Exploring Sparse Representation Classification for Speaker Verification in Realistic Environment

We address the problem of speaker verification (SV) by exploiting discriminative classification ability of the sparse representation. The proposed sparse representation based speaker verification (SR-SV) system uses dictionary created with the mean supervectors derived from adapted GMMs. For classification purpose, the sparse coefficients obtained by the l 1 minimization are used employing different scoring methods. The SV systems are developed using the speech data collected in realistic environments with multiple sensors. On comparing with standard 1024 mixture GMM-UBM system, we find that 128 mixture GMM based SR-SV system performs better for all of the four sensors data considered.