How to Extract an Acapella from a Song, and What Comes Out
A cappella is Italian for "in the manner of the chapel", and for four hundred years it meant singing with no instruments at all. Producers borrowed the word in the eighties for something narrower: the vocal of a record on its own, lead, doubles and harmonies, with the band stripped away, pressed on the B-side of a 12-inch so a DJ could lay it over another beat. The whole mashup craze ran on them. When the Freelance Hellraiser put Christina Aguilera's Genie in a Bottle vocal over the Strokes' Hard to Explain in 2001, the vocal came off an official a cappella, and so did most of what 2 Many DJs played the year after. Singers use them for a different reason, to hear the phrasing with nothing in the way, and anyone transcribing a vocal line wants one for the same reason. There are three ways to get one from a song you already have.
The three ways
| Method | Needs | Quality | Where it fails |
|---|---|---|---|
| Official a cappella | The label or the artist to have released one | A studio stem, nothing extracted | Dance singles, remix packs and DJ pools, and not much else |
| Phase cancellation | The full mix and the official instrumental of the same master | Clean when the two line up to the sample | A few milliseconds of offset, or a level mismatch, and it turns into a comb filter |
| Stem separation | The song file | The voices estimated as one stem, a faint band residue in the reverb | Live recordings, heavy reverb, a voice buried in distorted guitars |
The official one
Twelve-inch singles from the eighties and nineties often carried one, remix packs and DJ pools still do, and a few labels now sell stems outright. Outside dance music and pop remix culture, though, hardly any record has one, and the catalogue thins out fast once you leave the charts.
Cancelling the instrumental
If you have the full mix and the official instrumental from the same master, invert the polarity of the instrumental and play the two together: everything they share cancels, and the voice is what is left. It is the oldest trick in the remixer's kit and, when the files line up to the sample at the same level, it is cleaner than anything a model can do, because nothing is estimated. A streaming rip and a CD rarely match, a few milliseconds of offset turns the cancellation into a comb filter, and the instrumental has to exist in the first place, which brings the coverage problem straight back.
Separation
Since 2019, when Deezer published Spleeter and Facebook's research lab published Demucs, there has been a third way that needs nothing but the song. These are neural networks trained on multitrack recordings to listen to a mix and write out its parts; the six-source Demucs model on this site returns vocals, drums, bass, guitar, piano and a sixth stem for the rest. Solo the vocal stem and you have the a cappella. The acapella extractor page runs exactly this.
- Drop in the song, MP3, WAV, FLAC or M4A, up to 10 minutes. A lossless file gives a cleaner vocal than a low-bitrate MP3.
- Wait two to five minutes for the split on a cloud GPU.
- Solo the vocal stem; the other five mute.
- Export a 24-bit WAV at the sample rate of your file, or keep working in the browser.
What lands in the vocal stem
| In the vocal stem | Not in it |
|---|---|
| The lead vocal | Any instrument |
| Doubles and harmonies | A vocoder line, which the model usually files as an instrument |
| Ad-libs and spoken parts | A chopped vocal sample used as a hook, sometimes filed the same way |
| Delay throws and the reverb on the voice | |
| Breaths and mouth noise |
Every voice on the record goes into that one stem, so a stacked chorus comes out as the stack it was recorded as. Separating the lead from the harmonies needs the multitrack, and no model does it from a stereo mix, whatever the marketing says.
What stays in
The room, first. The reverb on the voice carries a faint print of the band, because the same reverb covered the whole mix; on a dry studio vocal it is barely there, on a live recording it is most of what you hear underneath. Then the bars where the voice sits in the same range as a piano or a clean guitar and shares harmonics with it, where a trace of the instrument stays under the vocal. And the breaths, which were always there but were covered by the band, and now are not, which for a remix is often exactly what you wanted and for a transcription is just noise. A dry, centred, well-recorded vocal from a modern studio mix comes out cleanest; a live album, a lo-fi recording or a voice buried in distorted guitars leaves more behind.
Studying it, not just downloading it
A download page gives you a file. A mixer gives you a few things to do with the vocal before it becomes one, and none of them needs a DAW.
| Move | What you hear |
|---|---|
| Slow it to 70 percent, pitch untouched | Each note in a run, on its own |
| Loop one line | The breath before it, the slide into the first word, the push ahead of the beat |
| Bring the piano back at a low level | Where the singer sits against the chord |
Using it
For practice, transcription and learning a part, the file is yours to use at home. For a remix, a mashup or a sample in a release, the rights belong to whoever owns the recording, and an extracted vocal is treated the same way a sample is: the Freelance Hellraiser's bootleg was never officially released for that reason, and the tool does not change the law either way.
Frequently asked questions
Is an acapella the same as isolated vocals?
Yes. Isolated vocals, vocal stem and a cappella all mean the voice of a record on its own. The spelling acapella, one word, is the one people type into search, which is why the tool page uses it.
Can I get only the lead vocal, without the harmonies?
No. The vocal stem holds every voice on the record, and no current model separates the lead from the harmonies from a stereo mix.
Why does the a cappella sound watery in places?
Where the model was unsure whether a sound was voice or instrument, it splits the difference, and that shows up as a soft, phasey texture in those bars. It is more audible on dense mixes and heavily reverbed vocals, and much less on dry studio vocals.
Which is better, cancellation or separation?
Cancellation against an official instrumental of the same master is cleaner when the files line up, because nothing is estimated. Separation needs only the song, works on any file and returns the other parts on their own faders as well. Since the instrumental is rarely available, separation is the one most people end up using, and the same split gives you the other half of the record through the instrumental maker.
Is it free?
Two songs a day per person, no account, no watermark, a full-length 24-bit WAV export, while the site is in beta. A site-wide daily cap applies on top, because separation runs on paid GPUs.