Microsoft AI on October 1, 2026 launched MAI-Transcribe-2-Streaming, its first streaming transcription mannequin, alongside two new text-to-speech fashions, MAI-Voice-2.1 and MAI-Voice-2.1-Flash, with all three accessible via Microsoft Foundry.
MAI-Transcribe-2-Streaming and the Synthetic Evaluation Benchmark
Microsoft describes MAI-Transcribe-2-Streaming as delivering low-latency, real-time transcripts in 60 languages with computerized, steady language detection. The corporate stated the mannequin ranks No. 1 for accuracy for each ultimate and partial transcripts on Synthetic Evaluation, and that it sits on the Pareto frontier of the benchmark’s accuracy-versus-latency analysis, that means increased accuracy doesn’t require a heavy latency tradeoff. The leaderboard chart within the publish, citing the Synthetic Evaluation streaming leaderboard dated September 28, 2026, reveals the mannequin at a 2.5 p.c ultimate word-error price, a 2.8 p.c first-partial price, and 0.13 seconds to ultimate transcription.
Fairly than ready for a speaker to complete earlier than returning textual content, the mannequin produces its first hypotheses, referred to as partials, in simply over 100 milliseconds of receiving audio, then revises them as extra context arrives earlier than committing a steady transcript. Microsoft stated this enables voice-enabled functions to behave on speech earlier than the speaker finishes: voice brokers can begin reasoning or calling instruments mid-sentence, and reside transcripts can seem as folks speak. For real-time dictation and subtitling, the corporate stated its inner evaluations present phrases showing within the transcript twice as quick as with its closest competitor.
Artificial Analysis states that its AA-WER Streaming index measures transcription accuracy for fashions the place audio streams in actual time, chunk by chunk, throughout roughly eight hours of audio from three datasets: AA-AgentTalk at 50 p.c, VoxPopuli at 25 p.c, and Earnings22 at 25 p.c. The datasets cowl real-world speech with numerous accents, domain-specific language, and difficult acoustic circumstances, and the benchmark’s Time to Ultimate and Time to First Partial measurements each begin on the finish of speech detected by the SileroVAD voice-activity detector.
MAI-Transcribe-2-Streaming is out there at an introductory value of $0.54 per hour of audio via the tip of the yr. The mannequin extends Microsoft’s MAI audio line, which already contains MAI-Transcribe-2, the sooner non-streaming speech recognition mannequin the corporate billed because the quickest, most correct and most cost-effective on the earth.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash
MAI-Voice-2.1 helps 23 languages and 26 locales, and Microsoft stated a single voice can use all of them with a local accent, preserving the identical speaker id when switching languages. A tutoring app, within the firm’s instance, can swap languages mid-lesson with out swapping academics, and a multilingual assistant can reply in no matter language it’s addressed in whereas nonetheless sounding like the identical voice. The mannequin is priced at $22 per 1M characters.
MAI-Voice-2.1-Flash helps the identical languages and cross-language audio system however is constructed for high-volume, latency-sensitive workloads. It might generate as much as 45 seconds of audio with an end-to-end latency of 150 milliseconds, and Microsoft stated it delivers 55 p.c quicker mannequin inference and is roughly 60 p.c cheaper than comparable fashions. It’s priced at $15 per 1M characters.
Each voice fashions help voice cloning throughout all supported languages utilizing a couple of seconds of reference audio, with built-in consent guardrails that Microsoft stated forestall misuse. In a 4,000-listener Turing take a look at combining the 2 new voice fashions, 50.3 p.c of listeners rated MAI-Voice as equally or extra human-like than human recordings, Microsoft stated.
Microsoft framed pairing MAI-Transcribe-2-Streaming with MAI-Voice-2.1-Flash as shopping for again time on each ends of a voice-agent loop, the sequence of listening to, understanding, deciding, and talking inside the window the place a human nonetheless experiences the interplay as a dialog. Listed developer use instances embody customer support brokers that transcribe requests as they’re spoken and reply in pure speech, multilingual assistants that detect the spoken language and reply in any of the 23 supported MAI-Voice languages, and interactive studying and media functions utilizing distinct audio system for tutoring, role-play, simulations, narration, and conversational content material.
Availability and the Chatter Demo
MAI-Voice-2.1 and MAI-Voice-2.1-Flash can be found via OpenRouter. All three fashions can be found via Microsoft Foundry, the MAI Playground, Vercel, and Azure Voice Stay, with LiveKit listed as coming quickly.
To indicate the fashions working collectively in a reside agent, Microsoft constructed Chatter, a brand new demo within the MAI Playground that lets customers speak to a voice assistant powered by the transcription and voice fashions.

