Mozilla Common Voice — Crowdsourced Open Speech Corpus Extending Voice Technology to 250+ Languages

Common Voice, launched by Mozilla in June 2017, is a crowdsourced project in which volunteers record sample sentences and review other people's recordings. The resulting datasets are released under a CC0 public-domain licence so that anyone can build speech-to-text or voice applications without cost, including for languages commercial vendors neglect.
The peer-reviewed LREC 2020 paper (Ardila et al.) reports 29 languages in the latest release and 38 collecting data as of November 2019, with over 50,000 individuals and about 2,500 hours of audio. In transfer-learning experiments with Mozilla's DeepSpeech across twelve target languages, starting from an English model gave an average character-error-rate improvement of 5.99 ± 5.48.
Growth documented on Wikipedia: about 500 hours of English from 20,000+ users in November 2017; roughly 1,400 hours from 42,000+ contributors in 18 languages in February 2019; 7,226 hours in 54 languages (5,591 validated) by July 2020; and over 250 languages by December 2025. Twi, from Ghana, became the 100th language in September 2022, and community partners such as Bengali.AI have used the platform to gather speech data.
Caveats: the 'largest public-domain corpus' claim is qualified by the authors as 'to our knowledge', transfer-learning experiments covered only twelve languages, and per-language quality and speaker diversity vary with the size of each volunteer community. This entry describes the data-collection model, not guaranteed recognition accuracy for any one language.

Read the full analysis: https://commonvoice.mozilla.org