Performance evaluation of speaker recognition system

Significance of accuracy in Speaker Recognition System (SRS) is still a challenging area of research. In this paper, the accuracy of a SRS based on Mel Frequency Cepstral Coefficient (MFCC) using Correlation is analysed on a limited amount of speech corpus. The experiments have been performed on 2 Bilingual Speech Corpus. A set of two different experiments has been carried out to study the impact of recording text, recording languages, duration of speech sample and recording devices on the accuracy of SRS. On the basis of the performed experiments, it has been concluded that recognition rate degrades due to acoustic mismatch in training and testing phase. The SRS achieved accuracy of 96.67% for text-dependent and device dependent while it degrades by 30% for Language independent and device independent.