Microsoft releases MAI-Transcribe-2-Streaming, its first real-time transcription model

Microsoft AI announced MAI-Transcribe-2-Streaming, its first streaming speech recognition model, on October 1, 2026. The model turns speech into text piece by piece while the speaker is still talking. The same day, the company introduced two new text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Together, the three models are aimed at developers building AI agents that hold spoken conversations with people. MAI-Transcribe-2-Streaming is in public preview on Microsoft Foundry and Azure Speech.

Transcribing while the speaker talks

According to its Microsoft Learn documentation page, MAI-Transcribe-2-Streaming is a low-latency speech-to-text model that takes audio as a continuous stream and returns incremental transcripts while the speaker talks. Intermediate results keep updating the current guess, and final results confirm each segment. Microsoft lists call centers, voice assistants, meeting and lecture captioning, voice-driven interfaces and real-time note taking as uses. The company earlier introduced MAI-Transcribe-2 for recorded audio; the new model is the first in that family to work on live audio.

According to Microsoft's announcement, the model supports 60 languages with automatic, continuous language detection. Turkish is on the language list in the documentation, as are Arabic and Chinese. If no language is set, the model runs in multilingual mode. The company says the model delivers its first partial results just over 100 milliseconds after it starts receiving audio, and that at announcement it ranked first on Artificial Analysis, an independent benchmarking site, for the accuracy of both partial and final transcripts.

Three models for voice agents

MAI-Voice-2.1 and MAI-Voice-2.1-Flash turn text into speech. According to the model page, MAI-Voice-2.1 is recommended where fidelity matters more than speed, such as audiobooks and voice-over, while the Flash version is meant for latency-sensitive work such as call center agents and voice assistants. Both support emotion control and voice cloning from a short reference clip, and Microsoft says they include consent guardrails to prevent misuse. Turkish is on the page's language list.

According to Microsoft, pairing MAI-Transcribe-2-Streaming with MAI-Voice-2.1-Flash saves time on both ends of the conversation, which gives an agent more room to reason, use tools and check its answer. The company also added Chatter, a demo that shows the models working together, to the MAI Playground.

How to access it

Developers deploy the model to their own resource in Microsoft Foundry and connect in one of two ways: a WebSocket connection compatible with the OpenAI Realtime API, or the Azure Speech SDK, which manages the connection. According to the documentation, the model is currently served from Sweden Central, Central US and South India, with East US 2 coming soon. Microsoft says the models are also available through the MAI Playground, Vercel and Azure Voice Live, with LiveKit support coming soon. Current prices are on Microsoft's site.

The model is in public preview. According to Microsoft's documentation, the preview comes without a service-level agreement (SLA), is not recommended for production workloads, and some features may be limited. There is also no server-side end-of-speech detection for now, so the application has to detect pauses itself and request the final transcript. It is worth testing the model in your own language and on your own recordings, and for a fuller picture, follow evaluations from independent researchers as well.

Sources: Microsoft AI: Our first streaming transcription model, Microsoft Learn: MAI-Transcribe-2-Streaming, Microsoft Learn: Using the Realtime API, Microsoft AI: MAI-Voice-2.1.