IndexTTS-2

IndexTTS-2 is Bilibili's industrial-grade zero-shot text-to-speech system, released in September 2025 as the successor to IndexTTS 1.x. It pairs a transformer text-to-semantic module with a non-autoregressive semantic-to-mel module, and is billed as the first autoregressive TTS model combining precise synthesis duration control with expressive emotional speech synthesis. Timbre and emotion are disentangled, so a voice reference and an emotion reference can come from different speakers.

The system supports text, audio, and duration inputs, handles Chinese polyphonic characters through hybrid character-plus-pinyin modeling, and was trained on tens of thousands of hours of Bilibili data. A follow-up IndexTTS 2.5 extended multilingual coverage and inference speed. The project is open source; commercial use requires contacting the Bilibili team.