Meta launches Muse Voice Transcribe at $0.18 per hour

• Muse Voice Transcribe combines streaming ASR, speaker diarization for 20-plus speakers, and endpointing in one model, as the launch announcement describes.
• The model varies when it commits words instead of using a fixed delay. The architecture thread says it can keep listening for difficult words and emit easier ones sooner.
• Meta is charging $3 per 1,000 audio minutes, or $0.18 per hour. The benchmark post places its final-transcript result at 3.1% WER and 0.16 seconds after speech ends.
• The model was trained across 70-plus languages, with 25 validated at launch, according to alexandr_wang's launch thread.
• It is shipping through the Meta Model API, Meta AI for Mac, and Muse Code, according to the availability post.

Meta's technical announcement names the tokens used to mark turns, speakers, and speech boundaries. The speech-to-text API guide exposes a WebSocket endpoint whose authentication must go in the handshake frame, not the usual Authorization header.

Meta calls Muse an autoregressive multimodal model in the Muse Spark family in its technical announcement. It converts each 80ms audio chunk into one soft token, then decides at every chunk whether to emit text or continue listening.

The training objective combines word-error-rate and delay rewards. Meta says that lets the model trade latency against accuracy word by word rather than finalizing every word on the same schedule.

Muse emits structure alongside text. Meta's technical announcement describes these special tokens:

• <|start_of_turn|> marks a new speaker turn.
• <|speaker_A|>-style tags identify speakers, including across separated segments from the same speaker.
• <|speech_onset|> and <|speech_endpoint|> mark the beginning and end of speech.

The launch also claims native mid-sentence code-switching, language, keyword, and context biasing, plus hour-plus audio and 20-plus speakers without post-processing. The language count and validation split came from alexandr_wang's launch thread.

Artificial Analysis measures both WER and the time from SileroVAD-detected end of speech to a partial or final transcript in its methodology. Its streaming set has about eight hours of audio drawn from AA-AgentTalk, VoxPopuli, and Earnings22, as the leaderboard explains.

• Final transcription reached 3.1% WER at 0.16 seconds after end of speech, according to the benchmark post; Cartesia Ink-2 with semantic endpoints posted 3.4% at 0.43 seconds, while ElevenLabs Scribe v2 Realtime posted 3.6% at 0.14 seconds.
• First partial transcription reached 3.6% WER at 0.13 seconds, where the benchmark post placed it just ahead of ElevenLabs on both accuracy and latency.
• Meta's diarization chart put Muse at 17.5% DER, versus 21.1% for AssemblyAI U3.5 Pro offline, according to WesRoth's benchmark summary.

Meta's API guide documents two transports for the same model:

• wss://api.meta.ai/v1/asr/realtime for live audio in voice agents, dictation, and captions. Authentication goes in the WebSocket handshake frame, and the Authorization header is ignored.
• POST https://api.meta.ai/v1/asr/transcribe for recorded audio through multipart upload, with the Authorization header used normally.

The model is already behind dictation in Meta's desktop app and voice input in Muse Code, while WesRoth's overview reports a zero-data-retention tier on the API. Availability remains regional: the Model API landing page says the API is not available in every country.