Microsoft unveils mai-transcribe-2, its most advanced AI speech-to-text model

Microsoft announced on September 3, 2026, the release of MAI-Transcribe-2, claiming it to be its fastest, most accurate, and most cost-effective AI speech-to-text model. The model is priced at $0.10 per hour for the first six months of its launch, with the rate remaining in effect until the end of 2026. MAI-Transcribe-2 supports 60 languages and can handle conversations in mixed languages without requiring users to predefine the language. In the FLEURS test set, it achieved an average word error rate of 5.2%, outperforming competitors like Gemini 3.1 Pro, GPT-Transcribe, and Whisper v3-Large. The model also features speaker separation, word-level timestamps, and customizable transcription styles. Developers can access MAI-Transcribe-2 through Microsoft Foundry, MAI Playground, and OpenRouter.