MAI-Transcribe-2

Microsoft has been building its own family of MAI models, and MAI-Transcribe-2 is its second-generation answer for turning recorded speech into text. Its appeal is breadth: one model covers 60 languages, copes with recordings that switch between languages, and returns speaker labels and timestamps that developers would otherwise have to add separately.

• Automatic language detection, including for mixed-language audio
• Speaker diarization and word-level timestamps built into the output
• Keyword biasing to improve accuracy on product names, people and technical terms
• A clean output style that strips disfluencies for readable captions and transcripts
• Available inside existing Azure Speech workflows through the Fast Transcription API

Microsoft says the model is much faster than several rival transcription models in its own testing, so it is worth benchmarking on your own audio. It is best for developers and teams already on Azure who need multilingual transcription at scale. Since it is still a public preview, keep a fallback model configured for production workloads until it reaches general availability, and use keyword biasing for any names your audio mentions often.