Voice agents live inside a hard physical constraint: human ears tolerate only a few hundred milliseconds of silence. Hearing, understanding, deciding, speaking — every component either buys time in that window or spends it. On October 1, Microsoft released three voice models that push both ends of the loop toward hundred-millisecond territory: the streaming transcription model MAI-Transcribe-2-Streaming, the TTS model MAI-Voice-2.1, and the high-volume variant MAI-Voice-2.1-Flash.

Transcription: first drafts in about 100ms

MAI-Transcribe-2-Streaming is Microsoft's first streaming transcription model. It supports 60 languages with automatic, continuous language detection, and it does not wait for the speaker to finish: it emits its first partial hypotheses just over 100ms after receiving audio, then keeps revising them as context rolls in and commits a stable transcript promptly. Microsoft says it ranks no. 1 for accuracy on both final and partial transcripts in Artificial Analysis' streaming evaluation, sits on the accuracy-versus-latency Pareto frontier, and — per internal evaluations — puts words on screen 2x faster than its closest competitor in real-time dictation. Introductory pricing is $0.54 per hour of audio through the end of the year.

Speech: one voice across 23 languages

MAI-Voice-2.1 is Microsoft's strongest multilingual TTS yet, covering 23 languages and 26 locales. A single voice can speak English, then Mandarin, then German, picking up a native accent in each rather than dragging one accent across languages. It is priced at $22 per 1M characters. The Flash variant targets high-volume, latency-sensitive workloads: it generates 45 seconds of audio with an end-to-end latency of just 150ms, runs 55% faster inference and is roughly 60% cheaper than comparable models, at $15 per 1M characters. Both voice models support voice cloning across all supported languages from a few seconds of reference audio, with built-in consent guardrails.

Cross-checking the leaderboard: whose "no. 1" is it?

Official claims deserve a cross-check. On Artificial Analysis's public non-streaming leaderboard, MAI-Transcribe-2 posts a 2.0% word error rate — third place, behind two Chinese models, StepFun's StepAudio 3 ASR and Alibaba's Fun-Realtime-ASR-preview, both at 1.7%. Microsoft's "no. 1 on both boards" claim is scoped to the streaming benchmark. The generational jump is real, though: the previous MAI-Transcribe-1.5 scored 2.4% WER with a speed factor of 194, while the new model hits 2.0% at 372 — and its listed price on the same board dropped from $6.00 to $1.67 per 1,000 minutes, roughly 70% cheaper at nearly double the speed. The pure speed crown still belongs to Deepgram's Nova-3 at 578x real-time.

So what: milliseconds saved at the edges are reasoning time in the middle

Microsoft frames a voice agent as a loop — hear, understand, decide, speak — where every component either buys or spends time. The real story of this launch is not any single benchmark: once the fixed overhead at both ends drops to the hundred-millisecond scale, the window left for the model in the middle to reason, call tools, and check its answer gets wider. All three models are available through Microsoft Foundry, the MAI Playground and Vercel; the voice models are also on OpenRouter, with LiveKit support coming soon. For teams building voice agents, $0.54-per-hour transcription makes "listening in real time" nearly free — the real bottleneck has moved back to the model that does the thinking.

References: Microsoft announcement https://microsoft.ai/news/our-first-streaming-transcription-model/ ; Artificial Analysis speech-to-text leaderboard https://artificialanalysis.ai/speech-to-text