Meta’s Superintelligence Lab released Muse Voice Transcribe, a real‑time audio model that can separate more than 20 speakers and switch among several languages in a single stream.
Real‑time speaker tracking and multilingual support
The model debuted on September 1, 2026, when Mark Zuckerberg posted a short video on X showing the system label each voice and flip between languages without pause. The demo included a conversation where participants moved from English to Spanish and back, and the transcription kept pace, even catching code‑switching mid‑sentence.
According to the announcement, the system performs speaker diarization and endpoint detection natively, meaning it does not rely on separate components to identify who is talking or when a phrase ends. It also uses an “adaptive delay” strategy: the model lingers on difficult words while moving quickly through simpler ones, a tweak meant to raise overall accuracy.
Related: Starlink Mini X Kit Adds Extra Router, But Is It Necessary
Training covered more than 70 languages, though only 25 were formally validated at launch. The developers say the model remains robust on noisy recordings and can sustain hour‑long sessions without losing track of individual contributors.
Access, pricing and early integrations
Users can try Muse Voice Transcribe through the Meta AI Mac app, which also powers voice features in other desktop applications. The model is exposed to developers via the Muse Code environment and Meta’s Model API, with a usage fee of $3 for every 1,000 audio minutes.
A demo version is posted on Meta’s research blog, offering a glimpse of the interface and sample outputs. The same technology already drives dictation inside the Meta desktop app and the Muse Code editor, showing how quickly the lab is moving from prototype to product.
By automatically attributing remarks to the right speaker, the model cuts down on post‑call editing, which often eats up hours of staff time. Teams that rely on lengthy conference calls—such as remote support centers or multinational project groups—could generate searchable transcripts without manual labeling.
Related: Meta Drops Plan to Cut Thousands of Jobs
For developers, the availability through an API opens the door to embed live captioning in custom tools, from video conferencing plugins to educational platforms. The modest price point suggests Meta aims to attract a broad user base rather than positioning the service as a premium add‑on.
External references note that speech‑to‑text technology has advanced significantly in the past decade, with many solutions now handling multiple speakers, though few combine that with real‑time language switching as described here.
Overall, the release adds another capability to Meta’s growing suite of AI utilities, which recently also included a dedicated coding agent and an open‑weight model. The lab’s rapid output hints at a strategy of quick iteration, though how these tools will mesh with the company’s core services remains to be seen.