Meta Superintelligence Labs Launches Muse Voice Transcribe, Unifying Real-Time ASR, Diarization and Endpointing

The report details that audio arrives in 80-millisecond chunks at 12.5 hertz, and each chunk is converted into a single soft token. After each chunk, the model makes a binary decision: either emit a <|next_audio|> token to keep listening or produce a text token. When the stream ends, an <|empty_audio|> token triggers the model to flush remaining text. Because listening and writing share one decoder loop, there is no separate alignment stage.

Speaker attribution and endpointing are handled by additional tokens in the same stream. A <|start_of_turn|> token marks a possible speaker switch, while a <|speaker_{A-Z}|> tag identifies the speaker. Endpointing uses <|speech_onset|> and <|speech_endpoint|> tokens. All three tasks are trained jointly, with reinforcement learning combining word error rate and delay rewards multiplicatively to create an adaptive per-word delay policy.

Meta, as cited in the report, reports that the model achieves a 3.1% final-transcript word error rate at 0.16 seconds after end of speech on Artificial Analysis AA-WER Streaming. That is ahead of Cartesia Ink-2 at 3.4% at 0.43 seconds and ElevenLabs Scribe v2 Realtime at 3.6% at 0.14 seconds. On diarization, it records a 17.5% average error rate across AMI-IHM, AMI-SDM, and VoxConverse. The model was trained on more than 70 languages, with 25 extensively verified at launch, and supports native code-switching and audio input exceeding one hour.

The report notes that the $3 per 1,000-minute price undercuts Cartesia Ink-2 at $4 and is less than half the $6.50 charged for ElevenLabs Scribe v2 Realime and Deepgram Flux. The model is hosted only, meaning users cannot run it locally.