Context—
IndexTTS2 is an autoregressive zero-shot text-to-speech model from the bilibili Index team with speech duration control: it either generates a set number of tokens for exact timing, as video dubbing needs, or speaks freely while reproducing the prosody of the prompt. It separates emotional expression from speaker identity, so timbre and emotion are controlled independently, and reconstructs the target timbre from a voice sample. It speaks Chinese and English.