Improvement in Speaker Verification Performance using an innovative combination scheme

Abstract

Performance of any machine recognition task can be improved by utilizing the benefit of multiple evidences lies in the combination schemes. In speaker verification tasks the score level combination scheme is widely used. In that scheme multiple features take decisions independently and later the overall decisions are modified based on individual strength accompanied with suitable weightage. The score level combination scheme provides interesting improvements in the overall performance, when evidences 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 speaker verification system. In the proposed scheme the individual adapted GMM-UBM models are built from different available features. The individual model parameters are padded together to build overall GMM-UBM adaptive models. The test features of different evidences are padded in similar pattern and placed before the system for verification. The experimental studies are made on Indian Institute of Technology Guwahati Multi-Variability (IITG-MV) speech database and well known NIST-2003 SRE database. Proposed method outperformed the score level combination scheme in both experiments which signifies the importance of the proposed method.