Microsoft AI announced on November 12 that MAI-Transcribe-2, its in-house speech-to-text model, is generally available in Microsoft Foundry for production use at $0.10 per hour of audio, a limited-time price through the end of 2026. The model entered public preview on September 3 and covers 60 languages in one model, with built-in speaker diarization, word-level timestamps and keyword biasing. Microsoft says it transcribes an hour of audio in about ten seconds of inference and averages 3.4% word error rate on the top 25 FLEURS languages.
Key Takeaways
- ✓From preview to production: public preview on September 3, general availability in Microsoft Foundry on November 12 (Microsoft Learn docs)
- ✓Price: $0.10 per hour of audio, which Microsoft labels a limited-time offer through December 31, 2026; check the official pricing page for later rates
- ✓One model, 60 languages: automatic language identification, code-switching such as Hinglish and Spanglish, and accuracy held in noisy audio
- ✓Built in rather than bolted on: speaker diarization, word-level timestamps, keyword biasing (phraseList) and clean or verbatim transcript styles
- ✓Speed and accuracy are Microsoft-reported: about 10 seconds of inference per hour of audio, 10x faster than GPT-Transcribe, 7x faster than Scribe v2 and 5x faster than Gemini 3.5 Transcribe; 5.2% average WER across all 60 FLEURS languages
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Meeting notes, call-center QA, clinical notes and voice agents all depend on speech-to-text, but developers have long traded off accuracy, latency and cost: the most accurate models were slow and billed per minute, and diarization or timestamps often meant stitching extra services together. Microsoft AI released MAI-Transcribe-2 in public preview on September 3, and on November 12 the official @MicrosoftAI account announced general availability in Microsoft Foundry for production use.
Architecture highlights
MAI-Transcribe-2 is built in-house by Microsoft AI and sits in the MAI family alongside MAI-Voice-2 and MAI-Image-2.6. Microsoft has not published its parameter count or architecture details. One model covers 60 languages with automatic language identification and handles mid-sentence code-switching. Speaker diarization, word-level timestamps and keyword biasing are built into the model, and output can be verbatim (keeping fillers) or clean.
Benchmarks and data
All figures are Microsoft-reported: 3.4% average word error rate on the top 25 FLEURS languages and 5.2% across all 60, which Microsoft says ranks first on FLEURS; it also says the model sits on the Artificial Analysis Pareto frontier for accuracy and latency. It transcribes an hour of audio in about ten seconds of inference, 10x faster than GPT-Transcribe, 7x faster than Scribe v2 and 5x faster than Gemini 3.5 Transcribe. No independent re-test was attached to the announcement.
Developer quick start
Create a Foundry Speech resource in Azure and call the Fast Transcription API (api-version 2025-10-15) with enhancedMode.enabled set to true and enhancedMode.model set to "MAI-Transcribe-2"; WAV, MP3 and FLAC are supported. Turn on diarization.enabled, set modelOptions.timestamps to "word", and pass product names or jargon through phraseList.phrases. The model can also be used for input transcription in the Voice Live API, and it is available in the MAI Playground and on OpenRouter. MAI-Transcribe-1 was deprecated on August 20, 2026, and the $0.10 rate is a limited-time price through year end.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.