Microsoft has released MAI-Transcribe-2, a speech recognition model that ranks first on the FLEURS benchmark across 60 languages with an average word error rate of 5.2%, and the company is pricing it at just $0.10 per hour of audio. The new model outperforms competing systems including OpenAI's Whisper V3-Large, Gemini 3.5 Transcribe, and ElevenLabs' ScribeV2 across a broader range of real-world audio scenarios.

How Does MAI-Transcribe-2 Compare to Other Speech Recognition Models?

Microsoft's latest transcription system delivers significant speed advantages over its competitors. According to the company's benchmarking, MAI-Transcribe-2 is 10 times faster than OpenAI's GPT-Transcribe, 7 times faster than ElevenLabs' Scribe v2, and 5 times faster than Gemini 3.5 Transcribe while delivering higher accuracy. The model achieves a 2.0% error rate and a speed factor of 403.6, meaning an hour of audio returns in about ten seconds, positioning it alone in the most attractive quadrant of the accuracy-versus-speed benchmark chart.

The multilingual performance is particularly noteworthy. Microsoft reported that MAI-Transcribe-2 holds consistently high accuracy across all 60 tested languages on the public FLEURS benchmark, making it accurate across more languages than any other model currently available. This means developers transcribing across multiple languages can rely on a single model, reducing complexity and potentially saving computing resources.

What Features Does the New Model Include?

MAI-Transcribe-2 comes equipped with several advanced capabilities designed for professional use cases. The model adds speaker diarization, which segments a recording by speaker and returns speaker-labeled segments with timing metadata. It also provides configurable transcription styles and word-level timestamps, allowing users to customize output for different needs.

The transcription style parameter offers flexibility for different workflows. A verbatim setting captures speech exactly as spoken, including filler words and false starts, which is useful for compliance, quality assurance, and analysis workloads. A clean setting removes fillers and auto-formats common speech patterns to produce more readable captions, notes, and published transcripts. The model also handles code switching automatically, meaning it can process blended language pairs such as Hinglish and Spanglish without requiring special configuration.

How to Use MAI-Transcribe-2 for Your Transcription Needs

• Select Your Transcription Style: Choose between verbatim mode for exact speech capture or clean mode for polished, readable output depending on whether you need compliance documentation or publishable content.
• Configure Speaker Diarization: Enable speaker identification to automatically segment recordings by speaker and label each segment with timing information, useful for meeting notes and interview transcripts.
• Use Phrase Lists for Accuracy: Supply domain-specific terminology, abbreviations, and proper nouns as hints to bias recognition toward your industry's specialized vocabulary without forcing exact output.
• Let Language Detection Work Automatically: Allow the model to detect the spoken language automatically unless auto-detection fails, which simplifies setup for multilingual content.
• Leverage Noise Robustness: Deploy the model confidently for audio recorded outside controlled environments, as noise robustness is built into the system inherently.

Microsoft positioned MAI-Transcribe-2 for workloads including clinical note-taking, legal documentation, accessibility services, and closed captioning. The model maintains transcription quality in noisy conditions outside controlled recording environments, making it suitable for real-world deployment scenarios.

What Are the Practical Implications for Businesses and Developers?

The aggressive pricing and performance combination signals a significant shift in the speech-to-text market. At $0.10 per hour, the limited-time offer running through the end of 2026 represents what Microsoft called "the most competitive price in the market". For organizations processing large volumes of audio, this pricing could translate to substantial cost savings compared to alternatives.

The speed advantage carries practical implications beyond raw performance metrics. Faster inference means lower latency, particularly for long-form audio, which enables real-time or near-real-time transcription workflows. This matters for live meeting transcription, customer service call analysis, and accessibility applications where users expect minimal delay between speech and text output.

MAI-Transcribe-2 is available in Azure Speech in public preview through the Fast Transcription API's enhanced mode. Audio input is limited to files under 300 megabytes in WAV, MP3, or FLAC format, and use requires an Azure subscription and a Microsoft Foundry resource for Speech. The model is also available to demo through Microsoft Foundry, the MAI Playground, and OpenRouter, allowing developers to test the system before committing to production deployment.

The release represents Microsoft's continued investment in building internal AI capabilities rather than relying solely on third-party models. By developing MAI-Transcribe-2 in-house, the company can optimize the model for its Azure ecosystem and offer competitive pricing that pressures the broader speech recognition market.