.google.com/site/shinnosuketakamichi/publication/jsut), [JVNV](https://sites.google.com/site/shinnosuketakamichi/research-topics/jvnv_corpus), [TEDxJP-10K](https://github.com/laboroai/TEDxJP-10K), [FLEURS](https://huggingface.co/datasets/google/fleurs) and [SpeechBSD](https://github.com/ku-nlp/speechBSD), with known label defects corrected or the samples removed. The [Haas-ja-v1.0](https://huggingface.co/third-intelligence/Haas-ja-v1.0) results are reported on it.
+
+This repository contains **no audio and no transcripts**: the upstream corpora have different licences, and several do not allow redistribution. It lists which upstream samples the benchmark uses, the corrections made to their transcripts, and hashes to verify them. [`rebuild.py`](rebuild.py) recreates the benchmark from the upstream corpora and checks every transcript against its hash.
+
+**Version: v1.1.** Report results with the version they were measured on.
+
+| subset | upstream | samples | hours |
+|---|---|---:|---:|
+| `jsut` | JSUT basic5000 | 5,000 | 6.78 |
+| `jvnv` | JVNV | 1,375 | 3.33 |
+| `tedxjp10k` | TEDxJP-10K | 9,200 | 8.21 |
+| `fleurs` | FLEURS (ja_jp, test) | 650 | 2.36 |
+| `speech-bsd` | SpeechBSD (ja, test) | 2,093 | 2.70 |
+| **total** | | **18,318** | **23.38** |
+
+## Differences from the upstream test sets
+
+| subset | upstream samples | not included (scope rule) | removed (label defect) | transcripts corrected | samples |
+|---|---:|---:|---:|---:|---:|
+| `jsut` | 5,000 | 0 | 0 | 3 | 5,000 |
+| `jvnv` | 1,615 | 240 | 0 | 0 | 1,375 |
+| `tedxjp10k` | 10,000 (9,233 reconstructed¹) | 10 | 23 | 13 | 9,200 |
+| `fleurs` | 650 | 0 | 0 | 2 | 650 |
+| `speech-bsd` | 2,120 | 1 | 26 | 23 | 2,093 |
+
+¹ TEDxJP-10K is reconstructed from YouTube. When this benchmark was built, 767 of its 10,000 utterances could no longer be recovered (videos or subtitles removed), so it uses the remaining ones.
+
+Why samples were removed (`excluded.jsonl`, field `reason`):
+
+- `out_of_scope`: JVNV's phrase-free session (`*_free_*`): its non-verbal vocalizations are improvised and have no transcript.
+- `label_too_long`: the transcript is longer than 15 characters per second of audio.
+- `wrong_audio`: the audio does not contain the transcribed utterance.
+- `no_speech`: the audio contains no speech.
+
+Corrected transcripts fix errors heard in the audio. In JSUT, two transcripts belonged to the neighbouring file, and one repeated a word.
+
+SpeechBSD transcripts were also normalized when the benchmark was built: full-width characters folded to half-width (NFKC), Latin letters lowercased, and 「」, ・ and spaces removed (540 samples). These changes are in the patches too, so a rebuild reproduces the benchmark exactly.
+
+## Files
+
+```
+data/<subset>/samples.jsonl    the benchmark's samples, in order
+data/<subset>/excluded.jsonl   upstream samples removed, with the reason
+rebuild.py                     upstream corpora -> the benchmark
+LICENSE.md                     which licence applies where
+```
+
+Each line of `samples.jsonl`:
+
+| field | |
+|---|---|
+| `id` | the upstream sample id: JSUT `BASIC5000_0001`, JVNV `F1_anger_regular_01`, the TEDxJP-10K utterance id, the FLEURS recording's file name without `.wav`, the SpeechBSD `ja_wav` stem |
+| `duration` | seconds of 16 kHz audio |
+| `upstream_sha256` | SHA-256 of the upstream transcript (UTF-8) |
+| `text_sha256` | SHA-256 of the benchmark transcript |
+| `patch` | only if the transcript was changed: `[[start, end, replacement], ...]`, code-point offsets into the upstream transcript, applied from the right. It holds only the replacement characters, never the upstream text. |
+
+## Rebuilding
+
+Get the upstream corpora:
+
+- **JSUT basic5000**: `jsut_ver1.1.zip` from the [JSUT page](https://sites.google.com/site/shinnosuketakamichi/publication/jsut) (pass the zip or its extracted directory as `--jsut`).
+- **JVNV**: `jvnv_ver1.zip` from the [JVNV page](https://sites.google.com/site/shinnosuketakamichi/research-topics/jvnv_corpus) (`--jvnv`).
+- **TEDxJP-10K**: Reconstruct TEDxJP-10K v1.1 with the [TEDxJP-10K repository](https://github.com/laboroai/TEDxJP-10K)'s instructions (download the listed YouTube videos and subtitles, run `compose_tedxjp10k.py`) and pass the resulting `TEDxJP-10K_v1.1` directory as `--tedxjp10k`. Videos and subtitles removed from YouTube cannot be recovered, so a reconstruction can have fewer utterances; `rebuild.py` lists any sample it cannot find.
+- **FLEURS (ja_jp, test)**: Read from the Hub (`google/fleurs`, `ja_jp`, pinned revision) by `rebuild.py`.
