Speaker Identification Based on Robust Sparse Coding with Limited Data

The sparse representation classifier has achieved interesting classification results in face recognition. In speaker identification task, we intend to form an over complete dictionary using the GMM supervector for the training data. Then, the sparse representation is shaped as a sparsity-restricted robust regression problem. By supposing that the representation residuary and the representation coefficient are respectively independent, we use robust sparse coding (RSC) based on maximum likelihood estimation (MLE) solution to solve the sparse representation problem. In RSC, the collaborative representation strategy, taking the training utterances from all the extra classes as the nonlocal utterances of one class, is quite suitable for speaker recognition with limited data. Finally, experiments were carried out to evaluate the RSC on the ELSDSR database. The results have shown the performance of the proposed algorithm is much effective than the state-of-the-art methods of speaker identification.