Microsoft launches streaming transcription model in public preview

Microsoft AI launched MAI-Transcribe-2-Streaming on October 1, 2026, and describes it as its first streaming transcription model, built in-house by the Microsoft AI team.[1][2][3] The model is designed to turn live audio into text as it is spoken.[3] It continuously processes audio, returning fast partial hypotheses before stable final results.[3] Microsoft says it supports 60 languages and is designed for varied accents, speaking styles and real-world audio conditions.[3] Listed uses include live captions, voice agents, dictation, call experiences and other interactive products.[3]

Microsoft lists the model's final word error rate at 2.5% and says it recorded a lower rate than the competitors in its Artificial Analysis comparison.[3] That comparison lists Grok Transcribe 2.0 at 2.7% average word error rate, compared with 2.5% for Microsoft's model.[3] SiliconReport reports that Microsoft cited a 2.5% final word-error rate across eight hours of test audio for the streaming benchmark.[1] Microsoft lists a price of $0.54 per audio hour through December 31, 2026.[3] The comparison is limited to the competitors Microsoft lists; the evidence does not establish a broader ranking.

SiliconReport says the model is available on Microsoft Foundry.[1] Microsoft's model card says developers can integrate it through an OpenAI Realtime-compatible API over WebSockets.[3] Microsoft Learn says both the Realtime API and Azure Speech SDK integration methods support the model and return intermediate and final transcription results.[4] SiliconReport reports that Microsoft also made two text-to-speech models available alongside the transcription model, and says the three models are available on Microsoft Foundry, Vercel, Azure Voice Live and MAI Playground, with LiveKit integration planned.[1] Microsoft lists East US 2 as a region where availability is coming soon.[3]

Microsoft Learn identifies the model as being in public preview and says the preview has no service-level agreement and is not recommended for production workloads.[4] Those terms matter for teams deciding whether to test the model or depend on it in a live service. SiliconReport reports that Microsoft is targeting contact center agents, multilingual assistants and interactive learning tools.[1] It also reports that the model produces initial partial hypotheses in just over 100 milliseconds and stable transcripts in 0.13 seconds.[1] Microsoft says continuous language detection lets voice agents prepare tool calls and reason mid-utterance.[1]