Improving the noise-robustness of Mel-Frequency Cepstral Coefficients for speaker verification

In this paper, we study the noise-robustness of Mel-Frequency Cepstral Coefficients (MFCC) for speaker verification systems and explore ways to improve their performance in noisy conditions. Improvements are based on new construction of filter bank which decouples filter bandwidth from the other filter bank parameters by employing the known relationship between filter center frequency and critical bandwidth of the human auditory system. The proposed method which incorporates human factor cepstral coefficients (HFCC) is shown to outperform the traditional front-end processing technique (i.e., MFCCs) in automatic speaker verification experiments.