Meta's Muse Voice Transcribe: real-time ASR at $0.18/hr, one model
One pass handles streaming speech-to-text, 20+ speaker diarization, and endpointing — 3.1% word error rate, 0.16s to final, 70+ languages, but no open weights.
Copy markdown
One model, three jobs
Muse Voice Transcribe folds streaming ASR, speaker diarization, and endpointing into a single real-time pass — Meta's first real-time audio perception model, from its Superintelligence Labs. No more gluing Whisper to a separate diarizer and a VAD.
3.1% WER, ranked #1
It tops Artificial Analysis' streaming leaderboard at a 3.1% word error rate, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 (3.6%), and GPT Live Transcribe (3.9%). Diarization error averages 17.5%, also first on public benchmarks.
80ms chunks, 0.16s to final
Audio streams in 80ms chunks with adaptive delay — hard words wait for context, easy words emit instantly — reaching 0.16s time-to-final. It live-tags 20+ speakers and spans 70+ languages (25 verified) with seamless code-switching.
$0.18/hour, but closed
Pricing is $3.00 per 1,000 audio minutes ($0.18/hr) through the Meta Model API, plus Meta AI for Mac and Muse Code. The catch: weights stay closed this round, so you can't self-host — every minute of audio runs through Meta.
What you could ship this weekend
Sub-second, speaker-labeled transcription makes live meeting notes that know who said what, real-time captions for streams, and voice agents that cut in cleanly on endpoints — all from one API call instead of a three-model pipeline.