Microsoft launches MAI-Transcribe-2-Streaming and MAI-Voice-2.1, 60-language streaming tops Artificial Analysis

Microsoft launched three new models on October 1 in Microsoft Foundry: MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash, all aimed at real-time conversational scenarios. MAI-Transcribe-2-Streaming supports 60 languages with continuous language detection, returns initial partials in just over 100 ms, and ranks first on Artificial Analysis for accuracy on both partial and final transcripts. The model also sits on the Pareto frontier for accuracy and latency; Microsoft says this combination is what enables voice agents to call tools mid-sentence and live transcripts to appear as people speak.

MAI-Voice-2.1 is Microsoft's most expressive multilingual TTS, supporting 23 languages and 26 locales with a single voice identity that adapts naturally across languages; MAI-Voice-2.1-Flash is optimized for high-volume, latency-sensitive use and can produce 45 seconds of audio within 150 ms end-to-end latency, with 55% faster inference and around 60% lower cost than comparable models. Both support cross-language voice cloning from a few seconds of reference audio and include consent guardrails against misuse.

All three models are available in Microsoft Foundry, Azure Speech, and the MAI Playground, with MAI-Voice-2.1 and Flash also on OpenRouter and Vercel; LiveKit integration is coming soon. Pricing: MAI-Transcribe-2-Streaming at $0.54 per audio hour through end of 2026, MAI-Voice-2.1 at $22 per 1M characters, and MAI-Voice-2.1-Flash at $15 per 1M characters. In a 4,000-person evaluation, 50.3% of listeners rated MAI-Voice outputs as indistinguishable from real recordings. The launch extends Microsoft's first-party MAI model lineup and is positioned to give Azure developers a full speech stack without depending on third-party transcription APIs.