Research On Vector Quantization For Linear Predictive Coefficients In Embedded Variable Bit Rate Speech Coding

In the last several years, with the development of network technology, there is a tremendous interest in the so-called "IP telephony", i.e., telephone calls transmitted through packet-switched data networks employing the Internet Protocol (IP). How to reduce the effect of packet losses in synthesis speech quality has become a more important issue in speech coding. The most available way is to use embedded coding algorithms. International Telecommunications Union (ITU-T SG 16) is studying a new speech coding standard from 2005, called EV-VBR (Embedded Variable Bit Rate). Recently, there is a great interest in developing embedded variable bit rate speech coders.Algebraic code excited linear prediction speech coding technique is always utilized in embedded variable bit rate speech coder, and the quantization for linear predictive coefficients becomes an important problem for this scheme. The performance of the quantizer for linear predictive coefficients will affect the whole codec. Some researches focused on vector quantization techniques for linear predictive coefficients in embedded variable bit rate speech coding are shown in this thesis.Before quantization, linear predictive coefficients are always transformed into linear spectral frequencies (LSF) parameters or immittance spectral frequencies (ISF) parameters. An improved fast codebook search approach for vector quantization of the LSF parameters based on Hadamard transform is presented in this thesis. The experiment results confirm that the calculating complexity of the proposed algorithm is lower than conventional full search algorithm. Aimed at the problems of previous quantization algorithm in packet losses and the characteristics of embedded variable bit rate speech coding, we proposed three quantization schemes used for quantizing the ISF parameters of wideband speech. Firstly, an improved switched split vector quantizer used for quantizing the ISF parameters of wideband speech is proposed in this thesis. Experimental results show that this memoryless quantization scheme achieves transparent coding at 42bits/frame. But the memory for the codebook storage is too big. Aimed at this problem, a new switched product code pyramid vector quantizer is proposed. The characteristics of this algorithm are low complexity and low memory. Experimental results show that this quantization scheme achieves transparent coding at 46bits/frame. Considering the scheme used for quantizing wideband ISF parameters in embedded variable bit rate speech coding, the quantizer must has good performance both in good and bad frames. Finally, we proposed a new unequal coefficient interframe predictive split vector quantizer. Experimental results show that this quantization scheme achieves transparent coding at 46bits/frame, and has low error propagatation in frame erasure.Finally, the 46bits/frame unequal coefficient interframe predictive split vector quantizer is applied to an 8~32 kb/s embedded variable bit rate speech codec. This codec is presented by Beijing University of Technology Speech and Audio Signal Processing Lab, and submitted to ITU-T by Huawei Company as a candidate for G.VBR codec. Experimental results show that the quality of synthesis speech can meet the terms of reference for ITU-T G.VBR codec in clean speech.