Acoustic binaural correspondence used for localization of natural acoustic signals

The left-right signal correspondence problem, that is considered as one of the most prominent problems by visual stereoscopic computational models, is much ignored by computational auditory stereophonic models. The correspondence problem, which is trivial if only one acoustic source is present, is highly complicated for a multiple sources environment. We present a computational model able to perform localization of natural complex acoustic signals (one or two human speakers). The model relies mainly on computing the cross-correlation functions of selected frequency channels arriving at the two ears, and performing a weighted integration on these functions. Thus, first attempts are made to establish a correspondence between acoustic features of the two channels. Preliminary results show that this model, which might be compared to "early vision" models in computational vision research, can serve as a first step in analyzing the acoustic scene.