Common Voice is a crowd-sourced, multilingual corpus of read speech created by the Mozilla Foundation to make voice technology more open and inclusive. Volunteers read short sentences aloud and validate one another’s recordings through a community up/down-vote process, so a clip is marked validated once it receives enough independent approvals. The project spans well over a hundred languages, including many low-resource ones that are underserved by commercial datasets.
What’s in it
Each versioned release bundles validated and unvalidated MP3 clips with their target transcripts and optional speaker demographic metadata such as age, gender, and accent. Predefined train, dev, and test splits are provided to support reproducible ASR experiments. The corpus grows continuously through periodic releases and now totals tens of thousands of recorded hours, of which a large validated subset is recommended for training.
Tasks, license & access
Common Voice is used primarily for automatic speech recognition (ASR), but its audio and transcripts also support speech synthesis, spoken language identification, and accent and fairness research. It is released under the CC0 public-domain dedication, making it one of the most permissively licensed large speech datasets; releases are downloadable per language from the project site and mirrors such as the Hugging Face Hub.
- Modality: speech audio with aligned transcripts and demographics
- Scale: 100+ languages; tens of thousands of hours, large validated subset
- Tasks: ASR, TTS, language and accent identification
- License: CC0 (public domain)
- Creator: Mozilla Foundation and its volunteer community
Its linguistic breadth and permissive license make Common Voice a cornerstone for building speech interfaces beyond a handful of high-resource languages.