Microsoft’s new speech models let agents begin work before a sentence ends

Microsoft launched MAI-Transcribe-2-Streaming and two voice-generation models on 1 October, with The Information covering the competitive challenge to ElevenLabs the following day.

The transcription service supports 60 languages and, Microsoft says, returns provisional text just over 100 milliseconds after receiving audio. It revises those partial transcripts as context arrives before producing stable text. The introductory price is $0.54 per audio hour through December 2026.

MAI-Voice-2.1 and its Flash variant support 23 languages, retaining a speaker’s voice across languages. They cost $22 and $15 per million characters respectively. Those output prices use a different billing unit from transcription; they are not directly comparable.

Engineering implication: provisional speech can trigger retrieval or preparation while somebody is still talking. Consequential actions need a stronger commitment boundary: a corrected word, negation or unfinished condition can change the request. Microsoft’s latency claims describe components, not a guarantee for an entire conversation involving reasoning, tools and network delays.