Simulation
Can you pull the voice out of this song?
Karaoke apps and video editors can now lift a singer out of a finished song. Here is a ten-second tune I built from pure maths, with a voice, drums, bass and chords. Your job: keep the voice, lose the rest.
Kit
Everything you hear on this page is synthesized in your browser. No recordings. Tap play when you're ready, and keep the volume low at first.
Spectrogram
Turn instruments on and off to see where each one lives in the picture.
Building the song…
Your mask
Quick mask: keep one range of pitches all the way through
How clean is it?
Turn off everything but the voice and look at its shape. Then switch to Your mask and drop what isn't voice.
How the real thing works
Masks, and how an AI learns to paint them
Your grid has 768 squares. A real separator paints hundreds of thousands per song, each with a value between keep and drop, and it does it in a second or two.
01
A picture of sound
A spectrogram cuts sound into thin slices of time and measures how loud each pitch is in every slice. Instruments leave different marks: drums are short vertical smears, bass sits at the bottom, a voice draws wavy stacked lines.
02
A mask is a stencil
Lay a stencil over the picture, keep what shows through, and turn the picture back into sound. Your mask can only say keep or drop for big squares. Where voice and chords share a square, you have to give up one of them.
03
Learning from separate tracks
To train a separator, people use songs where each instrument was recorded on its own track. The model sees only the mix, guesses a mask, and is corrected by the real voice track. After many thousands of songs it can guess well on songs it has never heard.
04
Where you meet it
Karaoke and music practice apps, video editors that clean up dialogue, and noise removal in video calls use the same family of ideas: find which parts of the sound belong to what you want, and turn the rest down.
Still not perfect
Real separators leave traces: a faint drum behind the voice, or a watery sound where two instruments overlapped. And pulling a voice out of a song doesn't give you the right to reuse it. The song still belongs to the people who made it.