MAI-Transcribe-1

What it is

MAI-Transcribe-1 is a speech-to-text model from Microsoft AI, released on 2 April 2026 and made available in Microsoft Foundry and the MAI Playground.^[1] Its job is the one behind every dictation box and caption track: audio goes in, text comes out. The model transcribes the 25 most-used languages, a list Microsoft defined by usage of its own products rather than by how many people speak each language worldwide.^[1]

Getting that list right matters more than it sounds. A transcription model is only useful to a product team if it covers the languages that team's users actually type and speak in, and a list built from product usage is a different, narrower set than a list built from global speaker counts.

How it was measured

Microsoft reported results on FLEURS, a multilingual speech benchmark. The company said the model ranks first in 11 core languages, beats Whisper-large-v3 on the other 14, and beats Gemini 3.1 Flash on 11 of those 14.^[1] The claim is therefore not a single headline number but a language-by-language comparison against two widely used alternatives, with the strongest results concentrated in the languages Microsoft treats as core.

Benchmark scores on read speech are one thing; the release materials emphasise a second, harder setting. Microsoft says the model is designed for noisy, real-world audio rather than clean studio recordings, which is where transcription systems tend to fail in practice — meeting rooms, phone calls, recordings made on a laptop microphone.^[1]

Speed, cost and price

Much of the pitch is about bulk work. For batch transcription — feeding a stack of recorded files through the system rather than captioning a live call — Microsoft says the model is 2.5 times faster than its existing Azure Fast offering.^[1] At launch, pricing started at $0.36 per hour of audio.^[1][2]

The Register reported Microsoft's claim of enterprise-grade accuracy across the 25 languages at roughly 50 percent lower GPU cost than leading alternatives.^[3] Cost per hour of audio and cost per GPU are two different numbers, and Microsoft cited both: one is what a customer pays, the other is what it costs Microsoft to serve the request.

Where it already runs

The model did not arrive as a standalone product waiting for customers. Microsoft said the models were already powering Copilot, Bing, PowerPoint and Azure Speech, and were then offered to developers through Foundry, with MAI-Transcribe-1 and MAI-Voice-1 also tryable via Azure Speech.^[3] The most visible consumer-facing use is Copilot's Voice Mode, which runs its transcription on this model.^[3]

That order — inside Microsoft's own products first, then sold to developers — is a common pattern for the company's in-house models, and it means the system had been handling real traffic before it had a price list.

The next version

The line moved quickly. On 2 June 2026 Microsoft followed with MAI-Transcribe-1.5, which it said supports 43 languages and is five times faster than competing models.^[4] Two months between versions, with the language count going from 25 to 43, suggests the first release was a starting point rather than a finished product.

Where it sits

The Register noted that Microsoft held an OpenAI stake valued at about $135 billion as of October 2025, which made the release look more like competition than investment.^[3] A company that both invests in a model developer and ships its own models in the same categories is hedging: it can keep using outside models where they are strongest while building in-house ones it controls the price, the hosting and the roadmap for.

For anyone using speech-to-text, the practical consequence is a third option alongside the established open and commercial systems, priced by the hour of audio and benchmarked language by language against them.