Cepstral Features and Text-Dependent Speaker Identification – A Comparative Study

In the study, the effectiveness of combinations of cepstral features, channel compensation techniques, and different local distances in the Dynamic Time Warping (DTW) algorithm is experimentally evaluated in the text-dependent speaker identification task. The training and the testing has been done with noisy telephone speech (short phrases in Bulgarian with length of about 2 seconds) selected from the BG-SRDat corpus. The employed cepstral features are – Linear Predictive Coding derived Cepstrum (LPCC), Mel-Frequency Cepstral Coefficients (MFCC), Adaptive Component Weighted Cepstrum (ACWC), Post-Filtered Cepstrum (PFC) and Perceptually Linear Predictive coding derived Cepstrum (PLPC). Two unsupervised techniques for channel compensation are applied – Cepstral Mean Subtraction (CMS) and Relative Spectral (RASTA) technique. In the DTW algorithm two cepstral distances are utilized – the Euclidean and the Root Power Sum (RPS) distance. The experiments have shown that the best recognition rate for available noisy speech data was obtained by using the combination of the MFCC, CMS and the DTW-RPS distance.