Meta's New AI Transcribes Live Meetings and Knows Who's Speaking

Meta introduced Muse Voice Transcribe, an AI that listens to live audio and writes out words as people speak. Older tools usually record first and transcribe later, which means you wait for the final text. Muse works in real time, can follow more than 20 speakers, and recognizes when someone switches languages mid-conversation. That makes it especially useful for meetings, phone calls, and voice commands in multilingual settings.

For regular users, the clearest benefit is already showing up in Meta AI for Mac, where Muse now powers dictation. Developers can also plug it into their own apps through Meta's API for about 18 cents per hour of audio. Because Muse combines transcription, speaker labels, and speech detection in one model, app builders may not need to stitch together several separate voice tools. That could mean faster, simpler voice apps and cheaper real-time captions.

The technology is clever about speed. It listens in tiny 80-millisecond chunks and decides how much context it needs before writing each word. Easy words get transcribed almost instantly, while harder ones wait for more audio. This avoids the usual trade-off between fast and accurate transcription, so live captioning can feel less laggy and more reliable.

The catch: Meta trained the model on more than 70 languages, but only 25 were thoroughly tested at launch. Hindi, Tamil, Telugu, Kannada, and Malayalam are included, but accuracy may vary with accents, background noise, or uncommon terminology. So while the promise is exciting, real-world performance will depend on testing it in your own conversations and settings.