How AI Stem Separation Actually Works
Splitting a song into vocals and instrumental is a 2-way split. Stem separation goes further: it pulls a full mix apart into four independent parts (vocals, drums, bass, and everything else) or six, with guitar and piano on their own. Same underlying idea as vocal removal, but a meaningfully harder problem, and one that opens up a different set of uses.
From 2 stems to 4 stems
A vocal remover only has to decide what's voice and what isn't: everything non-vocal gets lumped into one instrumental track. A stem splitter has to make that same vocal/non-vocal distinction, and then keep subdividing the non-vocal portion into drums, bass, and other. That's a harder task for the model: instead of one boundary to draw, it's drawing three, and some instruments sit closer to each other in character than a voice does to any of them.
Why bass and drums are the tricky pair
Vocals tend to separate cleanly regardless of stem count, because a voice has a distinctive harmonic and formant structure that doesn't closely resemble any instrument. Bass and low guitar are a different story: they often occupy an overlapping low-frequency range, so the model has less to distinguish them by. Programmed or heavily processed drums can also separate less cleanly than an acoustic kit, since processing can blur the transient characteristics the model relies on to identify a hit as "drums" rather than part of the other stem.
What "other" actually contains
The fourth stem isn't a leftover bucket for separation failures. It's a genuine category: guitars, keys, pads, synths, strings, anything that isn't vocals, drums, or bass. In a guitar-driven rock track, "other" might carry most of the melodic content. In an electronic track built around synth bass and drum programming, "other" might be comparatively sparse. What ends up in it depends entirely on how the source track is arranged.
From 4 stems to 6 stems
A 6-stem split takes guitar and piano out of "other" and gives each its own track. What is left in "other" is the rest: synths, pads, strings, brass. The six add back up to the original mix, so nothing is lost by splitting further. If a song has no piano, the piano stem is close to silent, which is the honest result rather than a fault.
On AudioForges the free result gives you four stems and Clean gives you six, at the same price as four.
Free vs. Clean
The free result runs a lighter separation model and finishes in under a minute on our GPU-accelerated infrastructure. Clean runs a heavier band-split transformer model, which produces noticeably cleaner separation across every stem, at the cost of 1 to 2 minutes instead of 20 seconds to 1 minute. The extra time buys real improvement specifically on the harder cases (bass/guitar overlap and busy drum programming), which is why it's worth it when a stem is headed into an actual mix rather than just a preview.
Stems vs. a plain instrumental
If all you need is vocals removed, a full stem split is more than the job requires: a Vocal Remover (or the YouTube Vocal Remover if you're starting from a link) does the same underlying separation and hands back one instrumental instead of three additional stems to sort through. Reach for stem separation specifically when you need to isolate or rebuild around drums, bass, or another instrument on its own: sampling a bassline, remixing with someone else's drum pattern, or studying a part note-for-note without the rest of the mix in the way.
Our AI Stem Splitter runs this exact process: upload a track and get back four stems free, or six in Clean, no account or software install needed. Working from a YouTube link instead of a file already saved on your device? The YouTube Stem Splitter runs the identical separation, just starting from a pasted link instead of an upload.