The new Meta Muse Voice Transcribe AI model has been introduced by Meta as its first real-time audio model designed for live dictation and transcription across multiple speakers and languages. Meta says the model can handle more than 20 speakers simultaneously, switch between languages automatically, and understand code-switching, where people blend words from different languages within the same sentence.
How the Meta Muse Voice Transcribe AI Model Works
Meta CEO Mark Zuckerberg demonstrated the model in a video showing it automatically identifying different speakers and switching between languages during a live conversation. According to Zuckerberg, Muse Voice Transcribe uses what the company calls adaptive delay to improve transcription accuracy, meaning it waits slightly longer before committing to difficult or ambiguous words while processing easier words more quickly, balancing speed against accuracy in real time.
The model was trained across more than 70 languages in total, with 25 of those languages validated at launch. Meta says the system can also handle noisy and complicated real-world audio conditions, mid-sentence language switching, and extended sessions lasting around an hour with more than 20 participating speakers, which is a notably demanding use case for a real-time transcription system.
Meta Takes on Google Gemini in Transcription
Muse Voice Transcribe arrives less than a week after Google introduced Gemini 3.5 Transcribe, which offers broadly similar audio transcription capabilities. Google has said it plans to integrate its own model into Android and eventually into Chrome, positioning transcription as a core feature across its major platforms rather than a standalone tool.
Meta has not yet said whether Muse Voice Transcribe will eventually become part of its major consumer-facing services such as WhatsApp, Instagram or Messenger, leaving open the question of how widely the technology will be deployed beyond its current developer-focused release.
Where the Technology Is Available
For now, users can access the technology through Meta’s recently launched Meta AI Mac app. Because the Mac app can provide voice features to other applications, Muse Voice Transcribe can also power dictation features inside other services through the app, extending its reach beyond Meta’s own products.
Developers can access the model directly through Muse Code and Meta’s Model API, with Meta pricing the service at $3 per 1,000 minutes of audio processed. A demonstration version of Muse Voice Transcribe is also available through Meta’s research blog for those who want to test its capabilities before committing to API integration.
Part of a Broader Push from Meta Superintelligence Lab
Muse Voice Transcribe is the latest release from Meta Superintelligence Lab. In recent weeks, the group has also introduced Meta’s first dedicated coding agent, an open-weight language model, and the Meta AI Mac app, suggesting the lab is moving through a rapid release cycle across multiple AI product categories simultaneously.
For more on Meta’s AI research initiatives, visit the Meta AI website. For more coverage of AI product launches, see our technology news section.
