Meta’s newest speech model commits to a transcription every 80 milliseconds, and the interesting part isn’t the speed. It’s what the model does with the pause.
Muse Voice Transcribe, the first real-time audio perception model out of Meta’s Superintelligence Labs, chops incoming audio into 80-millisecond chunks. After each one it makes a call: keep listening, or spit out the next word as text. That decision controls how much context it gathers before it commits.
Wait longer and you get better accuracy with longer delays. Meta says the model adjusts that wait time dynamically, word by word, based on how hard the word is. Easy words land fast. Harder ones buy themselves more listening time. Meta trained the behavior with reinforcement learning, rewarding low error rates and short delays at once, which is a genuinely awkward pair of objectives to optimize together.
The numbers hold up, but the lead is thin
Artificial Analysis confirms Meta’s claims in an independent evaluation. Muse Voice Transcribe hits a 3.1 percent word error rate on English in 0.16 seconds after a speaker stops talking.
ElevenLabs Scribe v2 Realtime lands at 3.6 percent and 0.14 seconds. So Meta is more accurate and slightly slower. AssemblyAI Universal-3.5 Pro Realtime comes in at 4.0 percent. Cartesia Ink-2 scores either 3.4 or 4.0 percent depending on whether the model detects utterance endings itself or leans on an external system, which tells you how much of this category depends on plumbing rather than the model.
Half a percentage point of word error rate is not a moat. OpenAI shipped GPT-Realtime-Whisper for the same job in May and cut transcription model prices in July.
One model doing three jobs
What separates this from a plain transcriber is that Meta folded the extra features into the same model instead of bolting on separate systems.
For speaker attribution, the model marks speaker changes inside the running text and tags each passage with an identifier from A to Z. For sentence boundaries, it marks where each utterance starts and ends. Both are trained jointly with the speech recognition.
The practical ceiling: more than 20 speakers told apart at once, and recordings over an hour with no post-processing. In a Meta demo with eight people in a room, the system assigns words to individual speakers in real time, according to the company.
Meta says it trained on more than 70 languages with 25 tested in depth. Code-switching is handled too, where a speaker jumps between two languages mid-sentence. Feeding it hints about language, keywords and context can push accuracy higher, which matters most for proper nouns like “Meta,” “Muse” or “Menlo Park.”
Cheaper than everyone, again
At $0.18 per hour, or $3 per 1,000 audio minutes, Meta undercuts the field. Cartesia Ink-2 runs $4. ElevenLabs Scribe v2 Realtime and Deepgram Flux both sit at $6.50, meaning Meta is charging less than half what the priciest rivals do for a model that beats them on word error rate.
That’s the same playbook as Muse Spark 1.1 and Muse Spark 1.2: compete on price, not on peak performance.
The disclosure is thin. Meta hasn’t published the parameter count, the training data volume or where the audio came from. And it isn’t releasing the weights, so nobody outside the company gets to check the work.
The glasses are the point
Meta ties the release to CEO Mark Zuckerberg’s vision of “personal superintelligence.” In a staged demo, Meta employees argue that reliable speech recognition is the foundation for personal AI agents that listen in on real conversations through AI glasses.
Which is where the speaker-attribution work stops being a technical footnote. A model that can separate 20-plus voices across an hour of audio isn’t built for dictation. It’s built for a device sitting on your face in a room full of people who didn’t agree to anything.
Germany recently discussed a ban on Meta’s camera glasses. The country’s Federal Network Agency decided not to pursue it.
Where you can actually use it
Muse Voice Transcribe now powers voice dictation in Meta AI and Muse Code, and it’s available through the Meta Model API. Hold the “Fn” key in any app and you’re talking to it.
Context on who built it: Meta reorganized its AI division in the summer of 2025 under the Superintelligence Labs umbrella, recruiting researchers from OpenAI, Google DeepMind and Apple with pay packages up to $300 million over four years. Not all of them stayed. Some went back to OpenAI after just a few weeks.
If you’re pricing a real-time transcription pipeline right now, the math is hard to argue with at $3 per 1,000 audio minutes against $6.50. Just go in knowing you’re buying a black box from a company that won’t tell you what audio it learned from.
