If you're building a product on open-source text-to-speech, the licence file deserves more of your attention than the demo audio. I learned this the slow way while picking models for a desktop app, and the specifics are worth writing down because the most popular option in the category is also the one you probably can't ship.
The three models worth knowing about.
Kokoro-82M — 82 million parameters, Apache 2.0, eight languages and fifty-four preset voices. It does not clone voices; you pick from its voice list. What it gives you instead is the ability to run on an ordinary CPU without a GPU anywhere in sight, which no other model here manages comfortably. For narration, app prompts, accessibility features, it's the obvious answer and the licence is the most permissive in the group.
XTTS-v2 — Coqui's model, seventeen languages, clones from a roughly six-second reference clip per its own model card. It is, by a distance, the one people search for and the one most tutorials cover. It's also the problem.
Chatterbox — Resemble AI, 500 million parameters, MIT licence, zero-shot cloning from a few seconds of audio. The base model is English only; a separate Multilingual v3 build covers twenty-three languages. Worth knowing before you ship: every output carries Resemble's PerTh watermark, which they describe as surviving MP3 compression and ordinary editing with near-perfect detection. It's inaudible and generated locally — nothing phones home — but it is detectable, and that's a product decision, not a footnote.
The licensing part.
Kokoro is Apache 2.0. Chatterbox is MIT. Both mean what you'd hope: use it, ship it, sell it, tell nobody.
XTTS-v2 is released under the Coqui Public Model License, which restricts use to non-commercial purposes. Not "attribution required", not "under a revenue threshold" — non-commercial. If you're charging for the thing you build, you're outside the licence.
Historically there was a way out: Coqui sold a commercial licence. Coqui Inc. shut down in January 2024. There is no sales team, no licensing portal, no entity to contract with. The commercial option doesn't exist at any price. So the most-downloaded, best-documented, most-searched open-source voice cloning model has no legal path into a paid product, and the reason isn't cost — it's that there's nobody left to ask.
Why this catches people.
Three things stack up. The weights are freely downloadable, which feels like permission. The ecosystem around it — community packages, wrappers, tutorials — is large and none of it leads with the licence. And "open source" does a lot of unearned work in people's heads; plenty of models described that way carry terms that aren't open source in any OSI sense.
What I'd actually do:
Read the LICENSE file in the exact repository you download from, not a summary and not a blog post. Variants, community re-uploads and quantised conversions don't reliably carry the original terms.
Check whether the licensor still exists. A permissive-sounding licence from a dissolved company with no successor is a different risk from the same licence backed by a live entity.
Do this at model-selection time. Swapping an inference engine after you've shipped is a couple of days of work; discovering the problem after you've built a year of marketing around a model's name is a different cost entirely.
Prefer Apache 2.0 or MIT when quality is close — and in this category it genuinely is close now, which wasn't true two years ago.
If you're running local TTS in anything commercial, the five minutes this takes is the cheapest five minutes in your stack.
https://getpocketcore.com/?utm_source=devto&utm_medium=article&utm_campaign=oct29_article10
Top comments (0)