ChatTTS — Expressive Text-to-Speech for Dialogue

What it is

ChatTTS is an open-source text-to-speech model designed specifically for dialogue scenarios. Unlike standard TTS systems that produce flat, robotic speech, ChatTTS generates expressive audio with laughter, pauses, interjections, and emotional variation. It supports both English and Chinese, and can produce speech that sounds like a natural conversation.

It targets developers building chatbots, voice assistants, podcast generators, and any application where AI-generated speech needs to sound human and conversational.

How it saves time or tokens

ChatTTS eliminates the need for expensive commercial TTS APIs for conversational use cases. The model runs locally, so there are no per-character costs or API rate limits. It generates audio from text in seconds, and the expressive controls (laughter, pauses) are embedded via text tokens rather than requiring separate audio processing pipelines.

How to use

1. Install and set up:

pip install ChatTTS

1. Generate speech:

import ChatTTS
import torchaudio

chat = ChatTTS.Chat()
chat.load(compile=False)  # Downloads model on first run

texts = ['Hey, have you tried this new AI tool? It is amazing.']
wavs = chat.infer(texts)
torchaudio.save('output.wav', wavs[0], 24000)

1. Add expressiveness with control tokens:

# Use special tokens for laughter, pauses, etc.
texts = ['So I tried to deploy it and [laugh] it actually worked on the first try.']
wavs = chat.infer(texts)
torchaudio.save('expressive.wav', wavs[0], 24000)

Example

import ChatTTS
import torchaudio
import torch

chat = ChatTTS.Chat()
chat.load(compile=False)

# Generate a dialogue with different speakers
speaker_a = chat.sample_random_speaker()
speaker_b = chat.sample_random_speaker()

lines = [
    ('What do you think about using AI for code review?', speaker_a),
    ('Honestly? [laugh] It catches things I miss all the time.', speaker_b),
    ('Same here. The false positive rate is still annoying though.', speaker_a),
]

for text, speaker in lines:
    params = ChatTTS.Chat.InferCodeParams(spk_emb=speaker)
    wav = chat.infer([text], params_infer_code=params)
    torchaudio.save(f'line_{lines.index((text, speaker))}.wav', wav[0], 24000)

Related on TokRepo

• AI tools for voice -- Voice synthesis and recognition tools
• AI tools for content -- Content creation and generation tools

Common pitfalls

• ChatTTS requires PyTorch and a GPU for fast inference. CPU inference works but is significantly slower. An NVIDIA GPU with at least 4GB VRAM is recommended.
• The model downloads on first run (several GB). Ensure you have adequate disk space and bandwidth for the initial setup.
• Audio quality varies by input text length and complexity. Very long texts should be split into sentences for best results.