论文部分内容阅读
本文讨论了自组织特征映射人工神经网络在语音矢量量化中应用时所涉及的两个重要问题,即码本训练和码本搜索的问题。根据语音反射系数的特点,提出了训练中初始码本的选择原则和实用训练算法。利用特征映射网络的聚类特性和语音相邻帧间的相关性,提出了码本搜索的两种快速算法——子域搜索法和邻域搜索法。大量实验结果表明,这两种快速搜索方法相结合,搜索时间减少为常用的LBG全搜索算法的1/4或1/10,同时保持精度不下降。本文提出的方法已在一种极低数据率的声码器中得到成功应用。
In this paper, two important problems involved in the application of speech vector quantization, ie, codebook training and codebook search, are discussed in this paper. According to the characteristics of speech reflection coefficient, the selection principle and practical training algorithm of initial codebook in training are proposed. Two fast algorithms of codebook search, sub-domain search and neighborhood search, are proposed by using the clustering characteristics of feature mapping network and the correlation between adjacent frames. Numerous experimental results show that the combination of these two fast search methods reduces the search time to 1/4 or 1/10 of the commonly used LBG full search algorithm while keeping the accuracy low. The proposed method has been successfully used in a very low data rate vocoder.