Abstract: This paper proposes a hierarchical speaker verification approach based on mixed Principal Component Analysis(PCA) classifier and Kernel Fisher Discriminant(KFD). PCA is utilized to reduce the dimension of registered speakers’ feature vectors, and Principal Component Space(PCS) and Truncation Error Space(TES) are obtained based on the transform matrix. A mixed-PCA classifier is proposed based on PCS and TES to select the most possible R target speakers fast. And the target speaker is found with KFD. Experimental results validate the effectiveness of the approach.

Key words: speaker verification, Principal Component Analysis(PCA), Principal Component Space(PCS), Truncation Error Space(TES), Kernel Fisher Discriminant(KFD)

摘要： 提出一种基于混合主成分分析(PCA)分类器和核Fisher判别(KFD)的多级说话人确认方法。利用PCA对注册说话人的特征向量进行降维，根据转换矩阵得到说话人特征向量的主成分空间和截断误差空间，结合这2个空间构造混合PCA分类器，用于快速判断最有可能的R个目标说话人，并采用KFD寻找最终目标说话人。仿真实验结果验证了该方法的有效性。

关键词: 说话人确认, 主成分分析, 主成分空间, 截断误差空间, 核Fisher判别

CLC Number: 

• TP391.42