On exploring the similarity and fusion of i-vector and sparse representation based speaker verification systems

The total variability based i-vector has become one of the most dominant approaches for speaker verification. In addition to this, recently the sparse representation (SR) based speaker verification approaches have also been proposed and are found to give comparable performance. In SR based approach, the dictionary used for sparse representation is either exemplar or learned from data using the KSVD algorithms and its variants. Recently the use of the total variability matrix of the i-vector system as the dictionary for the SR based approach has also been reported. Motivated by these, in this work, we first highlight the similarity between the i-vector and the learned dictionary SR based approaches for speaker verification. It is followed by the exploration about various kinds of learned dictionaries, their sizes and the sparsity constraint in context of SR based speaker verification. Further we have explored the feature level as well as the scores level fusions of these two approaches.