+- **SpeechBSD (ja, test)**: Read from the Hub (`ku-nlp/speech-bsd-hf`, pinned revision) by `rebuild.py`. The repository is gated: accept its terms on the Hub and log in (`hf auth login`) first.
+
+Then:
+
+```bash
+pip install numpy soundfile pyarrow huggingface_hub   # and ffmpeg on PATH
+python rebuild.py --jsut jsut_ver1.1.zip --jvnv jvnv_ver1.zip \
+    --tedxjp10k TEDxJP-10K_v1.1 --out asr-ja-bench-v1.1
+```
+
+This writes the benchmark data directory `asr-ja-bench-v1.1/`: `<subset>.jsonl` (`id`, `audio`, `text`, `duration`) and the audio as 16 kHz mono 16-bit FLAC. It is separate from this repository's own folder, which holds only the sample lists. `--subset` limits it to some subsets, and `--text-only` checks the transcripts without writing audio. It exits with an error if a transcript does not match its hash or a sample is missing.
+
+Transcripts are reproduced exactly. JSUT and JVNV are recorded at 48 kHz and are resampled to 16 kHz with ffmpeg's default resampler, which reproduces the benchmark audio bit for bit (tested with ffmpeg 7.1). FLEURS and SpeechBSD audio is used as is. TEDxJP-10K audio is whatever its reconstruction produces; reconstructions can differ by a few least-significant bits, because the upstream script converts the audio with sox, which dithers.
+
+## Licences
+
+Each `data/<subset>/` folder is derived from its source dataset and licensed under that dataset's licence, listed below; `rebuild.py` and this card are Apache-2.0 (see [`LICENSE.md`](LICENSE.md)). The audio and transcripts are not included: obtaining and using them is subject to each corpus's own terms.
+
+| subset | licence of `data/<subset>/` |
+|---|---|
+| `jsut` | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) |
+| `jvnv` | [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) |
+| `tedxjp10k` | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) |
+| `fleurs` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
+| `speech-bsd` | [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) |
+
+## Acknowledgements
+
+We thank the authors of JSUT, JVNV, TEDxJP-10K, FLEURS and SpeechBSD for creating these corpora and making them openly available. If you use asr-ja-bench, please also cite them:
+
+- **JSUT** ([page](https://sites.google.com/site/shinnosuketakamichi/publication/jsut)): Sonobe, R., Takamichi, S., & Saruwatari, H. (2017). JSUT corpus: Free large-scale Japanese speech corpus for end-to-end speech synthesis. *arXiv preprint arXiv:1711.00354*.
+- **JVNV** ([page](https://sites.google.com/site/shinnosuketakamichi/research-topics/jvnv_corpus)): Xin, D., Jiang, J., Takamichi, S., Saito, Y., Aizawa, A., & Saruwatari, H. (2023). JVNV: A corpus of Japanese emotional speech with verbal content and nonverbal expressions. *arXiv preprint arXiv:2310.06072*.
+- **TEDxJP-10K** ([repository](https://github.com/laboroai/TEDxJP-10K)): 安藤慎太郎, 藤原弘将 (2020). テレビ録画とその字幕を利用した大規模日本語音声コーパスの構築. *情報処理学会研究報告*, Vol.2020-SLP-134, No.8.
+- **FLEURS** ([dataset](https://huggingface.co/datasets/google/fleurs)): Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. (2022). FLEURS: Few-shot learning evaluation of universal representations of speech. *arXiv preprint arXiv:2205.12446*.
+- **SpeechBSD** ([repository](https://github.com/ku-nlp/speechBSD)): Shimizu, S., Chu, C., Li, S., & Kurohashi, S. (2023). Towards speech dialogue translation mediating speakers of different languages. In *Findings of the Association for Computational Linguistics: ACL 2023* (pp. 1122–1134). Association for Computational Linguistics. https://aclanthology.org/2023.findings-acl.72

data/jvnv/excluded.jsonl

@@ -0,0 +1,240 @@
+{"id": "F1_anger_free_01", "reason": "out_of_scope"}
+{"id": "F1_anger_free_02", "reason": "out_of_scope"}
+{"id": "F1_anger_free_03", "reason": "out_of_scope"}
+{"id": "F1_anger_free_04", "reason": "out_of_scope"}
+{"id": "F1_anger_free_05", "reason": "out_of_scope"}
+{"id": "F1_anger_free_06", "reason": "out_of_scope"}
+{"id": "F1_anger_free_07", "reason": "out_of_scope"}
+{"id": "F1_anger_free_08", "reason": "out_of_scope"}
+{"id": "F1_anger_free_09", "reason": "out_of_scope"}
+{"id": "F1_anger_free_10", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_01", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_02", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_03", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_04", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_05", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_06", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_07", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_08", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_09", "reason": "out_of_scope"}
+{"id": "F1_disgust_free_10", "reason": "out_of_scope"}
