Vocal Remover
Take the vocals out of any song for an instrumental or karaoke track — or keep only the vocals. Separated in your browser, so the song never leaves your device.
Split into
- Vocals
- Drums
- Bass
- Other
120 MB download · fastest
Drop a song here
MP3, WAV, M4A, FLAC or OGG, up to 10 minutes. It never leaves your device.
Your song stays in this browser and is never uploaded. The model (HT-Demucs (Meta AI), MIT) downloads from Hugging Face, and its runtime from jsDelivr, the first time you use it. Both see your IP address — the audio is not sent. With WebGPU a three-minute song takes about a minute with 4 stems and two with 6; without it, several times that. With 6 stems, guitar means acoustic and electric together, and synths stay in "other".
Table of contents
Why the old vocal-removal trick never really worked
The classic vocal remover — in old karaoke machines and in audio editors as "centre channel removal" — subtracts the right channel from the left. Lead vocals are usually mixed dead centre, identical in both speakers, so they cancel. So does everything else in the centre: the kick drum, the snare, the bass. What survives is the reverb on the voice, the stereo guitars and a hollow, phasey ghost of a band. It also folds the result to mono.
This page does something different in kind. A neural network trained on multitrack studio recordings listens to the whole song and estimates which parts of the sound are a voice, wherever they sit in the stereo field. The instrumental it leaves behind keeps the kick, the snare and the bass in the middle where they belong, and stays in stereo.
It is still an estimate, and it is worth being straight about that. A recording is a mix, and un-mixing it cannot be perfect: the room around a voice, a harmony doubled by a synth, a vocal effect that sounds like an instrument, can all land on the wrong side of the line. On most modern productions the result is clean enough to sing over; on a live or heavily reverberant recording, expect more of a ghost.
Making a backing track people can sing to
Check the quiet passages, not the chorus. Leftover vocal is masked when the whole band is playing and exposed in a sparse verse or an intro. Play those before you download — the waveforms show you where they are.
Keep a guide vocal for rehearsals. Instead of removing the vocal, take it down to 15–20%. Singers learning a part get the melody to lean on, and you can make the fully instrumental version from the same separation for the performance itself.
For worship and cover bands, remove what you will play live. The same separation gives you all four stems. Mute the vocals and the drums, say, and your singer and drummer have the rest of the band to rehearse over — the original recording, not a MIDI cover.
Use the best source file you can. An instrumental made from a lossless or high-bitrate file is noticeably cleaner than one made from audio ripped from a video, because compression throws away the detail the model uses to tell a voice from a guitar.
Frequently asked questions
Drop the song in and wait for it to separate — the "No vocals" mix is already selected on this page, so press play to check it, then Download instrumental. That WAV is the original drums, bass and everything else, with the vocal stem taken out.
Yes. Choose "Vocals only" and the download button saves the a cappella. The same separation produces both, so switching between them is instant; nothing runs again.
Mostly, yes — the model treats every sung voice as vocals, lead and backing alike. That is usually what karaoke wants. If you want to keep the harmonies, there is no way to split the lead from them here; they arrive in the same stem.
Reverb and delay on a voice spread it across the mix, and some of that tail ends up counted as "other". On most modern productions it is quiet enough to sing over. If it bothers you, take the other stem down a little rather than off — that hides the ghost while keeping the arrangement.
Not here, and on purpose: pitch-shifting a mixed track well is a separate, hard problem, and doing it badly ruins the instrumental this page just made. Download the WAV and shift it in a DAW or audio editor, which do it properly. For chord charts, the chord transposer does change the key.
No. The model runs in your browser on your own hardware; the song is decoded, separated and saved without ever leaving the tab. The only download is the model itself, about 120 MB, fetched once and then cached.
Free of charge and of watermarks, yes, and the tool asserts no rights over it. The song still belongs to whoever owns it, though: an instrumental of a commercial track is fine for singing along or practising, and needs permission to publish.
The Zeal Promise
Non-negotiable, on every tool we ship.
- No sign-up — ever
- No watermarks on any output
- No ads near the download button
- Works fully on mobile
- Privacy-first: processing stays in your browser
- Open source (MIT) — audit it, contribute, self-host
- Zero runtime dependencies in tool logic