What is ChatTTS?
ChatTTS is a text-to-speech model from 2noise built for spoken dialogue, such as the voice of an LLM assistant, in Chinese and English. You can insert control tokens such as [laugh], [uv_break] and [lbreak] into the text, or set sentence-level levels for orality, laughter and pauses, and sample a random speaker embedding to get a new voice. Output is 24 kHz audio.
The team says its main model was trained on more than 100,000 hours of Chinese and English audio, but the open release is a 40,000-hour base model without supervised fine-tuning. The weights on Hugging Face were last updated in October 2024. To discourage misuse, the authors added a small amount of high-frequency noise in training and compressed the audio as MP3, so the open model sounds deliberately rougher than it could.
The licences are the part to read first. The code is AGPLv3 or later, which requires you to publish your source if you offer a modified version as a network service. The model weights are CC BY-NC 4.0, and the README says they are for academic and research use only, not commercial use. The repo has about 40,000 stars; v0.2.5, released on 10 April 2026, is the latest release and the last commit.
Is ChatTTS worth it?
ChatTTS is a research tool for conversational speech, and its licence makes that a hard line rather than a suggestion. The CC BY-NC 4.0 weights forbid commercial use, the AGPLv3+ code imposes source-sharing on anyone who serves a modified version over a network, and 2noise publishes no commercial licence to buy. A startup building a paid voice assistant cannot ship it.
Within that boundary it does something many open TTS models still do poorly: it sounds like talk rather than reading. The prosody controls work at two levels. A sentence-level prompt such as [oral_2][laugh_0][break_6] sets how chatty, laughing and pause-prone the voice is, and word-level tokens drop a laugh or a break exactly where you want it. Random speaker sampling gives new voices, and the saved embedding brings the same voice back later.
The catch: the open model is not the one the team describes. It is a 40,000-hour base without fine-tuning, against a main model trained on more than 100,000 hours, and the authors added high-frequency noise and MP3 compression on purpose. The FAQ admits autoregressive instability, such as a voice switching speakers or audio quality dropping, and suggests generating several samples and picking one. Only English and Chinese are supported, and the README calls English experimental. The FAQ quotes at least 4GB of GPU memory for a 30-second clip and a real-time factor of about 0.3 on an RTX 4090.
Maintenance has slowed to upkeep. v0.2.5 in April 2026 fixed breakage with newer Transformers releases and added partial DirectML support; the weights have not changed since October 2024, and the multi-emotion control on the roadmap has not shipped. About 60 issues are open.
Our take: ChatTTS is a good reference for what conversational prosody should sound like, in the way a concept car shows where a design is going; anything you intend to sell needs a permissively licensed model instead.
Where it earns its place
A speech researcher at a university lab
Compare how prosody tokens change perceived naturalness in dialogue
Generate the same scripts with and without [laugh] and [uv_break] tokens, fix the speaker embedding so only prosody varies, and run a listening test. Free under CC BY-NC 4.0 for academic work; one 4GB GPU is enough.
A hobbyist building a local Chinese-speaking companion bot
Give an LLM's replies a voice that sounds like casual talk
Pipe the LLM's text into chat.infer with a saved speaker embedding and a low temperature for consistency. Non-commercial personal use only; expect to regenerate the occasional unstable clip.
Before you commit
- Setup
- pip install ChatTTS (or clone the repo and install requirements.txt, Python 3.11 in the conda example). chat.load() downloads the weights from Hugging Face; examples/web/webui.py starts a local web UI and examples/cmd/run.py writes MP3 files.
- Integrations
- Python API with batch inference; optional vLLM on Linux; streaming generation; weights on Hugging Face. Community front ends are listed in the Awesome-ChatTTS index, not the main repo.
- Data and privacy
- Runs fully locally; text is not sent anywhere. Code is AGPLv3+ and weights are CC BY-NC 4.0, for academic and research use per the README. The authors say they trained a detector for its audio but have not released it.
- Cost at scale
- No paid edition and no commercial licence, so there is no legal production cost to quote. Research use needs at least a 4GB GPU per the FAQ.
- Limits
- English and Chinese only, English experimental; only [laugh], [uv_break] and [lbreak] as word-level tokens; autoregressive instability means retries; real-time factor about 0.3 on an RTX 4090; TransformerEngine and FlashAttention-2 paths are flagged in the README as not to install.
What’s good
- Laugh, pause and orality control at sentence and word level
- Prosody tuned for dialogue, which the authors rank above most open TTS models
- Random speaker sampling with reusable voice embeddings
- pip install ChatTTS, plus a web UI and command-line example
What’s not
- Weights are CC BY-NC 4.0: no commercial use, no paid licence offered
- AGPLv3+ code obliges source release for modified network services
- Open model is a 40,000-hour base with deliberate noise and MP3 compression
- English is experimental; weights unchanged since October 2024
Best for
Researchers and hobbyists studying conversational speech in Chinese or English who will never put the output in a commercial product.
ChatTTS alternatives
ChatTTS: common questions
- Is ChatTTS worth it?
- ChatTTS still gives a research voice agent convincing laughs and pauses in Chinese and English, but its non-commercial weights, deliberate audio degradation and a model untouched since 2024 rule it out for products. We rate it 3.4 out of 5.
- How much does ChatTTS cost?
- Free. Code is AGPLv3+; the model weights are CC BY-NC 4.0, so no commercial use. No paid or commercial licence is published.
- What are the downsides of ChatTTS?
- Weights are CC BY-NC 4.0: no commercial use, no paid licence offered; aGPLv3+ code obliges source release for modified network services; open model is a 40,000-hour base with deliberate noise and MP3 compression; english is experimental; weights unchanged since October 2024.
- Who is ChatTTS best for?
- Researchers and hobbyists studying conversational speech in Chinese or English who will never put the output in a commercial product.
- How much does ChatTTS cost for a team or in production?
- No paid edition and no commercial licence, so there is no legal production cost to quote. Research use needs at least a 4GB GPU per the FAQ.
- Is ChatTTS safe for company data?
- Runs fully locally; text is not sent anywhere. Code is AGPLv3+ and weights are CC BY-NC 4.0, for academic and research use per the README. The authors say they trained a detector for its audio but have not released it.
- What are the best ChatTTS alternatives?
- The closest alternatives we have reviewed are VibeVoice and audio.cpp.
Answers checked Nov 14, 2026.