What happened?
Meta's Superintelligence Labs have released their first real-time audio perception model: Muse Voice Transcribe. During live conversation it transcribes speech, tells speakers apart and detects sentence boundaries — all in one model, without separate systems.
The Spark-family model breaks incoming audio into 80-millisecond chunks. After each one it decides whether to keep listening or to output the next word as text.
A wait time per word
The trade-off is the classic one: waiting longer improves accuracy but lengthens the delay. Muse Voice Transcribe adjusts the wait separately for each word, based on its difficulty — easy words come out fast, harder ones get more listening time. Meta trained this behaviour with reinforcement learning, rewarding the model for low error rates and short delays at the same time.
Speaker separation and sentence detection sit inside the same model, trained jointly with the speech recognition. It marks speaker changes in the running text and tags each passage with an identifier from A to Z; it can tell more than 20 speakers apart at once and handle recordings over an hour long without post-processing.
What does independent measurement say?
An independent evaluation by Artificial Analysis confirms Meta's claims. In English it hits a 3.1 percent word error rate 0.16 seconds after a speaker stops:
| Model | Word error rate | Delay |
|---|---|---|
| Muse Voice Transcribe | 3.1% | 0.16 s |
| ElevenLabs Scribe v2 Realtime | 3.6% | 0.14 s |
| Cartesia Ink-2 | 3.4% – 4.0% | — |
| AssemblyAI Universal-3.5 Pro | 4.0% | — |
The model was trained on more than 70 languages, with 25 tested in depth. It also handles code-switching, where speakers jump between two languages mid-sentence.
Price is the weapon again
Meta undercuts the field: $0.18 per hour, or $3 per 1,000 audio minutes. Cartesia Ink-2 charges $4 for the same work, while ElevenLabs Scribe v2 Realtime and Deepgram Flux ask $6.50 each. This continues the pattern seen with Muse Spark 1.1 and 1.2: Meta competes on price rather than peak performance.
The company has not disclosed the model's parameter count, training data volume or audio data sources, and it is not releasing the weights.
The real issue: an assistant that never stops listening
Meta ties the release directly to Mark Zuckerberg's vision of "personal superintelligence." In a staged demo, employees argue that reliable speech recognition is the foundation for personal agents that listen in on real conversations through AI glasses.
That sentence poses the question by itself: a model that can separate 20 speakers, run for hours uninterrupted and cost $0.18 an hour is as much a surveillance substrate as a technical achievement. One of the speakers it separates may not have consented — not everyone in the room is the person wearing the glasses. In Germany a ban on Meta's camera glasses was recently discussed, though the Federal Network Agency decided not to pursue it.
Availability
Muse Voice Transcribe now powers voice dictation in Meta AI and Muse Code, and is available through the Meta Model API. Users can try it by holding the "Fn" key in any app.