Well I have an idea... but the output may well be horrible. I do not have any audio processing experience at all so I cannot even determine whether is is viable until I have created it!
The basic premise is to take random 1 second samples from an audio corpus, lets say 1 million, normalise the samples then using k-mean clustering to create a set of 1024 filters. Compression would be, for each second, record the reverse transform to the normalised 1 second sample (using 8 bits) then choosing the best matching filter and recording its index (i.e. 10 bits). thus 18 bits per second which is …