Microsoft announced the release of MAI-Transcribe-2, claiming it is the fastest, most accurate, and most cost-effective AI speech-to-text model. In terms of pricing, the model is offered at a limited-time rate of $0.10 (approximately 0.67 RMB) per hour, with the promotion lasting until the end of 2026.
Leading in Accuracy and Speed, 10 Times Faster than GPT-Transcribe
In terms of performance, in the FLEURS test set, MAI-Transcribe-2 achieved an average word error rate of 5.2% in both forced language mode and automatic language detection mode. In the same test, Gemini 3.1 Pro achieved 5.3% and 5.8%, GPT-Transcribe achieved 10.4% and 10.6%, and Whisper v3-Large achieved 22.8% and 23.5%.
In terms of speed, Microsoft cited an Artificial Analysis evaluation stating that MAI-Transcribe-2 is 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2, and 5 times faster than Gemini 3.5 Transcribe.
Supports Speaker Separation and Word-Level Timestamps, Covering 60 Languages
In terms of features, MAI-Transcribe-2 adds speaker separation capabilities, which can distinguish different speakers within a recording and assign the text to the corresponding speaker; word-level timestamps allow precise annotation for each word, facilitating audio search, navigation, editing, and subtitle alignment.
In terms of style, the model offers configurable transcription styles: the verbatim mode preserves interjections and speech errors, suitable for compliance and analytical work; the clean mode removes interjections, generating more readable subtitles, notes, and published texts. Additionally, it supports keyword bias, automatic language recognition, language switching, and transcription in noisy environments. The model covers 60 languages and supports natural mixed-language dialogues, allowing users to not pre-specify the language, as the model can automatically detect the language of the recording.
Currently, developers can call or try the model on Microsoft Foundry, MAI Playground, and OpenRouter.
