Vocal Remover – Make a Karaoke Track or Acapella
Split any song into instrumental and vocals with AI, right in your browser. Runs on your device · nothing is uploaded.
Processed on your device · no uploadDrop a song here or choose a file
MP3, WAV, M4A, OGG, FLAC · stays on your device · or paste with Ctrl+V
What to do next
How to use Vocal Remover
Add your song
Drop an MP3, WAV, M4A, OGG or FLAC file. It is decoded on your device and never uploaded.
Pick a method
AI (Demucs) gives the cleanest result. The first run downloads the model once (about 180 MB) and keeps it for next time. Quick mode is instant but rougher.
Wait for the split
A progress bar shows each part of the song as it is separated, with the time left. You can cancel at any moment.
Compare and download
Switch between Original, Instrumental and Vocals while the song plays, then save the karaoke track, the acapella, or both as a ZIP in MP3 or WAV.
Frequently asked questions
Which AI model does this use?
Demucs v4 (HTDemucs) by Meta AI Research, released under the MIT licence, converted to ONNX and run with onnxruntime-web (MIT). The model file is hosted on the Hugging Face Hub (timcsy/demucs-web-onnx) and the runtime files on jsDelivr. The model is downloaded once (about 180 MB) and cached by your browser.
Is my song uploaded?
No. The song is decoded, separated and encoded in your browser. Only the model file is downloaded; your music never leaves your device.
How long does it take?
It depends on your computer. The AI model is heavy: a computer with WebGPU is fastest, otherwise it runs on the CPU and a three-minute song can take several minutes. The progress bar shows the time left. Quick mode takes a few seconds.
Will the vocals be removed completely?
AI separation is very good but not perfect: expect faint traces of reverb or backing vocals in the instrumental, and a few drum or synth sounds in the acapella. Quick mode only removes what sits in the centre of the stereo image, so it leaves more of the voice and can thin out the bass and kick.
Does it work on a phone?
The AI model needs a lot of memory, so it is meant for computers. Phones and low-memory devices start in Quick mode. Songs can be up to 12 minutes with AI and 30 minutes in Quick mode.
Can I use the results commercially?
The tool and the model are free to use, but the song itself is still owned by its artist and label. Check that you have the right to remix or perform it before you publish the result.