CONFIRMED: Meta Superintelligence Labs ships Muse Voice Transcribe

Editorial label: CONFIRMED. Basis: Meta Superintelligence Labs research blog, September 1, 2026. Muse Voice Transcribe is MSL's first real-time audio perception model. It does streaming ASR, speaker diarization for 20-plus speakers and endpointing in a single model from the Muse Spark family. Audio is processed in 80 ms chunks. Meta says it ranks first on Artificial Analysis streaming speech-to-text and on public diarization benchmarks as of September 1. Source: https://research.meta.ai/blog/introducing-muse-voice-transcribe

What we know

The model is live today on the Meta Model API, Meta AI for Mac and Muse Code. On Mac, holding Fn dictates into any app. It is trained on 70-plus languages, with 25 validated at launch, and handles mid-sentence code-switching plus language, keyword and context biasing. Sessions can last more than an hour with 20-plus speakers and no required post-processing. Meta's cookbook uses the slug muse-voice-transcribe-1.0, with a realtime WebSocket and a file transcription endpoint. Specialized press cites $3 per 1,000 audio minutes, about $0.18 per hour; that price is not in the research blog itself.

What we still do not know

Meta is not releasing open weights, unlike Muse Glimmer. The research post does not publish a full pricing table, Portuguese quality numbers or a latency SLA. Independent Artificial Analysis ranking is claimed by Meta as of launch day; third-party writeups cite about 3.1% WER on AA-WER Streaming, but that figure should be checked on the leaderboard, not treated as a Meta primary number.

Why it matters

MSL's first real product is not an LLM. It is an ear. That matters for voice agents, meetings and Mac dictation, and it lands days after Gemini 3.5 Transcribe. If you transcribe calls or build voice input, test Portuguese and messy multi-speaker audio before swapping Whisper or Gemini Live. Official sources: https://research.meta.ai/blog/introducing-muse-voice-transcribe and https://ai.developer.meta.com/docs/models/