论文部分内容阅读
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。
In order to improve the multi-level vector quantization performance of vocoder spectral line frequency parameters, an algorithm for quantizing the sub-modes based on the characteristics of codewords is proposed. The algorithm first obtains the code words (the number of code words is the number of patterns) of the representative pattern information of the current stage according to the criterion for minimizing the quantization error of the next level. Then, the proportion of the input vectors corresponding to each code word to the total vectors , And then get the number of codewords differentiated by each pattern codeword. Finally, according to the differentiation scheme, all the codewords of this level are obtained and the corresponding relationship between codewords and patterns is determined, so as to quantize the sub-patterns. The test results show that the proposed algorithm can reduce the mean spectral distortion (ASD) by 0.05 dB and increase the average opinion score (MOS) by about 0.02, compared with the simple scheme of mode allocation based on the average index of codeword.