Elamur Research
☰
home / papers / 2019 / Common Voice: A Massively-Multilingual Speech Corpus

Common Voice: A Massively-Multilingual Speech Corpus

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive licence to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/1912.06670

Summary

Most of the world's languages lack openly licensed transcribed speech for training speech recognizers, and existing multilingual collections are either restrictively licensed, have no quality control, or have no community behind them. Mozilla's Common Voice tackles this with crowdsourcing. Volunteers record themselves reading prompt sentences on a website or iPhone app, and other volunteers vote on whether each clip matches its text. Any community can add its language by localizing the interface and gathering a pool of public-domain sentences. The resulting corpus is released under CC0 and, by the authors' account, is the largest public-domain ASR corpus in both hours and number of languages. To show its usefulness, the authors fine-tune an English DeepSpeech model on twelve other languages. Transfer learning lowers Character Error Rate on average, and the majority of the languages tested had never before had end-to-end ASR results published. The project offers a scalable and sustainable route to speech data for low-resource languages.

Key points

  • Scale: The June 2019 release covers 29 languages, and 38 language communities were collecting data by November 2019. Contributors number over 50,000, with roughly 2,500 hours recorded and about 2,000 hours validated.
  • Validation:
    • Up to three listeners judge each clip; two up-votes make it valid and two down-votes make it invalid.
    • Train, dev and test splits are speaker-disjoint, with sizes set by power analysis (99% confidence, 1% margin).
    • Audio is distributed as 48kHz mono MP3.
  • Adding a language:
    • The roughly 610 interface strings are translated in Pontoon.
    • At least 5,000 sentences must be gathered, either through the Sentence Collector or by extraction from Wikipedia.
  • Experiments:
    • The model is DeepSpeech v0.3.0, a 6-layer CTC network with one LSTM layer.
    • Layers are transferred from an English model to 12 languages.
    • Training data ranges from under 1,000 clips (Slovenian) to over 65,000 clips (German).
  • Results: Mean CER improvement is 5.99 ± 5.48. Copying four English layers performed best overall. The largest gain was for Chuvash, where CER fell from 48.1 to 27.21.

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: Other

Cited by

531 papers of our library cite it (of the 140673 whose bibliographies we read); 14 of them have a page here, 517 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1912.06670 · arXiv 1912.06670

@misc{ardila2019common,
  title = {{Common Voice: A Massively-Multilingual Speech Corpus}},
  author = {Ardila, Rosana and Branson, Megan and Davis, Kelly and Henretty, Michael and Kohler, Michael and Meyer, Josh and Morais, Reuben and Saunders, Lindsay and Tyers, Francis M. and Weber, Gregor},
  year = {2019},
  doi = {10.48550/arxiv.1912.06670},
  eprint = {1912.06670},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1912.06670}
}

References

9 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Fiscus, J. G., Ajot, J., Radde, N., and Laprun, C. (2006). Multiple dimension levenshtein edit distance calculations for evaluating automatic speech recognition systems during simultaneous speech. In LREC, pages 803– 808. Citeseer. find
  2. Gales, M. J., Knill, K. M., Ragni, A., and Rath, S. P. (2014). Speech recognition and keyword spotting for low-resource languages: Babel project research at cued. In Spoken Language Technologies for Under-Resourced Languages. find
  3. Glorot, X. and Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. find
  4. Graves, A., Fernández, S., Gomez, F., and Schmidhuber, J. (2006). Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369–376. ACM. find
  5. Hannun, A. Y., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., Prenger, R., Satheesh, S., Sengupta, S., Coates, A., and Ng, A. Y. (2014). Deep Speech: Scaling up end-to-end speech recognition. CoRR, abs/1412.5567. find our card
  6. M-AILABS. (2019). The m-ailabs speech dataset. https://www.caito.de/2019/01/the-m-ailabs-speech-dataset/. accessed 11/25/2019. ↗
  7. Meyer, J. (2019). Multi-Task and Transfer Learning in Low-Resource Speech Recognition. PhD dissertation, The University of Arizona. find
  8. Roter, G. (2019). Sharing our common voices mozilla releases the largest to-date public domain transcribed voice dataset, Feb. https://blog.mozilla.org/blog/2019/02/28/sharing-our-common-voices-mozilla-releases-the-largest-to-date-public-domain-transcribed-voice-dataset/. ↗
  9. VoxForge. (2019). Voxforge. http://www. voxforge.org/. accessed 11/25/2019. find