An effective combination scheme for improving speaker verification performance

Abstract

In machine recognition the benefit of utilizing multiple evidence lies in the combination schemes employed. In speaker verification (SV) tasks the score level combination scheme is widely used. The score level combination scheme provides interesting improvements in the overall performance, but when evidence from different features are complementary in nature. It is conjecture that collectively contributed decisions may be more useful in achieving improved overall performance. Based on this idea, we propose a combination scheme for generic GMM-UBM based SV system. In the proposed scheme the individual adapted GMM-UBM models are built from different available features. The individual model parameters are normalized and padded together to build overall GMM-UBM adaptive models. The test features of different evidence are padded in similar pattern and placed before the system for verification. We made the investigation experimentally on combining the evidence from popular mel-frequency cepstral coefficients (MFCC) and excitation source information based residual MFCC (RMFCC) features. The experimental studies are made on Indian Institute of Technology Guwahati — Multi-Variability (IITG-MV) speech database that covers all kind of variabilities. The score level combination scheme provides an overall improvement of 18.76% as compared to 21.84% by the proposed scheme indicating its significance.