Questions and answers

Two files if you choose Vocals, Drums or Bass — the part on its own and the song without it. Eight files if you choose Everything. The full breakdown is on the stem splitter page.

No. It runs in the browser. The separation happens on our servers, so nothing is installed and your own hardware does not matter.

Most common audio formats work — including lossless ones. The ceiling is 50 MB per file and ten minutes of running time, which covers essentially all normal repertoire. Everything comes back to you as 320 kbps mp3.

Usually under a minute. The vocals-only option uses a lighter model and is the quickest; the four-part split does more work and takes longer, particularly on long tracks.

It is how the daily limit is tracked without asking anyone for payment details. You get a six-digit code, there is no password to create, and the browser stays signed in afterwards.

On a modern, well-mixed studio recording it is clean enough to perform or produce with. On live recordings, bootlegs, lo-fi sources and heavily effected vocals it is noticeably weaker. It is a prediction of what the separate parts were, not a recovery of them.

An MDX-Net model for the two-way voice-and-backing split, and Meta's Demucs for the four-part split. Both are open source. There is more detail on the Demucs page.

Usually reverb, which sits in the same space as the instruments around it, so removing all of it would take music with it. Frequency overlap is the other cause: cymbals spread across the whole spectrum, and the kick drum and bass guitar share the low end.

Yes, noticeably. The model can only work with detail that survived the original encode, so a wav or a 320 kbps file gives a better result than something ripped at 128 kbps.

Because the model separates four musical roles, not one file per instrument. Guitars, keys, synths and strings all arrive together in the "other" stem. The reasoning is explained on the isolate instruments page.

No. Your upload is deleted the moment separation finishes, and the separated files are erased once your download links expire. Nothing is kept afterwards.

No. Nothing you upload is used to train or fine-tune any model. The models are pre-trained and fixed.

Your email address, and basic job information such as the filename and track length. Nothing else. The privacy policy is two sentences long for a reason.

Five per device per day, with a wider cap per network. It exists because GPU time costs real money, not to push you toward a paid plan — there is no paid plan.

There is no trial, no card and no expiry. The daily cap is what makes free sustainable: separation runs on GPUs that spin up only while a song is processing, and nothing is stored afterwards.

The daily cap, and an email address — which also means occasional emails about new features, with a one-click unsubscribe. That is the whole arrangement.

Only with permission from whoever owns the original recording. Separating a track is free and takes seconds; it creates no rights in the underlying music and changes nothing about who owns it.

Yes. Practising along to a drumless track, studying an arrangement or transcribing a bassline is ordinary private use. Distributing the resulting files is the part that needs permission.

Playing recorded music in public is a licensing question between the venue and the local collecting society, and it works the same way whether you play the original or a version with the vocal removed.