New AI Splits Any Song Into Separate Instruments Automatically

Music source separation aims to pull stems out of a music mixture — useful for karaoke, remixing, and teaching. Most earlier systems targeted narrow sets of general stems, though recent work has tried to support broader source definitions. One approach lets you control what gets separated by supplying an audio query — but that's cumbersome, since you need audio examples whose features match the sources inside the mixture. This paper proposes Music Source Separation via Stem Discovery (MuS3D), a query-based framework that iteratively discovers the active sources from the mixture itself. On correctly detected sources, it matches manually queried baselines and beats state-of-the-art text-based models. In subjective evaluation, encoding artifacts came out as the limiting factor in current generative separation. The authors conclude that audio-based query representations offer an effective and automatable interface for source separation.