The song is split on your device. Nothing is uploaded, and it works offline once the page has loaded.
How it works
In most studio recordings the lead vocal is mixed equally into the left and right channels, while guitars, keys and backing parts are spread to the sides. The song is cut into short overlapping slices, and in each slice every frequency is checked for how alike the two channels are. Sound that is the same on both sides, within the range of the human voice, is lifted out into the vocal track. Everything else stays in the instrumental, and the lowest frequencies are left untouched so the bass and kick drum survive.
What to expect
Results are good on many pop, rock and hip-hop mixes and weaker on live recordings, songs with heavy reverb on the voice, or arrangements with a solo instrument in the centre. Some echo of the voice often remains in the instrumental, and the vocal track carries a little of the centred instruments. Machine learning stem splitters can do better but need a large model download and a powerful device; this tool stays small and fast.
Questions
Why does it need a stereo file?
The method compares the left and right channels. A mono recording has nothing to compare, so the voice cannot be told apart from the music.
MP3 or WAV?
MP3 at 192 kbps is small and plays everywhere. WAV is uncompressed and best for further editing.
Can I use the result?
Use it for practice, karaoke at home or study. Sharing a song you do not own the rights to may need the permission of the rights holder.