TTS built for dialogue
ChatTTS is a text-to-speech model tuned specifically for conversation rather than general narration. Where many TTS systems aim for a clean reading voice, ChatTTS targets the rhythm of dialogue, with natural prosody, laughter, and pauses, in Chinese and English including mixed input. That focus is the reason it became popular for voicing chat assistants and LLM agents. At 39,458 stars as of 2026-06, it is one of the most-starred open TTS projects.
The trade-off to know up front is that it is explicitly a research project with a restrictive license and deliberate anti-misuse measures, which shapes both how you can use it and how it sounds.
What it does
You drive it with text plus inline control tags for effects like laughter and breaks, and it samples a speaker rather than cloning an arbitrary target voice. The result is conversational and expressive, with prosody that suits assistant-style speech. Speaker sampling gives variety, though it also means reproducibility depends on fixing the random seed, which is a common source of confusion.
Install
pip install ChatTTSFor the development version, install from the repository directly with pip install git+https://github.com/2noise/ChatTTS. It needs a GPU with a few gigabytes of VRAM for comfortable inference, and the models download on first use.
Research-only, by design
This is the part that determines whether you can use it at all. The licensing is two-layered and restrictive: the code is under AGPLv3+, a strong copyleft that requires you to open-source your changes including over a network, and the model weights are under CC-BY-NC 4.0, which forbids commercial use. The README states plainly that the released model is for academic purposes only. On top of the license, the authors built in anti-misuse measures: high-frequency noise added during training, deliberate MP3-level quality reduction rather than raw audio output, and a detection model for generated speech. So ChatTTS is for research and education, and integrating it into a commercial product carries real legal risk on both the code and the weights.
Where it fits, and where it doesn’t
Reach for ChatTTS when you are researching or prototyping conversational speech, when you want natural dialogue prosody with laughter and pauses, and when Chinese and English cover your needs. For academic and personal work it is a strong, expressive choice.
It is the wrong tool for a commercial product, where the AGPL code and non-commercial weights are a hard blocker. The deliberate quality reduction also caps fidelity, stability can be inconsistent without a fixed seed, and maintenance has slowed, with long gaps between releases. If you need commercial-friendly licensing, the alternatives below differ sharply.
ChatTTS versus the other open TTS models
| ChatTTS | F5-TTS | fish-speech | VoxCPM | |
|---|---|---|---|---|
| Stars | 39,458 | 14,759 | 30,832 | 28,884 |
| Strength | conversational prosody | flow-matching cloning | multilingual cloning | voice design and cloning |
| Code license | AGPLv3 | MIT | Fish Audio Research | Apache-2.0 |
| Weights | CC-BY-NC | CC-BY-NC | research, non-commercial | Apache-2.0 |
| Commercial use | not permitted | retrain weights or license | separate license required | permitted |
Counts are from GitHub as of June 2026. F5-TTS has MIT code, though its weights are CC-BY-NC. fish-speech needs a separate commercial license. VoxCPM under Apache-2.0 is the most commercial-friendly. ChatTTS stands out for conversational quality, but it is the most restrictive on use of the group.
The star curve
The star curve spiked sharply when ChatTTS launched, as conversational TTS that sounded genuinely natural was new and shareable, and it kept climbing through the assistant boom. The recent flattening tracks the slowing release cadence, with long gaps between versions.
What the issue tracker warns you about
The recurring issues are about control and stability. Output can be inconsistent run to run unless you fix the random seed, and zero-shot voice similarity is limited because the model samples speakers rather than cloning. The control tags for laughter and pauses do not always take effect, and long text can be truncated or come out incomplete. VRAM is modest but acceleration paths like vLLM have been reported to degrade audio quality. And the slowing maintenance means these rough edges are unlikely to be smoothed soon.
Related
For MIT-code voice cloning, see F5-TTS. For multilingual cloning, see fish-speech. For an Apache-2.0, commercial-friendly model, see VoxCPM. For what is trending, see LLM tooling, the daily digest, and the weekly report.
FAQ
Can I use ChatTTS commercially? No. The code is AGPLv3, which forces you to open-source derivatives including networked use, and the weights are CC-BY-NC, which forbids commercial use. The README states the model is for academic purposes only. Commercial integration carries legal risk on both layers.
ChatTTS vs fish-speech, which should I use? ChatTTS is tuned for conversational prosody in Chinese and English but is research-only. fish-speech does multilingual voice cloning but needs a separate commercial license. For research, pick by whether you want dialogue prosody (ChatTTS) or voice cloning (fish-speech); for commercial use, look at VoxCPM instead.
Why is ChatTTS output inconsistent between runs? It samples a speaker rather than using a fixed voice, so results vary unless you fix the random seed. Set and reuse the seed for reproducible output.
Does ChatTTS do voice cloning? Not arbitrary cloning. It samples speakers and lets you reuse a seed for consistency, but it is not designed to clone a specific target voice the way F5-TTS or fish-speech are.
Is ChatTTS still actively maintained? It is maintained but has slowed, with long gaps between releases. The core works, but expect rough edges around control tags and stability to persist rather than get quick fixes.