Emergence of Vocal Developmental Sequences in a Predictive Coding Model of Speech Acquisition

Learning temporal patterns among primitive speech sequences and being able to control the motor apparatus for effective production of the learned patterns are imperative for speech acquisition in infants. In this paper, we develop a predictive coding model whose objective is to minimize the sensory (auditory) and proprioceptive prediction errors. Temporal patterns are learned by minimizing the former while control is learned by minimizing the latter. The model is learned using a set of synthetically generated syllables, as in other contemporary models. We show that the proposed model outperforms existing ones in learning vocalization classes. It also computes the control/muscle activation which is useful for determining the degree of easiness of vocalization.