Presenting A New Algorithm Based on GMM-UBM With Cochlear Filter- PNCC Feature for Speaker Verification

In this paper, an auditory-inspired feature extraction algorithm based on a recently published time-frequency transform, i.e., auditory transform (AT) and the power normalized cepstral coefficients (PNCC) is proposed. Usually, the performance of acoustic models trained in clean speech drops significantly when tested on noisy speech.The proposed feature, called Cochlear Filter PNCC (CFPNCC), has shown strong robustness in the acoustic mismatch situations. An important feature of the proposed algorithm is the combination of advantages of the cochlear filter with the advantages of the PNCC feature, which has the resistance to both stationary noise and non-stationary noise. As shown in our experiments, in a GMM-UBM speaker verification system, CFPNCC outperforms the original PNCC and achieves the best overall results on the SSC database compared to the conventional features such as MFCC and RASTA-PLP under noisy conditions.