Microsoft’s MAI Speech Models Compared: Transcription, Voice, Pricing and Access

Microsoft’s three new MAI models cover both sides of a spoken interaction: MAI-Transcribe-2-Streaming turns live speech into text, while MAI-Voice-2.1 and MAI-Voice-2.1-Flash generate speech from text. Announced for developers building voice agents, the models are available in public preview through Microsoft Foundry. 3639

At a glance

Model	Main role	Reported language support	Reported price
MAI-Transcribe-2-Streaming	Live speech-to-text with incremental transcript updates	60 languages, with automatic language detection	$0.54 per audio hour, an introductory rate through the end of 2026 435
MAI-Voice-2.1	Expressive text-to-speech	23 languages	$22 per million characters 3536
MAI-Voice-2.1-Flash	Faster text-to-speech for responsiveness-focused applications	23 languages reported for the voice models	$15 per million characters 3536

These are reported listings, not a guarantee that prices or availability will remain unchanged. In particular, Microsoft’s transcription price is described as introductory. 435

MAI-Transcribe-2-Streaming: live transcription

Unlike a transcription system that waits for someone to finish speaking, MAI-Transcribe-2-Streaming sends transcript updates while speech is still arriving. It accepts continuous audio over a WebSocket and automatically detects among 60 languages, according to reporting on its model listing. 1

The model’s latency figures vary by report and measurement. One account says it can begin producing partial results just over 100 milliseconds after receiving audio; another says words can begin appearing roughly 320 milliseconds after they are spoken. Those descriptions use different stated starting points, so they should not be read as a like-for-like latency comparison. 24

On the Artificial Analysis streaming-transcription benchmark, reports say the model debuted at No. 1 for accuracy. One report gives a final word-error rate of 2.5% among 38 models. That is a reported benchmark result, not a guarantee of accuracy for every language, recording condition or application. 34

The listed introductory price is $0.54 per hour of audio through the end of 2026. Developers evaluating the model should account for the fact that this is a time-limited promotional rate. 435

MAI-Voice-2.1 and Flash: speech generation

Both voice models turn text into spoken audio, but they are positioned for different priorities. MAI-Voice-2.1 is described as the more expressive option, while Flash is the faster variant for applications where keeping the wait between turns short matters. The reported language coverage is 23 languages for the voice models. 63536

Reported listings price MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15 per million characters. A report also describes Flash as generating 45 seconds of audio with 150 milliseconds of latency; treat that as a reported figure rather than a universal measure of end-to-end response time. 3536

There is no directly comparable public benchmark score for voice quality in the available reports. The transcription model’s accuracy ranking should not be taken as a ranking of either speech-generation model. 3234

What developers can access—and what the launch signals

All three models are reported in public preview through Microsoft Foundry. Public preview gives developers a way to test the models, but it is not the same as a statement that they are generally available. 36

For developers, the practical change is that Microsoft now offers its own speech-recognition and speech-generation components for a voice-agent workflow. That could make it easier to build more of the speech layer within Microsoft’s developer ecosystem and reduce the need to source those particular components elsewhere. It does not establish that a complete voice agent can work without other providers: reasoning, orchestration and other capabilities may still involve different models or services. 639

The bottom line

The clearest point of comparison is the transcription model: it supports 60 languages, returns incremental results, and has a reported top Artificial Analysis accuracy ranking, alongside a stated 2.5% final word-error rate. The two voice models offer an expressive-versus-speed choice, with reported coverage of 23 languages and different per-character prices. Developers should verify current preview access, pricing and latency in their intended setup before making a production decision. 1343536