Microsoft has released MAI-Transcribe-2-Streaming, a speech-to-text model that returns provisional words while someone is still speaking.
The model supports 60 languages and continuous automatic language detection, according to Microsoft. It sends an initial partial transcript shortly after audio starts arriving, then revises the words as more context comes in and marks each segment final when it is stable.
A voice assistant may begin interpreting a request while the user is talking, but it must account for revisions rather than treating early words as final.
Microsoft says the model is available through its Realtime API and Azure Speech SDK. Its documentation labels the service a public preview with no service-level agreement and says it is not recommended for production workloads.
Sources
- Microsoft AI announcement, October 1, 2026
- Microsoft Learn technical overview and preview terms
- Illustrative office-phone photograph by Panelxf, Wikimedia Commons, CC0 1.0.
