Microsoft's MAI-Realtime Aims at OpenAI's Voice Slot

Microsoft is testing a speech model that listens and talks at the same time — and if it ships, it fills the one slot in Microsoft's voice stack still occupied by OpenAI. The model, MAI-Realtime, surfaced as a hidden early-access entry in Microsoft's MAI Playground, spotted by TestingCatalog, with a small group of partners already able to talk to it.

Nothing about it is official. There is no model card, no pricing, no announced regions, and no entry in the public catalogue. Microsoft has not confirmed the model exists. What follows is what the preview exposes.

Full duplex is the whole feature

Today's voice assistants are a relay race: a recogniser turns your audio into text, a model reasons over the text, a synthesiser turns the answer back into audio. Each stage waits for the one before it, which is why you have to stop talking and why the pause afterwards feels like a pause. MAI-Realtime is described instead as processing a continuous audio stream in both directions and deciding, many times per second, whether to speak, keep listening, hold, or reach for a tool.

That is the same architecture OpenAI shipped as GPT-Live last month, and it is the reason interruption stops being a special case. If the model is already listening while it speaks, you talking over it is not an error condition to be detected and recovered from — it is just more input. Two voices are available in the preview, Victoria and Grant, and the reporting describes both as noticeably more natural than what Copilot's voice mode delivers today.

Two ways to decide whose turn it is

The most revealing part of the preview is that turn-taking is configurable, which is an admission that nobody has solved it. One mode, named Switchboard, uses a learned endpointer called MAI-Ears driven by inline control tokens — the model itself signals where a turn ends. The other is deterministic: silence-based endpointing paired with a Whisper semantic endpointer as a second opinion. The first is more fluid and less predictable; the second is duller and easier to reason about. Offering both is what you do when the right answer depends on whether the caller is ordering a pizza or authorising a payment.

Language coverage is seventeen — English, German, Spanish, French, Italian, Portuguese, Japanese, Korean, Chinese, Dutch, Hindi, Indonesian, Arabic, Russian, Turkish, Vietnamese and Thai — with automatic detection and mid-conversation switching, or a pinned language if you would rather it did not. The model can also drive tools, including web search. The stated limits are narrow and specific: it handles conversational speech only, so it will not sing or produce non-speech audio, and there is no sign of the zero-shot voice cloning MAI-Voice-2 offers.

The slot it is aiming at belongs to OpenAI

MAI-Realtime does not replace anything Microsoft already sells. MAI-Transcribe-1.5 stays for batch speech-to-text, MAI-Voice-2 stays for synthesis. What it would replace is the speech-to-speech component inside Microsoft's Voice Live service, and that component today is OpenAI's GPT-Realtime. Microsoft has consistently priced its own models below the OpenAI-based equivalent, so the likely outcome is not that GPT-Realtime disappears from Azure but that it becomes the more expensive default alongside a first-party option.

Read next to the news that Microsoft has been coaching its sales force to steer customers toward in-house models, the direction is not subtle. Microsoft still resells OpenAI, still hosts it, and is still its largest backer — while methodically building a substitute for every part of the stack it rents. Voice was among the last pieces missing.

The caveats are worth keeping in front of the enthusiasm. A hidden playground entry is not a launch, an impression of low latency is not a benchmark, and Microsoft has published no figure for time-to-first-audio or turn-transition latency — the numbers that decide whether a voice agent feels human or merely fast. Until there is a model card, MAI-Realtime is a credible signal of intent and nothing more.