I've seen visual demonstrations of PRNG bias before, but are the biases of typical implementations of Math.random "above the noise threshold" (pun intended) where humans could distinguish them from truly random streams of audio?
I don't think so, no. The way I understood op was to use the generated audio itself as a source of entropy, which would be a bad idea if not using a CSPRNG perfectly tuned to the sampling rate and lossless audio I think :)
but I'm just guessing there regarding the effects of quality.
