Adaptive filterbanks inspired by the auditory system for speech feature extraction

Document Type

Conference Proceeding

Date of Original Version



Using the human auditory system as a guide we propose a signal processing strategy for decomposing composite signals into bandpass signal components and extracting their features. We use two parallel filterbanks, one composed of a set of wideband overlapping filters and another consisting of a set of narrowband filters. The filterbanks cooperate to isolate regions of persistent and transient signal activity. The narrowband filters help identify spectral regions containing significant signal energy and groups of wideband filters in those regions are then optimally combined to isolate and track each of the bandpass signal components. Each group of wideband filters are combined to track one bandpass component while suppressing all other neighboring components. Narrowband filters in cascade with nonlinearities are used to characterize the transient components. ©2005 IEEE.

Publication Title

ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings