LPC vector quantization based on critical-band analysis

Abstract

Critical-band spectrum is known to represent the redundancies of the human auditory system. Vector quantization constitutes a minimal redundant code set for a given signal source. In this study, these techniques are combined in a low-rate speech analysis/resynthesis system. An important design criterion for a perceptually based source coder is to define an appropriate mapping between the analysis codebook in critical-band domain and the synthesis codebook in LPC domain. Several methods for generating and searching the synthesis codebooks based on the perceptual distance criterion are proposed. Their performance with other LPC-based distance measures (e.g., log area, LSP, cepstral) is compared. These results indicate a substantial improvement in the quality of synthetic speech for the perception-based system due to smoother spectral trajectories in the low-frequency regions.