Vocal Remover
Separate a song into vocals, drums, bass and other, and keep the instrumental. Everything runs on your device.
MP3, WAV, M4A, FLAC, OGG, Opus, AIFF, WMA, AMR and the sound of most videos. Files stay on your device.
How it works
- Choose a song.
- Pick AI separation (best quality, one-time model download) or the instant center-channel mode.
- Wait for separation to finish; progress and cancel are shown.
- Download the instrumental, or all four stems.
Two ways to remove vocals
AI separation uses HTDemucs, an open model by Meta (MIT licence), running in your browser. It splits a song into vocals, drums, bass and other, and the instrumental is everything except the vocals. The model is about 172 MB and is downloaded once, only after you agree. It is then kept in your browser cache, so the next visit does not download it again. You can delete it from the page at any time.
Instant mode needs no download. It cancels whatever is identical in the left and right channels, which is where lead vocals are usually placed. It is quick but crude: it only works on stereo mixes with a centered voice, it removes every centered instrument (kick, bass, snare and often a lead guitar) along with it, and the result is mono. The option to keep the low end brings back the bass and kick.
Speed, memory and quality
Separation is heavy. With WebGPU (recent Chromium browsers with a supported graphics card) it can run close to the speed of playback; on the CPU expect between half and several times the length of the song, depending on how many threads the browser can use. It needs about 1 GB of memory for the model plus roughly 130 MB for every minute of the song, so phones and very long files can fail. The page shows an estimate for your device and your file before you start, and you can cancel during the download and during separation.
The result is good but not perfect. On dense mixes some of the voice can remain, or the instrumental can sound a little thin or watery, and the stems do not add up exactly to the original. A very short clip (under a second) gives poor results.
Output
You get the instrumental and the vocals as separate files in the format you choose, and can also save drums, bass and other. Download them one by one or all together as a ZIP. If a stem would clip after separation, all stems are lowered by the same amount, and the page tells you by how much. Only use songs you have the right to edit.
Questions
Why does it ask me before downloading?
The AI model is large (about 172 MB). It is only downloaded after you tick the box and press the button, and you can cancel while it downloads. Nothing about your song is ever sent anywhere.
Will it work offline?
After the first download the model stays in your browser cache, so separation needs no new download. The page itself still has to load, and clearing the site data removes the model.
How do I free the space?
Press "Delete downloaded model" in the model panel, or clear this site’s data in your browser settings.
Which is better, AI or Instant?
AI separation is far better and works on mono files too. Instant mode is only for a quick try on a stereo song with a centered voice, or on devices that cannot run the model.
Why is some singing still audible?
Backing vocals, reverb and heavily processed voices are hard to separate. Try the AI mode if you used Instant, or check that the vocal is not very wide in the stereo field.
Is my audio uploaded?
No. Everything runs in your browser and the file never leaves your device.