Hierarchical speaker verification: Kernel fisher discriminant plus Mixed-PCA classifier and FCM clustering
Tan Ping, Xing Yujuan · 2015
In order to improve speaker verification accuracy, we proposed a new hierarchical speaker verification algorithm in this paper. In our algorithm, Mixed-PCA plus fuzzy c-means (FCM) clustering was combined with kernel fisher discriminant (KFD). In stage of feature extraction, we exploited PCA to reduce the feature vector dimensions, and then FCM was used to select more discriminant data and cluster training data set into some clusters. In stage of recognition, a novel MPCA classifier was proposed based on principal component space (PCS) and truncation error space (TES) to select the possible R target speakers fleetly. And then, KFD was adopted as final classifier to verify target speaker. The experimental results showed that the EER of our proposed method is 4.83%, meanwhile the minDCF is 0.0504. And our hierarchical classifier has shorter recognition time.