Warlaumont, A. S. 1, 2 , Westermann, G. 3 & Oller, D. K. 1, 4
1 University of Memphis
2 University of California, Merced
3 Lancaster University
4 Konrad Lorenz Institute for Evolution and Cognition Research
In the first few months of life human infants' non-cry vocalizations transition from short, quiet sounds to being longer and louder, demonstrating increased ability to actively control phonation, the generation of sounds using the vocal folds. Eventually, they come to have vowels and consonants resembling those in the infant’s native language, which requires precise control of the muscles of the upper vocal tract. The development of these abilities depends on a number of factors such as maturational changes in vocal tract morphology, perceptual development, caregiver imitation of produced sounds, and social responses that are contingent on the characteristics of child vocalizations.
Several connectionist models have addressed the adaptation of vowel productions to the ambient language, but none thus far have incorporated the role of social reinforcement in this adaptation, despite evidence from both experiments and naturalistic observation of human infants that support its role in infant vocal development. Furthermore, the development of phonatory control preceding vowel learning has not been addressed.
Here we present a connectionist model using a self-organizing map to control the muscle activation parameters of a realistic vocal-tract-simulating vocalization synthesizer. The model produces exploratory vocalizations by random activation of its neurons. Reinforcement is provided when the resulting sounds are phonated (i.e., not silent) and when sounds resemble vowels in a target language (English or Korean). On reinforced trials neuromotor connection weights are adapted so that similar productions become more likely in the future. No learning takes place on non-reinforced trials. We show that the model mimics infants’ developmental process, leading to articulations that are progressively more phonated and that come to resemble the vowel sounds in the native language. The model demonstrates a way in which reinforcement and self-organization may be combined in a connectionist model of speech development.