Transcribe as people speak
Use streaming speech input in an application.
Muse Voice
From spoken words to useful text.
Speech in. Text out.
Muse Voice Transcribe is Meta's speech-to-text model. Applications access it through Model API, an interface for sending audio and receiving text, with streaming and file transcription, speaker labels and detection of speech turns.
What Voice is good at
Work with recordings
Turn audio files into transcripts for review.
Follow speakers and turns
Use speaker labels and turn-level timing where supported.
Turn a conversation into an actionable draft.
See where transcription ends and application reasoning begins.
Speaker 1
Can we review the prototype tomorrow?
Speech is converted into text.
Speaker 2
Yes. Let’s make it ten in the morning.
A new turn can be attributed to another speaker.
Application
Illustrative transcript. No microphone recording or live transcription.Proposed task: review prototype tomorrow at 10:00.
An application can interpret the transcript. Voice Transcribe itself does not schedule the meeting.
Where it fits in Muse
Voice Transcribe turns audio into text. Spark or your application can interpret that text and choose an action. Voice Transcribe itself does not generate speech.
Real examples
Technical architecture
Audio is sent to Meta Model API over realtime WebSocket or file upload endpoints. The model returns transcript text with turn-level timing and optional speaker labels.
- 1Audio input
- 2Voice Transcribe
- 3Text + turns
- 4Your application
Specifications and limitations
- Model ID
- muse-voice-transcribe-1.0
- API pricing — checked September 29, 2026
- $0.18 per hour of processed audio; streaming and file transcription share the rate. Check current terms in the footer resources.
- Availability
- Meta Model API
Keep in mind
- Speech-to-text only — no text-to-speech in this model
- Turn-level timestamps, not word-level
- Separate from Muse realtime conversational voice and avatar experiences