Microsoft MAI-Transcribe-2: 10x Faster Speech at $0.10/Hour

Microsoft AI released MAI-Transcribe-2 on September 3 at $0.10 per audio hour, claiming it is 72% cheaper than OpenAI's GPT-Transcribe and 10x faster in non-streaming tests, with a 5.2% average word error rate on the FLEURS benchmark across 60 languages. The model, available via Azure's Fast Transcription API and OpenRouter, includes features like speaker diarization and word-level timestamps at no extra cost, but pricing after 2027 remains unannounced.

A 10-person team inside Microsoft AI just put OpenAI, Google, and ElevenLabs on notice. MAI-Transcribe-2 landed September 3 at $0.10 per audio hour — 72% cheaper than OpenAI’s GPT-Transcribe, 10x faster in non-streaming tests, and ranked first on the FLEURS multilingual benchmark across 60 languages. If you’re building voice features or transcription pipelines, this is the model worth evaluating before 2026 ends. The Benchmarks Hold Up MAI-Transcribe-2 posts a 5.2% average word error rate on FLEURS across 60 languages — 3.4% on the top 25. In non-streaming mode, WER drops to 2.0% at a speed factor of 410.7x real time. Independent coverage from Unite.AI https://www.unite.ai/mai-transcribe-2-tops-fleurs-benchmark-across-60-languages-microsoft-says/ and VentureBeat https://venturebeat.com/infrastructure/microsoft-ais-mai-transcribe-2-undercuts-openai-google-and-elevenlabs-on-price-and-speed confirms the rankings are real, not marketing. Speed comparisons from Microsoft’s benchmarks: - 10x faster than OpenAI GPT-Transcribe - 7.5x faster than ElevenLabs Scribe v2 - 4.6x faster than Google Gemini 3.5 Transcribe A note of honesty: composite marks are elite, but performance varies across individual languages. Chinese comes in at 4.5% WER, French at 2.8%. If your app targets a specific language, test it before migrating wholesale. Pricing: The Real Story Has a Catch At $0.10 per audio hour, MAI-Transcribe-2 sets a new price floor for production-grade speech recognition. Here’s how it compares to current alternatives: | Provider | Price/hour | Speed vs real-time | |---|---|---| | MAI-Transcribe-2 | $0.10 | 410x | | ElevenLabs Scribe v2 | $0.22 | ~55x | | OpenAI GPT-Transcribe | $0.36 | ~41x | | Gemini 3.5 Transcribe | ~$0.36 | ~89x | The caveat matters. Microsoft hasn’t announced what pricing looks like in 2027. At $0.10 this is a no-brainer to evaluate. However, at some unknown future price, the calculation changes — so the urgency is real: lock in your workflows now, pressure-test at scale, and have a fallback plan ready. What You Get Including Things That Cost Extra Elsewhere MAI-Transcribe-2 ships with a feature set that competitors typically charge for separately: - Speaker diarization — identifies and labels multiple speakers with offset and duration metadata - Word-level timestamps — per-word timing for alignment, search, and editing workflows - Transcript styles — verbatim fillers, false starts preserved or clean disfluencies removed - Keyword biasing — improves accuracy on domain-specific vocabulary - Automatic language ID — no need to specify the input language upfront - Code switching — handles mixed-language audio such as Hinglish and Spanglish Diarization runs $0.07 per hour extra on ElevenLabs. Here it’s included. That changes the cost calculus substantially for podcast transcription, meeting summaries, or call analysis use cases. Integrating MAI-Transcribe-2 in One Flag If you’re already on Azure, there’s no migration overhead. Enable MAI-Transcribe-2 through the Fast Transcription API’s enhancedMode property: curl --location 'https://YourResourceName.cognitiveservices.azure.com/speechtotext/transcriptions:transcribe?api-version=2025-10-15' --header 'Content-Type: multipart/form-data' --header 'Ocp-Apim-Subscription-Key: <YourSpeechResourceKey ' --form 'audio=@"YourAudioFile.wav"' --form 'definition={ "enhancedMode": { "enabled": true, "model": "MAI-Transcribe-2" } }' Add "diarization": {"enabled": true} to the definition object to turn on speaker separation. The full parameter reference lives in the Microsoft Learn documentation https://learn.microsoft.com/en-us/azure/ai-services/speech-service/mai-transcribe . Not on Azure? MAI-Transcribe-2 is also available via OpenRouter https://openrouter.ai/microsoft/mai-transcribe-2 without an Azure subscription. Microsoft AI Is Playing Its Own Game Now The 10-person team detail isn’t incidental. Microsoft AI — the division Mustafa Suleyman runs, distinct from the OpenAI partnership — shipped three speech models in five months: MAI-Transcribe-1 in April with 25 languages, MAI-Transcribe-1.5 in June with 43 languages, and now MAI-Transcribe-2 with 60. Each iteration meaningfully better and cheaper than the last. The implication is uncomfortable for the industry: Microsoft’s own team is now outperforming OpenAI’s flagship transcription product on Microsoft’s infrastructure. This follows a similar pattern with Microsoft AI’s growing independent model stack https://byteiota.com/microsoft-project-zenith-developer-pcs/ . MAI-Transcribe-2 isn’t a quiet internal tool — it’s a public model with an official Microsoft AI announcement https://microsoft.ai/news/mai-transcribe-2-is-the-fastest-most-accurate-and-cheapest-speech-recognition-model-in-the-world/ positioned explicitly against the competition. For developers, the takeaway is straightforward: evaluate MAI-Transcribe-2 now. Run your audio through it. The promotional pricing window is finite, and the feature set is already there.