Improved codebook definition for vector quantizers

An approach to the definition of the, full-search vector quantization codebook is presented. The proposed algorithm is based on successive nonlinear transformations of the training vectors, and a method of repeated clustering using a k-means type algorithm. The successive transformations of the training vectors eliminate the probability of convergence to a sub-optimum codebook of the k-means. The performance of the algorithm has been measured and compared for speech signals in distortion, memory requirements, and response time to that of the well-known LBG algorithm