Two questions come before anyone tries this: how long am I going to sit here waiting, and where does my music go while it happens. Both have short answers. The numbers below were measured on real phones rather than estimated, which is worth saying because almost nobody publishes them.
Separating a song is faster than playing it, on anything recent. Measured on an iPhone 14 Pro, a four minute song comes out in about twenty seconds. Measured on an iPhone XS, which is 2018 hardware, the same song takes close to three minutes. Every phone anyone is likely to be holding sits somewhere between those two, much nearer the first.
And it happens once. The separated backing is kept, so the second time you open that song there is nothing to wait for at all.
The song is separated in chunks, front to back, and playback starts as soon as there is enough of a head start. On a recent phone that is about half a second, so the twenty seconds above is not a wait you ever sit through. The question that matters is not how long the whole job takes, it is whether the job stays ahead of the playhead.
On a fast phone the lead grows the whole way through and you never think about it. On old hardware it is close. An XS separates a song only slightly faster than it plays it, so it is barely ahead, and a phone that has been working all day is slower than the same phone was that morning. One measured run on a warm XS took the whole six minutes to separate a six minute song, which is no lead at all, and the playback fell behind partway through. The separation itself finished and the whole song was kept, so the second play was perfect. It was the singing that got interrupted.
All of this runs on the phone. No file is sent anywhere, there is no account, and it works with the network off entirely.
That was not a principle applied on top of the design, it is what the measurement allowed. The original plan for this was a server, because that is how everyone does it, and a server means uploading your music, a bill per song, and somebody holding a copy of your library. Then the model turned out to separate a four minute song in about twelve seconds on a phone, and the whole server disappeared from the plan in an afternoon. The privacy is a consequence of the arithmetic.
It is worth being clear about what that means in practice. The music you put into this is mostly music you paid for, and it stays on the device you paid for it on.
It runs. It runs close to the line, which is a different thing from not running, and it is worth knowing which one you are in before you decide something is broken. On hardware from around 2018, expect to give a long song a moment before you start singing rather than pressing play and going straight in, and expect a warm phone to be slower than a cool one. After that first pass the song is kept and none of it applies again.
If you want the mechanics of what the separation is actually doing while you wait, that is a separate guide.
On the App Store, for any iPhone running iOS 17 or later. No account, and nothing is uploaded.
Get it on the App Store