Meta Unveils Muse Voice Transcribe, a Real-Time Speech AI Tool

MENLO PARK, Calif. —

Meta AI announced Muse Voice Transcribe on Monday, describing it as the first real-time audio perception model built by Meta Superintelligence Labs, the company's advanced AI research group.

According to Meta AI, the tool listens to live audio and turns it into text as a person speaks, rather than waiting for someone to finish before transcribing. The company said it can tell apart more than 20 different speakers in the same recording, mark when someone starts and stops talking, and handle audio that runs longer than an hour without any extra processing afterward.

How It Processes Speech

Meta AI said the model breaks incoming audio into tiny 80 millisecond chunks and decides, chunk by chunk, whether to keep listening or start writing down words. The company calls this ability to control its own listening time "adaptive delay," and said it was trained using a reinforcement learning system that rewards the model for both accuracy and speed at the same time. Meta AI said this lets the model wait longer on words that are harder to make out and move faster through easier ones, which the company said produces the best available tradeoff between speed and accuracy.

Meta AI also said the model was trained to recognize when speakers switch between languages mid-conversation, a skill known as code-switching, and to use context clues to catch specific names or keywords more accurately. The company said Muse Voice Transcribe was trained on more than 70 languages, though it is recommending 25 of those for now because they have been more thoroughly tested.

Meta AI said the model ranks first on Artificial Analysis, a third-party benchmarking site, for streaming speech-to-text and for public tests of speaker separation, based on rankings as of September 1, 2026. The company did not say how large the gap is between its model and competitors, or how the rankings might change as other companies release new tools.

Where It's Being Used

Meta AI said the model now powers voice dictation across its Meta AI assistant and Muse Code, a coding tool, letting users dictate by holding the "Fn" key on their keyboard. The company said Muse Voice Transcribe is available immediately through the Meta Model API, the Meta AI app for Mac computers, and Muse Code.

Meta AI did not say how much the service costs for outside developers or when support for its remaining, less-tested languages might expand. The company also did not say whether the model would come to its other products, including its smart glasses or WhatsApp.