Our first streaming transcription model debuts at no. 1 on Artificial Analysis | Microsoft AI
Short summary
Microsoft released three new models for building voice agents: a streaming transcription model that produces text in 100ms, and two text-to-speech models optimized for speed and cost. Together they're designed to handle the full loop of hearing, understanding, deciding, and speaking in real-time conversations.
- MAI-Transcribe-2-Streaming ranks #1 for accuracy on Artificial Analysis and delivers partial transcripts in just over 100ms, letting voice agents start acting mid-sentence rather than waiting for speakers to finish.
- MAI-Voice-2.1-Flash generates 45 seconds of audio with 150ms end-to-end latency and costs 60% less than comparable models at $15 per 1M characters, making it practical for high-volume applications.
- Both voice models support 23 languages with a single voice maintaining native accents across all of them, plus voice cloning from a few seconds of audio with built-in consent guardrails against misuse.
Full summary
Microsoft released three new AI models for voice work: MAI-Transcribe-2-Streaming for real-time transcription, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash for text-to-speech. The transcription model ranks first on Artificial Analysis for accuracy and hits the Pareto frontier on the speed-versus-accuracy tradeoff. It produces initial transcripts in just over 100 milliseconds and continuously refines them as more audio arrives, which means voice agents can start acting on what someone's saying before they finish speaking. This matters because it cuts the latency window—words appear roughly 2x faster than competitors' models. Pricing is $0.54 per hour of audio through the end of the year.
The voice models handle 23 languages across 26 locales, but the real trick is that a single voice can speak all of them with native accents rather than dragging one accent across languages. That means a tutoring app can switch languages mid-lesson without swapping speakers, or a customer service bot can respond in whatever language it's addressed in while keeping the same brand voice. MAI-Voice-2.1-Flash is the speed variant—it generates 45 seconds of audio with 150 milliseconds end-to-end latency, runs 55% faster than standard models, and costs about 60% less at $15 per million characters compared to $22 for the standard version. Both support voice cloning from just a few seconds of audio, though they have built-in safeguards against misuse.
The three models work together to close the latency loop that makes voice agents feel conversational. A voice agent has to hear, understand, decide, and speak within a window where humans still perceive it as natural dialogue. Pairing the fast transcription model with the Flash voice model buys back time on both ends, leaving room for the agent to reason and use tools without the conversation feeling sluggish. Microsoft built a demo called Chatter in their MAI Playground to show the models working together, and developers can access them through multiple platforms including Microsoft Foundry, Vercel, and LiveKit.