Take the voice out of a song, or split it into vocals, drums, bass and everything else. Runs on the server, nothing to install.
Drag & drop a song or video or click to browse
MP3, WAV, FLAC, M4A, AAC, OGG, Opus and any video ffmpeg reads — up to 200 MB
Upload a track and get the voice on one file and the backing on another — a karaoke version, an acapella, or the four separate stems if you want to work with them. The separation is done by htdemucs, a hybrid transformer model from Meta's Demucs project, running on our server rather than in your browser, because a four-minute song takes about four minutes of real computation and no phone is going to do that quietly.
| File | What is in it | Use it for |
|---|---|---|
| Vocals | The lead and backing voices, alone | Acapellas, sampling, checking a mix, transcribing lyrics |
| Instrumental (no vocals) | Everything except the voice | Karaoke, covers, backing tracks, video beds |
| Drums | Kick, snare, hats, percussion | Drum replacement, practising along, sampling a break |
| Bass | Bass guitar and synth bass | Learning a bass line, isolating the low end |
| Everything else | Guitars, keys, synths, strings — the rest | Rebuilding a mix, lifting a riff |
Source separation is very good now and it is not magic. What you should expect:
Yes, and it is the most common reason people come here. Choose the karaoke option and you get two files: the voice on its own, and the whole backing track with the voice taken out. The second one is your karaoke version. It is not the official instrumental the label released — it is the released mix with the singing removed — but for singing along, for a cover, or for a video bed it is usually indistinguishable.
On a clean modern pop or rock mix, good enough that most listeners would not notice anything was removed. Two things reliably leave traces. Reverb and delay on the voice are part of the voice, so a wet mix keeps a faint wash in the instrumental. And loud, dense, heavily-compressed masters give the model less to work with. A live recording of a whole band in one room is the hardest case of all, because the bleed between microphones is real sound, not a mixing decision, and it cannot be undone.
Take the instrumental if you want to sing over it, record a cover, or put music under a video without the original singer. Take the vocals if you want an acapella to sample, to check a mix, or to transcribe lyrics. If you are not sure, download both — they come from the same run and cost nothing extra.
Because that is what the model was trained to produce: vocals, drums, bass, and everything else. Six- and eight-stem models exist that also split out guitar and piano, and they are noticeably less accurate at each one. Four is the split that stays reliable, and "everything else" is a genuinely useful file rather than a leftovers bin.
Any audio file — MP3, WAV, FLAC, M4A, AAC, OGG, Opus — and any video file, since the sound is extracted first and the picture discarded. If ffmpeg can open it, so can this. You choose whether the stems come back as MP3 or as lossless WAV.
Roughly the length of the track itself, plus the upload. A four-minute song takes a little over four minutes. That is honest arithmetic rather than a hedge: separation runs at about real time on the processor this server has, which is why one track is handled at a time and why the page shows your place in the queue instead of a spinner.
Better not to. The job keeps running on the server, but the link to your files lives only in that tab — there is no account to come back to. If you reload or close it you will need to upload again.
Twelve minutes per file, and 200 MB. Free separations are capped at six minutes on top of the daily allowance. For anything longer, split the file first — most editors can export a section.
The audio is extracted, the original upload is deleted as soon as the stems exist, and the stems themselves are deleted from the server a few hours later so you have time to come back to the tab. Nothing is published, indexed or shared, and there is no account holding your files.
That depends on the song, not on this tool. Separating a recording does not create a new copyright and does not give you a licence to the old one: an instrumental made from a commercial release is still that release. Singing over it at home, practising, or studying an arrangement is one thing; publishing, monetising or distributing it is a question for the rights holder. Your own recordings, public-domain material and anything you have cleared are yours to do as you like with.