Split Vector Quantizer For Wideband ISF Parameters Based On Conditional GMM

Vector quantizing of parameter is always a hot spot in speech coding domain. At present, many international standard organizations, such as ITU-T,3GPP, adopt schemes based on split vector quantization in their speech coding standards. These algorithms have advantages as well as disadvantages. The good points are that training method is simple and calculate complexity is low, while the shortcoming is the too high bitrates. Foreign scholars have changed their focuses form the traditional scheme to new method based on Gaussian Mixture Model (GMM) gradually since 2000.The algorithms of this paper are also based on conditional GMM.It reviews the research history and status of vector quantization both at home and abroad, then the significance and characteristics of the parameter to be processed, ISF, are described. The characteristic of the traditional codebook training algorithm LBG and distortion measurement are also dwelt on. We then use the conditional GMM modeling the training data to exploit the interframe correlation effectively, after this step, train the better codebook with taking advantage of this correlation. We propose the integrated algorithm, SFSVQ and SSSVQ, for wideband ISF parameter. They make the most of correlation among the frame and subframe respectively.At last, this paper analyzes the performance of SFSVQ and SSSVQ in spectral distortion, computational complexity and memory requirements.