Meta launches Muse Voice Transcribe: streaming ASR and 20-plus-speaker diarization

Meta Superintelligence Labs launched Muse Voice Transcribe on 1 September 2026. The research post calls it the lab's first real-time audio perception model: real-time streaming ASR, diarization with 20-plus speakers, and endpointing, plus multilingual code-switching and language, keyword, and context biasing.

AIatMeta on X posted the same day: "Introducing Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs."

A follow-up from AIatMeta says: "Muse Voice Transcribe is an autoregressive multimodal LLM from the Muse Spark family." The research post calls it an autoregressive multimodal model in that family. Audio is processed in 80ms chunks (12.5 Hz). Each chunk becomes one soft token. At each chunk the model continues listening or emits a text token.

Key facts: - Meta Superintelligence Labs says this is its first real-time audio perception model. - Training includes RL with combined word error rate and delay rewards, per the same follow-up post. - Trained with 70-plus languages; 25 are extensively verified at launch. Meta recommends those 25 first. - The research post says the model natively supports long audio exceeding one hour and up to 20-plus speakers, with no required post-processing. - Meta says it is available today through the Meta Model API, Meta AI for Mac, and Muse Code. On Mac, Meta says users hold the Fn key to dictate into any application. - Meta says it ranks first on Artificial Analysis streaming speech-to-text and on public diarization benchmarks as of 1 September 2026.

How to read this: the architecture, language counts, hour-long context, Mac Fn dictation, and leaderboard ranks are Meta's launch claims. The X follow-up uses "LLM"; the research post says "multimodal model." Rankings can move after 1 September.

Frontier Flux is independent coverage. We did not train or host Muse Voice Transcribe.