What it does

ChatTTS is a generative speech model aimed at conversational use, for example reading out the replies of an LLM assistant. It supports English and Chinese, can produce speech for multiple speakers, and can predict and control fine-grained prosody such as laughter, pauses and interjections. The README claims better prosody than most open-source text-to-speech models. The main model was trained on more than 100,000 hours of Chinese and English audio, but the version published on Hugging Face is a 40,000-hour pretrained model without supervised fine-tuning. The README says English is still experimental and that, as with other autoregressive models, output can be unstable, with problems such as shifting speakers or poor audio quality; trying several samples helps. In the released model the only token-level controls are [laugh], [uv_break] and [lbreak]. The repository contains the algorithm code and simple examples: a Python package installable with pip, a command-line script and a local web UI. A 30-second clip needs at least 4GB of GPU memory, and on a 4090 GPU the README reports a real-time factor of about 0.3. The code is published under AGPLv3+ and the model under CC BY-NC 4.0. The README says the released model is meant for academic, educational and research purposes only, and rules out commercial or illegal use. To discourage misuse, the 40,000-hour model was trained with a small amount of added high-frequency noise and compressed audio.

Capabilities

Speech synthesis tuned for dialogue
English and Chinese support
Multiple speakers
Control over laughter, pauses and interjections
Python package, command-line script and local web UI

When to use it

You are researching or prototyping spoken output for a conversational assistant
You want an open model whose prosody you can steer with laughter and pause tokens
Your use is educational or research, since the model licence excludes commercial use