Self-organizing maps of stop consonant place from token-level locus equation inputs

Previously [D. E. Fruchter, J. Acoust. Soc. Am. 95, 2977 (1994)], identification curves were estimated for English /b,d,g/ using synthetic CV stimuli comprehensively sampling the F2-onset XF2-vowel acoustic space in the vicinity of Sussman’s /b,d,g/ locus equations. These results were used to delineate ‘‘identification surfaces’’ situated in locus equation space. The current research uses a biologically plausible neural network (the Kohonen algorithm) to model the above perception results. This algorithm is an abstraction of the local, unsupervised map-organizing process thought to occur in the brain. The Kohonen map forms a two-dimensional representation of stop consonant place categories from F2-onset and F2-vowel inputs. This emergent representation corresponds well with the experimentally observed identification surfaces and can be used to classify novel inputs and predict phoneme boundaries and confusability regions.