一种结构受限的异方差线性判别分析

异方差线性判别分析(HLDA)因在语音识别中起到了巨大的特征去相关作用而被广泛利用。然而在训练数据不足或特征维数较高时,HLDA易出现不稳定性和小样本问题。根据特征的矩阵表示形式,提出了一种结构受限的HLDA。首先用二维线性判别分析(2DLDA)压缩矩阵形式的特征,然后作一维的HLDA。通过分析我们指出,二维的特征变换实际上是一种结构受限的一维特征变换。在RM库上的实验,受限HLDA对常规HLDA的词识别错误相对下降12.39%;在TIMIT库上的实验,受限HLDA对常规HLDA的音素识别错误相对下降4.43%。

A Structure-Specific Heteroscedastic Linear Discriminant Analysis

Heteroscedastic linear discriminant analysis(HLDA) is applied widely in speech recognition due to its ability of feature de-correlation.To overcome its instability on high dimension features and the small sample issue on insufficient training samples,this paper proposes a structure-specific HLDA method to transform the feature matrix.The method adopts the two-dimensional linear discriminant analysis(2DLDA) to compress features in the matrix,and then,the one-dimensional HLDA is applied.It is revealed that two-dimensional feature transformation is actually a structure-constrained one-dimensional feature transformation.Experiments show that the proposed structure-specific HLDA achieves 12.39% word error rate(WER) reduction on RM database and 4.43% phone error rate(PER) reduction on TIMIT database compared with the traditional HLDA.