Learning feature hierarchies for musical audio signals