OmniVoice: Xiaomi Ships Zero-Shot TTS Supporting 600+ Languages

arXiv·medium signal

Xiaomi's team (led by Daniel Povey) released OmniVoice, a single-stage diffusion language model supporting 600+ languages with zero-shot voice cloning. The novel architecture directly maps text to multi-codebook acoustic tokens, bypassing the conventional two-stage pipeline. Trained on 581K hours of open-source multilingual data, it achieves the broadest language coverage in any TTS system to date. Code and weights are open-source.