The dorsal stream in speech processing: Model and theory

Abstract

The ability to produce and comprehend spoken language requires an internal understanding of the complex relations between articulatory gestures and their acoustic consequences. Recent theories of speech processing propose a division between the ventral stream, which involves the mapping of acoustic signals to lexical/semantic representations, and the dorsal stream, which mediates the mapping between incoming auditory signals and articulatory output. We present a connectionist model of the dorsal stream of speech processing that utilizes a novel schematic representation of time-varying acoustics and a featural mapping of articulation. The model successfully learns a large training vocabulary, accurately produces novel items and demonstrates patterns of perceptual errors highly similar to those observed in human subjects.


Back to Friday Papers