Stem Separation in plain English.

Stem separation is the use of AI to pull a mixed recording apart into its component tracks — vocals, drums, bass, and everything else — something that was essentially impossible before deep learning because the sounds overlap in the same frequencies. Models trained on huge libraries of multitrack recordings learned to recognise what a voice or a snare "looks like" in a spectrogram and isolate it. The technology powers karaoke and practice apps like Moises, remix and sampling workflows, restoration of old recordings, and the isolated-vocal features in music generators like ElevenLabs Music. Quality is now good enough for practice and creative use, with audible artefacts on dense mixes, and it raises copyright questions when the stems of commercial songs are reused without permission.