Exploiting Phase of Speech Signal for Speaker Recognition

Abstract

Abstract—Performance of speaker recognition system with feature based on temporal phase is presented in this paper. In the state of the art spectral feature, Mel Frequency Cepstal Coefficient (MFCC) only magnitude of the Fourier Transform of speech signal is considered while phase is ignored. The cepstral coefficients extracted from temporal phase (MFTPC) of speech signal are used as features for speaker recognition system. The performance of MFTPC feature is evaluated on IIT, Guvahati Multi Variability database (IITG-MV) and is compared with the baseline system using standard MFCC features. The performance of MFTPC feature is comparable to MFCC features in terms of identification rate, equal error rate and detection cost function. Fusion of MFTPC with MFCC features improves the performance of the system. Identification rate is improved by 0.10% and EER is reduced by 2% than MFCC only which shows the effectiveness of MFTPC feature.

Keywords —MFCC; MFTPC; IITG-MV database; GMM-UBM

Cite this Article

Pinky J. Brahmbhatt, Ami Gandhi, K. G. Maradia. Exploiting Phase of Speech Signal for Speaker Recognition. Journal of Communication Engineering & Systems. 2019; 9(2): 81–87p.