Microsoft Launches World's Most Accurate Transcription Model MAI-Transcribe-1 with 3.9% Word Error Rate

Microsoft recently launched its self-developed AI transcription model MAI-Transcribe-1, claiming it as the world's most accurate transcription model. In the FLEURS benchmark test covering 25 major languages, the model has an average word error rate of only 3.9%, topping the list in 11 core languages, beating OpenAI's Whisper-large-v3 in 14 languages, and surpassing Google's Gemini 3.1 Flash in 11 of them.

The model also has speed and cost advantages: its batch transcription speed is 2.5 times that of the existing Microsoft Azure Fast service, costing only $0.36 per hour, which is the most cost-effective among major cloud vendors. It is currently available on the Microsoft Foundry platform, but does not support real-time transcription and other functions yet; Microsoft plans to add these capabilities in future updates.