Microsoft's MAI-Transcribe-2 leads on streaming speed

Microsoft released MAI-Transcribe-2-Streaming, its first streaming transcription model, alongside MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The transcription model covers 60 languages with automatic continuous language detection and ranks first on Artificial Analysis for accuracy on both final and partial transcripts, also sitting on the accuracy versus latency Pareto frontier.

The latency is the part that matters for live captioning and voice agents. First partials arrive in just over 100 milliseconds of receiving audio, and Microsoft’s internal evaluations show words appearing twice as fast as the closest competitor. The $0.54 per hour introductory price runs through the end of the year.

On the speech side, MAI-Voice-2.1 does 23 languages across 26 locales in a single voice at $22 per million characters. MAI-Voice-2.1-Flash generates 45 seconds of audio at 150 milliseconds end-to-end latency, which Microsoft says is 55% faster inference and about 60% cheaper, at $15 per million characters.

Both voice models clone from a few seconds of audio and carry built-in consent guardrails, which is the right default given how voice cloning has been misused. Availability is through Microsoft Foundry, the MAI Playground, Azure Voice Live and OpenRouter, with LiveKit coming soon and a demo called Chatter in the playground.

The caveats: the $0.54 figure is introductory and explicitly ends with the year, the twice-faster claim comes from Microsoft’s own internal evaluation rather than an independent one, and the Artificial Analysis ranking is third-party but published without a date or snapshot reference.

(Original source: Microsoft AI)