Microsoft is testing a full-duplex voice model. TestingCatalog reported on August 2 that MAI-Realtime — a bidirectional model that listens and speaks at the same time — has appeared in a hidden early-access slot inside MAI Playground, open to a small partner group. Two voices so far, Victoria and Grant, both noticeably more natural than today's Copilot voice mode; 17 languages; and it can run web search and other tools mid-conversation. It's internal preview only: no public API, no benchmarks, no timeline. Microsoft Foundry is the likely developer surface, Copilot voice the consumer one.
Taken alone, that's a preview you can't touch. Taken together with the last two months, it's the more interesting story: full-duplex voice just became a platform default.
The set is closed#
A month ago, "listen and speak at the same time" was the headline of the week when OpenAI shipped GPT-Live on July 8 — full-duplex, but ChatGPT-only, with the API waitlisted. Google's Gemini Live already shipped native-audio full-duplex with a real WebSocket API. The open-source camp has Kyutai's Moshi. Microsoft was the conspicuous gap.
MAI-Realtime closes it. As of August 2026, all three US hyperscalers have a native full-duplex voice model. When a capability is present at every major platform, it stops being a reason to pick one vendor over another. It becomes a baseline — a checkbox, not a moat.
That matters most for the founders who were treating full-duplex as a differentiator to chase. It isn't one anymore. The demos are still uncanny, the barge-in is still the thing that makes voice feel alive — but "which model can interrupt itself" is no longer a question with a competitive answer. They all can.
Full-duplex was last month's headline. This month it's the floor. The question moved from who has it to who makes the hard parts around it manageable.
The tell: turn-taking became a setting#
The detail worth dwelling on is how MAI-Realtime handles turns. It exposes two configurable listener modes. Switchboard uses an MAI-Ears endpointer driven by inline control tokens — a learned, model-native signal for when a turn ends. The deterministic mode pairs classic silence-based endpointing with a Whisper semantic endpointer, so a turn closes only when there's both a pause and a sense that the sentence is finished.
If you've built a cascaded voice agent, you know that endpointing — deciding your turn is over — is the single most hand-tuned, most-blamed part of the stack. It's the VAD-versus-semantic-turn-detection fight, and getting the threshold wrong is the difference between an agent that interrupts you and one that leaves dead air. Microsoft is shipping that exact decision as a config flag on the model.
That's the real direction of travel. Not "voice models got a new feature" — the hard, artisanal parts of a voice pipeline are being absorbed into the model's configuration surface. Endpointing today; likely tool-orchestration and interruption-recovery next.
Where the moat actually went#
If the model layer is a commodity, the differentiation moves up the stack. For a voice-agent builder in August 2026, the questions that decide whether your product works are no longer about the model:
- Turn-taking tuning. A configurable endpointer is only useful if you tune it to your domain. A drive-through and a therapy line want opposite interruption behavior.
- Tool reliability under interruption. Full-duplex means a user can talk over a tool call. If mid-utterance self-correction ("book New York — actually, Boston") doesn't roll back state before the call fires, you ship wrong actions confidently.
- Real-audio evaluation. Turn-taking behavior does not show up in transcripts. If your voice-agent evals are text-based, they are blind to the exact thing full-duplex changes.
- Cost per minute. Native speech-to-speech is priced by audio tokens, and the gap between vendors is an order of magnitude. That, not naturalness, is what will decide margins.
None of those come free with the model. All of them are yours to build.
What to do this week#
Nothing, on MAI-Realtime specifically — and that's the point. It's an internal preview with no API. The same rule that applied to GPT-Live in July applies now: a model you can't call is a signal of direction, not a dependency you can build on. Don't reroute a roadmap around it.
What you can do is stop pricing full-duplex as a feature and start pricing the stack around it. Pick native listen-and-speak where interruptibility is the product — companions, tutors, hands-free — and keep a cascaded or orchestrated stack where tool reliability and observability matter more. Either way, the model is now the easy part. Budget your engineering for everything else.
