Since 2025-10-31 the datasets are offered through Mozilla Data Collective, and the contributor terms of the same date place every contribution under CC0, while the datasheet adds that a downloader agrees not to determine the identity of speakers.
This page reads the corpus from two sides, the people who donated their voices and the holders of speech data whose recordings compete with a free set.
Within AI training data licensing, Common Voice is the donated, unpaid counterpart to licensed speech, and the terms differ in each layer: the recording, the file and the distribution route.
Common Voice's Publisher, Paper and Distribution Route
Common Voice is run by Mozilla, whose terms say contributors "donate your voice, written sentences, and the other resources" to build an open-source voice database, and since 2025-10-31 its datasets are available only through Mozilla Data Collective. The table lists the identity facts as of 2026-11-16.
| Attribute | Value | Source |
|---|---|---|
| Run by | Mozilla; the terms do not name a legal entity | github.com/mozilla/legal-docs (common_voice_terms.md) |
| Stated purpose | an open-source voice database "that anyone can use to make innovative voice recognition apps" | Common Voice Legal Terms, 2025-10-31 |
| Paper | "Common Voice: A Massively-Multilingual Speech Corpus", arXiv 1912.06670, submitted 2019-12-13, accepted to LREC 2020 | arxiv.org |
| Distribution | exclusively through Mozilla Data Collective since 2025-10-31; publisher shown as "Mozilla Common Voice" | Legal Terms; mozilladatacollective.com |
| Distributor | Mozilla Data Collective Limited, Companies House 17054959 | mozilladatacollective.com/terms; Companies House |
| Release statistics | public repository common-voice/cv-dataset | github.com |
| Governing law of the contributor terms | California | Legal Terms s.9 |
The 2019 paper describes the project at an early stage: its abstract reports 29 languages in the then latest release, over 50,000 individuals and 2,500 hours, and says the corpus is intended for speech technology research and development.
Those are 2019 figures, stated with that date, and the release tables below supersede them. The platform that distributes the corpus is a separate company, Mozilla Data Collective Limited, incubated by the Mozilla Foundation, and its listing terms belong to Mozilla Data Collective's own record.
What Common Voice Holds: Releases, Hours, Languages and Record Fields {#common-voice-releases}
Common Voice Scripted Speech release 27.0 holds 42,593 hours of read speech in 295 languages, of which 29,295 hours are validated, with a data cut-off of 2026-09-11 and a release post of 2026-09-21. The repository explains that the date in a release name is the cut-off for data collection and validation, not the release date, so the two dates are never interchangeable.
The release post states 32,349,999 clips and "a little over 42,593 hours", and the repository counts 379,190 contributors, summed per language, so a contributor active in several languages is counted once per language.
Spontaneous Speech release 5.0 is a second, smaller corpus with the same cut-off and covers 80 languages and 89,754 clips, with 302 validated hours of 541 total. Each language is a separate file bundle, so a Persian or an Arabic slice is one locale of a release and not a separate dataset.
The three largest locales by hours in the 27.0 file are Pashto (5,580.9), Catalan (3,940.31) and English (3,795.9). The first table compares the two corpora as of 2026-11-16.
| Attribute | Scripted Speech 27.0 | Spontaneous Speech 5.0 | Source |
|---|---|---|---|
| Cut-off date / release post | 2026-09-11 / 2026-09-21 | 2026-09-11 / 2026-09-21 | github.com/common-voice/cv-dataset; community.mozilladatacollective.com |
| Languages | 295 | 80 | same |
| Hours | 42,593 total, 29,295 validated | 541 total, 302 validated | same |
| Clips | 32,349,999 | 89,754 | release post |
| Contributors | 379,190, summed per language | not stated here | cv-dataset README |
| What a speaker does | reads a written sentence | answers an open question in their own words | cv-dataset READMEs |
| Record fields | hashed client_id, path, sentence_id, sentence, sentence_domain, up_votes, down_votes, age, gender, accents, variant, locale, segment; demographics only if the speaker opted in | hashed client_id, audio_id and transcription fields | cv-dataset READMEs |
| Format | MP3 clips, TSV metadata, a datasheet per locale since 25.0 | MP3, TSV, datasheet per locale | cv-dataset READMEs |
The second table lists selected Scripted Speech releases in chronological order, with the cut-off dates as stated in the release summary, checked 2026-11-16.
| Release | Cut-off date | Languages | Total hours | Validated hours | Source |
|---|---|---|---|---|---|
| v1 | 2019-02-25 | 19 | 1,368 | 1,096 | cv-dataset scripted-speech README |
| v8.0 | 2022-01-19 | 87 | 18,243 | 14,122 | same |
| v17.0 | 2024-03-15 | 124 | 31,175 | 20,408 | same |
| v23.0 | 2025-09-05 | 286 | 35,921 | 24,600 | same |
| v27.0 | 2026-09-11 | 295 | 42,593 | 29,295 | same |
The corpus grew from 1,368 hours in 19 languages in v1 (2019-02-25) to 42,593 hours in 295 languages in v27.0. The jump from 137 languages in v22.0 (2025-06-20) to 286 came with v23.0 (2025-09-05). Common Voice is the speech entry among the major AI training datasets.
