Microsoft has begun early-access testing of MAI-Realtime, a new bidirectional voice AI model built in-house for natural, full-duplex conversation inside Copilot. Unlike turn-based voice assistants, the model listens and speaks simultaneously, letting users interrupt mid-sentence with clean, low-latency handling.
The model supports 17 languages, including English, German, Japanese, Korean, Chinese, Arabic and Hindi, with automatic detection and mid-conversation switching. Testers describe its two available voices, Victoria and Grant, as noticeably more natural than Copilot's current voice mode. It offers configurable turn-taking, either a silence-based endpointer or Microsoft's own semantic MAI-Ears system.
MAI-Realtime is currently limited to a small partner group inside the MAI Playground and has not been publicly launched. It slots alongside Microsoft's other in-house speech models, MAI-Transcribe-1 and MAI-Voice-1, marking a move away from dependence on OpenAI's GPT-Realtime model for voice functionality in Copilot.