Microsoft AI released MAI Transcribe-2 in 2026. This speech recognition model, according to the Microsoft AI lab, ranks as the best on FLEURS benchmark for 60 languages. It has an error rate average of just 5.2%. The following are some of the most common questions that arise when it comes to repairing your computer. its announcementMicrosoft describes the model as their most advanced transcription system and charges $0.10 for an hour of audio.
Microsoft stated that the MAI-Transcribe-2 model adds speaker diarization as well configurable transcription styles and timestamps at word level. The new version also performs better than competing models such as Gemini 3.5 Transcribe (GPT-Transcribe), Whisper V3 Large, and ScribeV2 on a wider range of audio. The model, according to Microsoft’s announcement, defines the Pareto-frontier for accuracy and latencies on Artificial Analysis. It also ranks second in the Artificial Analysis leaderboard on word error rates.
Microsoft has positioned its model as a solution for a variety of workloads, including legal documents, clinical notes, closed captioning, accessibility and accessibility. It reported that the company’s model had a 10x faster processing rate than its competitors, and a significantly lower latency. This was especially true for longer audio. The model also maintains transcription in noisy environments outside of controlled recording environment.
Benchmark Results
Microsoft cited evaluations by Artificial Analysis to claim that MAI-Transcribe-2 was 10 times faster, and more accurate, than OpenAI GPT-Transcribe. It is also 7 times faster, than ElevenLabs Scribe v2 or Gemini 3.5. The company said the model sits alone in the most attractive quadrant of the benchmark’s accuracy-versus-speed chart, at a 2.0% error rate and a speed factor of 403.6, meaning an hour of audio returns in about ten seconds.
Microsoft has reported on the FLEURS multilingual benchmark that MAI-Transcribe-2 consistently holds high accuracy standards across 60 languages. According to the company, this model is accurate for more languages and developers can use a single model when transcribing in multiple languages. This reduces complexity and could save GPU resources. Microsoft also stated that its model’s speed and efficiency allowed it to provide what they called the best price on the market. The $0.10 per hour rate will be available until the end the year.
Available Developer Features
It is important to note that the word “you” means “you”. Microsoft Learn documentation It lists MAI Transcribe-2 available as part of Azure Speech public preview. There is no service-level contract and it’s not recommended to be used for production workloads. Documentation describes MAI-Transcribe, a Microsoft AI-built speech-totext model, as covering tasks such as video transcription, meeting notes, clinical documentation, call center documentation and accessibility tools. The document also includes MAI-Transcribe-2 along with MAI-Transcribe-1 and MAI Transcribe 1.5, which were deprecated in 2026.
The enhanced mode of the Fast Transcription API is used to route requests. To select the model, set the enhanced mode property MAI-Transcribe-2. The audio input format is restricted to WAV files, MP3 or FLAC under 300MB. A Microsoft Foundry Resource for Speech and an Azure subscription are required.
The model can be controlled by optional parameters. Speaker diarization splits a recording into segments by speakers and provides metadata such as offset, duration and speaker labels. Segment option gives timing information per segment. None option does not provide timing. This parameter is used to bias recognition towards terms that are supplied, such as abbreviations or proper nouns.
For compliance, quality assurance, and analysis, the default transcription style is verbatim. This captures all speech, even filler words, false starts and other errors, exactly as it was spoken. Clean setting eliminates all fillers, auto-formats speech patterns, and produces more readable transcripts and captions. The model will detect the language automatically by default. Documentation advises that you only force a language if auto-detection does not work. The model handles code switching automatically for mixed language pairs, such as Hinglish or Spanglish. It also has noise-resistance for recordings outside of controlled environments.
Documentation also states that MAI-Transcribe is able to provide audio transcription input in Voice Live API via a field for session configuration. Microsoft says the demo model for the MAI Playground is now available through Microsoft Foundry. OpenRouter will be coming shortly.
