I. Introduction
Over the past two decades, source-system model based coding of speech has received considerable attention and has been incorporated in several commercial speech coding standards. Such a model presumes that the speech signal is synthesized by an autoregressive system excited by one of the following two excitation sources: train of impulses for voiced sounds, and random noise for unvoiced sounds. The autoregressive system can be represented by an all-pole synthesis filter whose parameters are obtained by linear prediction (LP) analysis performed on overlapping segments of the speech signal. A detailed tutorial review of LP coding (LPC) can be found in [1]. A major issue in LPC is the quantization of LP parameters.