Learning self-organized topology-preserving complex speech features at primary auditory cortex

Abstract

By applying independent component analysis (ICA) algorithm to auditory signals a computational model was developed for the speech feature extraction at the primary auditory cortex. Unlike the other ICA-based features with simple frequency selectivity at the basilar membrane and inner hair cells the learnt features represent complex signal characteristics at the primary auditory cortex such as onset/offset and frequency modulation in time. Also, the topology is preserved with the help of neighborhood coupling during the self-organization. The extracted complex features demonstrated good performance for the robust discrimination of speech phonemes. (c) 2004 Elsevier B.V. All rights reserved.