Remove Vocals

Remove vocals from a song with AI, free and private

This free vocal remover lets you remove vocals from a song and keep the music: a karaoke track for a party, an instrumental to practise singing over, or a backing for a cover. Drop an MP3, WAV, FLAC or M4A file — or a music video — and pick a method.

AI separation uses HTDemucs, the hybrid transformer model from Meta’s Demucs project, to split the recording into vocals, drums, bass and other instruments, then rebuilds the track without the voice. It copes with reverb, harmonies and even mono recordings. Classic mode is the instant karaoke trick: it cancels whatever sits in the centre of a stereo mix while keeping the bass and cymbals. Both run inside your browser, so the song never leaves your computer.

How to use the Remove Vocals

1. Drop a song on the upload box or click to choose one. Video files work too — their soundtrack is used.
2. Pick “AI separation” for the cleanest result, or “Classic (instant)” for a quick karaoke version with nothing to download.
3. For AI, choose “The whole track” or “First 30 seconds (quick test)”, press “Remove vocals” and confirm the one-time 166 MB model download. The progress bar shows each 8-second part and the time left.
4. Use “Listen and compare” to jump between the original, the instrumental and the removed vocals at the same point in the song.
5. Choose MP3, WAV or FLAC under “Save as” and press “Download instrumental”. “Download vocals” saves the removed voice too.

Remove Vocals features

How the AI vocal remover works

The song is converted to 44.1 kHz stereo and cut into 7.8-second parts that overlap by a quarter, so no syllable is split at a hard edge. HTDemucs hears each part twice — as a waveform and as a spectrogram — and transformer layers let the two views share what they hear. It returns vocals, drums, bass and other; the instrumental is drums + bass + other, cross-faded back into one track. Because the stems add up to the original, nothing else is thrown away.

The model is an ONNX conversion of Meta’s official weights (Ghilda/htdemucs-onnx, MIT), run by ONNX Runtime Web (MIT) with spectrogram code from demucs-web (MIT). In Chrome on a mid-range laptop with integrated AMD graphics we measured about 4.6 seconds per part on the graphics card and 20 seconds on the processor.

The classic centre-cancellation method

Studio mixes usually put the lead vocal in the middle, identical in both channels, while guitars, keys and reverb spread to the sides. The old karaoke trick subtracts one channel from the other; this tool does it frequency by frequency, using 4,096-point spectra 23 ms apart, and removes only what matches in left and right. Bass below 120 Hz and treble above 12 kHz are left alone by default, so the kick, bass guitar and cymbals survive.

In this page’s automatic test, a centred 440 Hz tone drops by about 60 dB while tones panned left, right or wide change by less than 0.1 dB. On real songs the voice gets much quieter rather than vanishing: other centre sounds drop too, and stereo reverb stays as a faint ghost.

Tips for a better karaoke track

• Start from the best copy you have — a lossless or 320 kbps file separates more cleanly than a low-bitrate rip — and run “First 30 seconds” before a long song on a slow computer.
• If a faint voice remains in classic mode, lower “Keep treble above” to about 8 kHz, or switch to AI.
• To sing in a comfortable key, shift the instrumental with the Audio Pitch Changer; the Key Finder tells you the original key.
• Want the voice instead? The Vocal Extractor saves an acapella and every stem.
• A backing track made from a commercial recording is still that recording: use it for personal practice, study and remixes you have rights to. Publishing it or performing it in public needs the rights holders’ permission.