Speech & Developer Tools
Microsoft details MAI-Transcribe-2 public preview
The speech model adds speaker labels, word timestamps and transcript styles, while Azure documentation explicitly limits its production readiness.
What happened
In a 3 September update, Microsoft described MAI-Transcribe-2 with support for 60 languages, speaker diarization, word-level timestamps and configurable clean or verbatim transcripts. Azure Speech documentation lists it as public preview, without a service-level agreement, and does not recommend it for production workloads.
Why it matters for AI
Speaker labels and timing can make transcripts easier to search and review. Choosing a cleaned transcript can also remove hesitations or false starts that matter when checking what was actually said.
AIRunsThis View
Our interpretation: test recognition errors, speaker attribution and omitted words together on representative recordings. Preview availability and vendor benchmark rankings are insufficient grounds for trusting customer-facing transcription.
What this could mean for us
A potential candidate for a bounded future evaluation; no customer audio was uploaded and no transcription service has been adopted.
Source and method
This item is based on a primary source. AIRunsThis separates the source’s factual announcement from its own interpretation. Vendor performance claims have not been independently reproduced unless stated.