ABSTRACT
We present a new variant of a simple recurrent network to model auditory word recognition in continuous speech and address the issue of lexical segmentation. Simulations based on small word sets show that the system provides a near-optimal solution to the opposite constraints of speed, which requires that lexical processing be immediate, and reliability, which imposes that identification decisions be postponed until unambiguous information is available. Contrary to an often-heard statement, the simulations show that the existence of embedded words is not incompatible with the notion of continuous on-line lexical processing.
