Early Human Vocalization Development: A Collection of Studies Utilizing Automated Analysis of Naturalistic Recordings and Neural Network Modeling

Understanding early human vocalization development is a key part of understanding the origins of human communication. What are the characteristics of early human vocalizations and how do they change over time? What mechanisms underlie these changes? This dissertation is a collection of three papers that take a computational approach to addressing these questions, using neural network simulation and automated analysis of naturalistic data.The first paper uses a self-organizing neural network to automatically derive holistic acoustic features characteristic of prelinguistic vocalizations. A supervised neural network is used to classify vocalizations into human-judged categories and to predict the age of the child vocalizing. The study represents a first step toward taking a data-driven approach to describing infant vocalizations. Its performance in classification represents progress toward developing automated analysis tools for coding infant vocalization types.The second paper is a computational model of early vocal motor learning. It adapts a popular type of neural network, the self-organizing map, in order to control a vocal tract simulator and in order to have learning be dependent on whether the model\u27s actions are reinforced. The model learns both to control production of sound at the larynx (phonation), an early-developing skill that is a prerequisite for speech, and to produce vowels that gravitate toward the vowels in a target language (either English or Korean) for which it is reinforced. The model provides a computationally-specified explanation for how neuromotor representations might be acquired in infancy through the combination of exploration, reinforcement, and self-organized learning.The third paper utilizes automated analysis to uncover patterns of vocal interaction between child and caregiver that unfold over the course of day-long, totally naturalistic recordings. The participants include 16- to 48-month-old children with and without autism. Results are consistent with the idea that there is a social feedback loop wherein children produce speech-related vocalizations, these are preferentially responded to by adults, and this contingency of adult response shapes future child vocalizations. Differences in components of this feedback loop are observed in autism, as well as with different maternal education levels