Microsoft Tests First Native Real-Time Voice Model MAI Realtime with 16 Languages

Microsoft is testing its first native real-time voice model, MAI Realtime, which supports 16 languages and 2 voices, allowing simultaneous listening and speaking with automatic language detection and switching. Testing is underway with select partners on the MAI Playground platform, but launch dates are unconfirmed.

MAI Realtime uses full-duplex interaction, so it can respond without waiting for the user to finish speaking, enabling real-time back-and-forth. Supported languages include English, Chinese, Japanese, and 13 others. Two voice styles, Victoria and Grant, are available, sounding more natural than current Copilot voices. The model doesn't sing or generate non-speech effects, focusing purely on conversation. A debug panel shows real-time latency and thinking steps, with sample sharing to be rolled out gradually.