Stem separation tools use machine-learning models to split a finished stereo song into parts, typically vocals, drums, bass and everything else. A few years ago this was barely possible. Today it is built into several DAWs and available as free, open-source software.
How it works
The models are trained on many songs where the separate parts are known. They learn what a voice, a kick drum or a bassline tends to look like in the audio spectrum, and estimate how to pull each one out of a mix. Open-source examples include Demucs, released by Meta's research team, and Spleeter, released by Deezer.
What it does well
- Practice tracks: mute the bass or drums and play along.
- Transcription: isolate a part to hear it clearly.
- Remixing and sampling your own songs when the original multitracks are lost.
- Cleaning up live recordings of your own band.
Where it falls short
Results depend heavily on the mix. Dense arrangements, heavy reverb and instruments that share frequencies, like a synth pad and a vocal, often leave artifacts: a watery, phasey sound or bits of other instruments bleeding through. Listen to each stem on its own before you use it in a release.
The rights question
Separating a stem doesn't change who owns the music. If you extract a vocal from someone else's song and release it in your track, you still need permission from the rights holders, just as with any sample. Using separated stems privately for practice or study is a different situation from publishing them. When in doubt, use your own recordings or properly licensed material.
