The dissertation examines the relationship between social and linguistic knowledge using a series of experiments eliciting linguistic convergence to Southern speech. I draw a terminological and theoretical distinction between previously observed input-driven convergence, in which speakers converge toward a linguistic form directly observed in the input, and expectation-driven convergence, in which speakers converge toward a linguistic form they only expect but do not observe in the immediate input. Using a novel Word Naming Game paradigm designed to elicit convergence toward expected rather than observed linguistic behavior, Experiment 1 finds experimental evidence for expectation-driven convergence, which had previously only been anecdotally observed; participants converge toward glide-weakened /ay/, a salient feature of Southern English, which they may expect but never directly observe from a Southern-accented model talker. The existence of expectation-driven convergence suggests that accounts of convergence relying on tight perception-production links where production is derived directly and automatically from the input cannot straightforwardly explain all instances of convergence. Experiment 2 investigates the perceptual underpinnings of input- and expectation-driven convergence using an auditory lexical decision task in which participants judge glide-weakened /ay/ items (e.g., "bribe" produced as "brahb") as words or non-words. I find higher word-endorsement rates for glide-weakened /ay/ words for participants who have recently heard a Southern-accented (compared to Midland-accented) talker, even if the Southern talker never produces the /ay/ vowel. Individual perception and production responses toward glide-weakened /ay/ show little evidence for strong individual perception-production links, though findings are consistent with an interpretation where perceptual shifts are a necessary (but not sufficient) precursor to production shifts. Finally, Experiment 3 uses a dialect-label manipulation version of the Word Naming Game and demonstrates that both top-down information about social categories and bottom-up acoustic cues independently contribute to expectation-driven shifts in production and perception. Further, reliance on these cues differs across dialect backgrounds, providing insights into the way sociolinguistic associations are formed and mentally represented. Taken together, results support a model of cognition in which social and linguistic information are tightly linked.