README.md

June 1, 2026 · View on GitHub

ChatTTS x OpenVoice

Enhance the authenticity of speech by utilizing ChatTTS for more natural voice generation, complemented with the voice timber simulation module from Openvoice for seamless tone transplantation.

Have a try on huggingface! https://huggingface.co/spaces/Hilley/ChatTTS-OpenVoice

Experimental LLM Integration

ChatTTS/experimental/llm.py provides an OpenAI-compatible LLM API wrapper for text pre-processing (e.g. text normalisation before TTS inference). Supported providers:

Provider	`base_url`	Prompt versions
Kimi (Moonshot AI)	`https://api.moonshot.cn/v1`	`kimi`
DeepSeek	`https://api.deepseek.com/v1`	`deepseek`, `deepseek_TN`
MiniMax	`https://api.minimax.io/v1`	`minimax`, `minimax_TN`

MiniMax quick start (models: MiniMax-M3, MiniMax-M2.7, MiniMax-M2.7-highspeed; all with 204K context):

import os
from ChatTTS.experimental.llm import create_minimax_client

client = create_minimax_client(os.environ["MINIMAX_API_KEY"])

# Conversational reply (TTS-friendly tone, ≤100 chars)
reply = client.call("今天北京天气怎么样？", prompt_version="minimax")

# Text normalisation before TTS
normalized = client.call("We paid \$123 for this desk.", prompt_version="minimax_TN")

Get a MiniMax API key at https://www.minimaxi.com/

Notice:

We need to download the OpenVoice Checkpoint and save it into the ./OpenVoice/checkpoint folder.

OpenVoice Checkpoint: https://huggingface.co/myshell-ai/OpenVoice/tree/main/checkpoints