Meta Launches Muse Voice Transcribe for Streaming Audio
Meta’s first real-time audio perception model processes speech in 80-millisecond chunks, supports more than 70 languages and adapts its delay to balance speed and accuracy.
Streaming rather than waiting
Meta introduced Muse Voice Transcribe, which it describes as the first real-time audio perception model developed by Meta Superintelligence Labs. The autoregressive multimodal model belongs to the Muse Spark family and is designed to produce streaming transcription while speech is still arriving, rather than waiting for a complete recording.
The model processes audio in 80-millisecond chunks, equivalent to 12.5 input steps per second. At each step it can either emit text or wait for more acoustic evidence. That adaptive-delay mechanism lets it postpone ambiguous words while returning clearer passages quickly, giving developers a tunable balance between transcription accuracy and the time required to finalize each phrase.
Meta says the model reaches the Pareto frontier for the speed–accuracy trade-off when measured by time to final transcription. The claim means that, within Meta’s comparison, competing systems did not improve one dimension without sacrificing the other. Meta says the model was trained on more than 70 languages, with 25 extensively verified at launch, and supports audio longer than an hour, more than 20 speakers and code-switching. Independent deployment evidence across accents, noisy environments and overlapping speech nevertheless remains limited.
Muse Voice Transcribe is available in public preview through the Meta Model API and also powers dictation in Meta AI for Mac and Muse Code. Meta did not announce downloadable weights, so this is a hosted release rather than an open-weight one. The developer API launched at $3 per 1,000 audio minutes, equivalent to $0.18 per hour.
Why it matters
Real-time transcription increasingly serves as infrastructure for voice agents, live assistants and accessibility tools. In those products, the final-word delay can matter as much as headline accuracy: hesitation makes an agent feel unresponsive, while premature text can trigger the wrong action. Meta’s per-chunk decision mechanism directly targets that tension. Its commercial importance will depend on how reliably the reported multilingual and multi-speaker capabilities hold up in real deployments.