Meta’s latest speech system decides whether to transcribe once every 80 milliseconds. The speed is not really the story, though. What matters is how the model handles the waiting.
Muse Voice Transcribe is the first real-time audio perception model to come out of Meta’s Superintelligence Labs, and it slices incoming sound into 80-millisecond segments. At the end of each segment it faces a choice: hold on and keep listening, or push out the next word as text. That single decision governs how much context the model accumulates before committing.
Hold out longer and accuracy improves, but so does latency. According to Meta, the model tunes that waiting period on the fly, word by word, depending on the difficulty of the word in question. Straightforward words are emitted quickly; trickier ones earn themselves extra listening time. The behavior came out of reinforcement learning, with the reward function pushing for low error rates and short delays simultaneously — a genuinely uncomfortable pair of goals to chase at the same time.
The benchmarks check out — but the margin is slim
An independent evaluation from Artificial Analysis backs up what Meta says. Muse Voice Transcribe posts a 3.1 percent word error rate on English, delivered 0.16 seconds after the speaker falls silent.
For comparison, ElevenLabs Scribe v2 Realtime records 3.6 percent at 0.14 seconds — meaning Meta wins on accuracy and gives up a sliver of speed. AssemblyAI Universal-3.5 Pro Realtime registers 4.0 percent. Cartesia Ink-2 lands at either 3.4 or 4.0 percent, depending on whether the model spots utterance endings on its own or relies on an outside system, an illustration of how much of this field comes down to plumbing rather than the model itself.
Half a percentage point on word error rate hardly constitutes a moat. OpenAI released GPT-Realtime-Whisper for the same purpose back in May, then reduced prices on its transcription models in July.
Three jobs, one model
The thing that sets this apart from a straightforward transcriber is that Meta baked the additional capabilities into a single model rather than attaching separate systems around it.
On the speaker attribution side, the model flags speaker changes within the running text and labels each stretch with an identifier ranging from A to Z. It also marks the beginning and end of every utterance for sentence boundaries. Both capabilities are trained jointly alongside the speech recognition itself.
As for practical limits: the system can distinguish more than 20 speakers simultaneously and handle recordings running over an hour without any post-processing. Meta showed a demo featuring eight people in one room where, the company says, the system attributes words to individual speakers as they talk.
Training covered more than 70 languages, per Meta, with 25 of those tested thoroughly. The model also copes with code-switching, when someone shifts between two languages partway through a sentence. Supplying hints about language, keywords and context can lift accuracy further — something that counts most for proper nouns such as “Meta,” “Muse” or “Menlo Park.”
Undercutting the field, once more
Priced at $0.18 per hour — $3 per 1,000 audio minutes — Meta comes in below everyone else. Cartesia Ink-2 costs $4. Both ElevenLabs Scribe v2 Realtime and Deepgram Flux are priced at $6.50, which puts Meta at under half the cost of the most expensive competitors while outperforming them on word error rate.
It’s the identical strategy behind Muse Spark 1.1 and Muse Spark 1.2: win on price rather than on peak capability.
What Meta has disclosed is minimal. There’s no published parameter count, no figure for training data volume, no word on where the audio originated. The weights aren’t being released either, so no one outside the company can verify any of it.
This was always about the glasses
Meta frames the launch around CEO Mark Zuckerberg’s concept of “personal superintelligence.” A staged demo has Meta employees making the case that dependable speech recognition underpins personal AI agents capable of listening to real conversations via AI glasses.
That’s precisely the point at which the speaker-attribution work stops looking like a technical detail. A model built to pull apart 20-plus voices over an hour of recording was not designed with dictation in mind. It was designed for hardware perched on your face in a room full of people who never consented to any of it.
A ban on Meta’s camera glasses was discussed in Germany not long ago. The country’s Federal Network Agency ultimately opted against pursuing one.
Where it’s live today
Voice dictation in both Meta AI and Muse Code now runs on Muse Voice Transcribe, and developers can reach it through the Meta Model API. Press and hold the “Fn” key in any application and you’re speaking to it.
Some background on the team behind it: Meta restructured its AI division in the summer of 2025 beneath the Superintelligence Labs banner, poaching researchers from OpenAI, Google DeepMind and Apple with compensation reaching $300 million across four years. Retention was imperfect — a few returned to OpenAI within weeks.
Anyone costing out a real-time transcription pipeline today will find $3 per 1,000 audio minutes tough to argue against when the alternative is $6.50. Just understand that what you’re purchasing is a black box from a company that won’t disclose the audio it was trained on.