Microsoft AI released MAI-Transcribe-2, an in-house speech recognition model, on September 3, 2026. It takes the top spot on the multilingual FLEURS benchmark with an average word error rate of 5.2 percent across 60 languages, and it is priced at 0.10 USD (about 16 yen) per hour of audio. Transcription has long forced a trade-off between accuracy, speed, and cost. This release goes after all three at once.
※1 USD = 156 JPY (as of September 4, 2026)
The 10-cent hour
The price is the first thing that stands out. Transcribing an hour of audio costs 0.10 USD, or under 20 yen. Run ten hour-long meeting recordings a day and the monthly bill still lands in the range of a few hundred yen. The catch is that this rate is a limited-time offer running through the end of the year, and Microsoft has not said what the standard price will be afterward. Anyone evaluating the model should plan to redo the math once regular pricing is published.
Microsoft attributes the pricing to throughput. If an hour of audio comes back in roughly ten seconds, a single GPU clears far more jobs per hour, and the cost per job falls accordingly.
An hour of audio in ten seconds
The speed claims rest on measurements from Artificial Analysis, an independent evaluation outfit. By those numbers, MAI-Transcribe-2 runs 10 times faster than OpenAI's GPT-Transcribe, 7 times faster than ElevenLabs' Scribe v2, and 5 times faster than Google's Gemini 3.5 Transcribe, while still scoring higher on accuracy.
The concrete figures are a 2.0 percent error rate and a speed factor of 403.6. A speed factor of 403.6 means processing runs about 400 times faster than real time, so an hour of audio returns in roughly ten seconds. On a chart plotting accuracy against speed, Microsoft says MAI-Transcribe-2 sits alone in the most favorable quadrant. On the Artificial Analysis word error rate leaderboard taken by itself, the model ranks second.
Accuracy that holds across 60 languages
The multilingual measure is FLEURS, a public benchmark. Microsoft reports that MAI-Transcribe-2 maintains high accuracy across all 60 languages tested, finishing first with an average word error rate of 5.2 percent. The company says no other model stays accurate across a language range this wide.
For developers who currently juggle a different model per language, that matters in practice. Multilingual audio pipelines usually detect the language first, then route to a matching model, and every branch adds implementation and operational weight. A single model removes the routing entirely. It also removes the need to reserve GPU capacity per language, which cuts idle time.
Specifying a language is optional; by default the model identifies the spoken language on its own. Code switching, where two languages mix inside one utterance as in Hinglish or Spanglish, is handled automatically. Microsoft's documentation advises pinning a specific language only when automatic detection fails.
Controls built for real workloads
The feature set is aimed squarely at business use. Turn on speaker diarization and the recording comes back segmented by speaker, with offset and duration metadata attached. Timestamps can be returned per word, per segment, or omitted.
There is also a phrase list. Supply domain terminology, abbreviations, or proper nouns in advance and recognition leans toward those terms. The documentation notes these act as hints rather than forced output.
Transcription style is configurable as well. The default is verbatim, which captures speech exactly as spoken, fillers and false starts included. That setting targets compliance review, quality assurance, and conversation analysis, where what was actually said is the point. The clean setting strips fillers and auto-formats the text into something readable, which suits captions, meeting notes, and published transcripts. Listed use cases include video captioning, meetings, clinical notes, call center documentation, accessibility, content creation, and voice agents.
Noise robustness is described as inherent to the model, so quality holds up on audio recorded outside controlled environments.
How to use it, and the current limits
The model ships in Azure Speech as a public preview. There is no service-level agreement, and Microsoft explicitly says it is not recommended for production workloads. Attractive pricing and performance aside, this is not yet the stage to wire it into a process that cannot go down.
Calls route through the Fast Transcription API's enhanced mode, with the model selected by setting the enhanced mode model property. Audio input must be under 300 MB in WAV, MP3, or FLAC format. Use requires an Azure subscription and a Microsoft Foundry resource for Speech.
enhanced mode model = MAI-Transcribe-2
The Voice Live API can also use MAI-Transcribe for input audio transcription through a session configuration field. For hands-on testing, demos are available through Microsoft Foundry, the MAI Playground, and OpenRouter.
Version management is moving along too. MAI-Transcribe-2 is offered alongside MAI-Transcribe-1.5, while the original MAI-Transcribe-1 was deprecated on August 20, 2026. Anyone running pipelines on the older version should check migration deadlines.
Summary
MAI-Transcribe-2 comes down to three points: accuracy that holds across 60 languages, an hour of audio returned in about ten seconds, and a 0.10 USD hourly rate that lasts through the end of the year. Collapsing a per-language model lineup into one model directly reduces implementation work for teams running multilingual services. On the other hand, it is a preview with no service-level agreement. The practical order is to measure accuracy and speed against your own audio in a test environment first, then decide on production once the standard price is known.
