I'm the author of software that, among other things, records audio. I got the requirement that the audio should be encrypted or at least obfuscated before saving to disk, so that someone using conventional audio software can't just listen to the output files and hear potentially sensitive information.
So I developed an algorithm to map any byte to a different byte, using a pseudo-random mapping, and then I can unmap by mapping in reverse. The byte mapping isn't totally random, but the correlation between input bytes and output bytes is very low, around 0.1.
Visualized as numbers, this really scrambles the bytes and produces a real mess of the data. And when you apply this to an audio file and listen to the result, it sounds like garbage, which was the goal.
But, to my astonishment, during testing someone did a recording of himself speaking, and in the scrambled file, you can still hear the original audio. It's distorted a bit and there's a lot of noise, but the audio is audible.
I took the waveforms of the two audio files and superimposed them like this:
Despite the lack of correlation of individual bytes, there does seem to be some correlation between the original audio (the inner, lighter color) and the scrambled audio (the dark green).
What could explain this?