Modeling the Acquisition of Intonation: A First Step

Abstract

Computational models of language learning focus primarily on the emergence of segmental categories to the exclusion of intonation [e.g. 1]. This runs counter to the considerable evidence that language learners rely as much on intonation as segmental categories while learning language. The current project adapts the Sensorimotor Integration Model, a popular model of language learning, to model the development of intonation. It builds on previous work that used reinforcement learning to model the development of phonation [2]. The learning simulations use a source-filter speech synthesizer to generate utterances that are then processed into intonational phrases, analyzed as f0 and amplitude. An utterance is reinforced if it is similar, as measured via distance in a self-organizing map, to a training set of infant-directed intonational phrases. Results demonstrate that, over time, the model learns to produce adult-like intonational phrases.