Spoken words are highly variable. A single word may never be uttered the same way twice. As listeners, we regularly encounter speakers of different ages, genders, and accents, increasing the amount of variation we face. How listeners understand spoken words as quickly and adeptly as they do despite this variation remains an issue central to linguistic theory. We propose that learned acoustic patterns are mapped simultaneously to linguistic representations and to social representations. In doing so, we illuminate a paradox that results in the literature from, we argue, the focus on representations and the peripheral treatment of word-level phonetic variation. We consider phonetic variation more fully and highlight a growing body of work that is problematic for current theory: words with different pronunciation variants are recognized equally well in immediate processing tasks, while an atypical, infrequent, but socially idealized form is remembered better in the long-term. We suggest that the perception of spoken words is socially weighted, resulting in sparse, but high-resolution clusters of socially idealized episodes that are robust in immediate processing and are more strongly encoded, predicting memory inequality. Our proposal includes a dual-route approach to speech perception in which listeners map acoustic patterns in speech to linguistic and social representations in tandem. This approach makes novel predictions about the extraction of information from the speech signal, and provides a framework with which we can ask new questions. We propose that language comprehension, broadly, results from the integration of both linguistic and social information.