What it does
IndexTTS is a zero-shot text-to-speech system that clones a voice from a single reference audio clip. The latest release described on the page, IndexTTS-2.5, supports Chinese, English, Japanese, Spanish and Arabic and is faster than IndexTTS-2. Delivery can be steered in several ways. Emotion can come from a separate emotional reference clip, an eight-value emotion vector (happy, angry, sad, afraid, disgusted, melancholic, surprised, calm) or the text itself. Speaking speed is set with a duration factor between 0.5x and 2.0x, and pronunciation can be fixed with Pinyin, CMU phonemes or Japanese Kana markup. To run it, you clone the repository, install dependencies with uv, download the weights from Hugging Face or ModelScope, and then use either a Python API or a local Web UI. The setup notes mention optional half-precision inference and DeepSpeed, an NVIDIA GPU with CUDA 12.8 or newer, and a vLLM recipe for production serving. The project is released under the bilibili Model Use License Agreement and carries a disclaimer to read before use. The README includes benchmark tables against other TTS models and says the GitHub repository is the only official channel.
Capabilities
- Zero-shot voice cloning from a single reference audio clip
- IndexTTS-2.5 covers Chinese, English, Japanese, Spanish and Arabic
- Emotion control through reference audio, an emotion vector or text
- Speaking-speed control from 0.5x to 2.0x duration
- Pinyin, CMU phoneme and Japanese Kana pronunciation control
- Python API, local Web UI and a vLLM recipe for serving
When to use it
- Generating speech in a cloned voice on your own hardware
- Adding controlled emotion or pacing to generated speech
- Correcting mispronounced words with phoneme or Pinyin markup
- Evaluating a multilingual TTS model against alternatives
