What kinds of representations underlie listeners knowledge of the way different types of speakers speak (sociophonetic variation)? Listeners store information about the social conditioning of phonetic variation, and use this knowledge to inform their perceptions of speech. But are expectations about sociophonetic variation stored on the level of the word, or do they also apply to never-heard words? A reaction-time experiment investigated whether listeners form the same representations of words for all speakers, or form different representations based on the social characteristics of the speaker. Participants reactions to real words were compared with reactions to nonce words, in order to investigate the level at which these representations differ. Results showed that social information influenced the processing of ambiguous nonce words the same way it influenced the processing of real words, suggesting that listeners form different representations of speech for speakers with different social characteristics at the level of the sub-lexical chunk. This finding about listeners knowledge of sociolinguistic variation supports the inclusion of a sub-lexical level of representation in exemplar theories of speech perception.