A series of six experiments addressed the question, what is the nature of sociolinguistic knowledge? Three sub-questions were investigated: First, how do listeners use facts about speech to inform their beliefs about speakers? Second, and conversely, how might listeners use facts about speakers to inform their perceptions of speech? Third, what is the nature of the representations that listeners form of the social conditioning of variation? Results of Experiments 1a and 1b demonstrate that listeners can infer characteristics of speakers from their use of individual sociolinguistic variables. Results of Experiments 2a and 2b show that listeners use social information about speakers to understand their speech. Results of Experiments 3a and 3b suggest that listeners form socially specified representations at some level during speech processing, but that their underlying phonological representations are the same for all speakers. Taken together, results of all six experiments show bidirectional influences of information in the speech stream on inferences about social characteristics of the speaker, and of social information on speech perception. I propose a Bayesian approach to integrating social information into a model of language comprehension, in which bidirectional influences between verbal and nonverbal factors emerge as a natural consequence of our cognitive capacities for learning and inference.