Automatic Speaker Recognition Dependency on Both the Shape of Auditory Critical Bands and Speaker Discriminative MFCCs

Abstract

Accuracy of an automatic speaker recognition system predominantly depends on speaker models and features that are used. An influence of the shape of auditory critical bands and a contribution of individual components of MFCC-based feature vectors are investigated in the paper and some experimental results are presented and showed their impact on the accuracy of automatic speaker recognition. The speaker-discrimination capability of the MFCCs was experimentally determined by comparing training and test models for the same speaker. The experiments are conducted with three speech databases and showed that 0th and 19th (the last one) MFCCs are non speaker discriminative. The values of MFCCs are determined by the type of applied auditory critical band. The exponential auditory critical bands based on the lower part of exponential function have outperformed the speaker recognition accuracy of other auditory critical bands such as rectangular or triangular shape.

Keywords

• automatic speaker recognition
• mel-frequency cepstral coefficients
• energy correction
• speaker discriminative
• exponential auditory critical bands