Meta released Muse Voice Transcribe, a speech recognition model that works as audio streams in. It labels which person is speaking and detects when a speaker has finished. Meta says it handles more than 20 speakers and audio over an hour long, without extra processing. Training covered over 70 languages, and Meta recommends 25 that it has verified. The model switches between languages inside a single sentence. Accuracy can be tuned with language, keyword and context hints. It processes audio in 80-millisecond chunks and decides per word how long to wait before transcribing. Meta trained that timing choice with reinforcement learning to balance speed against errors. Meta claims first place on Artificial Analysis streaming speech-to-text and on public speaker-labelling benchmarks, as of 1 September 2026. A browser demo transcribes from a microphone and does not store the audio.
What changed
Meta Superintelligence Labs had not released a real-time audio perception model.
What it unlocks
Live transcription that labels who is speaking across long multi-speaker recordings.
- 3.1% streaming word error rate
- 17.5% average diarization error rate
- 25 validated languages, 70+ trained
- 20+ speakers, 1+ hour audio
What you need to act on it
- access via Meta's developer speech-to-text docs
Sources