A Kernel Fisher Discriminant Classifier for Speaker Recognition
Feature extraction is the first important stage in an automatic speech recognition system. Traditional way to use different information sources is to concatenate different features into a long feature vector. Although it is simple to implement and works reasonably well, it has a few shortcomings.Firstly, not all the cepstral components contribute to the recognition accuracy. Secondly, the number of training vectors needed for robust density estimation increases exponentially with the dimensionality. In this paper, a kernel Fisher discriminant classifier was proposed. Instead of simply concatenating different sources features, we modeled each different feature set separately and made fusion of the KFD-based classifier output scores. A significant components selection procedure applying the kernel Fisher discriminant criterion, were presented to reduce the high dimensionality of the long feature vector before fed to the classifier. The speaker verification experiment indicates the superior recognition accuracy can be obtained in speech recognition.
kernel Fisher discriminant decision fusion speaker verification keywords spotting
X.Li Y.Zheng
Department of Electronic Science & Engineering, Nanjing University, Nanjing 210093 P.R.China;Nationa School of Computer Engineering and Science, Shanghai University, Shanghai 200072 P.R.China
国际会议
杭州
英文
344-348
2006-10-12（万方平台首次上网日期，不代表论文的发表时间）