A Computational Model of Spatial Hearing

The human auditory system performs remarkably at determining the positions of sound sources in an acoustic environment. While the localization ability of humans has been studied and quantified, there are no existing models capable of explaining many of the phenomena associated with spatial hearing. This thesis describes a spatial hearing model intended to reproduce human localization ability in both azimuth and elevation for a single sound source in an anechoic environment. The model consists of a front end, which extracts useful localization cues from the signals received at the eardrums, and a probabilistic position estimator, which operates on the extracted cues. The front end is based upon human physiology, performing frequency analysis independently at the two ears and estimating interaural difference cues from the resulting signals. The position estimator is based on the maximum-likelihood estimation technique. Several experiments designed to test the performance of the model are discussed, and the localization blur exhibited by the model is quantified. A "perceptual distance" metric is introduced, which allows direct localization comparisons between different stimuli. It is shown that the interaural intensity difference (IID) contains sufficient information, when considered as a function of frequency, to explain human localization performance in both azimuth and elevation, for free-field broad-band stimuli. Thesis Supervisor: Barry L. Vercoe Title: Professor of Media Arts and Sciences Thesis Supervisor: Patrick M. Zurek Title: Principal Research Scientist Research Laboratory of Electronics