A voice agent is only as quick as its ears. Microsoft AI has shipped MAI-Transcribe-2-Streaming, its first speech-to-text model built to transcribe live rather than after the fact, so an application can react before a speaker finishes a sentence.
Rather than wait for a whole utterance, the system offers a first guess, called a partial, just over 100 milliseconds after audio starts arriving. As more audio lands, the model reworks those guesses and then locks in a finished transcript. The payoff is an agent that can begin thinking, or reach for a tool, while the speaker is still mid-sentence.
The model listens in 60 languages and keeps detecting a switch on its own, without being told, as the audio continues. Independent evaluator Artificial Analysis places it at the top of a 38-model field for accuracy on both the final text and the very first partial. Its final word-error rate is 2.5 percent, reached 0.13 seconds after a speaker stops. Internal testing at Microsoft, the company says, has words surfacing at double the rate of the nearest rival.
Pricing for the transcription starts at an introductory $0.54 an hour, a rate Microsoft says holds through the end of 2026.
The same launch also carried MAI-Voice-2.1, one of two text-to-speech siblings, covering 26 locales across 23 languages and billed at $22 per 1M characters. The faster of the pair, MAI-Voice-2.1-Flash, can render up to 45 seconds of speech and runs with 150 milliseconds of end-to-end lag, priced at $15 per 1M characters. Both can clone a voice from a few seconds of reference audio under consent guardrails.
All three are within reach for developers, arriving through Microsoft’s Foundry platform, the MAI Playground, Azure Voice Live and Vercel.