+{"id": "F1_fear_free_01", "reason": "out_of_scope"}
+{"id": "F1_fear_free_02", "reason": "out_of_scope"}
+{"id": "F1_fear_free_03", "reason": "out_of_scope"}
+{"id": "F1_fear_free_04", "reason": "out_of_scope"}
+{"id": "F1_fear_free_05", "reason": "out_of_scope"}
+{"id": "F1_fear_free_06", "reason": "out_of_scope"}
+{"id": "F1_fear_free_07", "reason": "out_of_scope"}
+{"id": "F1_fear_free_08", "reason": "out_of_scope"}
+{"id": "F1_fear_free_09", "reason": "out_of_scope"}
+{"id": "F1_fear_free_10", "reason": "out_of_scope"}
+{"id": "F1_happy_free_01", "reason": "out_of_scope"}
+{"id": "F1_happy_free_02", "reason": "out_of_scope"}
+{"id": "F1_happy_free_03", "reason": "out_of_scope"}
+{"id": "F1_happy_free_04", "reason": "out_of_scope"}
+{"id": "F1_happy_free_05", "reason": "out_of_scope"}
+{"id": "F1_happy_free_06", "reason": "out_of_scope"}
+{"id": "F1_happy_free_07", "reason": "out_of_scope"}
+{"id": "F1_happy_free_08", "reason": "out_of_scope"}
+{"id": "F1_happy_free_09", "reason": "out_of_scope"}
+{"id": "F1_happy_free_10", "reason": "out_of_scope"}
+{"id": "F1_sad_free_01", "reason": "out_of_scope"}
+{"id": "F1_sad_free_02", "reason": "out_of_scope"}
+{"id": "F1_sad_free_03", "reason": "out_of_scope"}
+{"id": "F1_sad_free_04", "reason": "out_of_scope"}
+{"id": "F1_sad_free_05", "reason": "out_of_scope"}
+{"id": "F1_sad_free_06", "reason": "out_of_scope"}
+{"id": "F1_sad_free_07", "reason": "out_of_scope"}
+{"id": "F1_sad_free_08", "reason": "out_of_scope"}
+{"id": "F1_sad_free_09", "reason": "out_of_scope"}
+{"id": "F1_sad_free_10", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_01", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_02", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_03", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_04", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_05", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_06", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_07", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_08", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_09", "reason": "out_of_scope"}
+{"id": "F1_surprise_free_10", "reason": "out_of_scope"}
+{"id": "F2_anger_free_01", "reason": "out_of_scope"}
+{"id": "F2_anger_free_02", "reason": "out_of_scope"}
+{"id": "F2_anger_free_03", "reason": "out_of_scope"}
+{"id": "F2_anger_free_04", "reason": "out_of_scope"}
+{"id": "F2_anger_free_05", "reason": "out_of_scope"}
+{"id": "F2_anger_free_06", "reason": "out_of_scope"}
+{"id": "F2_anger_free_07", "reason": "out_of_scope"}
+{"id": "F2_anger_free_08", "reason": "out_of_scope"}
+{"id": "F2_anger_free_09", "reason": "out_of_scope"}
+{"id": "F2_anger_free_10", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_01", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_02", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_03", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_04", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_05", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_06", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_07", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_08", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_09", "reason": "out_of_scope"}
+{"id": "F2_disgust_free_10", "reason": "out_of_scope"}
+{"id": "F2_fear_free_01", "reason": "out_of_scope"}
+{"id": "F2_fear_free_02", "reason": "out_of_scope"}
+{"id": "F2_fear_free_03", "reason": "out_of_scope"}
+{"id": "F2_fear_free_04", "reason": "out_of_scope"}
+{"id": "F2_fear_free_05", "reason": "out_of_scope"}
+{"id": "F2_fear_free_06", "reason": "out_of_scope"}
+{"id": "F2_fear_free_07", "reason": "out_of_scope"}
+{"id": "F2_fear_free_08", "reason": "out_of_scope"}
+{"id": "F2_fear_free_09", "reason": "out_of_scope"}
+{"id": "F2_fear_free_10", "reason": "out_of_scope"}
+{"id": "F2_happy_free_01", "reason": "out_of_scope"}
+{"id": "F2_happy_free_02", "reason": "out_of_scope"}
+{"id": "F2_happy_free_03", "reason": "out_of_scope"}
+{"id": "F2_happy_free_04", "reason": "out_of_scope"}
+{"id": "F2_happy_free_05", "reason": "out_of_scope"}
+{"id": "F2_happy_free_06", "reason": "out_of_scope"}
+{"id": "F2_happy_free_07", "reason": "out_of_scope"}
+{"id": "F2_happy_free_08", "reason": "out_of_scope"}
+{"id": "F2_happy_free_09", "reason": "out_of_scope"}
+{"id": "F2_happy_free_10", "reason": "out_of_scope"}
+{"id": "F2_sad_free_01", "reason": "out_of_scope"}
+{"id": "F2_sad_free_02", "reason": "out_of_scope"}
+{"id": "F2_sad_free_03", "reason": "out_of_scope"}
+{"id": "F2_sad_free_04", "reason": "out_of_scope"}
+{"id": "F2_sad_free_05", "reason": "out_of_scope"}
+{"id": "F2_sad_free_06", "reason": "out_of_scope"}
+{"id": "F2_sad_free_07", "reason": "out_of_scope"}
+{"id": "F2_sad_free_08", "reason": "out_of_scope"}
+{"id": "F2_sad_free_09", "reason": "out_of_scope"}
