Well, luckily, we also got really good at automatically fingerprinting music, why not just two-stage it?

If the music has higher than a certain uniqueness threshold, bin it and train away from it.