Mozilla Common Voice

Common Voice is Mozilla's effort to build an open, public-domain speech corpus that anyone can use to train voice technology. Volunteers record themselves reading short sentences, and a dual-review process (two independent approvals) validates each clip before it enters a release. As of release 19.0 it had reached 32,584 hours of speech across 131 languages, with hours per language ranging from a handful to thousands, making it one of the largest openly licensed voice datasets around.

Each versioned release ships as `MP3` audio paired with `TSV` files that hold the transcripts and speaker metadata, split into validated, invalidated, and other buckets so you can pick how strict you want to be. You download per language rather than pulling everything at once. Since October 2025 the datasets are distributed only through the Mozilla Data Collective, not HuggingFace, so if you scripted downloads from HuggingFace in the past you will need to repoint them at the Data Collective.

For RAG specifically, Common Voice is not a text knowledge base you retrieve over. Its role is upstream: use it to train or fine-tune the automatic speech recognition (ASR, turning spoken audio into text) or text-to-speech models that sit at the front of a voice-driven RAG system, so a spoken question becomes clean text your retriever can search. The paired transcripts make it ready for exactly that kind of training.

Watch a few things. The clips are read speech, people reading prompts aloud, not spontaneous conversation, so a model trained only on Common Voice may struggle with natural, messy speech. Coverage is uneven across languages, and recording quality and accents vary because it is crowdsourced. On the plus side the licence is genuinely clean: CC0 means you can use the audio and transcripts for anything, including commercial products, with no attribution required.

It pairs well with LibriSpeech, which offers cleaner read English under CC BY 4.0, where Common Voice's strength is breadth of languages and accents.