Developing Models for Multi-Talker Listening Tasks using the EPIC Architecture: Wrong Turns and Lessons Learned

Abstract

This report describes the development of a series of computational cognitive architecture models for the multi-channel listening task studied in the fields of audition and human performance. The models can account for the phenomena in which humans can respond to a designated spoken message in the context of multiple simultaneous speech messages from multiple speakers - the so-called "cocktail party effect." They are the first models of a new class that combine psychoacoustic perceptual mechanisms with production-system cognitive processing to account for the end-to-end performance in an important empirical literature.