Deep feature for text-dependent speaker verification

•Utilized as identity vector: Similar to the normal i-vector idea, deep features extracted from the deep models can be directly used as the speaker identity representations. All of these deep features belonging to the same speaker are averaged to form the target identity vector. Besides four kinds of back-end classifying methods are adopted with these deep feature based identity vectors, including cosine distance, LDA, PLDA, PLDA+LDA; The EER of the best system using the proposed multi-task joint-learned identity vector is 0.10%, only one fifteenth of that in the GMM-UBM baseline.