Using high-level feature concentration for speaker identification

Abstract

Traditional and current speaker recognition systems primarily use low-level (physiological) features of speech that model the physical dimensions of the vocal tract. The popular MFCC is such a feature vector. There is a growing trend in the literature, however, that evidently supports the idea of improved systems by fusing low-level features with high-level (psychological) features like conversational, lexical, phonemic and prosodic patterns found in speech. In this work we investigated the performance of a speaker ID system evaluated on the NTIMIT database employing the popular MFCC feature vector concatenated with a high-level feature vector containing prosodic information, viz. voicing and pitch. The vector contains the maximum autocorrelation values of a segmented frame of speech and is accordingly named the MACV feature. This paper is an extension of the work done by Wildermoth and Paliwal [11] who reported on an improved speaker ID system that used a fused LPCC-MACV feature set instead of a LPCC-only system. Results presented in this paper showed an improvement from 82.74% to 85.32% for the fused system, a relative improvement of over 3% for the identification rate. This result corroborated with Wildermoth and Paliwal’s [11] performance (an increase from 78.4% to 86.8%) and supports literature on improved recognition systems due to high-level low-level feature fusion. The increase in performance on a popular, state-of-the-art feature vector, like the MFCC, further creates anticipation for promising results to future work on similar systems used on more challenging databases.