Microsoft released MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, alongside two new text-to-speech systems. The transcription model supports 60 languages with continuous language detection, a feature aimed squarely at real-time use cases like live captioning and multilingual meetings where switching languages mid-conversation previously broke transcription accuracy.
Streaming transcription is a meaningfully harder engineering problem than batch transcription, since the model has to commit to words before it has heard the full sentence, with no opportunity to revise once spoken. Microsoft's push into this specific capability lines up with its broader Copilot strategy, where real-time meeting transcription is one of the more concretely useful, lower-hype features enterprise customers have actually adopted at scale.