How Common Voice Recordings Are Collected and Validated
Common Voice recordings come from volunteers who read sentences on its website, and other volunteers vote on whether each recording matches its sentence. The steps, in order, are these.
- Sentences: the community playbook says works subject to copyright cannot be added, so sentences come from CC0 or public domain text.
- Recording: a volunteer reads a sentence aloud, and no account is needed to contribute.
- Validation: a clip with two or more votes and more up votes than down votes counts as validated.
- Bundling: from v25.0, over-length, corrupted or missing clips are excluded, and each locale ships a datasheet README.
All voice contributions are released as part of datasets regardless of validation status, so the files include invalidated and unvoted clips next to the validated ones.
A reader comparing the corpus with licensed speech should therefore count validated hours, because unvoted clips carry no check that the audio matches its sentence: for release 27.0 the repository shows 29,295 validated hours against 42,593 total. How datasheets and data cards work in general is covered on the page about dataset documentation and datasheets.
Common Voice's Terms: CC0 With a Condition on Speaker Identity
Mozilla releases Common Voice under the Creative Commons CC0 public domain dedication, and every datasheet adds that by downloading the data "you agree to not determine the identity of speakers in the dataset".
The community playbook says the datasets "are released under a CC0 'No Rights Reserved' License and are part of the public domain". The Common Voice listings on Mozilla Data Collective show CC0-1.0, checked 2026-11-16, while the v27 release post says only "a permissive licence". Both wordings are kept visible here.
The condition on speaker identity sits in the datasheet, not in CC0 itself, so it binds a user through that agreement and not through the dedication. Separately, the terms ask that Common Voice datasets not be posted, distributed or mirrored on other platforms and that access take place through Mozilla Data Collective.
That is a request in Mozilla's terms, and this page does not assess its legal effect. Any copy offered on another hosting site sits outside the route the terms describe. The table sets the layers side by side as of 2026-11-16.
| Layer | Terms | What it covers | What it leaves open | Source |
|---|---|---|---|---|
| The dataset | CC0 public domain dedication; CC0-1.0 on each listing | copyright and related rights, waived "to the extent we can under the law" | personal-data rules on the recordings | Legal Terms s.2; playbook; mozilladatacollective.com |
| The datasheet | datasheet: agreement "to not determine the identity of speakers" | the downloader's use of speaker identity | how it is enforced (not stated) | cv-datasheets templates/i18n/en.json |
| The distribution route | request not to post, distribute or mirror; access only through Mozilla Data Collective | where copies are offered | legal effect on third-party copies (not assessed) | Legal Terms s.2 |
CC0 waives copyright only "to the extent we can under the law", and it does not settle questions about personal data, which the legal standing section below takes up. How CC0 compares with CC BY, ODC and CDLA terms is covered on the page about dataset licenses for AI training.
What Common Voice Contributors Agree To
Every Common Voice contributor agrees that Mozilla may offer all their contributions, including text, recordings, validations and feedback, to the public under CC0, under Legal Terms effective 2025-10-31. The table lists the terms as of 2026-11-16.
| Term | What the documents say | Source |
|---|---|---|
| License of contributions | text, recordings, validations and feedback offered under CC0 | Legal Terms s.2, effective 2025-10-31 |
| Assurances | own creation; no third-party rights infringed; Acceptable Use Policy | Legal Terms s.2 |
| Eligibility | over 19; 19 or under with a parent's or guardian's consent and supervision | Legal Terms s.1 |
| Spontaneous Speech | no sensitive or personal data about oneself or others | Legal Terms s.2 |
| Demographics | optional (accent, age, gender), published with the recordings; not linked to submitted sentences | Legal Terms s.3; Privacy Notice |
| Email and username | email never published; leaderboard opt-out available | Legal Terms s.3 |
| Leaving | account deletion removes username and email; contributions remain public after termination of access; Spontaneous Speech clips removed on request | Privacy Notice; Legal Terms s.8; cv-dataset README |
| Liability | Mozilla's liability limited to USD 500 | Legal Terms s.5 |
The contributor gives three assurances: the contribution is their own creation, it infringes no third-party rights, and it complies with the Acceptable Use Policy. Eligibility is set by age: anyone over 19 may take part, and a contributor aged 19 or under needs a parent's or guardian's consent and supervision.
