ChatTTS

ChatTTS is a generative text-to-speech model built for dialogue scenarios such as LLM assistant voices rather than long-form narration. Trained on roughly 40,000 hours of Chinese and English speech, the autoregressive model produces conversational prosody with natural pauses and can insert paralinguistic elements like laughter through fine-grained control tokens at the word and sentence level. It supports multiple speakers, mixed language input, and streaming audio generation built on a DVAE-based pipeline, and it can synthesize a 30 second clip on a GPU with about 4 GB of VRAM. The code is published under AGPLv3+, while the released weights carry a CC BY-NC 4.0 license restricting them to non-commercial use; the authors state the open model is intended for academic purposes and note that compressed quality and embedded noise were added to deter misuse. Researchers and developers exploring conversational agents in Chinese and English make up its main audience.