Deep nonlinear metric learning for speaker verification

By applying Nonlinear Metric Learning to speaker recognition, a speaker verification algorithm based on deep independent subspace analysis network is proposed. Different from the traditional linear metric learning methods,the proposed method learns an explicit mapping from the original space to an optimal subspace by means of deep independent subspace analysis network. On the basis of this, the similarity between two i-vectors can be calculated in the optimal subspace in order to obtain a better speaker verification performance. The proposed method is evaluated on the NIST SRE 2008 dataset. Comparing with the traditional i-vector model with cosine distance metric, LDA and PLDA, the proposed method decreases the EER by 11.02%, 6.40% and 4.57%, respectively.