Speaker Recognition via Block Sparse Bayesian Learning

In order to demonstrate the effectiveness of sparse representation techniques for speaker recognition, a dictionary of feature vectors belonging to all speakers is constructed by total variability i-vectors. Each feature vector from unknown utterance is expressed as linear weighted sum of a dictionary. The weights are calculated using Block Sparse Bayesian Learning (BSBL) where the sparsest solution can be obtained. By exploiting the speech signal’s block structure and intra-block correlation, the system performance is improved. The experimental results validate that our method outperforms the baseline systems and the system using Orthogonal Matching Pursuit (OMP) algorithm on the typical corpus and realizes the identity validation function.