Microsoft launches MAI-Transcribe-2-Streaming, claiming the top spot on Artificial Analysis for real-time transcription

Start with what streaming actually changes. Microsoft's description: rather than waiting for someone to finish speaking before returning text, it produces its first hypotheses, known as partials, in just over 100 milliseconds of receiving audio, then revises them as more context rolls in and commits a stable transcript right away. The value is that voice applications can act before the speaker even finishes — a voice agent can start reasoning or calling tools mid-sentence, and live captions can appear as people talk. Microsoft says internal evaluations show words appearing in the transcript 2x faster than its closest competitor for cases such as real-time dictation or subtitling. On coverage, it delivers low-latency real-time transcripts in 60 languages and supports automatic, continuous language detection. Pricing: an introductory $0.54 per hour of audio through the end of the year.

On the synthesis side, MAI-Voice-2.1's pitch is one voice for every language. Microsoft says it is its strongest multilingual text-to-speech model yet: coverage expands to 23 languages and 26 locales, and a single voice can speak all of them while keeping a truly native accent — ask it to speak English, then Mandarin, then German, and the speaker stays unmistakably the same, naturally picking up the local parlance rather than dragging one accent across languages. Microsoft gives two concrete cases: a tutoring app can switch languages mid-lesson without swapping teachers, and a multilingual assistant can reply in whatever language it is addressed in while still sounding like the same voice. Pricing is $22 per 1M characters.

The third number is MAI-Voice-2.1-Flash, aimed at high-volume, latency-sensitive workloads. Microsoft says it supports the same languages and cross-language speakers as MAI-Voice-2.1 but is leveled up for throughput: it generates 45 seconds of audio with 150 milliseconds of end-to-end latency, delivers 55% faster model inference and is about 60% cheaper than comparable models at $15 per 1M characters. Microsoft frames the pairing clearly: combined with the transcription model, Flash buys time back on both the hearing and speaking ends, and the time saved gives an agent more room to reason, use tools and check its answer, while the conversation keeps moving at a human, conversational speed. Both voice models also support cloning across all supported languages using just a few seconds of reference audio, which Microsoft says makes it easier for customers to use their brands' voices; it emphasizes built-in consent guardrails to prevent misuse.

Microsoft also published a subjective test result and a set of deployment routes. It says that in a 4,000-listener Turing Test, 50.3% rated MAI-Voice as equally or more human-like than human recordings, while 49.7% rated the human recordings as more human-like; Microsoft notes the result combines MAI-Voice-2.1 and MAI-Voice-2.1 Flash. On availability, the two voice models are reachable through OpenRouter, and all three models are available through Microsoft Foundry, the MAI Playground, Vercel and Azure Voice Live, with LiveKit listed as coming soon. Microsoft also lists three ready use cases: customer service agents that transcribe requests as they are spoken and begin acting before the caller finishes before responding in natural speech; multilingual assistants that detect the spoken language automatically and reply in any of the 23 supported voice languages, in the same voice and with a native accent; and interactive learning and media that use distinct speakers for tutoring, role-play, simulations and narration. To show the three models working together in a live agent, Microsoft built a new demo in the MAI Playground called Chatter.