The data that I'm currently using is 1 bit sound, but it doesn't work because there are long stretches of either all 1's or all 0's, due to the frequencies being orders of magnitude lower than the sample rate. I'm reluctant to use higher bit rate sound because to make it binary I would have to splice together the bit vectors that represent each sample, causing the mutual information versus bit distance in the sequence to be choppy for small bit distances.
The reason that I need the property of smoothly decaying mutual information versus bit distance within the sequence is that I am training a generative model using a greedy DP algorithm.
[link][8 comments]