SAM Audio (Segment Anything in Audio)

Meta brought its Segment Anything idea to sound: point at a sound the way you would click an object in a photo, and SAM Audio pulls it out of the mix. A sound can be selected with a text prompt ("dog barking", "music"), by clicking the thing making it in the video frame, or by marking a stretch of time, and the model separates it from everything else. It builds on Meta's Perception Encoder, extended to audio and video (PE-AV), which Meta described as the "ears" to SAM Audio's "brain." Meta released code and weights and put it in the free Segment Anything Playground.

For AI filmmakers it became a rescue tool. Models that generate sound in the same pass, like Seedance 2.5, often add music nobody asked for. In October 2026 Tom Likes Robots showed the fix: isolate the music in SAM Audio and remove it, keeping the dialogue, sound effects and ambience, then lay in your own score, which also keeps the soundtrack consistent from one generation to the next. A shot that would have been thrown away for its audio could be kept.