Effectiveness of speaker-dependent feature score pruning in speaker verification

This paper presents investigations into the use of speech feature score pruning for enhancing the speaker verification accuracy. A new technique based on defining speaker-specific regions for discarding feature scores is proposed and experimentally investigated. The scope of the investigations covers both text-dependent and text-independent speaker verification. Based on the results, it is shown that considerable improvements can be obtained in the former scenario. The paper discusses the motivation for the proposed approach and details the experimental study.