Robust Auditory-Based Speech Feature Extraction Using Independent Subspace Method

Abstract

In recent years many approaches have been developed to address the problem of robust speaker recognition in adverse acoustical environments. In this paper we propose a robust auditory-based feature extraction method for speaker recognition according to the characteristics of the auditory periphery and cochlear nucleus. First, speech signals are represented based on frequency selectivity at basilar membrane and inner hair cells. Then, features are mapped into different linear subspaces using independent subspace analysis (ISA) method, which can represent some high order, invariant statistical features by maximizing the independence between norms of projections. Experiment results demonstrate that our method can considerably increase the speaker recognition accuracy specifically in noisy environments.