High quality music coders commonly use auditory masked thresholds to account for the characteristics of the human ear. Perceptual filters (based upon linear signal prediction used in speech coders) are compared to filters using masked thresholds. Using listening tests, we have noticed that the second method does not provide better perceptual results. A natural way of proceeding would be to define a better psychoacoustical model. However, an intermediate method is presented here which allows additional degrees of freedom in a standard technique. The roots of the whitening filter are treated individually.