Pole-zero code excited linear prediction using perceptually weighted error criterion

A new speech coding algorithm, pole-zero CELP (code excited linear prediction), is proposed in which the short-time spectral envelop of the speech wave is modeled with a pole-zero filter. Hence, zeros which occur in the speech spectrum (e.g., during nasalized speech sounds) are more accurately modeled than with the traditional all-pole filter. A linear prediction analysis of the speech signal to obtain a pole-zero filter representation of the spectral envelope results in a set of nonlinear simultaneous equations. Previously, complex iteration schemes have been used to solve for the filter coefficients from these equations. However, knowledge of the excitation in the pole-zero CELP coder leads to a linear but suboptimum solution for the filter coefficients. As expected, it shows improved spectral matching in regions of anti-resonant nasalized sounds. In addition the high frequency portion of the speech signal is matched better. Informal listening tests shown a slight preference for speech coded with the zero filter included and a perceptually weighted error criterion