Meta drops real-time voice model that tracks 20+ speakers

Real-time speech transcription usually forces an ugly compromise: you either get lightning-fast gibberish or painfully slow perfection.

Meta Superintelligence Labs just introduced Muse Voice Transcribe, its very first real-time audio perception model, designed to kill that trade-off once and for all.

Built on the autoregressive Muse Spark model family, it took the top spot on Artificial Analysis benchmarks for both streaming speech-to-text and public diarization as of September 1.

Listening in 80ms chunks

Instead of waiting for whole sentences, the model ingests audio in 80-millisecond slices, converting each into a single soft token.

At every step, the model decides for itself whether to keep listening for more context or emit a text token.

To balance latency and accuracy, Meta trained the model with reinforcement learning for adaptive delay. It spits out obvious words instantly while holding back on tricky phrases until it hears enough context.

Muse Voice Transcribe also bakes speaker diarization and endpoint detection directly into the model using special tokens, comfortably tracking over 20 distinct speakers without any external pipeline.

On the language side, it covers more than 70 languages with 25 extensively verified, handling mid-sentence code-switching seamlessly and processing continuous audio feeds longer than an hour.

The feature is rolling out immediately across the Meta Model API, Meta AI on Mac with a hold of the Fn key, and Muse Code.

Why it matters

We think Meta's move to treat all voice perception tasks as a single autoregressive sequence is remarkably elegant.

Building responsive voice agents traditionally requires duct-taping voice activity detectors, speech recognizers, and speaker diarizers together, losing accuracy and piling on latency at every handoff.

By unifying transcription, turn-taking, and speaker tracking into one unified model with adaptive delay, Meta gives developers a clean, production-ready foundation for natural voice interfaces.

For multilingual teams and remote workplaces, native intra-sentence code-switching and 20-plus speaker separation finally make live meeting transcription usable rather than embarrassing.

The open question is compute cost. Running high-frequency autoregressive inference on 80ms audio frames at scale won't be cheap, and Meta will eventually have to prove this architecture can squeeze onto edge devices.