Contextual Predictability and Phonetic Reduction

Abstract

Phonetic reduction is a process which alters the acoustic quality of a sound, often a vowel or word, to a perceived weaker or shorter state. Previous research suggests that the degree of reduction of a word is influenced by the contextual predictability of words in the context. However, the nature of how the context direction and size governs phonetic reduction has not been thoroughly explored. The advancement of self-supervised language models provides a means to assign meaningful estimates of word predictability conditioned on different contexts. This paper explores the effect of contextual predictability on phonetic reduction making use of such models. We train instances of GPT-2 on different context directions (past, future, and bidirectional) and context sizes (bigram vs. sentence) to provide measures of conditional word predictability, then use linear regression to quantify their correlation with a measure of phonetic reduction (word duration). Our results provide evidence suggesting that the contextual probability of a word given the following context correlates with word duration more strongly than the past context and the bidirectional contexts for both context sizes, suggesting that phonetic reduction may be a reliable indicator of reduced cognitive load in a speaker’s planning of the rest of an utterance.