Elamur Research⌕
☰
home / papers / 2025 / Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use

Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use

Paper · 2025 · Titouan Parcollet, Yuan Tseng, Shucong Zhang et al.

Summary

LibriSpeech has long served as the shared English ASR benchmark, but it contains only clean read audiobook speech, so systems now score close to zero errors on it while failing on real-world audio. Newer large corpora have their own problems: licences that forbid industrial use, noisy or misaligned transcripts, audio in the wrong language, or no defined evaluation splits. The authors built the Loquacious Set to address these gaps. It is a curated 25,000-hour English corpus drawn from six sources whose Creative Commons licences permit commercial use. They selected and filtered the audio, ran language identification on the YouTube-derived data, normalised all transcripts into one consistent style, and defined nested training subsets plus fixed dev and test sets. SpeechBrain conformer recipes trained on the corpus with as few as four GPUs beat considerably larger Whisper models on LibriSpeech test-other and on VoxPopuli. The result is a reproducible, legally safe benchmark that academic and industrial labs can both use, covering accents, noise, and spontaneous speech.

Key points

  • The six sources are CommonVoice 18.0, VoxPopuli, Libriheavy, People's Speech, YODAS and LibriSpeech, all under CC0, CC BY 3.0 or CC BY 4.0. The reproduction code is released under Apache 2.0 and hosted on Hugging Face as speechbrain/LoquaciousSet.
  • The splits are small (250 h), medium (2,500 h), large (25,000 h), clean (about 13,000 h) and dev/test (about 16.5 h each).
    • Small and medium take 50 h and 500 h from each of five sources.
    • The clean split excludes People's Speech and YODAS because their transcripts remain unreliable.
  • Clips shorter than 1 s or longer than 40 s are removed. For YODAS, the minimum is raised to 3 s because shorter clips were mostly noise.
    • The large split holds 9.4 million clips averaging 9.6 s.
  • Language identification on three YODAS English subsets kept only 9,200 of 16,500 hours. The authors estimate that roughly half of the "English" YODAS data is not English.
  • On LibriSpeech test-other, the authors' models compare with Whisper as follows:
    • Their 100M model scores 5.7% WER, versus 6.7% for Whisper Small.en (244M).
    • Their 480M model scores 4.6%, versus 5.1% for Whisper large v2 (1.5B).
    • On VoxPopuli, the 480M model scores 6.9% against Whisper's 7.5%.

Interesting ideas

  • I1. Training splits are nested so that each smaller set lies inside the larger one, and every source contributes an equal quota to the small sets; this makes cheap pilot runs comparable with full-scale runs and stops one domain from dominating.
  • I2. A separate "clean" split is restricted to sources whose labels were checked by humans before release, while noisier sources are used only in the large mix.
  • I3. Metadata is not trusted blindly: an independent classifier checks the claimed property of each item (here, the language), and spot checks by hand confirm the rejections, which revealed that about half of one corpus was mislabelled.
  • I4. The fixed evaluation sets combine clean, accented, and spontaneous material, each manually verified, so that one score reflects robustness rather than performance on easy data only.
  • I5. Every component's licence is audited, and only sources allowing commercial use are kept, while each retains its original licence.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Silver treebanks could be released as nested tiers with equal sentence quotas per genre (fables, school texts, news), so parser experiments at small scale predict behaviour at full scale.
  • I2 (learn): Witness-agreed or human-checked parses could form a distinct clean tier, kept apart from LLM-teacher output, so training and negative-control evaluation can be run on the trusted portion alone.
  • I3 (tree): Scraped or converted corpora should pass a language and sanity filter before parsing or treebank conversion, with manual review of a random sample of rejected items to measure the filter's precision.
  • I4 (answer): Test suites for readers could mix easy school texts with messy or informal ones in fixed, versioned splits, so that gains on clean fables are not mistaken for general competence.
  • I5 (world): Each compiled knowledge source (WordNet, FrameNet, Wikidata, VerbNet) should carry recorded licence metadata in the knowledge graph, so that derived artefacts can be redistributed safely.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksSpeech & spoken language · 480 works

Cite

Cite as Titouan Parcollet, Yuan Tseng, Shucong Zhang and Rogier van Dalen. 2025. Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use. arXiv:2505.21578.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.21578 · arXiv 2505.21578

