Effective Combination of Multiple Evidences for I-vector Based Limited Data Speaker Verification

Abstract

The performance of automatic speaker verification (ASV) system always depends upon the amount of information (speech sample) used in the process. ASV system's performance suffers when the information provided to the system is limited, even though the methodology is remain same. The issue of limited information can be resolved to some extend by using multiple evidences. In general, score level combination scheme is widely used to combine the effect of multiple evidences, where a decision is made based on the independent opinions of the evidences. We conjecture that the collectively contributed decisions may be more effective and propose a new combination scheme for limited data speaker verification task. In the proposed work, we have used mel frequency cepstral coefficient (MFCC) and residual MFCC (RMFCC) as representation of the vocal tract and excitation source information. The experiments are conducted with well-known NIST-2003 speaker recognition evaluation (SRE) database. The score level combination scheme provide a relative improvement of 14.93% in extremely limited data condition (≃ 2 sec), on an average 15.57% for all limited data conditions. In comparison, the proposed scheme provides 28.40% and 29.02%, respectively. Thus proposed method provides a relative gain of 13.47% for extremely limited data condition and on an average 13.42% for other limited data conditions. These experimental results signify the importance of using proposed combination scheme over the popular score level combination scheme.