256 layer 'impulse transform(??)' basically reconstructs sounds out of 256 bandlimited impulse (blip) trains. this output then gets convolved with the original signal. it's pretty sweet.