ChatTTS Conversational Voice Model: Fine-Grained Prosody Control, Speaker Cloning & FastAPI Serving

🗣️ ChatTTS core conversation rhythm mechanism

Unlike traditional mechanical TTS (text-to-speech) engines, ChatTTS is the first open source speech model that is deeply fine-tuned for real conversation scenarios. Its core features include:

Colloquial control tags: Supports implicit control tags such as [laugh] (laughter) and [break] (natural sentence break) embedded in the text.
Controllable speaker timbre Embedding: By fixing the spk_emb seed vector, a stable and unified voice image output is achieved.

⚡ Image pulling and FastAPI service encapsulation

Download ChatTTS official weight

export HF_ENDPOINT="https://hf-mirror.net"
huggingface-cli download 2noise/ChatTTS \
  --local-dir /data/models/ChatTTS \
  --local-dir-use-symlinks False

chat_tts_demo.py: production call script

import ChatTTS
import torch
import soundfile as sf

chat = ChatTTS.Chat()
chat.load(source="custom", custom_path="/data/models/ChatTTS", compile=False)

text = "Hello everyone, welcome to the HF-Mirror mirror acceleration station! [laugh] It’s really fast to pull large models here [break]!"
wavs = chat.infer([text], use_decoder=True)

sf.write("welcome_speech.wav", wavs[0][0], 24000)
print("Speech synthesis is successful and has been saved to welcome_speech.wav")