+{"id": "F2_sad_free_10", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_01", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_02", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_03", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_04", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_05", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_06", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_07", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_08", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_09", "reason": "out_of_scope"}
+{"id": "F2_surprise_free_10", "reason": "out_of_scope"}
+{"id": "M1_anger_free_01", "reason": "out_of_scope"}
+{"id": "M1_anger_free_02", "reason": "out_of_scope"}
+{"id": "M1_anger_free_03", "reason": "out_of_scope"}
+{"id": "M1_anger_free_04", "reason": "out_of_scope"}
+{"id": "M1_anger_free_05", "reason": "out_of_scope"}
+{"id": "M1_anger_free_06", "reason": "out_of_scope"}
+{"id": "M1_anger_free_07", "reason": "out_of_scope"}
+{"id": "M1_anger_free_08", "reason": "out_of_scope"}
+{"id": "M1_anger_free_09", "reason": "out_of_scope"}
+{"id": "M1_anger_free_10", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_01", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_02", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_03", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_04", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_05", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_06", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_07", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_08", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_09", "reason": "out_of_scope"}
+{"id": "M1_disgust_free_10", "reason": "out_of_scope"}
+{"id": "M1_fear_free_01", "reason": "out_of_scope"}
+{"id": "M1_fear_free_02", "reason": "out_of_scope"}
+{"id": "M1_fear_free_03", "reason": "out_of_scope"}
+{"id": "M1_fear_free_04", "reason": "out_of_scope"}
+{"id": "M1_fear_free_05", "reason": "out_of_scope"}
+{"id": "M1_fear_free_06", "reason": "out_of_scope"}
+{"id": "M1_fear_free_07", "reason": "out_of_scope"}
+{"id": "M1_fear_free_08", "reason": "out_of_scope"}
+{"id": "M1_fear_free_09", "reason": "out_of_scope"}
+{"id": "M1_fear_free_10", "reason": "out_of_scope"}
+{"id": "M1_happy_free_01", "reason": "out_of_scope"}
+{"id": "M1_happy_free_02", "reason": "out_of_scope"}
+{"id": "M1_happy_free_03", "reason": "out_of_scope"}
+{"id": "M1_happy_free_04", "reason": "out_of_scope"}
+{"id": "M1_happy_free_05", "reason": "out_of_scope"}
+{"id": "M1_happy_free_06", "reason": "out_of_scope"}
+{"id": "M1_happy_free_07", "reason": "out_of_scope"}
+{"id": "M1_happy_free_08", "reason": "out_of_scope"}
+{"id": "M1_happy_free_09", "reason": "out_of_scope"}
+{"id": "M1_happy_free_10", "reason": "out_of_scope"}
+{"id": "M1_sad_free_01", "reason": "out_of_scope"}
+{"id": "M1_sad_free_02", "reason": "out_of_scope"}
+{"id": "M1_sad_free_03", "reason": "out_of_scope"}
+{"id": "M1_sad_free_04", "reason": "out_of_scope"}
+{"id": "M1_sad_free_05", "reason": "out_of_scope"}
+{"id": "M1_sad_free_06", "reason": "out_of_scope"}
+{"id": "M1_sad_free_07", "reason": "out_of_scope"}
+{"id": "M1_sad_free_08", "reason": "out_of_scope"}
+{"id": "M1_sad_free_09", "reason": "out_of_scope"}
+{"id": "M1_sad_free_10", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_01", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_02", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_03", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_04", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_05", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_06", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_07", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_08", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_09", "reason": "out_of_scope"}
+{"id": "M1_surprise_free_10", "reason": "out_of_scope"}
+{"id": "M2_anger_free_01", "reason": "out_of_scope"}
+{"id": "M2_anger_free_02", "reason": "out_of_scope"}
+{"id": "M2_anger_free_03", "reason": "out_of_scope"}
+{"id": "M2_anger_free_04", "reason": "out_of_scope"}
+{"id": "M2_anger_free_05", "reason": "out_of_scope"}
+{"id": "M2_anger_free_06", "reason": "out_of_scope"}
+{"id": "M2_anger_free_07", "reason": "out_of_scope"}
+{"id": "M2_anger_free_08", "reason": "out_of_scope"}
+{"id": "M2_anger_free_09", "reason": "out_of_scope"}
+{"id": "M2_anger_free_10", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_01", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_02", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_03", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_04", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_05", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_06", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_07", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_08", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_09", "reason": "out_of_scope"}
