Discriminativetensor dictionaries and sparsity for speaker identification

Dictionary learning algorithms based upon matrices/vectors have been used for signal classification by incorporating different constraints such as sparsity, discrimination promoting terms or by learning a classifier along with the dictionary. However, because of the limitations of matrix based dictionary learning algorithms in capturing the underlying subspaces of the data presented in the literature, we learn tensor dictionaries with discriminative constraints and extract classifiers out of the dictionaries learned over each mode of the tensor. This algorithm, named as GT-D, is then used for the speaker identification. We compare classification performance of our proposed algorithm with other state-of-the-art tensor decomposition algorithms for the speaker identification problem. Our results show the supremacy of our proposed method over other approaches.