IndexTTS-2

IndexTTS-2 is an industrial-grade zero-shot text-to-speech system developed by Bilibili's speech team, achieving major breakthroughs in emotional expressiveness and precise duration control. As the first autoregressive TTS model to support precise duration control, it can control speech duration down to the millisecond level while also supporting a natural prosody generation mode.

Key Features

• Zero-shot voice cloning: Clone any speaker's timbre from just a few seconds of audio sample
• Decoupled emotion-timbre control: Independently control emotional expression and speaker timbre, with 8 emotion modes (happy, angry, sad, fear, disgust, melancholy, surprised, calm)
• Precise duration control: Explicitly specify the number of generated tokens to precisely control speech duration, perfectly fitting scenarios like video dubbing that require audio-video sync
• Natural language emotion guidance: Control emotional expression through text descriptions, using a soft instruction mechanism implemented with the Qwen3 model
• Pinyin pronunciation control: Supports precise control of Chinese pronunciation based on pinyin
• Multilingual support: Trained on 55,000 hours of multilingual corpus, supporting Chinese, English, and Japanese

Use Cases In video production and dubbing, creators can use IndexTTS-2 to achieve precise audio-video sync; content creators can generate expressive audiobooks and podcasts through emotion control; developers can integrate it into speech synthesis applications to build high-quality voice interaction systems. Community feedback calls it "voice quality so good you could watch an entire film or TV series dubbed with it."

Technical Advantages IndexTTS-2 uses a three-stage training paradigm to improve generation stability and integrates GPT latent representations to enhance speech clarity under high emotional expression. Experiments show it reaches industry-leading levels across multiple metrics including word error rate, speaker similarity, and emotional fidelity. The model supports FP16 inference and DeepSpeed acceleration, is fully open source (Apache 2.0 license), and can be deployed locally for commercial use.