Toward Optimal Mixture Model Based Vector Quantization

Gaussian mixture model (GMM) based vector quantization (VQ) using a data-dependent weighted Euclidean distortion measure is presented. It is shown how GMM-VQ can be improved by using GMMs that model the optimal VQ point density rather than the source probability density as is done in previous work. GMM training procedures as well as procedures for encoding and decoding that takes a weighted distortion measure into account are presented. The usefulness of the proposed procedures is demonstrated on a source derived from speech spectrum parameters