MAI-Transcribe is a family of speech-to-text models built in-house by the Microsoft AI team: a model takes recorded audio and returns a written transcript. The current version, MAI-Transcribe-2, transcribes 60 languages and detects the spoken language automatically. Microsoft reports that it ranks first on FLEURS, a public multilingual speech benchmark, with an average word error rate of 5.2%.
A program sends an audio file (WAV, MP3 or FLAC) to the Azure Speech REST API and names the model in the request. Options control the output: speaker diarization splits the transcript by speaker, word-level timestamps give the start and length of every word, keyword biasing takes a list of domain terms such as product names, and the transcription style is either verbatim, with fillers such as "um", or clean. The model also handles code switching, speech that moves between two languages. A separate model, MAI-Transcribe-2-Streaming, transcribes live audio in real time.
The models are offered as a cloud API through Microsoft Foundry (Azure Speech), with a web playground for tests, and are billed per hour of audio.