A scale-free distribution of false positives for a large class of audio similarity measures