+{"id": "M2_disgust_free_10", "reason": "out_of_scope"}
+{"id": "M2_fear_free_01", "reason": "out_of_scope"}
+{"id": "M2_fear_free_02", "reason": "out_of_scope"}
+{"id": "M2_fear_free_03", "reason": "out_of_scope"}
+{"id": "M2_fear_free_04", "reason": "out_of_scope"}
+{"id": "M2_fear_free_05", "reason": "out_of_scope"}
+{"id": "M2_fear_free_06", "reason": "out_of_scope"}
+{"id": "M2_fear_free_07", "reason": "out_of_scope"}
+{"id": "M2_fear_free_08", "reason": "out_of_scope"}
+{"id": "M2_fear_free_09", "reason": "out_of_scope"}
+{"id": "M2_fear_free_10", "reason": "out_of_scope"}
+{"id": "M2_happy_free_01", "reason": "out_of_scope"}
+{"id": "M2_happy_free_02", "reason": "out_of_scope"}
+{"id": "M2_happy_free_03", "reason": "out_of_scope"}
+{"id": "M2_happy_free_04", "reason": "out_of_scope"}
+{"id": "M2_happy_free_05", "reason": "out_of_scope"}
+{"id": "M2_happy_free_06", "reason": "out_of_scope"}
+{"id": "M2_happy_free_07", "reason": "out_of_scope"}
+{"id": "M2_happy_free_08", "reason": "out_of_scope"}
+{"id": "M2_happy_free_09", "reason": "out_of_scope"}
+{"id": "M2_happy_free_10", "reason": "out_of_scope"}
+{"id": "M2_sad_free_01", "reason": "out_of_scope"}
+{"id": "M2_sad_free_02", "reason": "out_of_scope"}
+{"id": "M2_sad_free_03", "reason": "out_of_scope"}
+{"id": "M2_sad_free_04", "reason": "out_of_scope"}
+{"id": "M2_sad_free_05", "reason": "out_of_scope"}
+{"id": "M2_sad_free_06", "reason": "out_of_scope"}
+{"id": "M2_sad_free_07", "reason": "out_of_scope"}
+{"id": "M2_sad_free_08", "reason": "out_of_scope"}
+{"id": "M2_sad_free_09", "reason": "out_of_scope"}
+{"id": "M2_sad_free_10", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_01", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_02", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_03", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_04", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_05", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_06", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_07", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_08", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_09", "reason": "out_of_scope"}
+{"id": "M2_surprise_free_10", "reason": "out_of_scope"}

data/speech-bsd/excluded.jsonl

@@ -0,0 +1,27 @@
+{"id": "190315_E004_02_spk0_no21_ja", "reason": "wrong_audio"}
+{"id": "190329_E06_01_spk0_no3_ja", "reason": "label_too_long"}
+{"id": "190329_E14_07_spk1_no3_ja", "reason": "no_speech"}
+{"id": "190329_E14_16_spk1_no3_ja", "reason": "no_speech"}
+{"id": "190329_E17_12_spk0_no18_ja", "reason": "wrong_audio"}
+{"id": "190329_E17_12_spk0_no19_ja", "reason": "wrong_audio"}
+{"id": "190329_J01_06_spk1_no6_ja", "reason": "no_speech"}
+{"id": "190329_J14_05_spk0_no1_ja", "reason": "wrong_audio"}
+{"id": "190329_J19_05_spk1_no2_ja", "reason": "wrong_audio"}
+{"id": "190329_E02_03_spk0_no25_ja", "reason": "wrong_audio"}
+{"id": "190329_E06_01_spk0_no27_ja", "reason": "wrong_audio"}
+{"id": "190329_E06_01_spk0_no4_ja", "reason": "no_speech"}
+{"id": "190329_E06_01_spk0_no6_ja", "reason": "no_speech"}
+{"id": "190329_E06_01_spk1_no7_ja", "reason": "wrong_audio"}
+{"id": "190329_E06_02_spk0_no10_ja", "reason": "wrong_audio"}
+{"id": "190329_E14_16_spk0_no28_ja", "reason": "wrong_audio"}
+{"id": "190329_E17_12_spk0_no21_ja", "reason": "wrong_audio"}
+{"id": "190329_E17_12_spk0_no22_ja", "reason": "wrong_audio"}
+{"id": "190329_E17_12_spk0_no23_ja", "reason": "wrong_audio"}
+{"id": "190329_E17_12_spk0_no4_ja", "reason": "wrong_audio"}
+{"id": "190329_J08_09_spk2_no39_ja", "reason": "wrong_audio"}
+{"id": "190329_J10_16_spk1_no37_ja", "reason": "wrong_audio"}
+{"id": "190329_J11_02_spk2_no10_ja", "reason": "no_speech"}
+{"id": "190329_J11_02_spk2_no6_ja", "reason": "no_speech"}
+{"id": "190329_J14_05_spk0_no15_ja", "reason": "wrong_audio"}
+{"id": "190329_J14_05_spk0_no8_ja", "reason": "wrong_audio"}
+{"id": "190329_J19_05_spk0_no30_ja", "reason": "no_speech"}

data/tedxjp10k/excluded.jsonl

@@ -0,0 +1,33 @@
+{"id": "0KTVqevvEjo-00023495-00023640", "reason": "wrong_audio"}
+{"id": "0KTVqevvEjo-00039568-00039827", "reason": "wrong_audio"}
+{"id": "0KTVqevvEjo-00072424-00072677", "reason": "wrong_audio"}
+{"id": "3cU5gRwbRiU-00060113-00060246", "reason": "label_too_long"}
+{"id": "46PxgGJiqqA-00034876-00035014", "reason": "wrong_audio"}
+{"id": "7GUhszccVEA-00096526-00096586", "reason": "label_too_long"}
+{"id": "9_QS74NKTEo-00104874-00104974", "reason": "label_too_long"}
+{"id": "Ab-KZT06gR0-00020962-00021235", "reason": "wrong_audio"}
