Use of complementary feature to enhance the performance of speaker identification system

Abstract

Nowadays, Bio-metric recognition systems are becoming popular due to cost-effectiveness. The Speaker identification (SI) system uses an audio signal of the speaker to identify the unknown speaker. There are various approaches to increase the performance of the SI system. That includes the use of the popular cepstral feature, MFCC (Mel frequency cepstral coefficient). In this paper, along with MFCC, the complementary features of MFCC, i.e., IMFCC (Inverse mel frequency cepstral coefficient) is used. Along with that, a recently used feature speech-signal-based frequency cepstral coefficient (SFCC) and its complementary feature inverse SFCC (ISFCC) are also used. The experiments are carried on the POLYCOST database. The performance of the SI system due to ISFCC is better than IMFCC. To enhance the SI system's accuracy, a score-level fusion of MFCC with IMFCC and SFCC with ISFCC is done. The relative improvement due to the fusion of SFCC with ISFCC is up to 4% and 7% over MFCC and IMFCC, respectively.