The ability of a binaural listener to concentrate on speech from a particul
ar location whilst preferentially suppressing speech from other locations h
as been termed the binaural "cocktail party" effect. The recently developed
unsupervised artificial neural network models for blind signal processing
have provided additional perspectives on the binaural cocktail party effect
. This paper presents a model of the cocktail party effect based on a tempo
ral anti-Hebbian adaptive maximum likelihood estimator. The complicating ef
fects of reverberation and physically non-stationary speech sources on unwa
nted speech suppression are empirically investigated. (C) 1998 Elsevier Sci
ence B.V. All rights reserved.