ChatTTS

ChatTTS is a generative text-to-speech model specifically designed to produce natural-sounding speech for dialogue and conversational contexts. Unlike traditional text-to-speech systems that produce robotic, uniform-sounding output, ChatTTS is trained to generate speech that sounds like a real person talking — with natural pauses, emphasis, laughter, and the kind of prosodic variation that makes conversation feel human. It supports both English and Chinese, and was trained on over 100,000 hours of audio data.

The model works by taking text as input and generating audio waveforms directly. It uses a generative architecture similar to how large language models generate text, but instead produces speech tokens that are decoded into audio. You can control aspects of the output such as the speaker's voice, speaking speed, the placement of laughter and pauses, and the emotional tone. Multiple speakers can be generated in a single inference call, making it suitable for scenarios where you want to simulate a dialogue between distinct voices.

You would use ChatTTS when building applications that need natural conversational speech output — for example, giving a voice to a conversational AI assistant, creating podcast-style audio content from text, building interactive voice interfaces, or generating training data for other speech models. It is particularly well-suited for use with LLM-based chatbots where the spoken output needs to feel genuinely conversational rather than robotic.

ChatTTS is written in Python and uses PyTorch as its deep learning framework. The pretrained model weights are hosted on Hugging Face and can be loaded with a few lines of Python. The code is licensed under AGPLv3, while the model weights are released under a Creative Commons non-commercial license, meaning the model is intended for research and educational use rather than commercial deployment.