Microsoft launches MAI-Transcribe-2, claims fastest and most accurate speech-to-text model

Microsoft released MAI-Transcribe-2 on September 3, claiming it's the fastest, most accurate, and cheapest AI speech-to-text model. Pricing is $0.10 per hour for a limited time, with the discount running through the end of 2026. In FLEURS tests, it achieved an average word error rate of 5.2%, outperforming Gemini 3.1 Pro, GPT-Transcribe, and Whisper v3-Large. It's 10x faster than GPT-Transcribe, 7x faster than Scribe v2, and 5x faster than Gemini 3.5 Transcribe. New features include speaker separation, word-level timestamps, and configurable transcription styles (verbatim and clean). It supports 60 languages with automatic language detection. Developers can try it now on Microsoft Foundry, MAI Playground, and OpenRouter.