Remove Vocals

AI separation runs in your browser with one of two models, trained to tell a singing voice from the instruments in a song. Best uses UVR-MDX-NET Inst HQ 5 from Ultimate Vocal Remover (MIT licence, about 56 MB). Fast uses Spleeter (MIT licence, by Deezer; about 38 MB, in the form the sherpa-onnx project publishes). The model you pick downloads once and then stays there. The vocals are the song minus the instrumental.

With Best, the voice is gone from most pop and rock songs, with faint traces at most in loud or reverb-heavy passages. Fast leaves more of the voice behind. With either, backing vocals and choirs may stay partly in the instrumental. The quick method is rougher and also removes anything else panned to the centre, such as snare drums or solo instruments.

On a fast computer, Fast separates a 4-minute song in about a minute and Best in about 9 minutes. Phones and computers without a graphics card the browser can use take longer. The page shows how long is left. Songs up to 8 minutes are separated (4 on phones).

For your own practice, karaoke or study, yes. Sharing or publishing separated tracks of someone else’s song needs the rights holder’s permission, as with the original recording.

No. Your songs are opened and processed on your device, and the results are saved straight back to it. Nothing is uploaded. How to check.

It is free to use and there is nothing to sign up for. There are no watermarks and no daily limits.