A Bayesian model of voice-onset time production.

Abstract

Talkers performed a listen-and-repeat task to investigate temporal detail in voice-onset time (VOT) productions of American English word-initial stop consonants. Experimental factors included linguistic context (isolation, carrier phrase, unfamiliar phrase, and familiar phrase), usage frequency (high and low), lexical status (word and non-word), training (baseline and posttraining), and posttraining generalization (test words and novel words). For each context, frequency, and lexical status, baseline VOT production estimates were collected, then a naive training regimen conducted, then posttraining estimates of both test words and novel words were obtained. Testing novel words explored whether the effect, if obtained, generalized throughout the lexicon. A Bayesian linear model (analogous to analysis of variance) was used to model VOT means as a function of these factors. Posterior distributions of modeled VOT means were compared across six talkers, with a focus on probing the relationships between lexical frequency and status, linguistic context, and training. Preliminary results suggest that a number of these experimental factors influence fine-grained control of VOT production. As expected from speech production data obtained from multiple talkers and thousands of productions, the overall model fit was imperfect, but the results indicate that a Bayesian model can be productively deployed for data exploration into temporal aspects of speech production.