A team lead messaged me last month with a simple question: we trained our acoustic model on LibriSpeech, can we ship it commercially? The honest answer was yes. The less honest answer, the one I see teams skip, is that "free" comes with a job attached, and it's a different job for each of the three corpora everyone reaches for first.
LibriSpeech, LJSpeech, and Mozilla Common Voice show up in nearly every speech dataset commercial use question I get from engineers scoping a new model. They're free, they're large enough to be useful, and they're everywhere in published baselines. But "free for commercial use" is doing a lot of quiet work in that sentence, and the three corpora don't mean the same thing by it. Our guide to free speech datasets and commercial use covers the landscape end to end. This is the narrower version: corpus by corpus, what the license actually obliges you to do.
LibriSpeech: CC BY 4.0, and attribution is not optional
LibriSpeech, distributed through OpenSLR as SLR12, is roughly a thousand hours of read English audiobook speech, segmented and aligned by Vassil Panayotov with Daniel Povey. It's released under Creative Commons Attribution 4.0 (CC BY 4.0). That license lets you use it commercially, modify it, redistribute derivatives, all without asking permission or paying a fee. What it requires in return is attribution: you credit the source, you don't imply endorsement, and you keep that credit attached through however many hops your pipeline takes.
The place this goes wrong isn't the initial download. It's three months later, when LibriSpeech has been blended into a larger internal training manifest alongside four other sources, the original folder structure is gone, and nobody remembers which utterances came from where. Attribution obligations don't expire when the data gets harder to trace, they just get harder to honor. The fix is boring and it works: log the corpus name, version, and license at ingestion time, not at release time. A line in a data card or a model card citing LibriSpeech and linking back to OpenSLR takes thirty seconds to write and closes the obligation for good.
LJSpeech: public domain, no obligation at all
LJSpeech is a different animal. It's roughly 24 hours of a single speaker reading passages from seven non-fiction books published between 1884 and 1964, all now in the public domain, recorded for the LibriVox project. Keith Ito's dataset page is explicit that both the text and the audio are public domain, meaning there's no copyright holder to attribute and no license clause to satisfy. You can use it, modify it, and ship a commercial TTS model trained on it without citing anything.
That said, most teams cite it anyway, and not out of legal caution. If a customer or auditor ever asks what your model trained on, being able to name LJSpeech cleanly, alongside its recording conditions and single-speaker limitation, is a five-minute conversation instead of an uncomfortable one. Provenance tracking is good practice even where it isn't a legal requirement.
Common Voice: CC0, but check the release you're actually using
Mozilla Common Voice dedicates its data to the public domain under CC0, "no rights reserved." As of the Common Voice 18 release, that's over 30,000 hours across 129 languages, all under the same waiver. Contributors donating clips sign a CC0 waiver themselves, so the chain of rights is clean by design.
The obligation here isn't legal, it's operational, and it's the one teams skip most: Common Voice releases change. Language subsets get revalidated, low-quality clips get pruned, new languages get added, and the corpus you trained on six months ago is not byte-identical to the one on the site today. If a result needs to be reproducible, or a customer asks what exact data underpins a shipped model, record the specific release version you pulled, not just "Common Voice." That's a data hygiene problem, not a licensing one, but it's the one that actually bites people.
Mirrors show a license field, the source shows the license
Most engineers don't pull these corpora from OpenSLR or keithito.com directly, they pull a Hugging Face mirror, because it slots straight into a dataloader. That's fine, but the license field on a mirror's dataset card is metadata someone typed in, not the license itself. I've seen mirrors list a CC BY-4.0 tag with the version number missing, or a stale description copied from an earlier corpus release before a maintainer updated terms upstream. It's usually right. It isn't always current.
Treat the mirror's license field as a pointer, not a source of truth. Before a model ships, click through to the primary page, OpenSLR for LibriSpeech, keithito.com for LJSpeech, commonvoice.mozilla.org for Common Voice, and confirm the license text matches what your data card says. It takes five minutes and it's the difference between citing what you actually agreed to and citing what a mirror maintainer assumed a year ago.
What the license doesn't tell you
Clean rights and fitness for purpose are two different questions, and a corpus can ace the first while failing the second. LibriSpeech is read audiobook speech from adult native English speakers in quiet conditions. LJSpeech is one speaker reading nonfiction prose. Common Voice's quality swings hard by language, since it depends on whoever showed up to record and validate that day. None of the three were built for noisy far-field audio, spontaneous conversational turns, or a specific accent your product actually needs to handle.
So here's the actual decision, not a hedge: if your production audio is clean, read-style, and in a language and register these corpora cover well, they can carry a real baseline and you should use them, full stop, that's what they're for. If your production audio is spontaneous, domain-specific, recorded in noisy or far-field conditions, or in a language or dialect these corpora don't cover at the quality you need, no amount of license clarity fixes an acoustic and linguistic mismatch. At that point you're not looking for a fourth free corpus, you're looking for data collected against your actual conditions.
A checklist, not legal advice
I'm not a lawyer and none of this is legal advice, so treat it as the operational checklist we actually run, not a substitute for your own counsel reading the license text. Before a model trained even partly on a public corpus ships: record the exact license and version at ingestion, not after the fact. Keep whatever citation the dataset's own page requests, even where it isn't strictly required, like LJSpeech. Check whether your product's customer contracts already require a data-sourcing disclosure, since some enterprise deals do. And re-check the license state before a major release, because dataset maintainers do occasionally change terms or restructure subsets, and the assumption that held true at project kickoff is worth confirming again before you ship.
When one of these three corpora genuinely fits your production conditions, use it, that's exactly the gap they're built to fill. When it doesn't, whether that's a language they don't cover well, a recording environment they weren't built for, or spontaneous speech where they only offer read speech, that's the point to stop stretching a public dataset past its design and either browse a licensed dataset built for your conditions or talk through a custom collection scoped to your actual audio.