How the Homegrown Open‑Source ChatTTS Model Scored 20K Stars in One Week

ChatTTS is a text‑to‑speech model specifically designed for dialogue scenarios, such as LLM assistant conversations. It supports both English and Chinese and the largest version was trained on more than 100,000 hours of bilingual data.

The version released on HuggingFace is an open‑source checkpoint trained on 40,000 hours and has not undergone supervised fine‑tuning.

Highlights

Dialogue‑oriented TTS : Optimized for conversational tasks, producing natural and fluent speech while supporting multiple speakers.

Fine‑grained control : The model can predict and manipulate detailed prosodic features, including laughter, pauses, and inserted words.

Improved prosody : In terms of rhythm and intonation, ChatTTS surpasses most existing open‑source TTS models and provides a pretrained checkpoint for further research.

Shortcomings

Difficulty recognizing Arabic numerals.

Long conversion time that prevents real‑time dialogue.

Occasional "read‑back" errors where the output becomes garbled.

Despite these issues, the model attracted significant community interest, accumulating 20,000 stars on GitHub within a week of release.