Improving the adaptive source model for CELP coding with long analysis frame size

Abstract

One way to lower the coding rate of CELP coders is to lengthen the excitation analysis frame size. For enhanced speech quality in such a case it is desirable to have the CELP excitation peaky (or sharpened). In this paper we first consider the relation between the LPC prediction residual and the CELP excitation, and show that the adaptive source of a CELP coder reconstructs the major pulse at glottal closure and the formant structure remained in the LPC residual, and that the stochastic source models the randomness of the LPC residual. Based on this observation, we propose a new adaptive source in which samples of the source have different gains according to their amplitudes by a two-tap pitch predictor. Simulation results show that peaky pulses at voiced onset and a burst of plosive sound are clearly reconstructed, and that in voiced sound the excitation has the desirable peaky pulse characteristic and the pitch periodicity is well reproduced.