Meta has released its first real-time audio model, Muse Voice Transcribe, developed by Meta Superintelligence Lab (MSI). It can recognize dictated text and transcribe conversations with over 20 speakers, and work with multiple languages simultaneously.
Meta has released its first real-time audio model, Muse Voice Transcribe, developed by Meta Superintelligence Lab (MSI). It can recognize dictated text and transcribe conversations with over 20 speakers, and work with multiple languages simultaneously.
Meta CEO Mark Zuckerberg posted an example of the model working, where it automatically identifies speakers and switches between languages. Going further, the system even recognizes “code switching” (mixing languages) and accurately transcribes sentences made up of words from different languages, Engadget writes .
“The model decides when to listen. It waits a little longer for complex words and processes simple ones faster, using adaptive latency to predict each token and improve accuracy,” Zuckerberg explained. “It does a great job with ‘dirty’, real-world audio: trained on over 70 languages (25 of which were validated at launch), handles language mixing within sentences, and works with hour-long sessions with 20+ speakers.”
Muse Voice Transcribe is MSL's first real-time audio perception model — rolling out today. SOTA in streaming speech-to-text, it handles speaker diarization, and endpointing natively in a single model. pic.twitter.com/LViMDSkbim
— Mark Zuckerberg (@finkd) September 1, 2026
Meta's release comes less than a week after Google announced Gemini 3.5 Transcribe with similar functionality. However, while Google is integrating its model directly into Android and Chrome, Meta's plans to implement Muse Voice Transcribe into its core services remain unknown.
3/ already powering dictation in the meta desktop app and muse code. live now via meta model api https://t.co/MFosERCV0E pic.twitter.com/UESTQhKJrH
— Alexandr Wang (@alexandr_wang) September 1, 2026
Currently, there are several ways to test the system's capabilities:
Meta AI app for Mac: Muse Voice Transcribe already provides dictation features in this desktop app;
for developers: the model is available through Muse Code and Meta's Model API at a price of $3 per 1,000 minutes of audio.
Demo version: you can learn about the technology at work on the Meta research blog.
