MAI-Transcribe-2-Streaming

MAI-Transcribe-2-Streaming is a real-time speech recognition model from Microsoft AI, the company's in-house model lab, released on 1 October 2026. The model covers 60 languages and keeps detecting which one is spoken, so a call can change language midway without reconfiguration. Without waiting for a speaker to finish, it returns partial results in just over 100 milliseconds, refines them as more audio arrives, and then commits a stable transcript.

Features

• Streaming transcripts: partial results are revised as audio arrives, then committed as a stable transcript.
• Language detection: continuous detection across 60 languages, so the language can change in the middle of a conversation.
• Agent timing: Microsoft says voice agents can start reasoning or calling tools before the caller finishes speaking.
• Access: Microsoft Foundry, MAI Playground, Vercel and Azure Voice Live are routes to it, and LiveKit is listed as coming soon.

Clients

Developers building voice agents and real-time transcription.

Key facts

• Released: 1 October 2026
• Relationship: a separate product from the non-streaming MAI-Transcribe-2
• Weights: not published

Microsoft says it ranks first for accuracy on Artificial Analysis and that words appear twice as fast as with its closest competitor. These are Microsoft's own statements, and no independent press report had been found.