A remix acapella is a stem, not a magic dry vocal
You get the voice the model could separate: lead, and usually the doubles and ad-libs that were sung with it. You do not get a studio multitrack, and you do not get a vocal with the original compressor and room printed out of it.
That is still the useful file. Drop it on a new drum bus, check the key, and treat the leftover bleed as something to EQ rather than as a failed export.
How clean it will be, by the source
Start from the best file you have. The model cannot invent detail an encoder deleted.
| Source | Vocal stem | Watch for |
|---|---|---|
| Dry modern lead | Usually the cleanest | A hat in the sibilance |
| Stacked pop vocal | Lead plus doubles | You cannot split one singer from another |
| Heavy reverb or delay | Room stays on the voice | Try a drier edit of the song if you have one |
| Rap with ad-libs | Main plus the answers | The beat should be gone, the doubles will not be |
| Lo-fi or distorted vocal | More crumbs | Do not expect a polished a cappella |
Putting it on a new beat
Download WAV, not a low-bitrate MP3, if you are about to process the vocal again. A second lossy encode on an already separated stem is where the swishy top comes from.
Small tempo moves belong on the BPM changer, in keep-pitch mode. A semitone of key belongs on the key changer. Warping a whole vocal by a lot will sound like a vocal, just a worse one. If you also want the original instrumental as a reference, it is the other three stems in the same session.
Private by default, still not a sample clearance
The song never uploads. That matters for unreleased vocals and client files. It does not clear the recording for a release. If the voice is not yours, a remix still needs permission.
The separation model is about 166 MB and stays on the device after the first download. Phones ask first. Keep the tab open while it runs.