Microsoft launches MAI‑Transcribe‑2‑Streaming at $0.54 an audio hour
Microsoft says its new real-time speech-to-text model returns partial transcripts about 100 milliseconds after audio arrives, covers 60 languages and ranks first for accuracy on Artificial Analysis.
Microsoft released MAI-Transcribe-2-Streaming on 1 October, a real-time speech-to-text model, TestingCatalog reported. Microsoft says it detects the language automatically and keeps updating partial hypotheses as it hears more audio. TestingCatalog is the only outlet reporting the launch so far, and the details below are the company's own claims as that outlet relays them.
Microsoft says partial transcripts appear just over 100 milliseconds after audio is received, and that words show up about twice as fast as with competing real-time dictation and subtitling systems. The model supports 60 languages, with language detection that runs continuously rather than once at the start of a call.
On accuracy, the report says the model ranks first on Artificial Analysis for both final and partial transcripts. That is a ranking Microsoft cites, and this report has not checked the leaderboard itself. Artificial Analysis publishes its own results, which readers can compare against the claim.
It is priced at $0.54 per audio hour through the end of 2026, an introductory rate, according to the report. Access runs through Microsoft Foundry, the MAI Playground, Vercel, Azure Voice Live and OpenRouter, with LiveKit support listed as coming soon. Developer documentation is on Azure AI Services.
Two voice models launched alongside it, TestingCatalog says. MAI-Voice-2.1 is text-to-speech at $22 per million characters across 23 languages. The Flash version costs $15 per million characters across 55 languages, with inference 55 per cent faster. Nobody outside Microsoft has published independent tests of any of the three yet.