Microsoft MAI-Transcribe-2-Streaming and MAI-Voice-2.1: The Numbers

Microsoft shipped three voice models in one day. One of them is now the most accurate streaming transcriber on the market.

On October 1, 2026, Microsoft AI released three models aimed at voice agents: MAI-Transcribe-2-Streaming for real-time speech recognition, and two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash.

The framing matters as much as the models. Microsoft describes a voice agent as a loop — hear, understand, decide, speak — that has to complete inside the window where a human still experiences it as a conversation. Every component either buys time in that window or spends it. These three models are the company's attempt to buy time back at both ends.

MAI-Transcribe-2-Streaming: 2.5% WER across 60 languages

The transcription numbers are unusually specific.

Accuracy: 2.5% word error rate on final transcripts, ranking #1 of 38 models on Artificial Analysis's AA-WER Streaming index
Latency: final transcript committed 0.13 seconds after the end of speech; first partial results arrive in just over 100 milliseconds
Languages: 60, with automatic continuous language detection
Price: $0.54 per hour of audio, introductory through the end of 2026

The detail that deserves attention is that the first partial transcript also scores 2.5% WER — identical to the final. That means an agent can begin reasoning or calling tools mid-sentence without trading away accuracy. Microsoft claims words appear in the transcript twice as fast as its closest competitor in internal dictation and subtitling tests.

The runners-up put the gap in perspective. Grok Voice Transcribe 2.0 sits at 2.7% WER and 0.49s to final. Muse Voice Transcribe is at 3.1% and 0.16s. Cartesia Ink-2 is faster still, returning finals in 0.07s, but its WER degrades to 4.0%.

Pricing is where the picture gets more complicated. Normalized, $0.54 per hour works out to $9.00 per 1,000 minutes. ElevenLabs and Deepgram Flux land near $6.50. Microsoft's own batch model, MAI-Transcribe-2, costs $0.10 per hour — roughly one-fifth the streaming rate. And there is no SLA: this is a public preview.

MAI-Voice-2.1: one voice, twenty-three languages

On the speech output side, MAI-Voice-2.1 is the flagship.

23 languages and 26 locales
A single "voice" uses every supported language with a native accent — English, then Mandarin, then German, and the speaker stays recognizably the same
Priced at $22 per 1 million characters
Model inference latency around 550 milliseconds
Voice cloning from a few seconds of reference audio, with built-in consent guardrails

That first property is the actual product claim. Conventional multilingual TTS has forced a choice: drag one accent across every language, or assign a different speaker per language. Microsoft is claiming the third option — switching language mid-lesson without swapping teachers, or answering a multilingual assistant in whatever language it was addressed in while still sounding like the same person.

Microsoft also reports that in one test, roughly half of 4,000 participants believed the generated voices belonged to a real person.

MAI-Voice-2.1-Flash: 150 ms and $15

The Flash variant trades a little fidelity for throughput.

Generates 45 seconds of audio at 150 milliseconds end-to-end latency
55% faster model inference and roughly 60% cheaper than comparable models
$15 per 1 million characters — about 32% below the flagship's $22
Same languages and cross-language speaker support as MAI-Voice-2.1

Paired together, Transcribe-2-Streaming spends 0.13s on hearing and Voice-2.1-Flash spends 0.15s on speaking. The remaining turn budget is what the agent gets for reasoning and tool calls — which is the entire argument for shipping them as a set.

The numbers in one place

MAI-Transcribe-2-Streaming: $0.54 per hour (introductory, through end of 2026)
MAI-Transcribe-2 batch: $0.10 per hour
MAI-Voice-2.1: $22 per 1M characters
MAI-Voice-2.1-Flash: $15 per 1M characters

Three caveats belong next to those figures. First, all three are public previews with no SLA and time-limited introductory pricing. Second, no weights are published — you cannot self-host or fine-tune these, unlike some competing voice models. Third, the streaming transcription rate is the highest among its direct competitors; teams optimizing purely for cost should compare it against the batch model and against ElevenLabs.

Access runs through Microsoft Foundry (Azure Speech), the MAI Playground, and — for the two voice models — OpenRouter. Microsoft also built Chatter, a live demo in the Playground that wires all three models into a single agent.

Why the partial-transcript result is the real news

The most consequential detail is not the #1 ranking. It is that the partial transcript now matches the final transcript in accuracy. Before this, agent builders faced a hard tradeoff: act on fast-but-wrong interim hypotheses, or wait for the accurate final and lose the conversational window. A 2.5% WER partial removes that tradeoff. It is what turns "respond while the caller is still mid-sentence" from a demo trick into a production design.

For non-English developers, one gap is worth flagging before you commit. Microsoft's announcement pages do not publish the full language list. If your users speak a language that falls outside the 60 transcription or 23 synthesis languages, you will not discover that at design time from the marketing page — check the Azure Speech documentation directly.

The reasoning-model layer has largely consolidated this year. The input and output layers — hearing and speaking — have not. Microsoft just made a serious bid for both.

Where these models sit against the alternatives

The voice stack is the one layer of the AI market that has not consolidated. OpenAI, Google, Anthropic, and DeepSeek all released decision models in October — none of them shipped a voice model in that window. Microsoft is competing against specialists rather than against its usual rivals, and the specialists are not standing still.

Transcription: ElevenLabs and Deepgram Flux undercut Microsoft on streaming price; Cartesia Ink-2 beats it on raw latency while losing on accuracy. Microsoft's edge is top-ranked accuracy at a latency that is still conversational.
Speech synthesis: the $15–$22 per million characters range is high against typical TTS pricing, and the absence of open weights removes the self-hosting path entirely. Teams that need to own their inference cannot use these models at all.
Ecosystem: Microsoft's leverage is distribution: Foundry, Azure Speech, Copilot, and the MAI Playground put these models in front of developers already on Azure. For teams outside, switching cost is a fresh integration.

Who should actually evaluate this

Worth a serious look in three situations.

First, if you are building a contact center or support agent where the caller speaks more than one language. The combination of automatic language detection on input and a single consistent voice across 23 languages on output is not something the open-weight alternatives do well today.

Second, if your product's perceived quality depends on low-latency turn-taking. The 0.13s-to-final transcription plus 150ms speech generation leaves roughly 1.2 seconds of a two-second turn for reasoning and tool calls — a workable budget, and the number to prototype against.

Third, if you need voice cloning for a brand voice with a documented consent framework. Microsoft's guardrails are a compliance argument, not a technical one, and for some buyers that decides it.

Skip it if you need open weights, if your users speak an unsupported language, or if the 5x gap between batch and streaming transcription dominates your cost model.

What to watch next

Three things will determine whether this launch mattered.

The first is whether Microsoft publishes a general-availability date and a permanent price. Introductory pricing that expires at the end of 2026 makes every cost model built on it provisional.

The second is the language list. A 60-language transcription and 23-language synthesis claim is strong, but the specific languages decide the addressable market. Publishing the list is the cheapest credibility Microsoft could buy.

The third is whether the partial-transcript accuracy holds up on real audio. The AA-WER Streaming index uses roughly eight hours of audio weighted toward agent conversations, earnings calls, and crowd-sourced speech. That is a good proxy, not a guarantee — accented speech, crosstalk, background noise, and domain jargon are where WER usually degrades first.

Honest summary: Microsoft has shipped the most accurate streaming transcriber on the public leaderboard and a multilingual synthesis model that solves a genuine problem, both behind a preview wall, neither downloadable. A real technical result attached to a real commercial caveat.