+{"id": "Ab-KZT06gR0-00031523-00031748", "reason": "wrong_audio"}
+{"id": "IBrUxfKQdEA-00026800-00026938", "reason": "wrong_audio"}
+{"id": "IcxboKkA9Xc-00034368-00034508", "reason": "label_too_long"}
+{"id": "LzviN1x-300-00025162-00025548", "reason": "wrong_audio"}
+{"id": "Mc044I55SCY-00004983-00005210", "reason": "wrong_audio"}
+{"id": "Sz4qMGRA9gs-00049730-00049839", "reason": "wrong_audio"}
+{"id": "Wqdb4uicNoc-00037781-00038068", "reason": "wrong_audio"}
+{"id": "Wqdb4uicNoc-00041894-00041967", "reason": "label_too_long"}
+{"id": "XL0SS8CChOQ-00014700-00014763", "reason": "label_too_long"}
+{"id": "XL0SS8CChOQ-00090799-00090891", "reason": "label_too_long"}
+{"id": "YGwoHfxDJi0-00044918-00045047", "reason": "wrong_audio"}
+{"id": "cIsIaJnX2BM-00009426-00010453", "reason": "no_speech"}
+{"id": "jgVBvO19VcE-00022060-00022209", "reason": "wrong_audio"}
+{"id": "k3ZteZsbKWY-00081908-00082189", "reason": "wrong_audio"}
+{"id": "k3ZteZsbKWY-00088979-00089217", "reason": "wrong_audio"}
+{"id": "kU9LcoHaFLo-00065084-00065309", "reason": "wrong_audio"}
+{"id": "knrMTJctBbg-00072037-00072169", "reason": "label_too_long"}
+{"id": "lL8aJE9Xp3Y-00030457-00030636", "reason": "wrong_audio"}
+{"id": "lqw9S4Kb5BE-00082677-00082871", "reason": "wrong_audio"}
+{"id": "lqw9S4Kb5BE-00085270-00085487", "reason": "wrong_audio"}
+{"id": "nXwLkxYZnew-00081693-00081850", "reason": "label_too_long"}
+{"id": "p0jKC4x3kgk-00027918-00028064", "reason": "wrong_audio"}
+{"id": "p0jKC4x3kgk-00034285-00034435", "reason": "wrong_audio"}
+{"id": "saHZ4bN3h28-00053729-00053874", "reason": "wrong_audio"}
+{"id": "t3MMsIG94cI-00059659-00059804", "reason": "label_too_long"}

rebuild.py

@@ -0,0 +1,298 @@
+#!/usr/bin/env python3
+"""Rebuild asr-ja-bench from the upstream corpora.
+
+asr-ja-bench ships no audio and no transcripts, only which upstream samples
+it uses and how their transcripts were corrected. This script reads each
+upstream corpus, takes the listed samples in order, applies the corrections,
+checks every transcript against its published hash, and writes the benchmark:
+
+    <out>/<subset>.jsonl          {"id", "audio", "text", "duration"} per sample
+    <out>/audio/<subset>/<id>.flac   16 kHz mono, 16-bit PCM
+
+Inputs (see the dataset card for where to get each):
+
+    --jsut        jsut_ver1.1.zip, or the directory it extracts to
+    --jvnv        jvnv_ver1.zip, or the directory it extracts to
+    --tedxjp10k   the TEDxJP-10K v1.1 directory written by Laboro.AI's
+                  compose_tedxjp10k.py (text, segments, wav/)
+    FLEURS and SpeechBSD are read from the Hugging Face Hub at pinned revisions
+    (SpeechBSD is gated: accept its terms on the Hub and log in first).
+
+    python rebuild.py --jsut jsut_ver1.1.zip --jvnv jvnv_ver1.zip \\
+        --tedxjp10k TEDxJP-10K_v1.1 --out asr-ja-bench-v1.1
+    python rebuild.py ... --subset fleurs --text-only     # transcripts only
+
+Requires: numpy, soundfile, pyarrow, huggingface_hub, and the ffmpeg command
+(JSUT and JVNV are 48 kHz; resampling them with ffmpeg's default resampler
+reproduces the benchmark audio bit for bit, tested with ffmpeg 7.1).
+Exits non-zero if any transcript fails its hash or a listed sample is missing.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import hashlib
+import io
+import json
+import shutil
+import subprocess
+import sys
+import tarfile
+import zipfile
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Callable, Iterator
+
+import numpy as np
+import soundfile as sf
+
+SUBSETS = ("jsut", "jvnv", "tedxjp10k", "fleurs", "speech-bsd")
+SAMPLE_RATE = 16000
+
+FLEURS = {"repo": "google/fleurs", "revision": "70bb2e84b976b7e960aa89f1c648e09c59f894dd",
+          "config": "ja_jp"}
+SPEECH_BSD = {"repo": "ku-nlp/speech-bsd-hf", "revision": "c925805c8607193b3a89fc705a61bd3647cd31f2"}
+
+
+@dataclass
+class Upstream:
+    """One upstream sample: its transcript and a loader for its 16 kHz mono audio."""
+    text: str
+    audio: Callable[[], np.ndarray]
+
+
+# ---------------------------------------------------------------------------
+# transcripts: hash and patch
+# ---------------------------------------------------------------------------
+def text_sha256(text: str) -> str:
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+def apply_patch(text: str, patch: list | None) -> str:
+    """Apply [[start, end, replacement], ...] (code-point offsets into the
+    upstream text, non-overlapping, ascending) from the right."""
+    for start, end, replacement in reversed(patch or []):
+        text = text[:start] + replacement + text[end:]
+    return text
+
+
+def make_patch(source: str, target: str) -> list | None:
+    """The patch apply_patch(source, patch) == target. Carries only the
+    replacement characters, never the source text."""
