Demucs
An open-source AI model that separates a finished song back into individual parts such as vocals, drums and bass.
Demucs is a source separation model, originally from Meta’s research group and now maintained as open source. Its job sounds impossible when you think about it: a finished song is just two channels of mixed sound, and the information about which instrument produced which vibration is gone. Demucs reconstructs it anyway, because it has learned what a snare drum or a human voice tends to look like when you draw sound as a picture of frequencies over time.
The newer variants split a track into as many as six parts: vocals, drums, bass, guitar, piano and everything else. Because the model is small by modern standards, it runs comfortably on an ordinary laptop, which is why most free offline stem separation tools are a friendly interface wrapped around it. Results are good, not perfect, and dense mixes still produce audible smearing.