A vector quantization schema for non-stationary signal distributions based on ML estimation of mixture densities

We show that by selecting an appropriate distortion measure for the encoding-decoding vector quantization schema of signals following an unknown probability density p(x), the process of minimizing the average distortion error over the training set is equivalent to the Maximum Likelihood (ML) estimation of the parameters of a Gaussian mixture model that approximates p(x). Non-stationary signal distributions can be handled by appropriately altering the parameters of the mixture kernels.