cohere launched transcribe on march 26 as its first dedicated speech recognition model. the release combines an open-source 2b asr model with hosted access through cohere's audio transcriptions api and model vault.
the practical angle is straightforward: a smaller team can test it cheaply, self-host it if needed, or move into managed deployment without changing product direction. that makes it more interesting than a pure research checkpoint.
cohere also framed the launch around strong open-asr leaderboard performance and a manageable inference footprint. if you need transcription for meetings, support, analytics, or game production notes, this is a serious baseline to compare against whisper, deepgram, and speechmatics.