YODAS v3: Over 1 Million Hours of High-Bandwidth, Stereophonic, Multilingual Speech
William Chen, Shinnosuke Takamichi, Sayaka Shiota, Satoru Fukayama, Samuele Cornell, Shinji Watanabe
cs.CL
Quick verdict
Claims largest open 1.1M-hour 48kHz stereo corpus, 147 languages; evidence supports scale and fidelity, but language tags and transcript quality go unverified.
Summary
YODAS v3 is an openly licensed (CC BY 3.0) corpus of over 1.1 million hours of web-mined speech spanning 147 YouTube locales, stored as 48kHz multichannel Opus with weak transcripts and English translations, built by crawling CC-licensed YouTube videos via language-specific Wikipedia keyword searches that prioritize recent uploads. The distribution is strikingly balanced: Arabic and Welsh lead at roughly 25K hours each, 22 languages exceed 10K hours and 73 exceed 5K, and English is under 2% of the total. The authors verify claimed audio quality: spectral analysis finds 92.5% of recordings above 32 kHz effective bandwidth and 71% genuinely multichannel. Baselines show ASR models trained without score filtering achieve the best results for six of seven CommonVoice languages, indicating out-of-the-box usability, and a 48kHz codec trained on only 1200 hours outperforms one trained on 33K hours. Key limitations: language labels rest on YouTube locale rather than verified language ID, the transcription-quality analysis promised in the abstract is not delivered, and the codec evaluation set's train/test separation is unstated; deduplication reduced the corpus from 1.4M to 1.1M hours after acceptance.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- YODAS v3 is the largest open speech dataset to date (over 1.1 million hours after post-acceptance deduplication) and the first truly large-scale corpus with high-fidelity stereo audio, at 48kHz across 147 languages under CC BY 3.0.
- A language-balanced crawl built from per-language Wikipedia keywords yields 22 languages with over 10K hours and 73 with over 5K hours, with English under 2% of the corpus.
- The corpus is genuinely high-fidelity: 92.5% of recordings exceed 32 kHz effective bandwidth and 71% are truly multichannel by spectral and channel-difference analysis.
- ASR models trained on the corpus without CTC-score filtering perform best on six of seven evaluated languages, showing the transcripts are usable out of the box, unlike YODAS v2 which required heavy filtering.
- A 48kHz neural codec trained on a 1200-hour subset outperforms models trained on larger or lower-rate corpora (LibriTTS, AMUSE) on both in-domain and LibriTTS evaluation.
Strengths
Verified audio fidelity, not just nominal specs
Sections 3.3-3.4 go beyond the 48kHz label: STFT-based bandwidth estimation (8192-Hann, 0.5% peak threshold, URGENT-2025-style mapping) shows 67.3% of recordings at 44.1 kHz effective bandwidth and 92.5% above 32 kHz; channel-difference RMS testing (threshold 1e-3) shows 71% (~780K hours) genuinely multichannel, including some 5.1 surround. They even cite URGENT 2024 findings that CommonVoice and DNS5 LibriVox are ~100% bandwidth-mismatched, positioning their audit as a contribution most dataset papers skip.
Unusual language balance for a web crawl
Language-specific Wikipedia keyword generation (Section 2.1: four syntactic filters, unigram-LM likelihood ranking) plus recency-prioritized search yields Arabic and Welsh on top at ~25K hours each, 22 languages over 10K hours, 73 over 5K, mean/median 7400/5180 hours per language, English under 2% (Section 3.1, Figure 2) — directly countering the English-dominance of YODAS v2, MLS, and Common Voice.
Out-of-the-box transcript usability demonstrated, not asserted
Strong codec result and fair, standardized tooling
Transparent error disclosure
Weaknesses
Language identity rests entirely on YouTube locale
The promised transcription-quality analysis never appears
Codec evaluation set provenance is undefined
ASR confound between language scale and the filtering trend
Headline shrank 21% after acceptance
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
YODAS v3 is a dataset paper with baseline experiments: a CC BY 3.0 corpus of over 1.1 million hours of web-mined speech across 147 YouTube locale labels, stored as original 48kHz multichannel Opus with weak transcripts (automatic and the rare manual subtitles — only 3.7% manual, down from 20% in YODAS v2), English translations, and descriptions. The headline claims are (i) largest open speech corpus to date and first large-scale stereo release, (ii) a language-balanced crawl with 22 languages over 10K hours and 73 over 5K, (iii) verified genuine fullband/multichannel quality, and (iv) baseline evidence of out-of-the-box usability for ASR and codec training. Footnote 1 candidly reports the total was corrected from 1.4M to 1.1M hours after acceptance-time deduplication.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is YODAS v3?
YODAS v3 is an openly licensed speech corpus of more than 1.1 million hours of audio drawn from Creative Commons-licensed YouTube videos in 147 languages. It is stored in the original 48kHz multichannel (mostly stereo) Opus format and ships with automatically generated transcripts, English translations, and video descriptions. It is the successor to YODAS v2 and, at release, the largest open speech dataset available, downloadable from Hugging Face (espnet/yodas3) under CC BY 3.0.
How was YODAS v3 collected?
The authors generated search keywords separately for each language from that language's Wikipedia dump, filtered them with simple quality rules, and ranked them with a unigram language model. They then searched YouTube for Creative Commons-licensed videos, prioritizing recently uploaded ones to avoid a view-count bias, and excluded any videos already in YODAS v2. Audio was downloaded at the maximum available quality (48kHz multichannel Opus) together with subtitles and descriptions.
Is YODAS v3 really stereo and fullband audio?
Largely yes. The paper measured actual spectral content rather than trusting format labels: 92.5% of recordings have an effective bandwidth above 32 kHz (67.3% at 44.1 kHz class), and about 71% of files (roughly 780K hours) have genuinely distinct channels rather than duplicated mono, including a small share of 5.1 surround material.
How balanced is YODAS v3 across languages?
Much more than typical web corpora: Arabic and Welsh are the largest at about 25K hours each, followed by Japanese (~21K) and English (~17K). 22 languages exceed 10K hours and 73 exceed 5K hours. One caveat from the paper itself: language labels come from the video's YouTube locale setting rather than verified language identification, so per-language counts inherit that uncertainty.
Can YODAS v3 be used for commercial purposes?
The corpus is distributed under CC BY 3.0, which permits commercial use with attribution. However, the underlying audio was crawled from Creative Commons-labeled YouTube videos, so downstream users should be aware of YouTube's terms of service and the usual caveats about CC labels being applied by uploaders rather than rights holders.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 (histograms, language distribution, bandwidth and channel charts) are referenced but their detailed contents cannot be verified from extracted text; judgments rely on the in-text numbers quoted for them.
- Table 1 of the paper (dataset comparison) is present in the extraction, but figure-level detail such as exact bin edges and per-language bars beyond the top languages named in text is unavailable.
- The Hugging Face repository was not inspected; statements about what the release contains (e.g., baseline checkpoints) rest solely on the paper text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.