Muse Voice Transcribe Explained: Meta's Real-Time Speech Model
Meta's real-time audio model handles transcription and speaker labels in a single pass.
Meta announced Muse Voice Transcribe on September 1, 2026. It transcribes live audio, labels who is speaking, and detects when a speaker has finished. Most stacks need two or three models to do all that.
The model runs on the Meta Model application programming interface (API) at $3.00 per 1,000 audio minutes, which is $0.18 per hour. Meta lists it as muse-voice-transcribe-1.0. The endpoint is compatible with the OpenAI software development kit (SDK), so existing transcription code usually needs only a new base URL and model name.
Meta reports a 3.1% word error rate (WER) on the Artificial Analysis streaming speech-to-text benchmark, as reported on September 1, 2026. The final transcript arrives about 0.16 seconds after a speaker stops. The same model handles 20 or more speakers and audio running past an hour.
What Muse Voice Transcribe Does
Muse Voice Transcribe converts speech to text as the audio arrives, and it attaches a speaker label to each stretch of that text while the call is still running. Most transcription stacks split those jobs across two models. One writes the words. A second listens to the finished recording and works out who said what.
Doing both in one pass removes the wait at the end. A meeting tool built on a two-model stack cannot show a correctly attributed transcript until the recording stops, because the speaker-labelling model needs the whole file. Muse Voice Transcribe attributes speech inside the stream, so a live view can name the speaker on screen while the person is still talking.
The model also decides when someone has finished a turn, which Meta calls end-of-speech detection. A voice agent needs that signal to know when to reply. Without it you either add a separate turn-detection model or guess from silence. Silence-guessing cuts off callers who paused to think.
Meta places the model in the Muse Spark family and says it processes audio in 80 millisecond chunks. Chunks that small are what let text appear a fraction of a second behind the speaker rather than at the end of a sentence. It also supports contextual and keyword biasing, which means you hand it a list of terms to expect. Feed it your product names, drug names, or case numbers, and it stops turning them into the nearest common English word.
- Streaming transcription: text appears while the audio is still arriving, so a live screen can show it.
- Live speaker attribution for 20 or more speakers, so a running transcript can name who is talking.
- End-of-speech detection built in, so a voice agent knows when to answer without a separate turn model.
- Keyword and contextual biasing, so names and jargon come through correctly without fine-tuning.
- Trained on more than 70 languages with 25 validated at launch, including code-switching inside a sentence.
Weighing Muse Voice Transcribe against another transcription API? Book a workflow audit and we will test both on your own call recordings.
Book a ConsultationWhat Meta Publishes on Accuracy and Latency
Meta reports a 3.1% word error rate for Muse Voice Transcribe on the Artificial Analysis streaming leaderboard, as reported on September 1, 2026. Word error rate counts the words a model gets wrong as a share of the words spoken. A 3.1% rate means roughly three mistakes in every hundred words.
The second number matters more for live use. Meta puts the final transcript about 0.16 seconds behind the end of speech. A caller hears that gap as the pause before an agent answers. A sixth of a second is shorter than the pause most people leave between turns.
For speaker labelling, Meta reports a 17.5% average diarization error rate across the AMI IHM and SDM benchmarks and VoxConverse. Diarization error rate measures how much audio time gets attributed to the wrong speaker. At 17.5%, about one minute in six of a hard multi-speaker recording carries a wrong name. Treat live labels as a strong hint rather than a record.
These are Meta's own published figures on public benchmarks. Benchmark audio is cleaner than a construction-site phone call, and your accuracy will depend on microphones, accents, and background noise. Run your own recordings through the model before you plan around either figure.
What Muse Voice Transcribe Costs
Meta publishes one rate of $3.00 per 1,000 audio minutes, or $0.18 per hour, on the Muse Voice Transcribe model page. Speaker labelling adds nothing, because it is not a separate model.
That flat structure is easy to plan against. A company running 500 hours of recorded calls a month pays $90. A single daily standup for a year, at half an hour a day, costs about $33.
The comparison worth making is against a stack that bills diarization separately. AssemblyAI publishes streaming transcription on Universal-3.5 Pro Realtime at $0.45 per hour and charges $0.12 per hour on top for streaming diarization, so $0.57 per hour for the same two jobs. Deepgram lists Nova-3 streaming at $0.0077 per minute at its standard rate, about $0.46 per hour, with a promotional rate of $0.0048 per minute in effect at the time of writing.
All three rates move, and promotional pricing expires. Confirm them on the Meta model page, the Deepgram pricing page, and the AssemblyAI pricing page before you build a budget.
Where You Can Use Muse Voice Transcribe Today
Muse Voice Transcribe is reachable three ways, and they suit different people. The three are listed at the end of this section.
The API path is the one that matters for a product. Because Meta made the endpoint compatible with the OpenAI SDK, a team already calling a transcription API can point the same client at Meta's base URL and change the model name. That keeps a switching test cheap, which is the main reason to run one.
The Mac dictation path is worth knowing about even if you never write a line of code. It puts the same recognition model behind system dictation in any app, and it costs nothing at that layer. Meta's developer site carries the current access details.
