Feature selection for fusion of speaker verification via Maximum Kullback-Leibler distance
Di Liu, Dongmei Sun, Zhengding Qiu · 2010
This paper proposes an optimal scheme of feature selection for the fusion technique of speaker verification by Maximum Kullback-Leibler Distance. Through the investigations of bi-feature fusion schemes by six acoustic features, the information content of each fusion scheme via the Maximum Kullback-Leibler distance are computed in turn. The advantage of this distance is to overcome the shortcoming of the asymmetry of conventional Kullback-Leibler distance. This can keep the stability and correctness of the computation of the information content. In the experimental section, NIST 2001 corpus is used for evaluation. From the computation results by a variety of fusion schemes, it is found that the fusion between MFCC and residual phase hold the most information content. It indicates this scheme is able to yield an excellent performance. To verify its correctness, the EER evaluations are conducted. From the evaluation results, the EER of the fusion between MFCC and residual phase outperforms other fusion schemes. Therefore, the Maximum Kullback-Leibler distance can be considered as an effective metric for the feature selection in the fusion of speaker verification.