What this actually does, and what it does not
A lead vocal is nearly always mixed dead centre, which means it appears identically in the left and right channels. Subtract one channel from the other and anything identical in both disappears. That is the whole method, it is older than digital audio, and on the right song it works startlingly well.
It is not separation. Nothing here identifies a voice or understands what a voice is — it removes whatever was centred, which on most records also means the kick drum, the bass, and often the snare. Tools that genuinely separate a vocal use a trained model of tens of megabytes and a server to run it on. This runs instantly on your own machine, and the difference in results is the price of that.
Keeping the bass
Losing the low end is what makes a naive centre-cancel sound hollow rather than instrumental, so the bass is added back from the original mix below a frequency you choose. Two hundred hertz suits most music: low enough that the voice does not come back with it, high enough to recover the kick and the bass line.
Raise it and the track gets fuller but the vocal starts to reappear as a low mumble. Lower it, or switch it off, when the singer is a bass or the song is mostly low male voice — there the two overlap and something has to give.
Which songs work
Wide stereo mixes with a clearly centred voice: most pop, rock and dance from the eighties onwards. What defeats it is a vocal with heavy stereo reverb, a double-tracked vocal panned apart, a mono recording, or anything mastered so hard that the channels are nearly identical anyway.
Preview before you render — it takes a second and immediately tells you whether this particular song is one of the ones it works on. If the result is silent, the file was mono; if the voice is still there, it was never centred.
Frequently asked questions
Why can I still hear the singer?
Because the voice was not centred, or not only centred. Stereo reverb, double-tracking and any panning leave a difference between the channels, and this method removes only what is identical in both.
Why did the drums and bass disappear too?
They were centred as well, which is normal. That is what the bass-keeping control is for — 200 Hz brings back the kick and the bass line without bringing the voice with them.
My file came out silent.
It was mono. Both channels were identical, so subtracting one from the other leaves nothing. The tool warns about this as soon as the file loads.
Is this the same as an AI stem splitter?
No, and it is worth being clear about. A stem splitter uses a trained model to recognise a voice; this cancels whatever is in the middle. The splitter gives better results and needs a large download and a server — this is instant and private, and sometimes that is the trade you want.
Is the result mono?
Yes. The output is the difference between the two channels, which is a single signal by definition. Centre cancellation always costs the stereo image.
Something wrong with this tool, or an idea for it? Tell us