Microsoft launches MAI-Transcribe-2-Streaming, claims No. 1 accuracy on Artificial Analysis

Microsoft released MAI-Transcribe-2-Streaming, a 60-language real-time transcription model at an introductory $0.54 per audio hour.

Microsoft has released MAI-Transcribe-2-Streaming, its first real-time speech-to-text model. It ships alongside two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Microsoft says the model returns first partial transcripts "in just over 100ms," covers 60 languages with continuous language detection, and ranks No. 1 on Artificial Analysis for both partial and final accuracy. If you build voice agents, partials matter because they let the agent start reasoning or calling tools before the user finishes speaking.

Treat the headline numbers as Microsoft's. Tech Times says the No. 1 rank was not yet confirmed on the public Artificial Analysis leaderboard when it checked, with Grok Voice Transcribe 2.0 still at the top. It notes the 2.5% WER is from Microsoft's internal benchmark. Microsoft's claim that words appear in the transcript "2x faster than with our closest competitor" comes from its internal evaluations. SiliconANGLE reports that Microsoft cites a 320ms average for first hypotheses and says network conditions affect latency.

Access runs over WebSocket through Microsoft Foundry, Azure AI Speech and Vercel AI Gateway. Tech Times says it is a public preview with no SLA. It also notes the streaming docs don't claim speaker diarization, which the batch MAI-Transcribe-2 has. The introductory price of $0.54 per hour holds through the end of 2026. Microsoft hasn't said what it will cost after that.

You can compare that price against these streaming alternatives:

Before you commit, test on your own audio.

Sources