Decomposition of the LPC excitation using the zinc basis functions

Abstract

A linear predictive coding (LPC) excitation signal composed of a set of orthogonal functions called zinc functions is introduced. These functions are shown to form a complete orthogonal set and have properties that are well suited for modeling the LPC residual signal. A benchmark comparison between Fourier series and zinc function modeling shows that the zinc function model for the residual is superior in the mean-squared-error sense. The zinc basis functions are used in two low-bit-rate speech coding systems targeted at the 4.8-9.6-kb/s range. The first is a zinc excited LPC (ZELPC) system, where the voiced excitation is modeled using the zinc functions while the unvoiced excitation is represented by the usual white noise source. The second system is a zinc multipulse LPC (ZMPLPC) system, where the LPC excitation is constructed using the zinc basis functions instead of the usual ideal impulses. Results show that, given a fixed segmental signal-to-noise ratio with similar computational complexity, the ZMPLPC system is more efficient than a conventional multipulse LPC (MPLPC) system. Subjective listening tests also indicate a preference for the ZMPLPC system. >