Xiaomi AI Lab Launches OmniVoice: A Top-tier Speech Synthesis Model Covering Hundreds of Languages

Xiaomi AI Lab's Kaldi team has launched OmniVoice, a new speech synthesis model. It delivers top performance in Chinese and English scenarios and covers hundreds of languages, making it the industry's first TTS model supporting multi-language voice cloning. The model has strong generalization ability for low-resource small languages, enabling high-quality speech synthesis for almost all languages.

OmniVoice uses a minimalist architecture: a single bidirectional Transformer network directly converts text to speech without extra steps. Its training efficiency is impressive—100,000 hours of training per day—and PyTorch inference speed reaches 40x real-time, adapting to various scenarios. Even for small languages with less than 10 hours of training data, it can produce high-quality speech.

Additionally, it has practical features: generate desired voices by describing attributes (gender, age, tone, etc.) without reference audio; automatically filter noise to clone clear speech from noisy environments; support expressive tones like laughter; and correct mispronunciations of polyphonic words and proper nouns.