Microsoft MAI-Realtime Leak Explained 2026: Full-Duplex Voice, 17 Languages (Not Official)

Let’s be clear up front: this is a leak / hidden preview, not an official announcement. On August 2, TestingCatalog spotted a hidden early-access entry called MAI-Realtime inside Microsoft’s MAI Playground. Microsoft has not announced it, published a model card, disclosed pricing, or named launch regions. So read this as “spotted,” not “confirmed.” Still, the direction is interesting enough to lay out.

What was spotted?

The core is full-duplex — it listens and speaks at the same time. Most voice AI today is turn-based (“I speak, then it replies”); this overlaps like a human conversation.

• Appears to be Microsoft’s first native full-duplex speech model
• Two voices, Victoria and Grant — reportedly noticeably more natural than Copilot’s current voice
• 17 languages (English, German, Spanish, French, Japanese, Korean, Chinese and more)

Microsoft’s earlier voice/image moves are covered in Microsoft MAI image and voice.

What else showed up?

From the leaked working listing:

Item	Detail
Mid-conversation language switch	Pinned or auto-detect, switches without losing its footing
Turn-taking modes	Two selectable modes
Latency	Live debug panel
Tools	Web search and more

This tracks with competition shifting from raw capability toward naturalness and real-time feel, a pattern also visible in AI rankings via OpenRouter.

So is it confirmed?

No. To repeat — there is no official Microsoft announcement. It surfaced as a hidden early-access entry in the MAI Playground; launch, pricing and regions are undecided and it’s not in the public catalogue. As a preview, final specs can change. The low-cost, low-latency race also shows in cases like Grok 4.5.

Frequently Asked Questions

Q. Can I try it now? Not as a general user. It was spotted as a hidden early-access entry in the MAI Playground; it’s not in the public catalogue and hasn’t officially launched.

Q. What makes full-duplex different? Existing voice AI replies after you finish (turn-taking). Full-duplex listens and speaks simultaneously, so it can interrupt and overlap like a human conversation.

Q. Does it support Korean? The spotted listing includes Korean among 17 languages. But this is a preview observation, not an official spec, so the final version may differ.