Xiaomi Open-Sources OmniVoice, a Speech Cloning TTS Model Covering Over 600 Languages

Xiaomi AI Lab has launched OmniVoice, a multilingual speech cloning TTS model that adopts a minimalist single bidirectional Transformer architecture. It supports speech synthesis in 646 languages, outperforming mainstream models in synthesis quality and inference speed for Chinese and English scenarios. The model is trained on approximately 580,000 hours of data from 50 open-source datasets and employs a dynamic upsampling strategy for low-resource languages. In tests across 24 and 102 languages, OmniVoice surpasses multiple commercial systems in speech similarity and intelligibility, with some metrics approaching or even exceeding real human speech. OmniVoice supports cross-language speech cloning, custom voice timbres, adaptation to noisy reference audio, sublanguage control, and pronunciation correction. The training, inference code, and model weights have been open-sourced on platforms such as Github and Huggingface.