On reducing pitch sensitivity of LPC parameters

Abstract

Mean‐squared prediction error criterion used in linear predictive coding of speech has a number of inherent shortcomings. For low pitch frequencies, the LPC analysis largely ignores the pitch‐related fine structure of the speech spectrum. However, the sensitivity of the LPC parameters increases rapidly as a function of the pitch frequency. This sensitivity can be directly traced to the mean‐squared prediction error criterion used in LPC analysis. Moreover, in the presence of noise and errors resulting from the assumption of an all‐pole model, the minimization of prediction error leads to a perceptually suboptimal solution. In this paper, the LPC analysis is considered as a short‐time spectral envelope matching problem. Using the LPC‐derived parameters as initial values, a search procedure is used to refine the LPC parameter estimates. The new procedure minimizes a perceptual distance metric between the spectrum based on LPC parameters and the samples of the speech spectrum at spectral peaks.