A methodology for speaker-dependent acoustic features based on a simplified cortical response for speaker verification

摘要

Recently, the incorporation of the research done in biologically inspired systems has shown satisfactory results. A promising research area is the understanding of the human auditory systems and its performance under noisy conditions. Moreover, the incorporation of brain functions (cortical response) as an active part of the auditory system seems a viable alternative to increase the robustness of speech and speaker recognition systems. In this study, we propose a simplified model of the mammalian central auditory system for speaker verification systems. This model is based on a dimension-expansion representation, attempting to capture the response of the cortical cells. Then, by means of the Principal Component Analysis (PCA) approach, we reduce the dimensionality and create speaker-dependent features. Our results showed that by using our modeling technique, we were able to improve the performance of speaker verification systems.