+    import difflib
+    if source == target:
+        return None
+    ops = [[i1, i2, target[j1:j2]]
+           for tag, i1, i2, j1, j2 in difflib.SequenceMatcher(None, source, target, autojunk=False).get_opcodes()
+           if tag != "equal"]
+    if sum(len(r) for _, _, r in ops) > len(target) / 2:
+        ops = [[0, len(source), target]]  # mostly rewritten: one replacement reads better
+    assert apply_patch(source, ops) == target
+    return ops
+
+
+# ---------------------------------------------------------------------------
+# upstream readers: {upstream_id: Upstream}
+# ---------------------------------------------------------------------------
+class Archive:
+    """A corpus zip or the directory it extracts to, read the same way."""
+
+    def __init__(self, path: Path):
+        self.path = path
+        self.zip = zipfile.ZipFile(path) if path.is_file() else None
+
+    def read(self, name: str) -> bytes:
+        if self.zip:
+            return self.zip.read(name)
+        return (self.path / name).read_bytes()
+
+    def names(self) -> set[str]:
+        if self.zip:
+            return set(self.zip.namelist())
+        return {str(p.relative_to(self.path)) for p in self.path.rglob("*")}
+
+
+def _decode(data: bytes) -> np.ndarray:
+    """16 kHz mono float32. Other rates and multi-channel audio go through ffmpeg,
+    which is how the benchmark audio was made."""
+    info = sf.info(io.BytesIO(data))
+    if info.samplerate == SAMPLE_RATE and info.channels == 1:
+        return sf.read(io.BytesIO(data), dtype="float32")[0]
+    if shutil.which("ffmpeg") is None:
+        raise SystemExit("ffmpeg not found on PATH (needed to resample to 16 kHz)")
+    out = subprocess.run(["ffmpeg", "-v", "error", "-i", "-", "-ar", str(SAMPLE_RATE), "-ac", "1",
+                          "-f", "f32le", "-"], input=data, capture_output=True, check=True).stdout
+    return np.frombuffer(out, dtype="float32")
+
+
+def read_jsut(path: Path) -> dict[str, Upstream]:
+    a = Archive(path)
+    root = "jsut_ver1.1/" if a.zip else ""
+    out = {}
+    for line in a.read(f"{root}basic5000/transcript_utf8.txt").decode("utf-8").splitlines():
+        if line.strip():
+            uid, text = line.split(":", 1)
+            out[uid] = Upstream(text, lambda u=uid: _decode(a.read(f"{root}basic5000/wav/{u}.wav")))
+    return out
+
+
+def read_jvnv(path: Path) -> dict[str, Upstream]:
+    # transcription.csv rows are `<emotion>_<session>_<n>|<nonverbal phrase>|<full text>`,
+    # shared by the four speakers; the transcript is the full text.
+    a = Archive(path)
+    root = "jvnv_v1/" if a.zip else ""
+    scripts = {}
+    for line in a.read(f"{root}transcription.csv").decode("utf-8").splitlines():
+        if line.strip():
+            key, _nonverbal, text = line.split("|")
+            scripts[key] = text.strip()
+    names = a.names()
+    out = {}
+    for spk in ("F1", "F2", "M1", "M2"):
+        for key, text in scripts.items():
+            emotion, session, _n = key.split("_")
+            uid = f"{spk}_{key}"
+            name = f"{root}{spk}/{emotion}/{session}/{uid}.wav"
+            if name in names:              # not every speaker recorded every script line
+                out[uid] = Upstream(text, lambda n=name: _decode(a.read(n)))
+    return out
+
+
+def read_tedxjp10k(path: Path) -> dict[str, Upstream]:
+    texts = {}
+    for line in (path / "text").read_text(encoding="utf-8").splitlines():
+        parts = line.split(maxsplit=1)
+        if len(parts) == 2:
+            texts[parts[0]] = parts[1]
+    cache: dict[str, np.ndarray] = {}
+
+    def cut(video: str, start: float, end: float) -> np.ndarray:
+        if video not in cache:
+            cache.clear()                       # one video at a time; samples come in video order
+            cache[video] = sf.read(path / "wav" / f"{video}.16k.wav", dtype="int16")[0]
+        x = cache[video]
+        a, b = int(start * SAMPLE_RATE), min(int(end * SAMPLE_RATE), len(x))
+        return x[a:b].astype("float32") / 32768
+
+    out = {}
+    for line in (path / "segments").read_text(encoding="utf-8").splitlines():
+        uid, video, start, end = line.split()
+        if uid in texts:
+            out[uid] = Upstream(texts[uid], lambda v=video, s=float(start), e=float(end): cut(v, s, e))
+    return out
+
+
+def _hub(repo: str, filename: str, revision: str) -> str:
+    from huggingface_hub import hf_hub_download
+    return hf_hub_download(repo, filename, repo_type="dataset", revision=revision)
+
+
+def read_fleurs() -> dict[str, Upstream]:
+    rev, cfg = FLEURS["revision"], FLEURS["config"]
+    texts = {}
+    with open(_hub(FLEURS["repo"], f"data/{cfg}/test.tsv", rev), encoding="utf-8") as f:
+        for row in csv.reader(f, delimiter="\t", quoting=csv.QUOTE_NONE):
+            texts[row[1].removesuffix(".wav")] = row[3]      # file name -> transcription
+    audio = {}
