Speaker Recognition System for Limited Speech Data Using High-Level Speaker Specific Features and Support Vector Machines

摘要

High-level speaker-specific features (HLSSFs), such as the style of pronunciation of words, their use, phonotactics and prosody, form the main subjects of state-of-the-art research on automatic speaker recognition (ASR). In this paper, we experimentally verify HLSSF extraction and support vector machine (SVM)-based modelling techniques. The HLSSF extraction produces patterns of symbols for each speaker during ASR training. The strategy involves changing these patterns during the training and testing of ASR using frequencies (n-gram) for a given voice sample. We used SVM and n-gram frequencies to implement ASR, where the application consisted of a new kernel based on the linear log-probability proportional scoring framework. This approach yielded impressive outcomes on an assortment of abnormal state highlights in ASR. We showed that the proposed ASR based on the linear log-probability proportional scoring framework is superior to other standard log-probability frameworks. The equal error rate (EER) of our ASR method was 2.5% with a 2% improvement over the standard method.