New directions in signal processing inspired by the auditory system

University Of Rhode Island, Kingston RI

Investigators

Abstract

Proposal #0105499 U of Rhode Island PI: Kumaresan, Ramdas Many natural and man-made sounds have a time-varying, broad-band spectrum. Although a number of computer processing methods have been developed to analyze such signals, the auditory systems of humans and animals seem to process such signals and make inferences on them much more successfully. What kind of signal analysis underlies this phenomenon? In this research auditory scientists and signal processing engineers collaborate to attempt to answer this question, and, in the process advance the state of the art in computer-based signal analysis. The potential applications of this research include improved feature extraction for sound/speech recognition/classification, separation of overlapping sounds/speech, sound source localization and better understanding of the auditory system. The auditory system, in addition to performing some form of spectral analysis, also extracts temporal modulation information from the acoustic signals. This information is encoded in the timing patterns of the spike trains that originate in the inner ear. In this research, with the help of auditory scientists, the above observations are translated into mathematically tractable signal processing problems. Specifically, the following basic question is explored: how can band-pass signals be represented by timing information only, as opposed to traditional Nyquist-rate amplitude sampling (as in Shannon's sampling theorem)? Recent results by the investigators indicate that the information about the phase and envelope modulations of arbitrary band-pass signals can be represented by certain zero-crossings alone, if appropriate adaptive preprocessing(demodulation) is performed on the signal. Based on these ideas an analysis-synthesis procedures is being developed in which time-varying signals such as speech can be effectively represented by a small number of modulated components using only timing information. The robustness of such procedures will be examined and the analysis-synthesis procedures will be applied to a number of applications mentioned in the above paragraph.