Studies on perceptual learning are motivated by phonetic variation that listeners encounter across speakers, items, and context. In this study, the authors investigate what control the listener has over the perceptual learning of ambiguous /s/ pronunciations through inducing changes in their attentional set. Listeners' attention is manipulated during a lexical decision exposure task such that their attention is directed at the word-level for comprehension-oriented listening or toward the signal for perception-oriented listening. In a categorization task with novel words, listeners in the condition that maximally biased listeners toward comprehension-oriented attentional sets showed the most perceptual learning. Focus on higher levels of linguistic meaning facilitated generalization to new words. These results suggest that the way in which listeners attend to the speech stream affects how linguistic categories are updated, providing insight into the qualitative differences in perceptual learning between the psychophysics and language-focused literatures.