Stem Separation and What to Do with Stems

For most of recording history, a finished stereo mix was final: if you didn't have the multitrack session, you couldn't get the vocal on its own. AI stem separation changed that. A model trained on thousands of songs where the individual parts are known learns what a voice, a drum kit or a bass guitar sounds like, and given a full mix it estimates each part separately.

What you get

A standard separation returns four stems:

• Vocals — lead and usually backing vocals.
• Drums — the whole kit, as one stem.
• Bass — bass guitar, synth bass, sub.
• Other — everything else: guitars, keys, pads, strings, effects.

In SoundGen's Studio, Stem Separation lands each part on its own track so you can mix them independently. Deep mode runs a larger six-source model that additionally pulls out piano and guitar; the drums still come back as one stem. Expect more bleed in the piano stem than the others — the model's own authors flag piano as its weakest source. Isolate Vocals uses the same engine and keeps only the vocal.

For a quick vocal-free version of a song, the free vocal remover runs in your browser with nothing uploaded.

What stems are good for

Practising and learning

Mute the bass stem and play along with the band. Solo the bass stem to hear exactly what the bass player is doing under a busy chorus. Slow it down with the pitch shifter & tempo changer without changing its pitch.

Remixing your own songs

If you've lost a session, or only have a bounce of an old track, separation gives you parts to work with again: replace the drums, re-arrange the structure, put a new bass line under the old vocal.

Sampling your own material

Pull a vocal phrase from a demo, a drum break from a jam you recorded, or a guitar part from a rough take, and build something new on it.

Rebalancing a mix you can't reopen

Turning a buried vocal up two decibels, or taming an overbearing kick, is possible when the parts are separate — something no EQ on the full mix can do cleanly.

Getting MIDI from a single part

Transcribing a full mix to MIDI is hard because everything plays at once. A clean bass stem transcribes far better (see Audio to MIDI).

The artifacts to expect

Separation is very good but not perfect. Know what to listen for:

• Bleed. Faint hi-hats or guitar in the vocal stem, or a breath of the vocal left in the instrumental. Loud, dense choruses bleed the most.
• Watery or warbly sound. Where the model can't decide which source a sound belongs to — a sustained synth in the same register as a vocal is the classic case — the result can sound phasey or underwater.
• Reverb splits. Part of a vocal's reverb tail may end up in "other" or the instrumental.
• Backing vocals and doubles may land with the lead, in "other", or split between them.
• Transient smearing on drums, especially when cymbals overlap with bright instruments.

Cleaner, sparser mixes separate better than heavily compressed, distorted masters. Results also vary between songs; try a few before judging what the tool can do.

Working with imperfect stems

• Use stems in a mix, not soloed. Artifacts that are obvious in solo usually disappear once other parts play over them.
• EQ out the bleed. A high-pass on the vocal stem can remove kick and bass bleed; a gentle low-pass on the bass stem removes hi-hat leakage.
• Gate or edit. Cut silence between vocal phrases to remove quiet bleed.
• Layer to hide gaps. If an isolated drum stem sounds thin, layer clean drum samples on top and blend.

Whose songs to separate

Separating a commercial song to practise, study or make a private karaoke track is a common use. Releasing anything built from someone else's recording — a remix, a sample, a cover backing — needs permission from the rights holders, however clean the separation. That's covered in Rights and Responsible Use.