Meta: Muse Voice Transcribe 1.0

Muse Voice Transcribe 1.0 is Meta’s audio-perception model for transcribing speech, detecting utterance boundaries and attributing turns to speakers. It belongs to the Muse Spark model family and combines recognition, diarization and endpointing within the model. Its output is recognised text with the timing and speaker information available in the selected mode.

The model processes audio in 80-millisecond chunks and decides when to continue listening or emit text. Its adaptive delay allows it to wait for additional audio when a word is more difficult to resolve. This produces partial text while speech is arriving, followed by completed transcription results. Treat the partial hypothesis and the completed turn as different stages of the listening experience.

Choose among three modes. Push-to-talk is the default for a single controlled utterance, such as a voice query or a short dictated passage. Endpointing lets the model detect speech-turn boundaries and return one completed transcript per detected segment. Diarization attributes the recognised speech to speaker labels and is intended for multi-speaker recordings rather than the fastest voice-command turn-taking.

For a recorded clip, the file-transcription workflow returns a complete transcript in one request. The default result is JSON containing the transcript and audio duration, with a turns array in endpointing and diarization modes. Each turn can carry an identifier, start and end in milliseconds and its recognised text. Speaker labels are present in diarization mode. Push-to-talk produces the complete transcript without detected turns in that array. The manufacturer’s recorded-file interface can also return plain text or a stream of transcription events. Plain text keeps detected turns on separate lines and can include speaker labels; an event stream ends after its final event without a separate JSON summary. Its delta partial mode returns incremental text, while the final transcript event contains the complete wording.

Prepare a RIFF/WAVE file containing mono, 16-bit integer PCM audio at 16 or 24 kHz. The model’s native rate is 24 kHz, and the interface also accepts 16 kHz. Convert other containers or encodings into that supported WAV form before supplying a recording. The file-transcription profile allows audio up to 10 minutes and a total request body up to 32 MB. Keep the original recording and any conversion associated with the returned transcript for later listening review.

The live-audio workflow uses raw signed 16-bit little-endian mono PCM, with 16 or 24 kHz audio paced approximately in real time. Its documented session window is 60 minutes. Meta describes native audio input exceeding one hour; the documented file and live-session limits remain separate interface boundaries. Select the recorded-file or live workflow according to whether the audio is already complete or still being captured, and keep results from separate sessions associated with their own input segments.

Muse Voice Transcribe is trained across more than 70 languages, while the documented recognition interface enumerates 25 supported languages. These include English, French, German, Spanish, Arabic, Mandarin Chinese, Hindi, Japanese, Korean, Italian and Portuguese. It recognises code-switching within or between sentences. A language hint can name the languages expected in the audio, but it steers recognition rather than forcing the transcript into one language.

Vocabulary hints help with product terms, acronyms, names or places that are important in a recording. Supply the actual words and spellings you expect to hear. The keyword list biases recognition without guaranteeing an exact spelling, so review decisive terms against the audio. Language hints and keywords are selected at the start of the transcription; keep them appropriate to that recording rather than changing their meaning while a session is running.

The model supports diarization for conversations with more than 20 speakers. Labels such as A and B identify speakers within one transcription. They are not verified names and do not establish that a label refers to the same person in another session. For an original meeting-analysis task, request speaker-attributed turns, retain the recording and use the labels to organise the discussion before assigning any real participant names through the surrounding workflow.

Live partial transcripts can change as more speech arrives. In cumulative mode, each partial replaces the preceding hypothesis instead of adding another copy of it. A detected end of speech marks a boundary, while the completed-turn result supplies the final text and can differ from the last partial. Keep overlapping turns associated with their turn identifiers, and consume the completed results before treating a transcript as finished. When live input ends, signal end of input and keep receiving until the producer finishes flushing pending results. Closing the connection immediately can discard those results; a detected speech boundary ends a turn rather than the whole session.

Timing is at the turn level. This profile does not supply word-level timestamps or confidence scores, and audio-processing progress is not a timestamp for a particular word. Sound-event detection, emotion detection and transcript reformatting are also outside the documented output profile. Use the returned turn intervals for a speaker timeline or segment-level review, and retain the recognised words if a separate editing step later creates cleaner notes.

Useful artifacts include a completed dictated passage, a voice-query transcript, a speaker-attributed conversation or an utterance timeline. Compare names, numbers and technical terms with the original recording, preserve the mode and timing associations, and keep the complete turns available for review. The transcription establishes what was recognised in the audio; answering a spoken question, translating it or turning it into meeting notes belongs to a subsequent application step.