@misc{parcollet2025loquacious,
  title = {{Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use}},
  author = {Titouan Parcollet, Yuan Tseng, Shucong Zhang and Rogier {van Dalen}},
  year = {2025},
  doi = {10.48550/arxiv.2505.21578},
  eprint = {2505.21578},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.21578}
}

References

17 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. D. B. Paul and J. Baker, “The design for the wall street journal-based csr corpus,” in Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992. find
  2. J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett, “Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,” NASA STI/Recon technical report n, vol. 93, p. 27403, 1993. find
  3. V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210. find
  4. A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning. PMLR, 2023, pp. 28 492–28 518. ↗
  5. K. Zhang, L. Chkhetiani, F. M. Ramirez, Y. Khare, A. Vanzo, M. Liang, S. R. Martin, G. Oexle, R. Bousbib, T. Peyash et al., “Conformer-1: Robust asr via large-scale semisupervised bootstrapping,” arXiv preprint arXiv:2404.07341, 2024. ↗
  6. M. Gaido, S. Papi, L. Bentivogli, A. Brutti, M. Cettolo, R. Gretter, M. Matassoni, M. Nabih, and M. Negri, “Mosel: 950,000 hours of speech data for open-source speech foundation model training on eu languages,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 13 934–13 947. ↗
  7. X. Li, S. Takamichi, T. Saeki, W. Chen, S. Shiota, and S. Watanabe, “Yodas: Youtube-oriented dataset for audio and speech,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8. find
  8. W. Kang, X. Yang, Z. Yao, F. Kuang, Y. Yang, L. Guo, L. Lin, and D. Povey, “Libriheavy: a 50,000 hours asr corpus with punctuation casing and context,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 991–10 995. find
  9. D. Galvez, G. Diamos, J. Ciro, J. F. Cerón, K. Achorn, A. Gopi, D. Kanter, M. Lam, M. Mazumder, and V. J. Reddi, “The people’s speech: A large-scale diverse english speech recognition dataset for commercial usage,” arXiv preprint arXiv:2111.09344, 2021. ↗
  10. W. Chan, D. Park, C. Lee, Y. Zhang, Q. Le, and M. Norouzi, “Speechstew: Simply mix all available speech recognition data to train one large neural network,” arXiv preprint arXiv:2104.02133, 2021. ↗
  11. Y. Peng, J. Tian, B. Yan, D. Berrebbi, X. Chang, X. Li, J. Shi, S. Arora, W. Chen, R. Sharma et al., “Reproducing whisper-style training using an open-source toolkit and publicly available data,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8. find
  12. G. Chen, S. Chai, G. Wang, J. Du, W.-Q. Zhang, C. Weng, D. Su, D. Povey, J. Trmal, J. Zhang et al., “Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio,” arXiv preprint arXiv:2106.06909, 2021. ↗
  13. R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” arXiv preprint arXiv:1912.06670, 2019. ↗
  14. M. Ravanelli, T. Parcollet, A. Moumen, S. de Langen, C. Subakan, P. Plantinga, Y. Wang, P. Mousavi, L. D. Libera, A. Ploujnikov, F. Paissan, D. Borra, S. Zaiem, Z. Zhao, S. Zhang, G. Karakasidis, S.-L. Yeh, P. Champion, A. Rouhe, R. Braun, F. Mai, J. Zuluaga-Gomez, S. M. Mousavi, A. Nautsch, H. Nguyen, X. Liu, S. Sagar, J. Duret, S. Mdhaffar, G. Laperrière, M. Rouvier, R. D. Mori, and Y. Estève, “Open-source conversational ai with speechbrain 1.0,” Journal of Machine Learning Research, vol. 25, no. 333, 2024. [Online]. Available: http://jmlr.org/papers/v25/24-0991.html ↗
  15. V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” arXiv preprint arXiv:2012.03411, 2020. ↗
  16. C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haziza, M. Williamson, J. Pino, and E. Dupoux, “Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021, pp. 993–1003. ↗
  17. S. Kim, T. Hori, and S. Watanabe, “Joint ctc-attention based end-to-end speech recognition using multi-task learning,” in 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, pp. 4835–4839. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/parcollet-2025-loquacious-set-25-000-hours/