Meta Superintelligence Labs just announced Muse Voice Transcribe, and the pitch is refreshingly direct: one autoregressive model that does streaming ASR, speaker diarization for 20+ speakers, and endpointing in a single pass. No stitching together three separate systems. No post-processing required. Audio arrives in 80ms chunks, each transformed into a soft token, and the model makes one binary choice after every chunk — keep listening or emit text.
What makes this different from the usual benchmark flex is the delay policy. Instead of fixing a latency-accuracy trade-off, Meta trained it with reinforcement learning. A word error rate reward and a delay reward combine multiplicatively, producing a per-word adaptive delay. Harder words get more audio context; easier ones stream faster. The result: 3.1% final-transcript WER at 0.16 seconds after end of speech on Artificial Analysis, beating Cartesia Ink-2 (3.4% at 0.43s) and ElevenLabs Scribe v2 Realtime (3.6% at 0.14s). Meta claims the Pareto front for speed against accuracy now belongs to Muse, displacing the previous frontier formed by Soniox, Cartesia, and ElevenLabs.
Diarization and endpointing are handled with special tokens in the same stream — <|start_of_turn|> marks potential speaker switches, <|speaker_{A-Z}|> tags identify speakers, and <|speech_endpoint|> signals when the user finished. Meta reports a 17.5% average diarization error rate across AMI-IHM, AMI-SDM, and VoxConverse, compared to 21.1% to 28.6% for five other systems. Trained on 70+ languages with 25 extensively verified, the model handles code-switching natively — a practical win for bilingual speakers who mix languages mid-sentence.
The catch: no open weights. This is API-only via the Meta Model API at $3.00 per 1,000 audio minutes ($0.18 per hour), undercutting Cartesia’s $4.00 and less than half of ElevenLabs and Deepgram’s $6.50. It already powers dictation in Meta AI for Mac and Muse Code. For developers who need self-hosted models, that’s a dealbreaker. But for teams already building on hosted APIs, the price-performance combination is hard to ignore. The real question is whether Meta keeps this closed or follows the Llama playbook and releases weights later. Given the competitive pressure from Cartesia and ElevenLabs, an open release would reset the market. For now, it’s a hosted-only flex.
