基于码字特征的多模式多级矢量量化算法  

Multi-mode multi-stage vector quantization algorithm based on codeword characteristics

作　　者：徐敬德[1] 常亮[1] 计哲[1] 崔慧娟[1] 唐昆[1] 

机构地区：[1]清华大学电子工程系,微波与数字通信国家重点实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2011年第2期172-175,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(60572081)

摘　　要：为提高声码器中线谱频率参数多级矢量量化的性能,提出了一种根据码字特征进行分模式量化的算法。该算法首先根据下一级量化误差最小化的准则,通过训练得到本级代表模式信息的码字(码字数目为模式数目);然后统计与各个码字相对应的输入矢量占总矢量的比重,继而得到各模式码字所分化的码字个数;最后根据该分化方案训练得到本级所有码字并确定码字与模式的对应关系,从而进行分模式量化。测试结果表明:相比于根据本级码字索引平均进行模式分配的简单方案,该算法可以使平均谱失真(ASD)降低0.05 dB,而平均意见得分(MOS)提高0.02左右。The multi-stage vector quantization（MSVQ） performance of the linear spectral frequency（LSF） parameter in speech coding is improved by a multi-mode quantization algorithm based on the codeword characteristics.This algorithm first determines the codebook representing the mode of the current stage by minimizing the quantization distortion of the next stage.The proportions of the input vectors corresponding to each codeword in the current stage are calculated to get the actual number of codewords for each mode.Then,the overall codebook is trained and the relationship between the mode and the codeword is determined in the multi-mode quantization algorithm.Simulations show that the current method reduces the average spectral distortion（ASD） by about 0.05 dB compared to a simple mode-based quantization algorithm and increases the mean opinion score（MOS） of the synthesized speech by 0.02.

关 键 词：语音编码 线谱频率 多级矢量量化 多模式 

分 类 号：TN912.32[电子电信—通信与信息系统]