Microsoft adds streaming transcription and two voice models to its MAI line

Microsoft AI announced three models on Thursday: MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. According to the announcement as relayed by Runtime Wire, the streaming transcriber takes continuous audio over a WebSocket and sends incremental updates before it confirms a final segment.

Microsoft says first transcript hypotheses arrive within the low hundreds of milliseconds, with words appearing as early as 320 milliseconds in most cases. The model supports 60 languages and detects the language automatically, so a session that switches mid-stream need not declare it up front. Streaming transcription costs $0.54 per audio hour.

Runtime Wire's report lists the voice models at $22 per million characters for MAI-Voice-2.1 and $15 per million for the Flash version. Microsoft positions the standard model as higher-fidelity and more expressive. It says Flash favours responsiveness and cost for interactive uses. Both cover 23 languages, across 26 locales for the standard model.

Secondary coverage says Microsoft's transcriber ranks first of 38 models on Artificial Analysis's streaming word-error-rate benchmark, at 2.5 percent. We have not opened that leaderboard, so the ranking is unconfirmed here, and the framing is Microsoft's.

The models follow Microsoft's April launch of MAI-Transcribe-1 and MAI-Voice-1, when its AI chief Mustafa Suleyman said Microsoft was pricing them to be the cheapest of any hyperscaler. No independent tests of the new models have been published that we found.