Meta's new model tells 20 speakers apart while they talk at once.
Meta released Muse Voice Transcribe on 1 September, its first real-time audio model. One pass writes the words, names the speaker and marks when someone stops, 0.16 seconds after speech ends. Meta sells it only as an API at three dollars per thousand audio minutes, with no downloadable weights.
Why it mattersLive captions and meeting notes have needed three stitched models to do this job. Folding it into one makes always-on transcription cheap enough to leave running.
Read in the app — free, in 9 languages
Related stories
A breach at coding agent OpenCode exposed users' saved AI keys.2026-10-18AI bug hunting helped push the Postgres 19 database to November.
2026-10-17Perplexity gave away a sandbox that no AI escaped in its tests.
2026-10-16Amazon's new $75 wristband turns your conversations into to-do lists.
2026-10-15Apple says its AI is approved in China, with a test due this month.
2026-10-15