Redistributed

General

Mozilla Common Voice

Published March 8, 2026

dataresearchinfrastructure

An open dataset of human voices in 100+ languages, crowdsourced and freely available to train speech recognition systems.

What It Is

Common Voice is Mozilla's answer to a simple problem: voice assistants work well in English and major commercial languages, but fail most of the world. The project collects voice recordings through a web interface where anyone can donate their voice by reading sentences aloud. Other volunteers validate recordings for accuracy.

The result is a speech dataset that:

  • Covers 100+ languages — including many with no commercial speech recognition at all
  • Is CC0 licensed — true public domain, not "open with restrictions"
  • Includes demographic metadata — age, gender, accent (voluntarily provided) for bias research
  • Grows continuously — volunteers add recordings daily

Since launching in 2017, Common Voice has collected over 20,000 hours of validated speech across more than 100 languages. For many of these languages, it's the only substantial open speech dataset that exists.

Sustainability Model

Mozilla Foundation funds the infrastructure and a small staff. The actual data collection depends entirely on volunteer labor — people donating their voices and validating others' recordings.

This creates a foundation + volunteer collective hybrid: professional infrastructure, community-generated content. Mozilla provides the platform, hosting, and coordination; communities provide the voices and quality control.

The CC0 license means anyone can use the data, including commercial entities. Mozilla accepts this tradeoff — broader adoption of open voice data outweighs exclusive control.

Why It Matters for Redistribution

Who benefits?

  • Speakers of underserved languages — voice tech becomes possible for their languages
  • Researchers — access to multilingual speech data without corporate partnerships
  • Open source projects — can build speech recognition without proprietary datasets
  • Communities — control over how their language is represented in AI

Who pays?

  • Mozilla Foundation covers infrastructure costs
  • Volunteers donate time (recording and validating)
  • Communities do the work of building linguistic resources

What changes? The default pattern in voice AI: companies build for profitable language markets, ignore the rest. English voice assistants work well because training data is abundant. Welsh, Yoruba, Basque? Not commercially viable.

Common Voice inverts this. A community that wants voice recognition for their language can build toward it — not by waiting for tech companies to notice them, but by contributing recordings directly. The dataset is theirs to use.

This is practical linguistic sovereignty. When a language community controls its own speech data, outside entities can't extract it without giving back. The CC0 license means the data is open to all — but it's also open to the community itself.

As Ruha Benjamin argues in Race After Technology, the design of technical systems embeds social choices (Benjamin, 2019). Common Voice makes different choices visible: what if voice AI served linguistic diversity rather than market concentration?

The project connects to several redistribution currents. It directly addresses The Digital Divide in AI — not through policy argument but through infrastructure. It tackles Who Owns the Training Data by ensuring the answer is "everyone" via CC0 licensing. And it demonstrates what Public AI can look like in practice: publicly funded infrastructure, community-generated content, universally available outputs. When the Gender Shades study revealed demographic biases in commercial AI, Common Voice's inclusion of demographic metadata gave researchers the tools to measure and address similar biases in speech recognition.

Current Status

Active and growing. Key milestones:

  • 20,000+ hours of validated recordings
  • 100+ languages represented
  • Major speech models trained on Common Voice data, including Mozilla's own DeepSpeech and community projects like Coqui TTS

Recent focus areas include African languages, Indigenous languages, and dialect coverage. The project runs periodic "contribution sprints" to accelerate data collection for specific languages.

Challenges remain: some languages have thousands of hours, others have minutes. Volunteer contribution is uneven, and sustaining momentum for lower-resource languages requires ongoing community organizing.

How to Get Involved

  • Donate your voice: commonvoice.mozilla.org — record sentences in your language
  • Validate recordings: Listen to others' contributions and confirm accuracy
  • Download the dataset: Free access for research and development
  • Organize locally: Start contribution drives for your language community
  • Build with it: Train speech recognition models on the data

Sources

  • Mozilla. "Common Voice." commonvoice.mozilla.org
  • Ardila, Rosana et al. "Common Voice: A Massively-Multilingual Speech Corpus." LREC, 2020. arXiv
  • Benjamin, Ruha. Race After Technology. Polity Press, 2019. Publisher
  • Mozilla Foundation. "Common Voice Dataset." GitHub
← Back to all articles
On this page