Comparative analysis of vector quantization methods used in speech processing

Vector quantization techniques play a vital role in compression of speech signals. There are a variety of vector quantization techniques. Each technique has its own advantages and disadvantages and there is no vector quantization technique presenting perfect results in all aspects till now. This thesis deals with enhancing the performance of the existing vector quantization techniques by using new methods. In this thesis hybrid vector quantization techniques which are produced from the existing methods are proposed. The performance of the designed vector quantizers are evaluated in terms of the spectral distortion measured, computational complexity and memory requirements. In the scope of this thesis, Multistage vector quantization (MSVQ), Split Vector Quantization (SVQ), Residual Vector Quantization (RVQ), Residual Multistage Vector Quantization (R-MSVQ), Residual Split Vector Quantization (R_SVQ) and voiced/unvoiced Residual Multistage Vector Quantization methods (VUV_RMSVQ) are analyzed. Because the VUV_RMSVQ method gave the better test results, further research is directed to find an optimum performance for codebook design with this method. Then, the overall performance of the proposed vector quantization techniques is compared with the existing vector quantization techniques. Whole work is carried out using the standard TIMIT database and both clean and noisy data are tested to evaluate the performance of the designed codebooks against noise. A linear predictive coding (LPC) based codebook generation algorithm is designed for each vector quantization method. Vector quantization is the process done in between LPC analysis and synthesis. The speech parameters required for vector quantization are the line spectral frequencies (LSF) and are obtained from the LPC coefficients. At the beginning of the thesis study, we designed codebooks with MSVQ and SVQ methods and we compared them in terms of spectral distortion. We found that the codebooks with MSVQ method gave better performance. Then, we used the RMSVQ and RSVQ methods to design codebooks. It is seen that the best result was given by RMSVQ. As a result, we continued with RMSVQ and we combined the voiced and unvoiced decision method and RSMVQ technique to achieve better result for spectral distortion. According to the results, it is seen that the best performance is achieved with VUV_RMSVQ method