You very much do need ML for that, because the EQ distribution of the music in a video is going to be completely different from the original, and both volume and EQ distribution will vary even within fractions of a second as the sonic environment changes as objects move.

The only time it's closer to trivial to filter out is when the music has been edited in as an audio track, and no additional filters/EQ have been applied. But even then, subtracting a lossless original will leave funky-sounding audio compression artifacts, a "ghost" of the song left behind.