Microsoft's MAI speech lineup documents real-time use only on the voice-output side

A live voice agent has two places where users feel delay. On the way in, the caller's speech has to become text while they are still talking. On the way out, the reply has to become audio before the pause gets awkward. Microsoft has a documented model for the outbound side: MAI-Voice-2-Flash, the lower-latency variant it intends for real-time voice-agent scenarios ^[6]. For the inbound side, a dev.to review of Microsoft's documentation found no native real-time streaming transcription advertised for MAI-Transcribe-2 ^[5]. Of the three documented models, the only one framed for real time is on the output side ^[1].

MAI-Transcribe-2 fits work where the audio already exists. The review says it may suit recorded audio, or workflows where transcription runs once an audio segment is available ^[10]. Diarization shows who said what. Word-level timestamps let a reviewer find the exact point in a long recording where a statement was made ^[12]. The review names post-call analysis, meeting notes, quality review and searchable audio archives as uses for those features ^[14].

I think putting transcription and voice response in separate models is the right design. The review treats them as separate implementation decisions with different latency and quality requirements ^[13]. A pipeline that reviews yesterday's support calls can wait for a batch job. A voice agent in the middle of a conversation cannot.

The obvious workaround for live use is to cut incoming audio into short segments, on a fixed interval or on silence, and submit each one as it closes. That is batch transcription on a timer. The caller-side delay can be no shorter than one segment plus one request. Whether speaker labels stay consistent across separately submitted segments needs measuring before anyone builds on it. The review's own advice is narrower: do not design a live transcription experience around an assumed MAI-Transcribe-2 streaming endpoint until Microsoft documents one ^[9].

The 60-language figure from the catalog entry is a list of supported languages ^[3]. Support and accuracy are separate claims. For the number to mean anything for a given contact centre, the model has to hold up on that centre's own calls. The review tells teams to test representative audio, languages, accents, noise conditions and conversation formats before relying on the output in customer-facing or operational processes ^[11].