双向二维加权LPP语音特征降维算法

Research on(2D)2WLPP for Speech Feature Vector Reducion

提出一种双向二维加权局部保持投影算法(Two-directional Two-dimensional Weighted Locality Preserving Projections,(2D)2WLPP)用于语音特征提取后维度的降低,考虑到普通的二维降维算法只能从一个方向进行特征降维且所降至的维数选择非常受限,该方法能够从水平和垂直两个方向对语音矩阵进行降维处理,这样可以大大降低提取后的语音特征数目;考虑到不同投影向量对保持局部结构的重要程度不同,进而对各个特征赋予不同的权重系数.实验证明,该算法运算速度快,与已有的二维局部保持投影相比,获得了更高的识别率.

A novel Two-directional Two-dimensional Weighted Locality Preserving Projections（（2D）2WLPP）,which can effectively reduce feature dimension is applied to speech feature extraction.Feature matrix is obtained by reducing speech matrix horizontally and vertically,while common two-dimensional reduction method can only reduce the dimension by one direction.Considering that projecting vectors are not of the same importance to preserve local structure,different vectors are given different values.Experimental result shows that this algorithm is efficient and can achieve higher recognition rate than 2DLPP.