Microsoft's MAI-Realtime Leaks in Internal Preview: Full-Duplex Voice AI with Tool Use

One-line summary: Microsoft is internally testing MAI-Realtime, its first full-duplex voice model capable of simultaneous speech and listening, tool calls, and 18-language support — the company's answer to GPT Live and Gemini Live.

Key Points

• Spotted as a hidden entry in the MAI Playground on August 2; no official launch date or public benchmarks released
• Full-duplex design: handles speaking and listening simultaneously rather than taking turns, producing more natural conversation than current Copilot voice mode
• Two voice personas (Victoria, Grant) described as noticeably more natural than MAI Voice 2; supports web search and external tool calling
• Covers 18 languages including English, Korean, Japanese, Chinese, and Arabic, with automatic language detection and mid-conversation switching

Why It Matters

Microsoft has long depended on OpenAI for Copilot's voice capabilities. A native, tool-enabled full-duplex model signals a deliberate move toward supply-chain independence. If MAI-Realtime ships publicly, voice AI becomes a three-way race between Microsoft, OpenAI, and Google — with each company's model now capable of acting on the world, not just speaking into it.

Read More

• Exclusive: Microsoft tests new MAI Realtime voice model — TestingCatalog
• Microsoft previews MAI-Realtime bidirectional voice model — CryptoBriefing