ChatTTS is a generative text-to-speech model built for dialogue scenarios such as LLM assistants. It supports English and Chinese, including mixed-language input, and produces multi-speaker output for interactive conversations. The model predicts and controls prosodic features such as laughter, pauses, and interjections. The main model is trained on more than 100,000 hours of Chinese and English audio; the release on Hugging Face is a 40,000-hour pretrained model without supervised fine-tuning, intended for academic use.
The Python API loads the model with ChatTTS.Chat, and chat.infer synthesizes a list of texts to 24 kHz waveforms. You can sample a random speaker embedding and reuse it to keep a consistent voice, tune temperature, top-P, and top-K decoding, and steer prosody at sentence level with [oral0-9], [laugh0-2], and [break0-7] prompt tokens or at word level with inline [laugh], [uvbreak], and [lbreak] markers. A WebUI and a command-line example are included, and streaming audio generation, a DVAE encoder, and zero-shot inference code are available. A 30-second clip needs at least 4 GB of GPU memory; on an RTX 4090 the real-time factor is around 0.3. The 40,000-hour model has added high-frequency noise and MP3 compression applied during training to limit misuse.
Features
- Conversational synthesis: speech tuned for dialogue with multi-speaker support
- Languages: English and Chinese, with mixed-language input
- Prosody tokens: sentence-level oral, laugh, and break levels plus word-level laugh and pause markers
- Speaker control: sample and persist speaker embeddings for a repeatable timbre
- Decoding parameters: temperature, top-P, and top-K settings per inference call
- Streaming generation: audio can be produced as a stream
- Zero-shot inference: DVAE encoder and zero-shot code are published
- Interfaces: Python package, WebUI example, and command-line script; optional vLLM backend on Linux
