A controllable zero-shot text-to-speech system that separates speaker identity from emotional expression. It requires local execution with an NVIDIA GPU and CUDA Toolkit 12.8 or newer.
Project overview
It decouples timbre and emotion controls, addressing duration and expressiveness limitations in autoregressive speech models.
Project type
Model Runtime · Audio & Speech
Use cases
Audio & Speech
Deployment
Refer to project documentation
License
License pending
Best for
Developers and AI engineers who need independent control over voice timbre and emotion in synthesized speech and have local NVIDIA GPU hardware.
Key capabilities
Synthesizes speech by cloning a voice from a single short reference audio clip.
Controls speech emotion using a separate reference audio file, an 8-float emotion vector, or an emotion text script independently from the speaker's voice.
Provides a local WebUI for interactive text-to-speech synthesis with adjustable inference and acceleration settings.
Provides a Python library and CLI to run text-to-speech synthesis and integrate it into custom scripts.
Limitations and risks
The pinyin control feature for pronunciation only works for valid Chinese Pinyin cases.
Users face security and accuracy risks when using unofficial websites or services claiming to be IndexTTS.
Getting started
The setup process requires installing git and git-lfs, cloning the repository, installing the uv package manager, running uv sync --all-extras, downloading the IndexTeam/IndexTTS-2 models, and running the webui.py file or a custom Python script.
Evidence and sources
GitHub project description: An Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System
GitHub repository metadata: Python
README: Existing autoregressive large-scale text-to-speech (TTS) models have advantages in speech naturalness, but their token-by-token generation mechanism makes it difficult to precisel…
README: 2025/09/08 🔥🔥🔥 We release IndexTTS-2 to the world!
README: Synthesize new speech with a single reference audio file (voice cloning):