In this dissertation, I ask how systematic patterns of pronunciation variation in speech production reveal speakers' awareness of abstract structure and usage patterns during the planning and articulation of an utterance. I explore how structure and usage combine to affect speech production at the level of the sentence and the level of the word. In particular, I focus on subject verb agreement, because computing agreement draws on sentence-level mechanisms and word-level mechanisms of speech production. At the sentence level, agreement implicates structural knowledge because speakers must know which phrases are subjects and which words are verbs in order to inflect the verb appropriately. It implicates usage patterns because, in some constructions, multiple agreement forms may be used, each with its own probability of being selected. The contextual probability of using a particular form in such constructions is thus the sentence-level intersection of structure and usage. At the word level, agreement implicates structural knowledge through paradigmatic relations between inflectional forms of a lexeme. Speakers must know that the relationship between speak and spoke, for example, is the same as the relationship between talk and talked, despite the lack of any phonological similarity between the two alternations. Usage patterns come into play at the word level because different forms within a given lexeme's paradigm are used with vastly different frequencies, and the shape of these frequency distributions affects how people retrieve words (Bien et al., 2011; Kuperman et al., 2007). The paradigmatic probability of using a particular form from an inflectional paradigm therefore represents the word-level intersection of structure and usage. In three production experiments in English and Russian, I ask how contextual and paradigmatic probability of using a particular form affect the pronunciation of that form.

In both English and Russian, I find that higher contextual probability of producing a particular agreeing form results in some type of phonetic reduction, while higher paradigmatic probability results in some type of phonetic enhancement. The phonetic feature that shows reduction or enhancement, however, depends on the language. This featural specificity leads me to propose the Contrast Dependent Pronunciation Variation hypothesis (CDPV). According to this hypothesis, structure and usage combine to restrict the types of probabilistic pronunciation variation that a speaker employs. By CDPV, pronunciation is not simply "reduced' when certain forms are contextually probable (Jurafsky et al., 2001) or "enhanced" when certain forms are paradigmatically probable (Kuperman et al., 2007). Rather, the phonetic features that vary with respect to contextual or paradigmatic probability are exactly those features which encode salient contrasts between competing forms. Phonetic "reduction" and "enhancement" are not general processes that weaken or strengthen the articulation in predictable, universal ways. Rather, they are targeted adjustments -- reductions and enhancement -- of the contrasts themselves, and are therefore sensitive to language-specific and perhaps even construction-specific properties.