What are stems — and can you really unmix a finished song?

Stems, properly speaking

A stem is a grouped submix kept separate from the rest: all the drums on one, all the vocals on another, and so on. They exist so that a mastering engineer, a remixer or a live sound engineer can work on one part without touching the others.

Real stems are simply files somebody kept. If you have them, nothing needs to be estimated — they were never combined in the first place.

Separation is not the same thing

Once a track is mixed, every part has been summed into two channels of audio. That addition is not reversible; the individual parts no longer exist anywhere in the file.

So a separation model does not recover them. It estimates: given this mixture, what did each part most likely sound like on its own? It has heard enough music to be very good at that guess, but a guess is what it is.

This is why the results have a particular character. The parts are usually convincing on their own, with a faint trace of the others still audible, and dense passages can take on a slightly watery, smeared quality. Those are not bugs to be fixed — they are what estimation sounds like.

What you get back

Four tracks: vocals, drums, bass, and everything else. That last one is not a mistake in the model — it is the category for guitars, keys, strings, synths and anything that is not one of the other three. If your song is built on a piano, the piano is in "other".

Every stem is the full length of the original and lines up with it exactly, so you can load them into any editor and they will sit in sync.

How well does it work? We measured it

"It ran and produced four files" proves nothing at all — four files of plausible-sounding mush would look identical. So we built a mixture whose ingredients we knew exactly: a speech recording, a 60 Hz sine wave standing in for bass, and short noise bursts standing in for drums. Then we correlated each output stem against each known ingredient.

Each stem matched its own ingredient at almost exactly 1.0 and the others at almost exactly 0. The drum stem contained the noise bursts and essentially nothing else; the bass stem contained the sine wave and essentially nothing else.

That test proves the separation is real and correctly wired. It does not prove any given song will come apart cleanly — real music is much harder than a synthetic mixture, and a heavily layered or heavily compressed master is harder still.

It is slow, and that is a fact about the method

Separation runs the whole waveform through a neural network. We measured 7.8 seconds of audio taking 15.1 seconds to process — roughly twice real time. A full song therefore takes more than six minutes.

That is why the tool works on a 30-second preview and tells you the number before you press anything, rather than starting a six-minute job in a browser tab and hoping you stay.

Separating something does not give you the right to use it

Pulling a vocal out of a commercial record does not change who owns the recording. Practising over an instrumental, studying an arrangement or making something for yourself is one situation; publishing or distributing the result is a different one, and separation does not affect that at all.

Nothing you load here leaves your machine — but that is a statement about privacy, not about licensing.