摘要
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。
Two issues relating to applications of self-organizing feature mapping nets to VQ of speecl training and searching codebook, are discussed in this paper. Based upon properties of reflect coeficients of speech, a principle of choosing initial codebook and a realistic training procedure used to populate the codebook are presented. Taking advantage of the clustering properties of feature mapping nets and the correlation between two adjacent frames of speech, two fast-search algorithms-sub-region search and neighbouring-region search algorithms-are developed. Experimental results demonstrate that one ean combine the two fast-search algorithms together to reduce the search-time by 1/4 or 1/10 of that of the popular LBG full-search algorithm, without reducing accuracy. The proposed methods in this paper have-been successfully nsed in a very-low-rate vocoder.
出处
《通信学报》
EI
CSCD
北大核心
1992年第5期70-75,共6页
Journal on Communications
关键词
神经网络
矢量量化
语音处理
Neural network, Vector quantization, Speech processing, Self-organizing feature mapping.
