MAI-Transcribe-2-Streaming: features, pricing, and who it is for

MAI-Transcribe-2-Streaming is Microsoft's real-time speech recognition variant in the Transcribe 2 family. Microsoft describes about 120 milliseconds to the first partial result, roughly 128 milliseconds end-to-end latency, and coverage of 60 languages. Its intended jobs include live captions, voice interfaces, and call analysis where waiting for a completed recording is unacceptable.

The version distinction matters. Microsoft's own comparison says the streaming variant does not include word-level timestamps, contextual keyword biasing, diarization, or clean-versus-verbatim style selection, while the non-streaming MAI-Transcribe-2 does. A team that needs speaker labels or specialized vocabulary may choose batch Transcribe 2 despite the delay, or add other pipeline stages. The published accuracy ranking is a vendor-cited leaderboard result; test representative noisy calls and accents before making service promises.