Vol. 2, Issue 4 (2016)
Noise robust text-independent speaker identification using GFCC
Author(s): Pardeep Sangwan
Abstract: Automated speaker recognition performs efficiently in matched conditions but performance degrades in noisy conditions. Recent research shows that a relatively new feature, Gammatone Frequency Cepstral Coefficients (GFCC), more noise robust as compared to generally used Mel-Frequency Cepstral Coefficients (MFCC). To understand the intrinsic robustness of GFCC as compared to MFCC, speaker identification experiments are done to analyze their (dis)similarities systematically. This study reveals that the nonlinear rectification accounts for the noise robustness differences primarily. The present research proposes a novel paradigm which utilizes the strong pattern matching capability of ANNs for identification of speakers. Here ten speech samples are collected from 40 different speakers. Gammatone Frequency Cepstral Coefficients (GFCCs) are extracted for all the speakers and these coefficients are used to train ANN and then test signals are validated and verified for ANN. The results of identification under noisy conditions are very encouraging.