Vocal remover: take the vocals out of any song in your browser

A vocal remover used to mean a phase trick: flip one stereo channel, sum the two, and whatever sat dead-center — usually the voice — cancels out. It also cancels the kick drum, the bass and anything else mixed down the middle, which is why those instrumentals sound hollow. SplitVocals does not filter anything. It runs Demucs v4, a hybrid-transformer model trained to recognise the singing voice as its own sound source inside a finished mix, and rebuilds the vocal and the rest of the band as separate audio streams.

Because the voice is separated rather than cancelled, the instrumental keeps its full low end and its center-panned instruments, and the vocal comes out as a usable stem of its own. One pass gives you both: download the instrumental, the isolated vocal, or every stem the model produced (vocals, drums, bass and other).

Both pages run the same separation; they are written for different people. The karaoke maker is for singers who want a backing track to sing over tonight. This page is for everyone else who needs the vocals gone: producers who want a clean instrumental to sample, chop or remix; video editors who need a song's music bed without its lyrics under a voice-over; DJs building an edit that drops the vocal for eight bars; guitarists and pianists who want to play the melody themselves over the original rhythm section. If you need the opposite — just the voice — the acapella extractor pulls that from the same model.

The model removes every voice it finds, lead and backing vocals alike, so a harmony stack goes with the lead rather than staying in the instrumental. What it cannot do is undo the recording. A long reverb or delay tail on the voice is part of the vocal source and mostly leaves with it, but on very wet mixes a faint wash can remain. Live albums are the hardest case: the singer bleeds into the drum overheads and the crowd sings along, and both are voices. Heavily compressed masters and low-bitrate MP3s also give the model less to work with.

If an instrumental does not come out clean, start from the studio version rather than a live or remastered edit, and feed the model the best file you have — a WAV or FLAC over a 128k MP3. The HD Pack and Pro run a larger model on a server for the tracks that need it, with noticeably less bleed.

Every step — decoding the file, running the model, mixing the instrumental — happens inside your browser tab through WebGPU or WebAssembly. Nothing is uploaded, so there is no upload time for a large WAV, no queue behind other people's jobs, and no per-minute meter. You can check for yourself: open the Network tab of your browser's developer tools during a separation and you will see no request carrying your audio.

The first run downloads the ~170 MB model into the browser cache; after that, separations start immediately, even offline. Chrome and Edge use your graphics card and finish a typical song in one to three minutes; Safari and Firefox take the CPU path, which is slower but produces the same output.