A new trellis vector residual quantizer with applications to speech and image coding

Abstract

Summary form only given. We present a new trellis coded vector residual quantizer (TCVRQ) that combines trellis coding and vector residual quantization. Our TCVRQ is a general-purpose sub-optimal vector quantizer with low computational costs and small memory requirement that permits high memory savings when compared to traditional quantizers. Our experiments confirm that TCVRQ is a good compromise between memory/speed requirements and quality and that it is not sensitive to codebook design errors. We propose a method for computing quantization levels and experimentally analyze the performance of our TCVRQ when applied to speech coding at very low bit rates and to direct image coding. We employed our TCVRQ in a linear prediction based speech codec for the quantization of the LP parameters. Several experiments were performed using both SNR and a perceptive measure of distortion known as cepstral distance. The results obtained and some informal listening tests show that nearly transparent quantization can be performed at a rate of 1.9 bits per parameter. The experiments in image coding were performed encoding some 256 gray levels, 512/spl times/512 pixel images using blocks of 3/spl times/3 pixels. Our TCVRQ were compared, on the same training and test sets, to an exhaustive search vector quantizer (built using the generalized Lloyd algorithm) and to a tree quantizer for different coding rates ranging from 3 to 10 bits per block.