How Muse Voice Transcribe Compares to the Streaming Field
Benchmarks do not settle this. The decision is whether bundled diarization is worth switching for. The published rates for streaming transcription with speaker labels line up like this, using each company's own figures.
| Capability | Muse Voice Transcribe | Deepgram, AssemblyAI and Whisper |
|---|---|---|
| Streaming rate | $0.18/hour ($3.00 per 1,000 minutes) | AssemblyAI $0.45/hour; Deepgram Nova-3 about $0.46/hour at its standard rate |
| Speaker labels | Included in the same model, live | AssemblyAI adds $0.12/hour for streaming diarization; Deepgram supports it on streaming and includes it on pre-recorded audio |
| End-of-speech detection | Built into the recognition model | Usually a separate turn-detection step in the stack |
| Published streaming accuracy | 3.1% word error rate (Artificial Analysis, September 1, 2026) | Deepgram and AssemblyAI publish their own figures on their own test sets |
| Self-hosting | Not offered; hosted API only | Whisper can be self-hosted on your own hardware |
- Choose Muse Voice Transcribe when you need live speaker labels and want one bill instead of two line items.
- Choose Deepgram when you need per-request control over endpointing and its promotional rate holds.
- Choose AssemblyAI when you want summaries and topic detection from the same audio and can absorb the diarization add-on.
- Choose self-hosted Whisper when audio cannot leave your own infrastructure and you accept running the hardware.
Who Should Not Use Muse Voice Transcribe
Three groups should look elsewhere, and they are listed below.
Speaker labels need a human check before they carry weight. At a 17.5% average diarization error rate on benchmark audio, a label is good enough to route a support ticket. It is not good enough to attribute a quote in a deposition without review.
Three things would change that answer: a self-hosted or on-premise option from Meta, a validated language list covering all 70-plus, or a competitor bundling live diarization at a lower all-in rate. Any of the three moves the answer for a large group of readers.
- Teams that cannot send audio to a third-party API at all, for data-residency or contractual reasons. Self-hosted Whisper is the usual answer, because its weights are open.
- Teams working in one of the languages Meta trained on but has not validated. Only 25 of the 70-plus carry a validation claim, so test yours heavily first.
- Teams transcribing finished files with no live element. If nothing reads the transcript while the audio plays, the streaming advantage buys you nothing and a batch rate elsewhere may be cheaper.
What This Changes for Teams Running Transcription
The bundled architecture matters more than the price. At Layer3Labs, we build and run automation inside other people's businesses, and transcription is usually one link in a longer chain rather than the product. In the client intake workflows we have automated for law firms, the recurring failure was never raw accuracy. Speaker attribution arrived only after the call ended, so anything the workflow wanted to do live had to wait for a file.
A model that attributes speech in the stream removes that wait. It does not remove the review step. Plan for a person to correct labels on anything carrying legal or clinical weight, and treat live labels as a way to route and prompt rather than as a record.
One practical note for anyone testing it. Because the endpoint speaks the OpenAI SDK, a bake-off costs an afternoon. If you already pay for a transcription API, run the same hour of your own worst audio through it and through Muse Voice Transcribe. Read both transcripts by hand and count the errors that would have reached your team. If you want help scoping that test against your own call recordings, book a workflow audit and we will run it with you.
Frequently Asked Questions
- Muse Voice Transcribe is a real-time speech-to-text model Meta announced on September 1, 2026. It transcribes live audio, labels which speaker is talking, and detects when a speaker has finished, all in a single model rather than a chain of separate models.
- Meta publishes a rate of $3.00 per 1,000 audio minutes, which is $0.18 per hour, on the Meta Model API. Speaker labelling is not billed separately. Confirm the current rate on the Meta model page linked above, since rates change.
- Meta reports a 3.1% word error rate on the Artificial Analysis streaming benchmark, as reported on September 1, 2026. For speaker labelling it reports a 17.5% average diarization error rate across the AMI IHM and SDM benchmarks and VoxConverse. Both are vendor-published figures on public test sets, so run your own audio before planning around them.
- Meta states the model supports 20 or more speakers, with attribution happening live in the stream rather than in a pass after the audio ends. It also handles recordings longer than one hour.
- Meta says the model was trained on more than 70 languages and that 25 were extensively validated at launch. It also supports code-switching, meaning a speaker can change language inside a sentence. Test your own languages before relying on any that fall outside the validated 25.
- No. Meta offers it through the Meta Model API, inside Meta AI for Mac, and in Muse Code, but does not publish weights for self-hosting. If audio cannot leave your own infrastructure, an open model such as Whisper is the usual alternative.
- Meta says the endpoint is compatible with the OpenAI SDK, so a team already calling a transcription API can usually change the base URL and the model name rather than rewriting the client. Run your own audio through both before you switch.
Testing Muse Voice Transcribe on your own calls?
Book a workflow audit and we will run your recordings through Muse Voice Transcribe and your current provider, then show you the errors each one sends downstream.
Book a Consultation