Research2026-10-01
Microsoft AI released MAI-Transcribe-2-Streaming. It took the top spot for Final Transcript accuracy on AA-WER Streaming. It also leads First Partial Transcript accuracy on the same benchmark. The model achieved a 2.5% word error rate. It reached this at 0.13 seconds after the end of speech. The model is built entirely in-house by Microsoft AI. Owning both top spots is a meaningful result for real-time transcription. It reduces Microsoft’s dependence on external suppliers for this capability.
What changed
Previous top speech-to-text models held lower accuracy on AA-WER Streaming.
What it unlocks
Achieving top accuracy for both final and first partial transcripts in real-time speech-to-text.
- 2.5% word error rate
- 0.13s after end of speech
Sources