Split a song into stems

The four stems

The split comes from Demucs, an open-source model from Meta's AI researchers, trained on these four kinds of stem. Guitars and piano share the last part and cannot be pulled apart from each other.

Practicing along with a band

The most common use is practice. A bass player turns the Bass part off and plays along with the rest of the band; a drummer does the same with Drums. Set the speed to 0.75× to learn a hard part without changing its pitch (change audio speed); the parts are separate sounds, so give each the same speed.

Parts that are off stay on their tracks, silent, so you can bring them back at any time. Export leaves them out of the file.

Remixes and samples

Producers export single stems as WAV or FLAC to build on in a music program: turn every part off but one, press Export, and repeat for the next. Each stem lasts as long as the song and starts where it starts, so the files line up when you drop them in side by side.

The four stems are the model's guess at each instrument and do not add up to the original exactly; small traces of one can sit in another, most often the hi-hats in the voice or the bass in the drums. Choosing Voice and music instead gives two parts that do add up to the original exactly, if the voice is all you need to separate: see the vocal remover.