MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash

Microsoft AI released three speech models on 1 October 2026, all aimed at voice agents: software that listens to a person, decides what to do and answers out loud, fast enough to feel like a conversation. MAI-Transcribe-2-Streaming turns live speech into text as it is spoken. MAI-Voice-2.1 and MAI-Voice-2.1-Flash turn text into speech, in 23 languages.

“A voice agent is a loop. It has to hear, understand, decide, and speak,” the launch post says. “And do it all within the window where a human still experiences the interaction as a conversation.” All three run through Azure Speech in Microsoft Foundry, in public preview.

MAI-Transcribe-2-Streaming writes while you talk

MAI-Transcribe-2-Streaming transcribes audio as a continuous stream, returning partial transcripts that update while the speaker talks and final transcripts that confirm each segment, in 60 languages with automatic language detection. It is the streaming sibling of MAI-Transcribe-2, the batch model Microsoft AI released on 3 September 2026, and its first partials arrive “in just over 100ms of receiving audio”, Microsoft says.

Developers connect through the Azure Speech SDK or over a WebSocket protocol compatible with OpenAI’s Realtime API. On the WebSocket route they send 16-bit mono audio in small chunks, and each session can last up to an hour, according to Microsoft Learn. Word-level timestamps, keyword biasing and speaker labels stay with the batch model, Microsoft’s comparison table shows.

It ranks first for accuracy on the streaming board

Artificial Analysis ranks MAI-Transcribe-2-Streaming first of 38 streaming models for accuracy, at a 2.51% word error rate on final transcripts and 2.52% on first partials, on its board read on 2 October 2026. The index mixes three test sets, around eight hours of voice-agent speech, European Parliament proceedings and company earnings calls, according to its methodology. Artificial Analysis reported it “taking the #1 spot” on 1 October.

Accuracy is the claim it wins. On the same board ten models return a final transcript sooner, led by Deepgram Flux at 0.021 seconds, and its $9.00 per 1,000 minutes is one of the higher prices: 28 of the 38 models cost less. Microsoft says it sits “on the Pareto frontier” for accuracy against latency.

Transcription costs $0.54 an hour

MAI-Transcribe-2-Streaming costs $0.54 per hour of audio, an introductory price that runs to 31 December 2026, according to its model card. That is 5.4 times the $0.10 an hour, also an introductory price, that Microsoft charges for the batch MAI-Transcribe-2, which waits for a recording to end before it transcribes.

At those rates, an hour of a voice agent listening costs $0.54 to transcribe, and 100,000 characters of spoken reply cost $1.50 on Flash.

Two voices that keep their identity across 23 languages

MAI-Voice-2.1 and MAI-Voice-2.1-Flash speak 23 languages, and one voice keeps the same identity in each of them with a native accent, Microsoft says. “Just ask it to speak English… then Mandarin… then German… and the speaker stays unmistakably the same,” the launch post says. Microsoft Learn lists 97 prebuilt voices, 50 male and 47 female, all usable with both models, and styles such as joy, excitement and empathy set through speech markup.

Flash streams its output, so a caller can interrupt it mid-sentence. The quality figures are Microsoft’s own. In blind side-by-side comparisons over 5,032 judgments, listeners preferred MAI-Voice-2.1 with new prompts 58.9% of the time against MAI-Voice-2 with old ones. In a test with 4,000 listeners, 50.3% rated the MAI voices as human-like as or more human-like than human recordings, with 2.1 and 2.1-Flash results combined, the model page says. Flash delivers “55% faster model inference” and is about 60% cheaper than comparable models, Microsoft adds, in a comparison against ElevenLabs’ published model documentation.

Voice cloning is gated. Microsoft’s documentation describes “instant voice cloning” from a 5 to 60 second reference clip, available only after approval through Azure’s limited access review and with a recorded consent statement from the speaker.

Where to use them

All three models run through Azure Speech in Microsoft Foundry, in public preview, and the voices are served from 14 Azure regions. The two voice models are also on OpenRouter, at the same prices. Microsoft lists Vercel, LiveKit (marked coming soon) and Azure Voice Live among the other routes to the models, and its MAI Playground runs a demo called Chatter on them.

A live transcriber joins Microsoft AI’s voice models

Microsoft AI already had speech models of its own: MAI-Voice-1 reached Microsoft Foundry alongside the new MAI-Transcribe-1 on 2 April 2026, and MAI-Voice-2 followed on 2 June. MAI-Transcribe-2-Streaming is its first model that listens live. “Pairing MAI-Transcribe-2-Streaming with MAI-Voice-2.1-Flash buys time back on both ends,” the launch post says. The transcription side has an independent first place on Artificial Analysis’s accuracy ranking; the voice side’s quality figures are Microsoft’s own. For the batch model released a month earlier, see MAI-Transcribe-2; for the rival voice platform Microsoft compares Flash with, ElevenLabs.