asr-ja-bench

A Japanese speech recognition benchmark built from five public test sets: JSUT, JVNV, TEDxJP-10K, FLEURS and SpeechBSD, with known label defects corrected or the samples removed. The 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 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 (pass the zip or its extracted directory as --jsut).
• JVNV: jvnv_ver1.zip from the JVNV page (--jvnv).
• TEDxJP-10K: Reconstruct TEDxJP-10K v1.1 with the TEDxJP-10K repository'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:

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). 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
jvnv	CC BY-SA 4.0
tedxjp10k	Apache-2.0
fleurs	CC BY 4.0
speech-bsd	CC 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): 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): 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): 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): 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): 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