What it is
Meta Superintelligence Labs released Muse Voice Transcribe on 1 September, its first real-time audio perception model. It performs three tasks that speech pipelines usually assign to separate components: streaming automatic speech recognition, speaker diarisation for more than 20 voices, and endpointing — deciding when a speaker has actually finished.
The model, muse-voice-transcribe-1.0, is a single autoregressive multimodal system that processes audio in 80-millisecond chunks at 12.5 Hz, turning each into a soft token, according to MarkTechPost’s technical account. An adaptive delay mechanism, trained with reinforcement learning, trades accuracy against latency word by word.
It was trained on more than 70 languages, with 25 extensively verified at launch, and handles code-switching mid-sentence and sessions running over an hour.
The numbers
Meta says the model ranks first on Artificial Analysis for both streaming speech-to-text and diarisation. It reports a streaming word error rate of 3.1% at 0.16 seconds after speech ends, against 3.4% to 4.0% for the models it names as competitors, and an average diarisation error rate of 17.5% where rivals fall between 21.1% and 28.6%.
Artificial Analysis is an independent evaluator, which puts these figures in a different category from a vendor’s internal benchmark. The ranking is nonetheless a snapshot of a leaderboard that moves, and the competitor set is Meta’s selection.
Why one model instead of three
The engineering argument is the interesting part. A conventional pipeline runs recognition, then diarisation, then a separate endpoint detector, each adding latency and each able to disagree with the others about where a word or a speaker boundary sits. Folding them into one model removes the post-processing step and lets a single set of representations answer all three questions.
Meta is explicit about why it wants this: the stated design target is conversational hardware. A quote in the announcement frames the requirement as needing to “listen like a human on AI glasses in real conversations, not just voice commands”, which is a harder problem than dictation and a different one from meeting transcription.
Availability and the missing weights
The model is API-only. It is served through the Meta Model API with a zero data retention tier, and already powers dictation in Meta AI for Mac and in Muse Code. Meta says audio is processed to produce the transcript and not stored. MarkTechPost puts the price at $3.00 per 1,000 audio minutes, about $0.18 an hour; Meta’s own announcement does not publish pricing.
No open weights were released, which is notable for a company whose position on open models has been a defining part of its public argument. A real-time audio perception model is exactly the kind of artefact that would be most useful on-device and least controllable once released.
The market context is that transcription is being commoditised quickly from several directions at once, and the interesting comparison over the next quarter is not word error rate but whether any of these models end up in hardware people actually wear.
