MODELING THE PERCEPTUAL MAGNET EFFECT AND CATEGORICAL PERCEPTION USING SELF-ORGANIZING NEURAL NETWORKS

Bernd J. Kröger1, Peter Birkholz2, Jim Kannampuzha3 & Christiane Neuschaefer-Rube3
1Department of Phoniatrics, Pedaudiology, and Communication Disorders, UKAachen and Aachen University; 2Department of Computer Science, University Rostock; 3Department of Phoniatrics, Pedaudiology, and Communication Disorders, UKAachen and Aachen University

ID 1013
[full paper]

Purpose: A neural model of speech production based on self-organizing neural networks is introduced. The model is capable of describing speech acquisition stages as well as speech perception effects. Method: 20 instances of the neural model were trained imitating early stages of speech acquisition (babbling and imitation) in order to create 20 different virtual toddlers. Perceptual experiments were performed using these virtual listeners. Results: Typical effects of speech perception occur by performing identification experiments on vocalic and consonantal acoustic stimulus continua. Consonantal categorical perception directly occurs during babbling while the perceptual magnet effect occurs later on during language specific imitation training. Conclusion: This neural model of speech production using self-organizing neural networks is capable (a) of illustrating the close relationship between production and perception of speech and (b) of elucidating the formation of speech perception effects during speech acquisition.