Nonnegative dictionary learning in the exponential noise model for adaptive music signal representation