MAI Realtime: Microsoft Tests Its First Bidirectional Voice Model
On 2 August 2026 a model called MAI Realtime surfaced as a hidden early-access entry in Microsoft’s MAI Playground, described in reporting as a bidirectional full-duplex system with two voices, seventeen languages, two selectable turn-taking modes and partner access already granted. Microsoft has confirmed none of it: no model card, no benchmark, no price, no availability date. This guide separates what was actually reported from what is inference, explains why full duplex is an architectural change rather than a latency improvement, and sets out where MAI Realtime would sit against MAI-Voice-2, MAI-Transcribe-1.5 and the GPT-Realtime model currently doing the speech-to-speech work inside Azure’s Voice Live API. It covers the OpenAI dependency this closes, what the existing MAI price points imply, the 200-millisecond turn-taking budget that decides whether a voice agent feels human, where voice deployments actually fail in production, the EU AI Act transparency obligations that became applicable the same day, voice as biometric data under GDPR, the TCPA position on synthetic voices, and a ninety-day plan whose value does not depend on MAI Realtime ever shipping.
