MetamodelMuse Voice Transcribe
A real-time speech-to-text model that also handles speaker separation and endpointing in a single pass, on the Meta Model API at $0.18 per hour of audio.
Details
- Performs streaming transcription, speaker diarization for more than 20 speakers, voice-activity detection, and endpointing in one model instead of a stack of separate ones
- Audio enters as 80ms chunks at 12.5 Hz, one soft token each; at every chunk the model chooses to emit a text token or to keep listening, predicting a next-audio placeholder that is swapped for the real following chunk
- Reinforcement-trained on a combined word-error-rate and delay reward, so it waits longer on hard words and commits sooner on easy ones, adapting latency word by word rather than using a fixed window
- Covers 70+ languages in training with 25 extensively validated, handles code-switching mid-sentence, and supports language, keyword, and context biasing
- Scores 3.1% word error rate on the Artificial Analysis streaming speech-to-text benchmark, ahead of Cartesia Ink-2 at 3.4% and ElevenLabs Scribe v2 Realtime at 3.6%, with roughly 17.5% average diarization error
- Available on the Meta Model API as muse-voice-transcribe-1.0 with a zero-data-retention tier, in Meta AI for Mac where it powers system dictation, and inside Muse Code; priced at $0.18 per hour of audio
- Weights are closed, with no open-weight release or self-host path, unlike Meta's earlier open speech work
Best for
Live transcription of multi-speaker audio where you need to know who said what as it happens
When to use it
You are transcribing calls, interviews, or meetings and a fixed-latency stack is either too slow or too error-prone
Who benefits
Teams building meeting notes, call review, or interview tooling who currently stitch together separate transcription and diarization services
Join the newsletter to get ahead with AI.
I share strategies, new modules, free guides, and tools to get you better at using AI in your work.
Leading a team? See team programs →