1/3
The mixing desk: a song has 4 tracks
Studios record each part separately, then mix them into one file. Press play, then switch tracks on and off to hear each part.
2/3
Spectrogram and mask
The machine turns sound into a spectrogram: time across, pitch up the side. The model gives every cell a number from 0 to 1 called the mask: the voice's share of that cell.
voicemusicmixed
Tap a cell on the spectrogram to see how the machine splits it.
Quick quiz: A cell has energy 8 and a vocal mask of 0.5. How much goes to the voice?
3/3
The splitter: make your own karaoke
The model learns from songs that come with separate tracks. More training, sharper masks. Pick what to hear, drag the training level, then try making it harder.
Three takeaways
The machine cuts sound into thousands of cells by pitch and time.
The model guesses the voice's share of each cell, learned from songs with separate tracks.
Singing at home isn't posting online. Only split recordings you have the rights to.
Try it: record yourself singing over background music, open the vocal-separation feature in an audio app, and listen for what leaks into the vocal track.
Watch the video
Found this useful? Send it to a friend
The song on this page is a made-up tune your own browser generates; no real song or singer is used. The splitter here is a simulation of how masks work, not a real AI; the cell numbers are the examples from the video. Nothing is sent anywhere.