Transcription drops to 10 cents an hour

Microsoft AI made MAI-Transcribe-2 available on Microsoft Foundry and MAI Playground on September 3, at an early-bird rate of $0.10 per hour of audio. That rate covers speaker diarization, word-level timestamps, keyword biasing, a choice of verbatim or clean output, and handling of speech that switches language mid-sentence. In other words, the features that turn raw text into a usable transcript are part of the base product.

As VentureBeat reports, this is the third model in the line in five months. MAI-Transcribe-1 launched on April 2 with 25 languages at $0.36 an hour. MAI-Transcribe-1.5 followed on June 2 with 43 languages and keyword biasing. Version 2 covers 60 languages and cuts the price by roughly 72%. VentureBeat’s example is a bank or telecom running 100,000 hours of call-center audio a year: its bill falls from $36,000 to $10,000. Microsoft calls the $0.10 rate a launch offer, though it has named neither an end date nor a standard price.

What the base rate buys

All of these features are request parameters. According to Microsoft’s Speech service documentation, you call the model through enhancedMode, with enabled: true and model: "MAI-Transcribe-2". Diarization, timestamps, keyword lists and output style are each set separately.

Diarization splits a recording by speaker, so a meeting transcript comes back as turns attributed to people rather than one block of text. It works on shorter recordings than plain transcription does, so long files may have to be cut up before you ask for speaker labels. Timestamps can be set to word, segment or none. The word level gives a time marker for every word, which is what search, editing and video alignment need.

Keyword biasing goes through phraseList.phrases. You give the model a list of drug names, product codes or employee names, and it leans toward those spellings. The documentation describes the terms as hints, so the model is not forced to output them. They are meant for proper nouns and abbreviations that context alone cannot resolve.

transcribeStyle takes verbatim or clean. Verbatim keeps every filler, false start and stutter for compliance, QA and legal review, and it is the default. Clean drops them for captions, notes and published transcripts. The model detects the language automatically unless you set locales, which the documentation describes as a very strong hint. Code switching covers conversations that move between languages within a sentence. Microsoft names Hinglish and Spanglish, mixes that VentureBeat notes can come up a dozen times in a single customer-service call.

When MAI-Transcribe-1 shipped, its Foundry catalog entry promised three upcoming capabilities: real-time transcription, diarization and context biasing. The last two have now shipped. The release is about batch throughput and long-form audio and does not cover real-time streaming, which voice agents and live captioning depend on. On Artificial Analysis’s separate streaming chart, Meta’s Muse Voice Transcribe led in September with a 3.1% word error rate.

Three benchmarks, three yardsticks

Each of Microsoft’s three performance claims uses a different measurement. The first is a number-one ranking on FLEURS across 60 languages, with an average word error rate of 5.2%. Google researchers published FLEURS in 2022. Native speakers read about 2,000 sentences in each of 102 languages, around 12 hours of speech per language, so a model’s Swahili and Swedish can be scored on the same content. Word error rate counts substitutions, insertions and deletions against a human reference, so 5.2% works out to about one word in 20 wrong.

FLEURS is read speech, though, and the average went up from the 3.7% Microsoft reported for MAI-Transcribe-1.5 across 43 languages in June. VentureBeat puts the rise down to the wider language set, which now includes low-resource languages where every model does worse. That reading also means one average could hide 3% on major languages and 12% on others, so the outlet suggests asking for the per-language breakdown.

The second claim is second place on the Artificial Analysis word-error-rate leaderboard. Artificial Analysis is an independent firm that tests models through their public APIs. Its test set mixes simulated agent conversations, European Parliament speeches and corporate earnings calls, and leans heavily toward English business speech. In June it ranked MAI-Transcribe-1.5 third at 2.4%, behind Alibaba’s Fun-Realtime-ASR-preview and ElevenLabs’ Scribe v2, and called it the fastest model in the top 10. Moving up to second would put Microsoft ahead of ElevenLabs. Microsoft does not claim to beat Alibaba on accuracy.

Microsoft also says the model sits on that firm’s accuracy-latency Pareto frontier. That means no rival is more accurate without also being slower, and none is faster without also being less accurate.

Neither figure measures speaker attribution. A transcript can have near-perfect word accuracy and still assign every other sentence to the wrong person, and Microsoft has published no diarization error rate.

Why speed sets the price

The third claim is raw speed. In Artificial Analysis evaluations, MAI-Transcribe-2 ran 10 times faster than OpenAI’s GPT-Transcribe, seven times faster than ElevenLabs’ Scribe v2 and five times faster than Google’s Gemini 3.5 Transcribe. In batch work nobody is waiting on the result, so speed shows up as cost. A model that processes audio at 300 times real-time needs a fraction of the GPU-hours of one running at 30 times. VentureBeat sees that efficiency as what lets Microsoft charge a dime and presumably still make money.

Mustafa Suleyman, Microsoft AI’s chief executive, made the same point in April, when he told The Verge the first model ran at “half the GPU cost of the other state-of-the-art models.” He said it was built by “a small, focused 10-person team” that had been “liberated from any of the bureaucracy,” with a larger group handling vendor management and data acquisition.

Microsoft has its own workloads that could move to the model: meeting audio from Teams, clinical documentation at Nuance, and the Azure speech services that thousands of enterprises already call. Bloomberg reported in July that Microsoft had started answering some Word and Excel prompts with MAI models, in products it had previously advertised as powered by OpenAI and Anthropic.

Teams already on the line need to plan a move. Microsoft Learn lists MAI-Transcribe-2, MAI-Transcribe-1.5 and MAI-Transcribe-1 as supported, but the $0.36 model from April has been marked deprecated since August 20.