How to use Vocal remover
- 1
Drop a song or a video
Pick an MP3, WAV, M4A, FLAC or OGG file, or a video whose soundtrack you want. Rarer formats such as AVI or WMA are decoded by the built-in ffmpeg engine instead of the browser.
- 2
Choose 2 or 4 stems
2 stems gives a vocal track and an instrumental. 4 stems splits the music further into drums, bass and the rest. The first run downloads the model once: 39 MB or 79 MB.
- 3
Press Separate
The song passes through the model in overlapping 12-second blocks. The bar shows how far it has got and roughly how long is left, and Cancel stops the work at once.
- 4
Listen, mute and solo
The instrumental plays first with the vocals muted. Mute or solo any stem, or click a waveform to jump there; all stems stay in sync while you switch.
- 5
Save WAV or MP3
Download the instrumental, any single stem, or every stem in one ZIP, as lossless 16-bit WAV or as a 320 kbps MP3.
Limits and Pro
Songs up to 10 minutes and 500 MB; from a longer file the first 10 minutes are separated. Phones and devices with 4 GB of memory or less separate up to 5 minutes at a time (3 minutes with 4 stems).
Up to 20 minutes per song with 2 stems and 12 minutes with 4 stems on a computer. See Pro
About Vocal remover
Vocal remover splits a finished song into separate tracks, called stems, with Spleeter, a source-separation model that Deezer's research team trained on its own licensed music catalogue and released under the MIT licence. In 2-stem mode you get the vocals and everything else, which is the instrumental or karaoke version of the song. In 4-stem mode the music is split again into drums, bass and the remaining instruments. Stems come out at 44.1 kHz in stereo (a mono recording stays mono), and played together they add up to the original recording exactly.
People use it for karaoke and singing practice over the original backing, a "minus one" track to play their own instrument along with, a cappella vocals for a remix or mash-up, and picking out a bass line or drum pattern by ear.
The model works on a spectrogram: for every slice of time and pitch it estimates how much belongs to each stem. It decides up to about 11 kHz; above that each stem gets its average share, so cymbals and the airy top of the mix are kept rather than cut off. Results depend on the recording. A clear lead singer over a band separates well. Reverb tails, doubled backing vocals, choirs and heavily processed voices often leave a faint ghost in the instrumental, and instruments in the vocal range, like a lead guitar or a saxophone, can bleed into the vocal track. It cannot take one singer out of a duet, change the key or tempo, or remove a single instrument other than drums or bass.
Privacy
The song is decoded and separated by a background worker inside this browser tab, so the audio never leaves your device and no account is needed. Only program files are downloaded: the ONNX Runtime engine from jsDelivr and the Spleeter model from Hugging Face (huggingface.co). The model stays in this browser's storage for the next song until you press Remove; the stems exist only in this tab until you save them.
Features
Separated on your device
Spleeter runs through ONNX Runtime in a Web Worker, on the graphics card with WebGPU or on the processor with WebAssembly.
Mute and solo while it plays
Every stem has its own waveform, mute and solo button, so you hear the karaoke version, the vocals alone or any combination before saving.
WAV, MP3 or a ZIP of all stems
Save lossless 16-bit WAV for editing, or 320 kbps MP3 for phones and karaoke apps.