ChatTTS: Generate Expressive Speech for Dialogue

Overview

ChatTTS is a generative text-to-speech model designed for conversational audio, such as speech from an LLM assistant. It targets dialogue that benefits from expressive delivery and multiple speaker voices rather than only straightforward narration.

The repository provides model infrastructure and examples for running inference. It may suit developers and researchers exploring conversational speech synthesis, especially when Chinese or English support and fine-grained delivery cues are useful.

Key Features

• Synthesizes speech from text for dialogue-oriented applications.
• Supports Chinese and English.
• Generates audio with multiple speaker options.
• Offers controls for delivery features including pauses, laughter, and interjections.
• Supports streaming audio generation.
• Includes command-line and web UI examples, plus Python inference usage.
• Provides pretrained model resources for research and development.

Use Cases

• Add spoken responses to an LLM assistant or conversational application.
• Prototype Chinese- or English-language dialogue audio without building a speech model from scratch.
• Explore multi-speaker synthesis for interactive demonstrations.
• Test how pauses, laughter, and other delivery cues affect generated dialogue.
• Conduct academic research on generative speech systems and conversational prosody.

Getting Started

Install the package from PyPI:

pip install ChatTTS

See the README for setup options, examples, and inference details.

Alternatives

• Spark-TTS: Spark-TTS uses Qwen2.5 and focuses on efficient bilingual synthesis and zero-shot voice cloning, rather than dialogue-oriented speech controls.
• GPT-SoVITS: GPT-SoVITS combines voice cloning and speech conversion with TTS, including fine-tuning from short voice samples.
• VoxCPM: VoxCPM is tokenizer-free and emphasizes multilingual speech, voice design, and cloning rather than ChatTTS's Chinese-English dialogue focus.
• MOSS-TTS: MOSS-TTS covers long-form narration, sound effects, streaming, and voice design in addition to dialogue and voice cloning.

Considerations

• The released model is described as intended for academic and research use. The model is licensed separately under CC BY-NC 4.0, so it is not for commercial use. The code uses AGPLv3+.
• The README reports that a 30-second clip requires at least 4 GB of GPU memory; hardware needs may affect whether local inference is practical.
• The project notes that output stability can vary, including speaker consistency and audio quality, and suggests trying multiple samples.
• The released model has limited token-level controls. Additional emotional controls are listed as future work.
• English is described as experimental, and the repository warns against installing certain optional acceleration packages because they may not work correctly or may slow generation.