ElevenLabs and Western voice APIs vs CosyVoice 2: the 39,836 star alternative worth checking

TL;DR — Everyone reaches for ElevenLabs when they need a voice pipeline. The open-source alternative that actually rivals it is CosyVoice by Alibaba's QwenAudio team — Apache-2.0, free, runs on your own machine. But the second-most-hyped Chinese TTS, ChatTTS, is not what its GitHub badge says it is: the repository shows AGPL-3.0, while the model weights ship under CC BY-NC 4.0 — non-commercial. You cannot ship it. Verified via the GitHub API on 2026-09-14.

Why it matters

Text-to-speech stopped being a research demo and became a line item. If your product reads anything aloud — support replies, course narration, dubbing, accessibility, an agent that talks back — you are paying per character, per month, forever, to a vendor that holds your audio.

That last part is the part teams under-weight. A hosted TTS API means every sentence your users hear leaves your infrastructure and lands on someone else's. For a generic assistant voice, nobody cares. For a cloned voice — a customer's own voice, a presenter's, a patient's — you have just shipped biometric-adjacent data to a third party and written it into a contract you did not read.

The Chinese open-source ecosystem solved this earlier and more completely than the Western default did. CosyVoice is the clearest example: a permissively licensed, locally runnable TTS stack from the same organisation that maintains the Qwen models many Western developers already run in production. The reason to switch is not nationalism or novelty. It is that the cost curve and the data boundary both move in your favour at the same time.

The "because Z"

1. Cloned voice and multilingual synthesis run on your own hardware, so cost per minute falls to zero against a monthly subscription.

This is the argument that closes the deal for anyone with volume. ElevenLabs is priced as a subscription; CosyVoice is priced as electricity. Once the model is on a machine you own, the marginal cost of the ten-thousandth minute is the same as the first: nothing. The break-even is not a subtle calculation — it is a fixed hardware purchase against a recurring bill that scales with your success.

CosyVoice 3 supports voice design and controllable cloning, and there is an MLX port for Apple Silicon, which means the "own hardware" in question can be a laptop rather than a rack. The authors publish quality comparisons against commercial systems; treat those as a claim published by the authors, not an independent measurement. What is not a claim is the architecture: the inference happens on your machine, and no meter is running.

2. No audio is sent to a third party — the point that matters when the voice is a client's.

A hosted API is a data-processing relationship. You send raw audio or raw text, you receive audio back, and in between it sits on infrastructure you do not control, governed by terms you did not negotiate. For a synthetic narrator voice, that is a procurement detail. For a cloned voice, it is the whole risk.

When the voice belongs to a customer, a presenter, or anyone who signed something, "we upload it to a US vendor's API on every request" is a sentence you do not want to write in a security review. Local inference deletes the sentence. The audio never crosses a network boundary you do not own, because there is no network boundary. This is not a performance argument and it is not a cost argument — it is the argument that decides whether the feature ships at all in regulated or privacy-sensitive contexts.

3. Licence is the real selection criterion, and it is not where you look for it.

Developers pick TTS by star count and demo quality. Both are the wrong signal. The signal that determines whether your product survives contact with a legal review is the licence on the weights, and that is frequently not the licence GitHub shows you.

CosyVoice is Apache-2.0 end to end — code and weights. That is a standard permissive open-source licence: you can ship it inside a commercial product, modify it, and keep your changes private. ChatTTS is not. Its repository badge reads AGPL-3.0, which is already a copyleft licence with real consequences for a networked service. But the badge is not even the full story, which is the trap below.

The data

	ElevenLabs	CosyVoice (QwenAudio/CosyVoice)	ChatTTS (2noise/ChatTTS)
Licence	Proprietary	Apache-2.0 ✅	AGPL-3.0 on the repo badge 🔴 — weights are CC BY-NC 4.0
Price	Subscription (tiers not published in this data set)	Free (self-host)	Free (self-host)
Runs locally	No	Yes (MLX port for Apple Silicon)	Yes
Stars	—	23,612	39,836
Last commit	—	2026-05-25 (112 days before verification)	2026-04-10 (157 days before verification)
Archived	—	No	No
Verified	—	2026-09-14 (GitHub API)	2026-09-14 (GitHub API)

Two things to read off this table. First, ChatTTS has roughly 16,000 more stars than CosyVoice and is the worse choice for anyone shipping a product — popularity and usability are not the same axis. Second, both projects are maintained but neither is moving fast right now: 112 and 157 days since the last commit at verification time. That is normal for a model release that has stabilised, but it is a fact you should know before you bet a roadmap on either.

The trap to avoid

GitHub displays exactly one licence per repository. That single field is not a complete description of what you are allowed to do.

The GitHub API reports AGPL-3.0 for ChatTTS. Read that alone and you conclude the project is open source, copyleft, awkward for a closed product but at least usable. That conclusion is wrong.

The repository actually publishes two separate regimes: the code under AGPLv3+, and the model weights under CC BY-NC 4.0, with the non-commercial restriction stated explicitly — not for any commercial purpose. The weights are the part you need. A TTS model without its weights is a build script. So the operative licence for anyone who wants to generate speech is the non-commercial one, and the GitHub badge never mentions it.

CC BY-NC 4.0 is not an open-source licence in the sense that matters commercially. It is a source-available licence with a field-of-use restriction. The consequence for someone shipping a product is concrete: you may not use the weights in a commercial product, and "we're just prototyping" stops being a defence the moment the prototype becomes the feature. There is no threshold, no revenue cap, no small-team exemption in the licence text.

This is why the column exists. A licence check that reads the GitHub badge and stops there will clear ChatTTS for production and be wrong. The check has to go one level deeper — into the model card, the weights repository, the LICENSE files — because that is where the restriction lives.

And licences in this ecosystem move. MinerU went AGPL to Apache-2.0 in 2026; projects flip the other way too. Re-verify on every update rather than trusting a note you wrote six months ago.

The verdict

• You need a commercial voice pipeline → CosyVoice. Apache-2.0 on code and weights, runs locally, no per-minute cost, no audio leaving your infrastructure. This is the default answer for anyone with volume or anyone handling a voice that belongs to someone else.
• You are prototyping and want polish with zero setup → keep ElevenLabs. It remains the fastest path to a working demo and the API is genuinely good. Just be clear that you are renting, and that the rent scales with usage.
• Never build production on a CC BY-NC model, however good the demos sound and however many stars the repository has. Check the weights licence, not the badge.

Data verified 2026-09-14 · Sources: GitHub API (QwenAudio/CosyVoice, 2noise/ChatTTS), project pages. Pricing tiers for the commercial option are not published in this data set and are therefore not quoted.

References

1. CosyVoice 2 — Apache-2.0, 23612 stars, last commit 2026-05-25, primary language Python.
2. ChatTTS — AGPL-3.0, 39836 stars, last commit 2026-04-10, primary language Python.
3. ChatTTS model licence (CC BY-NC 4.0).
4. ChatTTS README, Licenses section.

All figures were read from the GitHub API (api.github.com) on 2026-09-14. They move: re-check before you ship.