Microsoft released MAI-Transcribe-2 on September 3, 2026, and the numbers make a statement. The model tops the FLEURS benchmark across 60 languages with an average word error rate of 5.2%, beating Gemini 3.5 Transcribe, OpenAI’s GPT-Transcribe, Whisper V3-Large, and ElevenLabs’ ScribeV2. And it does it at $0.10 per audio hour — an introductory price that holds through December 31, 2026.

That is not a minor upgrade. It is a direct shot at three competitors who have been charging significantly more for worse results.

What MAI-Transcribe-2 Actually Does

MAI-Transcribe-2 is available now in public preview through Azure Speech. It handles 60 languages and brings three features that matter for real production use:

Diarization. The model can identify and separate speakers in multi-person audio, which makes it useful for meeting transcription, call centre analysis, and interview processing — not just single-speaker dictation.

Configurable transcription styles. Businesses can tune output formatting for different use cases, from verbatim transcripts to cleaned-up summaries.

Word-level timestamps. Every word is tagged with a timestamp, which makes the output searchable and usable for downstream analysis rather than just readable.

The speed improvement is the other headline: up to 10 times faster processing than leading competitors, particularly for long-form audio. For businesses processing hours of recordings — sales calls, customer support sessions, executive meetings — that gap compounds quickly.

Why Microsoft Is Building This

There is a strategic reason Microsoft spent the engineering effort to build MAI-Transcribe-2 in-house rather than continuing to license speech capability from third parties.

Microsoft’s commercial Azure voice service (Voice Live) currently runs on a model Microsoft licenses rather than owns. Every customer conversation it handles is a transaction Microsoft pays a third party to enable. Owning the model converts that from a licensing cost into infrastructure.

MAI-Transcribe-2 follows MAI-Transcribe-1 (released April 2026), MAI-Realtime (a full-duplex voice model currently in limited preview), and a growing suite of in-house AI models that replace OpenAI and Google dependencies across Microsoft’s product stack. The pattern is consistent: Microsoft is rebuilding its AI layer on models it controls.

What This Means for Business

If you are building a voice application or buying voice AI services, the competitive dynamics just shifted.

Pricing pressure will spread. At $0.10 per hour, MAI-Transcribe-2 sets a benchmark that other providers will have to respond to. OpenAI, Google, and ElevenLabs all charge more for comparable or lower-quality transcription. The intro pricing through December only accelerates that pressure — the question is where prices land in 2027 when the promotional period ends.

Accuracy gaps affect everything downstream. A 5.2% word error rate versus a competitor’s 7-8% sounds like a small difference. In practice, for automated workflows that act on transcribed content — routing a support call, extracting data from an interview, triggering an agent based on spoken input — every misrecognised word is a potential error in the output. Accuracy compounds.

Diarization makes voice AI more useful in real business contexts. Most enterprise voice scenarios involve multiple speakers: a sales rep and a prospect, a support agent and a customer, a leadership team in a meeting room. Models that handle single speakers well but struggle with speaker separation are limited to narrow use cases. Diarization that actually works opens the full range.

Multi-language capability matters for global operations. 60 languages with genuine accuracy (rather than the degraded performance most models show outside English) makes transcription infrastructure viable for international teams and customer bases.

The Broader Voice AI Race

2026 has been a year of significant investment in voice AI infrastructure. xAI released Grok Voice Think Fast 2.0 in July with sub-second response times and 1.4x better accuracy. Microsoft is testing MAI-Realtime for full-duplex communication (listening and speaking simultaneously). Gemini crossed 1 billion monthly active users with voice as the dominant interaction mode.

The market is moving from voice AI as a novelty to voice AI as standard enterprise infrastructure — something every contact centre, meeting platform, and customer-facing application needs to get right. MAI-Transcribe-2’s combination of accuracy, speed, and price point makes the technology more accessible for businesses that were previously priced out or put off by quality limitations.

For businesses evaluating voice AI infrastructure — whether for internal operations, customer-facing applications, or agent-driven workflows — the calculus has changed. The best-performing transcription model is now also the cheapest.

---

Enterprise DNA builds Omni Voice, an AI voice employee service for businesses that need to handle inbound enquiries, internal knowledge discovery, and operational reporting without adding headcount. Learn more about Omni Voice or book a discovery call to see what it can do for your team.