Partial words in a tenth of a second
Microsoft has released its first streaming speech-to-text model, MAI-Transcribe-2-Streaming, alongside two text-to-speech models, MAI-Voice-2.1 and a faster variant called MAI-Voice-2.1-Flash. The transcription model starts emitting provisional words — what the field calls partials — just over 100 milliseconds after it begins receiving audio, then revises them as more speech arrives.
That revision behaviour is the whole point of a streaming model. A batch transcriber waits for the speaker to stop; a streaming one commits to a guess early and corrects itself, which is what makes a voice agent able to interrupt and be interrupted. Microsoft says the model supports 60 languages with automatic, continuous language detection, so a speaker switching mid-sentence does not have to be announced.
The claim, and whose claim it is
Microsoft says the model “ranks no. 1 for accuracy for both final and partial transcripts on Artificial Analysis”, and that words appear twice as fast as from its closest competitor. Artificial Analysis is an independent evaluator, but the ranking here is Microsoft’s description of where its model placed, published by Microsoft on the day of launch. It is a vendor’s account of a third-party result, not an independent write-up, and the company does not name the competitor it doubled.
The voice side, and the price
MAI-Voice-2.1 covers 23 languages and 26 locales, and Microsoft says a single synthetic voice can move between them without sounding like a different speaker — the failure mode that makes multilingual TTS obvious. Both voice models can clone a voice in any supported language from a few seconds of reference audio, with what Microsoft calls built-in consent guardrails to prevent misuse.
The prices are the aggressive part. MAI-Voice-2.1 is $22 per million characters and the Flash variant is $15. The streaming transcription model is $0.54 per hour of audio as an introductory rate through the end of 2026.
A crowded fortnight for speech
This lands in the middle of the busiest stretch the speech market has had. ElevenLabs shipped a new speech architecture that takes stage directions on 30 September, and doubled its valuation without raising new capital in the same week. Fermion Research opened a 164-megabyte transcription model that runs on a laptop. Suno added spoken voice to its music generator on Thursday.
What separates Microsoft’s entry is where it is aimed. The company has said it wants these models powering call transcription in Teams and clinical transcription in its Dragon Copilot product for health providers — captive, high-volume workloads where a tenth of a second and a fifth of a cent compound fast.
Watch the independent leaderboards rather than the launch post. Artificial Analysis publishes its own tables, and a vendor’s claim to a top position holds or does not within a week of everyone else submitting.
