UC Berkeley

In this dissertation, I investigate how word predictability in context modulates the listener’s attention to phonetic details, and how this in turn affects sound change. Three sets of experiments are designed to investigate these questions: In the first set of experiments, involving discriminability tasks, I demonstrate that (1) contextual predictability affects speech perception, and that listeners attend more to the phonetic details of unpredictable speech. In the second set of experiments I use the phonetic accommodation paradigm to show that (2) the effect of contextual predictability on speech perception in turn affects speech production. This by itself suggests relevance in sound change. In the third set of experiments I apply the model to a specific example of sound change: the reduction of function words. Using an error detection task I show that (3) listeners attend to the details of content words more than function words (with all other variables controlled for) which is linked to their differences in contextual predictability. I then propose a two-step model of sound change involving the propagation of contextually-modulated variation with a perceptual (rather than production) bias followed by the acquisition of new variants.

The results build and expand on several strands of literature which have not been fully connected previously. The findings for the effect of predictability on speech perception corroborate a number of past experiments showing that higher level linguistic information can have the effect of aiding speech recognition (Miller, Heise & Lichten 1951, Pollack & Pickett 1963), perceptually restoring missing information (Warren 1970, Marslen-Wilson, & Welsh 1978, Samuel 1981), or generally diverting attention from the raw auditory signal (Cole, Jakimik, & Cooper 1978, Ganong 1980). Additionally, this research considers dual-processing models of speech perception (Norris & Cutler 1979, Lindblom et al. 1995, Hickok and Poeppel 2004, 2007) in a broader context, considering how word predictability and expectancy modulate the type of listening used. The findings also add to the literature on exemplar theory (Johnson 1997, Pierrehumbert 2002, Goldinger 2007), particularly to hybrid models including both abstractions and exemplar clouds within the lexicon. Finally, I propose a new model of perception-based sound change driven by contextual predictability that can account for cross-linguistically common patterns of function word and morpheme reduction (Bell et al. 2001, Jurafsky et al. 2001, Beckman 1998) that does not rely on teleological production-based accounts of reduction (Lindblom 1990, Alyett & Turk 2004).