Published: 2 September 2026
Meta Superintelligence Labs released Muse Voice Transcribe on 1 September 2026. It transcribes speech, tells speakers apart and works out when someone has stopped talking, all in one model instead of three. It costs $3.00 per 1,000 audio minutes, and there are no open weights.
What can Muse Voice Transcribe do?
Most voice products today stitch three systems together. One writes the text, a second works out who spoke, and a third decides when the speaker finished. Every hand-off adds delay and one more thing that can break. Muse Voice Transcribe does all three jobs in a single pass, with no clean-up step afterwards.
- Speakers: labels more than twenty voices in the same recording.
- Audio length: handles input longer than an hour without splitting it up.
- Languages: trained on more than seventy, with twenty-five verified at launch.
- Code-switching: handles a speaker changing language inside one sentence.
- Availability: the Meta Model API as muse-voice-transcribe-1.0, with a zero data retention tier.
- Already live in Meta AI dictation on Mac and in Muse Code.
How the model decides when to listen and when to write
Audio arrives in short chunks. After each one the model makes a single choice: keep listening, or write the next word. Because it controls that choice, it also controls how much audio sits behind each word. More audio means a better guess but a slower answer.
Meta did not fix that trade-off by hand. It trained the model with reinforcement learning, rewarding both accuracy and speed together, so the model waits longer only on words it finds hard. Speaker labels and end-of-turn markers are just extra tokens in the same stream, which is why no second model is needed.
Muse Voice Transcribe benchmarks and price
The interesting comparison is not accuracy alone, but accuracy against speed and price at the same time.
| What is measured | Muse Voice Transcribe | Cartesia Ink-2 |
|---|---|---|
| Word error rate on the final transcript (lower is better) | 3.1% | 3.4% |
| Delay after the speaker stops | 0.16 seconds | 0.43 seconds |
| Price per 1,000 audio minutes | $3.00 | $4.00 |
| Average speaker labelling error, three public sets | 17.5% | rival systems ran 21.1% to 28.6% |
| Speaker labelling inside the same model | yes | no |
| Weights you can download and self-host | no | no |
The speech figures come from Artificial Analysis, the speaker labelling from Meta, and the prices from each vendor’s own pricing page. ElevenLabs Scribe v2 Realtime answers slightly faster, in 0.14 seconds, but is less accurate and costs $6.50 for the same minutes. Meta scores its own diarization, so treat that half of the table with care.
What Meta is not saying
There are no weights. Meta has released open models this year, including Muse Glimmer, so the decision to hold these back is a choice rather than a habit. If your reason for wanting a single-model stack is that recordings cannot leave your building, this release does not help you.
Meta also names twenty-five verified languages out of more than seventy trained, without saying which twenty-five, or how the other forty-five behave. Nor does it publish a word error rate per language. For a European team, that is the number that decides whether this works, and it is missing.
What this means
Muse Voice Transcribe is worth testing now if you run meeting notes, call analytics or live captions and currently pay three vendors to do it. The price is roughly half what ElevenLabs charges, and collapsing three systems into one removes two places where your pipeline can drift out of sync. A small product team shipping a voice feature will feel that simplification more than the accuracy gain.
Safe to ignore if you are regulated, on-premise, or already committed to open weights. There is no self-hosted path and no sign of one. The accuracy lead over Cartesia is three tenths of a percentage point, which alone does not justify a migration. The price and the simpler stack are the real arguments, so test on your own audio first.
For more information, visit the official announcement of Muse Voice Transcribe on the Meta AI Research blog.
