Microsoft previews MAI-Realtime, a bidirectional voice model that uses tools mid-conversation
Microsoft previewed MAI-Realtime, a new bidirectional voice model it says is more natural than MAI Voice 2 and can run web search and other tools during a conversation. Voice is shifting from a text-to-speech layer bolted on top of a model to a native, tool-using, real-time capability of the model itself.
Bidirectional real-time voice is a different architecture than what came before. Earlier voice assistants transcribed speech, ran a text model, and spoke the answer — a pipeline with latency and seams. A model that natively handles a live two-way voice stream, interrupting and being interrupted, closes those seams and makes the interaction feel like conversation rather than turn-taking with a machine.
Tool use mid-conversation is the consequential part. A voice model that can run a web search or call a tool while it talks is a voice agent, not a voice interface — it can act on the world in the middle of speaking, which is the difference between an assistant that answers and one that does. That turns the voice channel into a full agentic surface.
Microsoft building its own MAI voice line, rather than relying solely on partner models, is a strategic tell. Owning the voice model that will sit inside Windows, Teams, and Copilot means owning the most personal interface to AI, and MAI-Realtime is Microsoft investing in that surface directly rather than renting it — the same first-party push visible across its agent stack.
Crypto Integrated — AI news, August 3, 2026 → · LLM-Stats — AI updates today (August 2026) →