chattts

ChatTTS is a generative text-to-speech model designed specifically for dialogue scenarios such as LLM assistants. The repository provides the algorithm infrastructure, pretrained model weights, and simple examples for running inference. Supported languages are English and Chinese, with additional languages listed as coming soon. The model is trained on Chinese and English audio data, and the open-source release on HuggingFace is described as a 40,000-hour pre-trained model without SFT. Key capabilities described in the README: - Conversational TTS optimized for dialogue-based tasks, with support for multiple speakers. - Fine-grained control over prosodic features, including laughter, pauses and interjections. - Token-level control units such as [laugh], [uv_break] and [lbreak], plus sentence-level control via prompts like [oral_2][laugh_0][break_6]. - Speaker sampling from a Gaussian distribution, with the sampled speaker embedding saved for later timbre recovery. - Adjustable decoding parameters including temperature, top_P and top_K. - Streaming audio generation. Getting started options include cloning the repository and installing requirements via pip or conda, installing the package from PyPI or GitHub, and running either a WebUI example or a command-line inference script. A basic Python usage example loads the model, runs inference on a list of texts, and saves output WAV files at 24 kHz. Advanced examples cover speaker sampling, sentence-level and word-level control, and a self-introduction sample. The README notes hardware expectations: at least 4GB of GPU memory for a 30-second audio clip, and on a 4090 GPU an RTF of around 0.3. It also notes that autoregressive models can show instability such as multi-speaker output or poor audio quality, and suggests trying multiple samples. Licensing: the code is published under AGPLv3+, while the model is published under CC BY-NC 4.0 for educational and research use, not for commercial or illegal purposes. The README states that a small amount of high-frequency noise was added during training of the 40,000-hour model and audio quality was compressed using MP3 format to limit potential misuse, and that an internally trained detection model is planned for open-source release. The project acknowledges prior work including bark, XTTSv2, valle, fish-speech and vocos, and points to a community-maintained index repo, Awesome-ChatTTS, for extended end-user products.