MAI-Transcribe-2 adds speaker attribution and word timestamps

Microsoft announced MAI-Transcribe-2 on September 3, adding speaker diarization, word-level timestamps and configurable transcription styles to its speech-recognition family. The release post described coverage across 60 languages.

More structure around recognized words

The model offers a choice between a verbatim transcript, which retains fillers and false starts, and a cleaned version intended for readability. Microsoft also described keyword biasing, automatic language identification and support for mixed-language speech. The post linked demonstration paths through Foundry, the MAI Playground and OpenRouter.

Microsoft made strong comparative claims in the announcement. This brief focuses on the concrete feature additions rather than treating those rankings as independently established conclusions.

Choose transcript semantics deliberately

Our engineering recommendation is to make transcript style part of the data contract. A cleaned transcript may suit a reading interface, while an audit or research workflow may need the original spoken sequence. Do not let that difference become an undocumented default hidden in a model call.

Store the selected style with the output. If the application produces both a readable transcript and an archival record, label them clearly and preserve the relationship to the same source audio.

Verify alignment and attribution separately

Word recognition, timing and speaker labels should receive separate checks. A correct sentence assigned to the wrong speaker is a different defect from a misspelled word, and each may affect downstream summaries differently.

Use a test set with interruptions, short interjections and speakers returning after long gaps. Check whether clicking a timestamp reaches the expected audio and whether edits preserve that alignment. These tests turn the new structural features into observable application behavior instead of simply adding more fields to a stored transcript.