通用码本语音矢量编码算法研究

Research on speech vector coding algorithm for general codebooks

基于预测编码、SOM自主神经网络矢量编码和Huffman编码的联合编码算法(PV算法)压缩效果虽然较好,但它在对每段语音编码时,都需要利用该段语音信号,通过SOM自主神经网络训练得到码本,算法复杂、耗时。为此文中提出从具有一般特征的多段语音信号中通过SOM自主神经网络训练提取码本,所有的语音信号段PV编码都统一用该码本,不需要对每一段语音信号编码都做一次提取码本的运算,这样不仅节省了每段语音PV编码时用于训练码本的时间,也节省了需要编码的专用码本的信息,减小了码率。实验结果显示,通用码本的PV编码算法在保证一定语音质量的条件下,是可行的。文中提出的编码算法在语言压缩编码方面具有较高的研究价值和很好的应用前景。

The joint encoding algorithm based on the predictive coding,SOM autonomous neural network vector coding and Huffman coding(PV algorithm)has a good combination effect,but is complex and time-consuming when used to obtain the codebook by means of the SOM autonomous neural network training since the speech signal segment needs to be used during the encoding of each speech segment. Therefore,the SOM autonomous neural network training is proposed in this paper to extract the codebook from multiple speech signal segments with general features. The codebook is used for PV coding of all speech signal segments. There is no need to perform a codebook extraction operation for encoding of each speech signal segment,which not only saves the codebook training time for PV coding of each speech segment,but also saves the information of specific codebooks that need to encode,and reduces the bit rate. The experimental results show that the PV coding algorithm of the general codebook is feasible under the condition of guaranteeing a certain speech quality,and the coding algorithm proposed in this paper has a high research value and good application prospect in the aspect of language compression coding.