Most production voice systems still stitch together three separate models: one for transcription, one for speaker separation, and one for detecting when the user stops talking. Each hand‑off adds latency and creates another point where errors can creep in. Muse Voice Transcribe, introduced by Meta Superintelligence Labs, collapses those three functions into a single autoregressive model that works in real time. Audio is processed in 80 ms chunks, each turned into a soft token. After every chunk the model decides whether to ask for the next audio piece or to emit a text token, so listening and writing share one decoder loop and there is no separate alignment stage that can drift.
Because the model controls when it listens, it also controls the amount of audio context behind each word—a variable delay that is tuned with reinforcement learning. The policy trades word error rate against latency on a per‑word basis, placing Muse Voice Transcribe on the Pareto front for speed versus accuracy. In addition to streaming ASR, the model emits special tokens for speaker turns and speech endpoints, so diarization and endpointing are handled in the same pass without a second model.
Meta reports a final‑transcript word error rate of 3.1 % with 0.16 s latency after speech ends, beating alternatives such as Cartesia Ink‑2 (3.4 % at 0.43 s) and ElevenLabs Scribe v2 Realtime (3.6 % at 0.14 s). Average diarization error is 17.5 %, better than the 21‑28 % range of competing systems. Priced at $3.00 per 1 000 audio minutes ($0.18 per hour), it is cheaper than Cartesia Ink‑2 at $4.00 and less than half the cost of ElevenLabs Scribe v2 Realtime and Deepgram Flux.
The model is available today as a hosted API (muse‑voice‑transcribe‑1.0) on the Meta Model API; no weights have been released, so self‑hosting is not an option. It already powers dictation in Meta AI for Mac and Muse Code, supports more than 70 languages with 25 verified at launch, handles code‑switching natively, and can process audio longer than one hour with 20+ speakers without extra post‑processing.
#AI #Product #VoiceAI #ASR #Meta #RealTime