Whispered speech speaker identification based on SVM and FA

This paper presents Chinese whispered speech speaker identification in the channel mismatch circumstance. In this paper, a simple factor analysis is used to deal with channel variability problem. We put the speaker factor vector to input SVM, similarly the speaker´s new feature. It´s found that this integration has achieved improvement to some extent. Then two channel compensation techniques, WCCN and LDA, are combined with SVM-FA to remove the channel informatin. It proves that the average recognition rate of 8 channels is about 92% in the best condtion. It also has a significant speedup in training and testing process.