ECMCS ' 99 EURASIP Conference DSP for Multimedia

Vector quantization is an eecient compression technique for which many variants are known. Product code vector quantizers use multiple codebooks for coding separately features of a vector. In shape-gain and mean-shape-gain vector quantizers, the bottleneck in the encoder is a nearest neighbor search on a hypersphere. We deene an angular constraint for speeding up the search in shape-gain and mean-shape-gain vector quantizers, based on a spherical triangular inequality. This constraint may be combined with other known techniques to give signiicant acceleration ratios. Experimental results are provided, showing the number of codewords compared and the total search time for three diierent methods.