That 20,000 bytes might not be a real difference, as I understand OP. It sounds like it's the XOR or Hamming distance, byte or bitwise - in which case you could

That 20,000 bytes might not be a real difference, as I understand OP. It sounds like it's the XOR or Hamming distance, byte or bitwise - in which case you could get very large differences in the number by very small errors like dropping 1 bit (thereby putting all following bits out of alignment) yet the perceptual difference would be unhearable (and maybe not exist at all depending on how the encoding works). eg 0101010 vs 101010. You'd want to calculate some sort of edit distance which takes into account the possible bit flips and burst errors to see how big the real delta is. And of course, as pointed out in the other comments, for all of OP's hard work, there's no way to know which source is right or how many errors remain in his final version.