Muse Voice Transcribe listens before deciding when to write

Muse Voice Transcribe combines real-time transcription, separation of more than 20 speakers, and end-of-speech detection in a single model priced at $0.18 per hour.

A conversation can be transcribed as it unfolds, with markers showing when the speaker changes and when each person finishes talking. Muse Voice Transcribe combines these three operations in a single model instead of assigning them to separate services.

The release is the first real-time audio perception model developed by Meta Superintelligence Labs. It handles automatic speech recognition, diarization—the process of distinguishing between speakers—and detection of when each person starts and stops talking.

That final capability does more than add punctuation to a transcript. In a voice assistant, it determines when the system can begin responding. Detecting an endpoint too early can interrupt the user during a pause, while waiting too long makes the assistant appear unresponsive.

Muse Voice Transcribe belongs to the Muse Spark family. It divides the audio stream into 80-millisecond segments, or 12.5 segments per second. After each one, the model decides whether to keep listening or start producing the corresponding text.

The delay is not the same for every word. A familiar expression may be transcribed quickly, while a name, an ambiguous phrase, or a difficult passage may receive more audio context. Meta calls this process “adaptive delay.”

Training combines a reward tied to the word error rate with another focused on waiting time. The model must therefore balance accuracy and responsiveness instead of applying the same fixed delay to an entire conversation.

Diarization relies on markers inserted directly into the transcription stream. When a voice change is detected, the model opens a new speaking turn and assigns it a label such as “Speaker A” or “Speaker B.” It can distinguish between speakers, but it does not necessarily know their real identities. An application must still connect those labels to names when that information is available.

Meta says the model can handle more than 20 people in a single conversation and audio sessions exceeding one hour without requiring additional processing. One demonstration runs for slightly more than an hour with 11 participants, while another features eight people speaking in the same room.

The model was trained on more than 70 languages. Meta considers 25 of them extensively validated for the initial release, including English, French, Chinese, Spanish, Hindi, Japanese, and Vietnamese. Support for the remaining languages does not mean that they have all received the same level of testing.

Speakers can switch languages between sentences or in the middle of a single statement. This capability is intended for bilingual conversations in which English terms, product names, or professional expressions are inserted into another language without an explicit transition.

Language, keyword, and contextual guidance can also be supplied to improve recognition. This should help the system transcribe names, places, brands, or terminology specific to an organization. Its effectiveness still depends on the information provided by the application and does not guarantee that every unusual term will be captured correctly.

On the independent Artificial Analysis leaderboard, Muse Voice Transcribe records a 3.1% final-transcription word error rate, with an average delay of 0.16 seconds after the detected end of speech. The word error rate measures substitutions, insertions, and deletions against a reference transcript after normalization.

Cartesia Ink-2 reaches 3.4% using semantic endpointing, with a delay of 0.43 seconds. ElevenLabs Scribe v2 Realtime records 3.6% at 0.14 seconds. Muse is therefore more accurate in this evaluation, although it is not consistently the fastest under every comparison mode.

The Artificial Analysis testing methodology uses approximately eight hours of audio from three sources: conversations designed for voice-agent use cases, European Parliament proceedings, and corporate earnings calls. The material is primarily representative of English and cannot confirm equivalent performance across all 25 languages validated by Meta.

For speaker separation, Meta reports an average diarization error rate of 17.5% across AMI-IHM, AMI-SDM, and VoxConverse. The other systems shown in its chart range from 21.1% to 28.6%. Unlike the transcription ranking, this comparison is presented by Meta using public datasets and does not independently validate every condition claimed for the model.

The diarization error rate includes missed speech, unnecessary detections, and segments assigned to the wrong person. A result of 17.5% remains high enough to require human review in sensitive settings such as official minutes, interviews, medical records, or legal conversations.

Real-world discussions may also include overlapping voices, background noise, distant microphones, accents, and changes in volume. Meta’s demonstrations provide an overview of how the system behaves, but they cannot replace tests using each organization’s own equipment and recording environments.

Muse Voice Transcribe is available as `muse-voice-transcribe-1.0` through the Meta Model API. Developers can use a WebSocket connection for live transcription or a separate endpoint for prerecorded files.

The listed price is $3 per 1,000 minutes of audio, equivalent to $0.18 per hour. At launch, this is below the $4 per 1,000 minutes listed for Cartesia Ink-2 and the $6.50 reported by Artificial Analysis for ElevenLabs Scribe v2 Realtime and Deepgram Flux. These comparisons should still account for the features included, volume commitments, and commercial terms of each provider.

Meta has not announced downloadable model weights. Muse Voice Transcribe currently operates as a service hosted by the company, making it easier to integrate but preventing organizations from running it entirely on their own infrastructure.

The technology already powers dictation in Meta AI for Mac and Muse Code. On macOS, users can hold the `Fn` key to dictate into different applications and use voice commands to assign a task to Meta AI based on