Daniel E. CALLAN, Kiyoshi HONDA, Shinobu MASAKI, Ray D. KENT, Frank H. GUENTHER and Houri K. VORPERIAN
Robustness of an Auditory-to-Articulatory
Mapping for Vowel Production by the DIVA
Model to Subsequent Developmental Changes in
Vocal Tract Dimensions
Abstract:There are considerable changes in the size, shape, and the corresponding acoustical properties of the vocal tract throughout the course of development. It is necessary for a model of speech production either to adapt to these changes or to be robust with respect to them. This study explores the robustness of an auditory-to-articulatory directional map to drive vowel production during the course of development. A robust mapping is likely to be better able to accommodate for developmental alterations in muscle and sensory responses by means of co-registration within the same reference space for planning articulation than would be a mapping that is continually updated throughout development. The robustness of an auditory-to-articulatory map is investigated by using a modified version of the DIVA neural network model [Guenther et al., Psych. Rev. (1998)] in which the dimensions of the articulatory model are modified to reflect developmental changes in the dimensions of the vocal tract. Experiments using two implementations of the DIVA model were conducted: a modifiable and a static implementation. In the modifiable implementation the weights of the auditory-to-articulatory directional map are allowed to adapt throughout development by using auditory feedback as a training signal. In the static implementation weights learned during early development are used to test production performance throughout development. The results of the simulations show that, for most vowels, formant values characteristic of child speech, are produced by both the modifiable and static implementations. The performance of the static implementation demonstrates the robustness of an auditory-to-articulatory directional map to drive vowel production throughout development. This is an important finding because it demonstrates that a robust auditory-to-articulatory mapping could plausibly be used by the neural control system for speech production. Furthermore, a robust auditory-to-articulatory mapping may be able to accommodate for developmental alterations in muscle and sensory responses by modification of articulatory maps regarding changes in muscle length and innervation patterns as well as orosensory information based on adaptive feedback of self produced speech.