+    with tarfile.open(_hub(FLEURS["repo"], f"data/{cfg}/audio/test.tar.gz", rev)) as tf:
+        for m in tf:
+            if m.isfile() and m.name.endswith(".wav"):
+                audio[m.name.rsplit("/", 1)[-1].removesuffix(".wav")] = tf.extractfile(m).read()
+    return {uid: Upstream(text, lambda d=audio[uid]: _decode(d)) for uid, text in texts.items() if uid in audio}
+
+
+def read_speech_bsd() -> dict[str, Upstream]:
+    import pyarrow.parquet as pq
+    from huggingface_hub import HfApi
+
+    rev = SPEECH_BSD["revision"]
+    files = sorted(f for f in HfApi().list_repo_files(SPEECH_BSD["repo"], repo_type="dataset", revision=rev)
+                   if f.startswith("data/test-"))
+    out = {}
+    for f in files:
+        table = pq.read_table(_hub(SPEECH_BSD["repo"], f, rev), columns=["ja_wav", "ja_sentence", "audio"])
+        for row in table.to_pylist():
+            if row["ja_wav"]:
+                uid = Path(row["ja_wav"]).stem
+                out[uid] = Upstream(row["ja_sentence"], lambda d=row["audio"]["bytes"]: _decode(d))
+    return out
+
+
+def read_upstream(subset: str, args) -> dict[str, Upstream]:
+    if subset in ("jsut", "jvnv", "tedxjp10k"):
+        path = getattr(args, subset)
+        if path is None:
+            raise SystemExit(f"{subset}: pass --{subset} (see the dataset card for where to get it)")
+        return {"jsut": read_jsut, "jvnv": read_jvnv, "tedxjp10k": read_tedxjp10k}[subset](path)
+    return read_fleurs() if subset == "fleurs" else read_speech_bsd()
+
+
+# ---------------------------------------------------------------------------
+# rebuild
+# ---------------------------------------------------------------------------
+def rebuild_subset(subset: str, samples: list[dict], upstream: dict[str, Upstream],
+                   out: Path | None) -> tuple[list[dict], list[str]]:
+    rows, problems = [], []
+    for s in samples:
+        u = upstream.get(s["id"])
+        if u is None:
+            problems.append(f"{s['id']}: not found upstream")
+            continue
+        if text_sha256(u.text) != s["upstream_sha256"]:
+            problems.append(f"{s['id']}: upstream transcript differs from the one this "
+                            f"benchmark was built from")
+            continue
+        text = apply_patch(u.text, s.get("patch"))
+        if text_sha256(text) != s["text_sha256"]:
+            problems.append(f"{s['id']}: corrected transcript fails its hash")
+            continue
+        row = {"id": s["id"], "text": text, "duration": s["duration"]}
+        if out is not None:
+            x = u.audio()
+            rel = f"audio/{subset}/{s['id']}.flac"
+            sf.write(out / rel, x, SAMPLE_RATE, format="FLAC", subtype="PCM_16")
+            row = {"id": s["id"], "audio": rel, "text": text, "duration": round(len(x) / SAMPLE_RATE, 4)}
+        rows.append(row)
+    return rows, problems
+
+
+def main(argv=None) -> int:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--data", type=Path, default=Path(__file__).resolve().parent / "data",
+                    help="this dataset's data/ directory (default: next to this script)")
+    ap.add_argument("--out", type=Path, help="output directory (required unless --text-only)")
+    ap.add_argument("--subset", action="append", choices=SUBSETS, help="repeatable; default: all")
+    ap.add_argument("--text-only", action="store_true", help="check transcripts, write no audio")
+    ap.add_argument("--jsut", type=Path)
+    ap.add_argument("--jvnv", type=Path)
+    ap.add_argument("--tedxjp10k", type=Path)
+    args = ap.parse_args(argv)
+    if args.out is None and not args.text_only:
+        ap.error("--out is required unless --text-only")
+
+    failed = 0
+    for subset in args.subset or SUBSETS:
+        with open(args.data / subset / "samples.jsonl", encoding="utf-8") as f:
+            samples = [json.loads(line) for line in f if line.strip()]
+        print(f"[{subset}] reading upstream ...", flush=True)
+        upstream = read_upstream(subset, args)
+        audio_out = None if args.text_only else args.out
+        if audio_out is not None:
+            (audio_out / "audio" / subset).mkdir(parents=True, exist_ok=True)
+        rows, problems = rebuild_subset(subset, samples, upstream, audio_out)
+        if args.out is not None:
+            args.out.mkdir(parents=True, exist_ok=True)
+            with open(args.out / f"{subset}.jsonl", "w", encoding="utf-8") as f:
+                for r in rows:
+                    f.write(json.dumps(r, ensure_ascii=False) + "\n")
+        for p in problems[:10]:
+            print(f"[{subset}]   {p}", file=sys.stderr)
+        print(f"[{subset}] {len(rows)}/{len(samples)} samples rebuilt"
+              + (f", {len(problems)} problems" if problems else ", all transcripts verified"), flush=True)
+        failed += bool(problems)
+    return 1 if failed else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())