An LPC Excitation Model using Wavelets

This paper presents a new model of linear predictive coding (LPC) excitation using wavelets for speech signals. The LPC excitation becomes a linear combination of a set of self- similar, orthonormal, band-pass signals with time localization and constant bandwidth in a logarithmic scale. Thus, the set of the coefficients in the linear combination represents the LPC excitation. The discrete wavelet transform (DWT) obtains the coefficients, having several asymmetrical and non-uniform distribution properties that are attractive for speech processing and compression. The properties include magnitude dependent sensitivity, scale dependent sensitivity, and limited frame length, which can be used for having low bit-rate speech. We show that eliminating 8.97% highest magnitude coefficients degrades speech quality down to 1.49dB SNR, while eliminating 27.51% lowest magnitude coefficient maintain speech quality at a level of 27.42 dB SNR. Furthermore eliminating 6.25% coefficients located at a scale associated with 175-630 Hz band severely degrades speech quality down to 4.20 dB SNR. Finally, our results show that optimal frame length for telephony applications is among 32, 64, or 128 samples.