A naturalistic approach to the cocktail party problem.

While studies of simple acoustic features have provided excellent bases for models of spatial hearing, we are seeking, here, to create a new paradigm for examination of shared attention and scene analysis in natural environments, where the listener is confronted with semantic information from multiple sources. In this new simulation of the cocktail party problem, a subject (S) is questioned, on-line, about information heard in multiple simultaneous stories spoken by different talkers. Questions based on brief passages in the stories are presented visually for manual response. To ensure that responses are based on semantic information rather than just keywords, the latter are replaced in the questions with synonyms. Pay is for performance, and S knows that while a majority of the questions come from a "primary talker," there is potential value in obtaining information from secondary sources. Results, to date, suggest that obtaining semantic information from separate stories is limited by two spatial factors, an exclusive filter that protects information from the attended talker and an inclusive filter that incorporates information from secondary talkers.