固定语音检索中特征降维算法研究  

Research on Dimensionality Reduction for Speech Retrieval

作　　者：陈鸿昶[1] 齐晓倩[1] 黄海[1] 

机构地区：[1]信息工程大学信息工程学院,河南郑州450002

出　　处：《信息工程大学学报》2011年第4期441-446,共6页Journal of Information Engineering University

基　　金：国家863计划资助项目(2008AA011002)

摘　　要：将非线性流形学习的思想引入到语音特征降维中,设计完成了局部投影(Locality Preser-ving Projections,LPP)算法,并针对该算法对降至的目的维度敏感的缺点提出了一种加权局部投影(Weighted Locality Preserving Projections,WLPP)算法。对LPP算法及WLPP算法进行了实验,实验证明LPP算法在对语音特征降维后也能有较高的准确率,同时,也证明了WLPP算法可以大大降低LPP算法对目的维度的敏感度。An algorithm based on nonlinear manifold is proposed in this paper. LPP algorithm is proved effective but very sensitive to embedding dimension, so WLPP algorithm is designed to overcome such a drawback. Several contrast experiments are designed and performed to evaluate the effectiveness of LPP and WLPP algorithms. The experimental data supports the method of using LPP algorithm and WLPP algorithm,and also shows that WLPP is superior to LPP in insensitivity to embedding dimension.

关 键 词：线性降维 流形学习 局部保持映射 加权局部保持映射 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]