Relaxing model-imposed constraints based on decoder analysis

It is well established that the linear-predictive model of order 10 performs well for the coding of speech signals. However, the relatively low model order constrains the performance when the coder is used for more complex signals such as music. We find that this limitation results in annoying coding noise in sharp spectral valleys. We show that this effect can be reduced significantly by means of a novel postfiltering structure which attenuates coding noise in the sharp spectral valleys. High-resolution spectral analysis of the decoded signal is used to find the relevant frequency regions. In contrast to conventional postfiltering methods, the new method does not emphasize the LP model structure, but instead enhances the signal structure found in the LP excitation signal, effectively relaxing the LP model constraints