Meta Superintelligence Labs Releases Muse Voice Transcribe — A Single Model That Replaces Entire Voice Stacks

Meta has launched Muse Voice Transcribe, its first real-time audio perception model from Meta Superintelligence Labs. The single autoregressive model handles streaming speech-to-text, speaker diarization for 20+ speakers, and endpointing in one pass — tasks that typically require three separate systems. It achieves 3.1% word error rate at just $0.18 per audio hour, undercutting competitors by up to 80%.

Key Takeaways

Meta has launched Muse Voice Transcribe, its first real-time audio perception model from Meta Superintelligence Labs.
The single autoregressive model handles streaming speech-to-text, speaker diarization for 20+ speakers, and endpointing in one pass — tasks that typically require three separate systems.
It achieves 3.1% word error rate at just $0.18 per audio hour, undercutting competitors by up to 80%.

AI synthesis3 sources5

Meta Superintelligence Labs on September 1 released Muse Voice Transcribe, what it calls its first real-time audio perception model. The system collapses three typically separate jobs — streaming automatic speech recognition, speaker diarization, and endpointing — into a single autoregressive model from the Muse Spark family.

Audio arrives in 80ms chunks at 12.5 Hz, each transformed into a single soft token. After every chunk the model makes one binary choice: predict a next-audio token and keep listening, or emit a text token. Listening and writing share one decoder loop, eliminating alignment drift.

The model uses reinforcement learning to train an adaptive delay policy — varying how much audio context sits behind each word based on difficulty. Meta reports this puts Muse Voice Transcribe on the Pareto front for speed versus accuracy, ahead of systems from Soniox, Cartesia, and ElevenLabs.

Benchmarks show 3.1% final-transcript word error rate at 0.16 seconds after end of speech on Artificial Analysis AA-WER Streaming, compared to 3.4% for Cartesia Ink-2 and 3.6% for ElevenLabs Scribe v2. On diarization, it achieves 17.5% average error rate across standard benchmarks, versus 21-29% for competitors.

The model is trained on 70+ languages with 25 extensively verified at launch, and supports native code-switching. It is available as an API on Meta Model API at $3.00 per 1,000 audio minutes ($0.18/hour), undercutting Cartesia and costing less than half of ElevenLabs and Deepgram. No weights have been released.