You cannot make a separation model perfect, but a few choices before and after splitting make a noticeable difference.
Start with the best file you have
Use WAV or a high-bitrate MP3 from a source you legally own. Heavily compressed audio has already lost detail the model would use to tell instruments apart.
Avoid files that were re-encoded several times. Each round of lossy compression adds artifacts that the model can mistake for part of an instrument.
Choose the original release over a live recording, a bootleg, or a version with applause and room noise, when you have the option.
Know which songs separate well
Some music is easy for a model and some is hard.
Easier: clearly arranged songs where each instrument has its own space, such as pop and hip hop productions with a clean vocal, punchy drums, and a distinct bass.
Harder: dense or distorted arrangements, wall-of-sound rock, heavy reverb, orchestral recordings, and music where several instruments play the same notes together.
Very loud masters can be tricky because limiting and clipping squash the differences between sounds.
Expect some bleed
A little leakage between stems is normal. A faint hi-hat in the vocal stem or a trace of vocal in the "other" stem is the model making a judgement call, not a sign that something went wrong. Our guide to how AI stem separation works explains why.
Clean up with the mixer
Use EQ to hide bleed. If a bass stem has a bit of cymbal hiss, pull the high band down. If a vocal has low rumble, pull the low band down.
Let the full mix mask it. Bleed is far less noticeable when the stem plays alongside the others than when it is soloed. Judge a stem in context, not just on its own.
Lower the stem slightly if its artifacts stand out. A small gain reduction often hides them.
Check the "other" stem. Instruments like keys, strings, and effects can end up there. Listen to it before you decide a part is missing.
Pick the right stems for the job
Different stems tolerate separation differently. Drums and bass usually come out strong and are the easiest to reuse. Vocals are the most revealing because any watery or phasey sound stands out. For a mashup, you may get the most polished result by using a vocal under a full instrumental from another track, rather than soloing the vocal by itself.
When results are not good enough
If a particular song never separates cleanly, try a different version of the track, such as the radio edit or the original studio mix, before giving up. Different masters of the same song can separate differently.