Last updated 25 August 2026
Taking the singer out of a finished recording used to be either a studio job or a trick that only worked on a handful of songs. It is now something a phone can do in the time it takes to read this page. Here is what it really does, what it cannot do, and how to try it on a song of yours.
A finished song is one stereo file. The voice is not stored separately in it. It has been mixed together with the drums, the bass and everything else, and once mixed, it cannot be subtracted back out by arithmetic.
What can be done is separation: a model that has heard an enormous number of songs, and their unmixed parts, guesses what the voice must have been and what the rest must have been. It gets two new files out of one. One of them is the backing track you can sing over.
This is a guess, not a recovery. It is a very good guess on most music, and a poor one on some. That is the honest frame for everything below.
The classic advice is to subtract one stereo channel from the other, "left minus right", sometimes called centre cancellation. Lead vocals are usually mixed dead centre, so they appear equally in both channels and cancel out.
It does work, a bit, and it costs nothing. But it takes out everything in the centre along with the voice: the kick drum, the snare, the bass, often the main guitar. What is left is a thin, hollow-sounding track with no bottom end. And the moment a song has any reverb on the voice, which is nearly all of them, the reverb tails are not centred, so they stay behind and the singer haunts the result.
A separation model has none of those problems, because it is not doing arithmetic on channels. It is identifying a voice.
Most vocal removers on the web work by uploading your file to a server, running the model there and sending the result back. That is a reasonable engineering choice and a bad privacy one: your music, and whatever else is in that file, goes to somebody else's computer.
It is no longer necessary. The same class of model runs on the phone itself. In Homeoke the whole thing happens on the device:
On a recent iPhone a four-minute song is fully separated in around twenty seconds. On an iPhone XS, the oldest phone this runs on at all, it is closer to real time, so it keeps up with playback but not much more. Either way the result is kept, so a song you have sung before starts instantly the next time.
Nothing about the audio leaves the phone at any point. The only network request is the song's title and artist, sent to a public lyrics catalogue to find the words.
If you want a good result, this is the part worth knowing. Separation is easiest when the voice is easy to tell apart from everything else.
There is no setting that fixes a hard song. If a track comes out badly, the useful move is to try a different version of it: a live recording, an acoustic version, or a different master, will often separate much better than the one you started with.
Homeoke gives you the backing to sing over, and uses the separated voice internally for two things: working out the range the song asks for, and pulling your own voice back out of a recording made without headphones. It does not export an isolated vocal track as a file; it is built for singing, not for remixing.
Separating a song you already have, to sing along with at home, is the ordinary private use that copyright has always left alone. Publishing the result, whether that is the instrumental or a cover recorded over it, is a different question with a different answer, and the answer depends on where you are and on the rights holder. Worth knowing before uploading anything.
Free public beta, on any iPhone running iOS 17 or later. No account.
Join the beta on TestFlight