A Biologically Plausible Mechanism for Phonology Acquisition in Human Infants

Abstract

Infants go through a serial developmental process in language acquisition, which is observable as specific linguistic milestones: by around six weeks of age, infants coo; by about six months they begin to babble, entering the reduplicated babbling stage at about eight months and non-reduplicated babbling at 11 months, with the first word not long after that. There are several theories of how infants do this, but few of them are end-to-end testable, often either falling short of being end-to-end, or else being too vague in certain details. This thesis presents a new theory of phonology acquisition, called the Evolving Signal Chain (ESC) theory of speech acquisition, which is extensible to lexical acquisition and which is end-to-end testable by means of computer simulation. Such a simulation is described and results from key portions of it are given. Key results include an analysis of deep convolutional autoencoders as a plausible means of learning an unsupervised encoding of raw speech and the use of that encoding in learning to reproduce natural speech. Results show that continuous natural speech can be encoded in at least two time scales and that these encodings can be used to recreate the beginnings of marginal babbling. Original contributions include the use of deep learning techniques in a computer model of primary language acquisition, and the use of an autoencoder instead of a self-organizing map, which allows for learning an optimal encoding space without specifying many a priori features.