An acapella you cannot beatmatch is a sound effect. The vocal comes out of the mix with the key, the tempo and the grid already attached, so it behaves like a proper record the moment it lands in your library.
A separated vocal is the hardest thing in a DJ library for software to analyse. There is no kick to lock onto, the transients are soft, and long held notes give a beat tracker almost nothing to work with. So the automatic grid lands wrong far more often than it does on a full record, and you find out mid-set.
The fix is not to analyse the acapella at all. The key, the tempo and the beat positions are measured on the full mix, where all the rhythmic information still exists, and then written into the separated stems. The vocal inherits a grid derived from the drums it was sitting on top of.
Separation runs on htdemucs_ft, an open model — the same technology behind the
tools you already know. On a sparse, clean mix the vocal comes out remarkably intact. On a
dense, heavily-compressed master there will be artefacts, and no service on the market avoids
that. Feed it the most lossless copy of the record you own — a 320 MP3 or better — because the
model cannot recover detail the encoder already discarded.
Pull the vocal from a record you actually play. No account, no subscription — €0.10 per minute.
Open the toolUpload the track, and the vocal is separated out and handed back as a lossless AIFF and an MP3 320, already carrying the song's key, tempo and a Serato beatgrid. Credit is €0.10 per minute of the source track.
Separating someone else's record does not give you rights to it. For a bootleg or an edit you play out, that is between you and the venue; for anything you release or monetise you need clearance from the rights holders, exactly as with any sample.
Any common audio file. You get AIFF and MP3 320 back, because those are the two formats Serato reads tags from reliably — writing tags into a WAV corrupts it.