Linde, Buzo, and Gray (LBG) proposed a VQ design algorithm based on a training sequence. The use of a training sequence bypasses the need for multi-dimensional integration. The LBG algorithm is of iterative type and in each iteration a large set of vectors, generally referred to as training set, is needed to be processed. Usually, vectors sampled from a group of typical signals to be encoded altogether construct a training set T={x1,x 2,?.x M} ,where xi represents a sampled training vector and M represents the size of training set which is far greater than the codebook size N.