"Free" and "free to use in a commercial product" are two different questions, and a lot of procurement research collapses them into one. LibriSpeech, LJSpeech, Common Voice, and AliMeeting are all free to download. Whether you can train a model on them and sell that model is a separate question, answered by the licence attached to each release, not by the fact that no invoice arrived.
This guide walks through the licence types you will actually meet, CC0, CC BY 4.0, CC BY-NC, CC BY-SA, and research-only terms, verifies what each of those four named corpora is actually licensed under, and lays out the risks a clean licence does not cover, alongside the cases where free data is genuinely the right call.
The short answer
Some free speech datasets are cleared for commercial use. Others are not, and the fee, or lack of one, tells you nothing about which is which. The variable that decides it is the licence attached to the release, not the price tag.
Take the query that sends people to this page more than any other: is there a free Greek ASR dataset cleared for commercial use? Mozilla Common Voice includes a Greek dataset released under a CC0 waiver, which is about as clean a yes as licensing gets: no attribution requirement, no share-alike clause, no research-only carve-out. But Common Voice's Greek coverage is thin and crowdsourced, so a clean licence does not mean the hours or consistency are production-ready. That gap, between legally clear and actually sufficient, runs through almost every free corpus, and it is what the rest of this guide works through.
The licences you will actually meet
Five licence patterns cover almost every open speech corpus you will run into. Knowing which one you are looking at, before a data pipeline gets built on top of it, is the entire exercise.
| Licence | What it permits for a shipped commercial model | Obligation it carries |
|---|---|---|
| CC0 / public domain | Full commercial use: training, fine-tuning, and redistribution, with no restriction. | None. No attribution, no share-alike. |
| CC BY 4.0 | Full commercial use, including training a model you sell or ship. | Attribution to the original creator, typically in documentation or a model card. |
| CC BY-NC | Non-commercial use only. Training a model you monetise, directly or indirectly, falls outside the grant. | Attribution, plus the non-commercial condition itself. |
| CC BY-SA | Commercial use is allowed, but derivative works must carry the same licence. | Attribution and share-alike, which is hard to reconcile with a proprietary model. |
| Research-only or custom terms | Usually restricted to non-commercial research, evaluation, or benchmarking; read the specific grant. | Whatever the custom agreement states. Assume nothing beyond it. |
The two that cause the most damage are CC BY-NC and CC BY-SA, because both look permissive at a glance and are not. Our guide to speech data licensing goes deeper into how these terms interact with model weights, which is a genuinely unsettled question even among IP lawyers.
The major open corpora and what their licences allow
Four corpora account for most of the free ASR dataset search traffic, and each sits in a different place on the table above.
LibriSpeech, hosted on OpenSLR as resource SLR12, is released under CC BY 4.0. It contains roughly 1,000 hours of read English speech drawn from LibriVox public-domain audiobooks, split into clean and noisier training, development, and test sets. CC BY 4.0 clears it for commercial use, model training included, provided attribution carries through to your documentation.
LJSpeech is public domain in the US: "no restrictions on its use," in the creator's own words. It is a single English speaker reading roughly 24 hours across 13,100 clips, drawn from seven public-domain non-fiction books and recorded by the LibriVox project. Attribution is requested, not required, which makes it one of the cleanest licences on this list, though its single-speaker, single-domain nature limits what it is good for beyond TTS work.
Mozilla Common Voice is released under CC0, a full public-domain dedication with no attribution requirement at all. It is also the widest in language coverage, spanning well over 100 languages, Greek and Vietnamese both included, because it is built by volunteer contributors reading and validating sentences rather than one fixed recording effort. That crowdsourced model is also its limit: depth per language varies enormously, and a clean licence does not guarantee sufficient hours or acoustic diversity for any given language.
AliMeeting, an Alibaba-released Mandarin meeting corpus hosted on OpenSLR as resource SLR119, is licensed CC BY-SA 4.0, not research-only as it is sometimes described. That is commercially usable, but the share-alike clause means a derivative you distribute may need to carry the same licence, a real complication if the derivative in question is a proprietary model. Our guide to open speech datasets catalogues these and other corpora in more depth, and our guide to ASR training data covers how corpora like these fit into a broader training pipeline.
The four risks free data carries into production
A clean licence answers a narrower question than most teams assume: it says the corpus owner will not pursue you for copyright infringement. It does not resolve everything else a legal or compliance review will ask.
- Attribution obligations. CC BY licences require crediting the source, and that requirement does not expire once a model ships. If attribution lives only in a training script nobody maintains, it quietly falls out of your documentation and becomes a compliance gap the next time someone audits data lineage.
- Share-alike contamination. CC BY-SA corpora like AliMeeting require derivative works to carry the same licence. Whether a trained model counts as a derivative work is genuinely unresolved, and legal teams read it differently. The conservative position, and the one most counsel land on, is to treat CC BY-SA data as carrying real risk for a proprietary model.
- Unclear speaker consent. A Creative Commons licence governs copyright in the recording and transcript. It says nothing about whether the person speaking agreed to have their voice used to train a commercial product, a separate question the licence cannot answer either way.
- No indemnity. Open datasets ship with no warranty and no one to call if a rights claim surfaces later. A commercial vendor contract typically gives you a counterparty and some contractual protection; a public download gives you neither.
When free data is genuinely the right call
None of this makes free data a poor choice. It makes it the right choice for a narrower set of jobs than the search volume around it suggests.
Prototyping is the clearest case. Before you know whether an architecture or approach works at all, spending a procurement cycle on licensed data is premature. LibriSpeech or Common Voice will show whether the model trains before you spend a cent.
Benchmarking is another. LibriSpeech's test-clean and test-other sets are the standard yardstick for English ASR word error rate precisely because they are free, fixed, and widely reported against, which makes your numbers comparable to everyone else's.
Augmentation is the third, and the one teams underuse. Blending a slice of Common Voice or LibriSpeech into a larger licensed set can add acoustic diversity, different microphones, rooms, and accents, more cheaply than commissioning more of the same condition from scratch.
All three are legitimate engineering practice, not a workaround. The distinction that matters is whether the output of that work ships to paying customers, or stays inside a lab.
When you need licensed data instead
The moment a model trained on this data reaches a paying customer, the questions change: not "is this legally downloadable" but "did the speakers consent to this specific use," "is there anyone to call if a claim surfaces," and "does coverage in this language or domain actually exist at production depth." Free corpora answer the first question well and the other two poorly or not at all.
That is also where free data thins out fastest. Outside a handful of well-served languages, hours are shallow, recording conditions are narrow, mostly clean read speech, and specific domains such as call-centre audio, in-car noise, or a named dialect rarely exist at all. Our guide to buying AI training data walks through how to scope that gap before it becomes a launch blocker.
To be direct about the trade-off: if the only need is a benchmark number or a working prototype, free data is the right tool, there is no reason to pay for what a public test set already provides. Licensed data earns its cost when consent, indemnity, and language depth actually matter to what ships.