Muse Voice Transcribe, explained
Meta’s first real-time audio model turns speech into text as people talk, tells speakers apart and notices when someone stops. It’s a model for apps and developers, not a feature you switch on inside the Muse agent. Here’s what Meta says it does, where you can use it, and what it costs.
Quick answer
Muse Voice Transcribe is Meta’s real-time speech-to-text model, announced September 1, 2026, by Meta Superintelligence Labs. One model transcribes speech as it streams, labels 20+ speakers and detects when a speaker is done. It’s available on the Meta Model API as muse-voice-transcribe-1.0 for $0.18 per audio hour, and powers voice dictation in Meta AI for Mac and Muse Code. Meta lists 25 verified languages.
How we checked
Not testedWe haven’t used the model ourselves. Everything here comes from Meta’s announcement, API documentation and pricing page, read on October 5, 2026. Benchmark figures are Meta’s own. See how we label evidence.
What it does
Meta calls it its first “real-time audio perception model.” It combines three jobs that usually need separate systems:1
- Streaming transcription
- Turns speech into text as it arrives, processing audio in 80-millisecond chunks, with an “adaptive delay” per word.
- Diarization
- Labels who is speaking, for 20 or more speakers.
- Endpointing
- Detects when someone has finished talking.
- Code-switching
- Handles speakers who switch languages mid-sentence.
- Biasing
- Accepts language hints and keywords (product names, people, places, acronyms) to improve accuracy.
Where you can use it
| Where | How |
|---|---|
| Meta Model API | Model ID muse-voice-transcribe-1.0; realtime (WebSocket) or file transcription |
| Meta AI for Mac | Voice dictation |
| Muse Code | Meta’s terminal coding agent |
| Muse agent (muse.ai) | Not listed by Meta |
Sources: Meta’s announcement and API docs.12 Note that Muse Voice Transcribe is a different product from the Muse agent, even though both carry the Muse name. How Meta’s Muse model family fits together.
Price and limits
- Price
- $0.18 per hour of audio; streaming and file transcription cost the same; billed in whole seconds.23
- Tiers
- Zero data retention (ZDR) costs the same as Standard; free-tier credits apply. The cheaper tier that lets Meta train on your data isn’t offered for this model at launch.3
- Rate limits
- 128 concurrent streams and 16,000 streams per hour, shared by realtime and file transcription.3
- File length
- Files over 10 minutes are rejected; split them first. Meta says the model itself handles audio “exceeding one hour.”12
- Audio format
- Mono, 16-bit PCM at 16 or 24 kHz (WAV).2
Languages
Trained on more than 70 languages, with 25 extensively verified.1 The API docs list: Arabic, Bengali, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Mandarin Chinese, Marathi, Polish, Portuguese, Spanish, Tagalog, Tamil, Telugu, Thai, Turkish and Vietnamese.2
How accurate is it?
Meta reports a 3.1% word error rate on streaming speech-to-text, against 3.4% to 4.0% for the competitors it compared, and an average diarization error rate of 17.5% against 21.1% to 28.6%. It says the model ranks “first on Artificial Analysis on streaming speech-to-text and on public diarization benchmarks.”1 These are Meta’s own figures; real-world accuracy depends on audio quality and accents. Meta’s own note: transcripts are “AI-generated and may be inaccurate.”1
What it doesn’t do (yet)
The API docs say word-level timestamps, confidence scores, sound-event detection, emotion detection and transcript reformatting “are not available.”2
What happens to your audio
Meta’s announcement says: “Your audio is processed to produce the transcript and is not stored.”1 Developers can also choose the zero-data-retention tier on the API.3
Sources
- Introducing Muse Voice Transcribe Meta AI Research · Sept 1, 2026 · read Oct 5, 2026
- Speech-to-text (Muse Voice Transcribe) documentation Meta Model API docs · read Oct 5, 2026
- Pricing and rate limits Meta Model API docs · read Oct 5, 2026
Last checked: 2026-10-05. Prices and availability can change; tell us if you spot a difference.
Frequently asked questions
What is Muse Voice Transcribe?
Meta’s real-time speech-to-text model, announced September 1, 2026. Meta says one model handles streaming transcription, telling 20+ speakers apart (diarization) and detecting when someone has stopped talking (endpointing).1
How much does Muse Voice Transcribe cost?
$0.18 per hour of audio on the Meta Model API, the same for streaming and file transcription. Meta says platform free-tier credits apply.3
Can I use Muse Voice Transcribe in the Muse app?
Meta doesn’t list the Muse agent app. It lists the Meta Model API, voice dictation in Meta AI for Mac, and Muse Code.1