Demographics are optional and published with the recordings, while the email address is never published, and opting out of the leaderboard still leaves the recordings public. Spontaneous Speech contributors also agree not to include sensitive or personal data about themselves or others in the recordings.
On leaving, the privacy notice says an account can be deleted at any time and the username and email are removed. The terms say contributions "will remain publicly available as part of Common Voice, even if we terminate or suspend your access". The Spontaneous Speech documentation says clips are removed only at the contributor's request.
For Scripted Speech, the documents read describe no route to withdraw recordings from releases already downloaded, and this page makes no claim about how Mozilla answers such requests. The release post calls the corpus one of "consented, known provenance", which is Mozilla Data Collective's own wording.
Common Voice's Legal Standing: No Court Record and Voice as Personal Data
No court ruling, pending case or licensing deal on file names Common Voice, as of 2026-11-16. The search covered the court records and deals tracked on this site, and no dispute over Common Voice's recordings, terms or contributors appears in them.
Copyright in the recordings and sentences is waived under CC0 to the extent the law allows.
A voice recording with a hashed contributor ID and optional age, gender and accent can still be personal data in many jurisdictions. Under Article 9(1) of Regulation (EU) 2016/679, which applies from 2018-05-25, biometric data processed to uniquely identify a natural person is a special category.
The downloader's agreement not to identify speakers works alongside such rules and does not replace them. How the regulation applies to training sets is covered on the page about GDPR and AI training data.
The same recording can also carry voice and likeness interests and biometric-law questions, which this page does not assess.
A Speech Data Holder and Common Voice: Benefits, Risks, Downsides and Alternatives
For an organization or community holding speech recordings, Common Voice is both a free competitor to licensed speech and a route for donating recordings openly, and donating brings no payment and cannot be undone for copies already downloaded. The four lists separate what the route offers from what it costs.
Benefits
- Known terms for contributors and users: CC0 for contributions and the datasheet condition for downloaders.
- Wide reach as the speech entry among public corpora, with 295 languages in release 27.0.
- Per-locale datasheets that document what was collected.
Risks
- Contributions stay public once released, even if access is ended.
- The corpus carries personal data whatever the downloader condition says.
- Copies offered on other platforms would sit outside the distribution route the terms describe.
Downsides
- No payment, since contributions are donated under CC0.
- Free hours compete with paid speech, so a holder's licensed set must offer what the open corpus lacks, such as consent records for commercial use or validated hours in its language.
Alternatives
- License recordings directly to AI developers, as described in the page on selling speech and audio data for AI training.
- List them on a dataset platform at a set price.
- Keep the recordings internal.
- Not license or donate at all.
Nothing in the contributor terms creates a claim to payment, and not licensing is a valid outcome.
Speech Licensing, Voice Rights and Dataset Terms Beyond This Page
This page covers the Common Voice corpus and its terms; selling a voice, voice rights, biometric rules and other open corpora each have their own page.
- Selling your voice to AI: an individual's voice, licensed.
- Voice and likeness rights in AI training data: voice rights beyond copyright.
- Illinois BIPA and AI training data: a state biometric statute.
- Privacy law and AI training data: personal data by jurisdiction.
- Nonprofits licensing data to AI companies: foundations and communities as holders.
- FineWeb: an open text corpus.
- Licensed vs scraped training data: the licensed alternative.
- Alternatives to selling data, including not selling at all: not licensing at all.
- Biometric data in AI training: voiceprints under biometric law.
- Multilingual and low-resource language data: language coverage beyond speech.
- Public domain and openly licensed material: public domain material in general.
- Removing content from AI training datasets: removal routes across datasets.
- SpeechData.ai: a commercial speech dataset vendor.
Frequently Asked Questions About Mozilla Common Voice
Is Mozilla Common Voice free?
Yes, in price and in copyright terms: Common Voice is released under CC0 and offered through Mozilla Data Collective. Downloading comes with the datasheet's agreement not to determine the identity of speakers, so free does not mean free of conditions.
How big is the Common Voice dataset?
Scripted Speech 27.0 holds 42,593 hours, 29,295 of them validated, in 295 languages, with a cut-off of 2026-09-11 and a release post of 2026-09-21. Spontaneous Speech 5.0 holds 302 validated hours in 80 languages. The two are never added together.
What license does Common Voice use?
Common Voice uses the CC0 public domain dedication according to Mozilla's terms and community playbook. The v27 release post says "a permissive licence", and the platform's listings show CC0-1.0.
Can contributors remove their recordings from Common Voice?
Deleting an account removes the username and email, and Spontaneous Speech clips are removed at the contributor's request. The documents read describe no withdrawal of Scripted Speech clips from releases already downloaded. This page is information, not legal advice.
Agents and advisers can use the introducer form.