A widely held view for many years was that listeners derive an abstract phonetic representation during speech processing and, in so doing, discard detailed information about the signal. However, more recent research has shown that the representations of speech are richer than this emphasis on abstract categories would suggest, and that listeners retain in memory a substantial amount of fine-grained information. One line of evidence for this new view comes from research showing that phonetic categories are internally structured in a graded fashion, with some members of the category perceived as better exemplars than others. In this talk I will briefly review some findings that demonstrate basic characteristics of such structure and then present results from two new sets of studies that further examine this structure and its role in processing. In both sets, we examined the syllable-initial voicing distinction specified by voice-onset time (VOT), with a specific focus on the voiceless category. The first set of studies, which used a category-goodness rating task, investigated how a higher-order linguistic contextual variable, lexical status, affects the internal structure of the category. We knew from previous work that acoustic-phonetic contextual factors (e.g., speaking rate) alter not only perception of stimuli in the category boundary region, but also which stimuli are perceived as the best category exemplars. The new results showed that lexical status has a qualitatively different effect: Although a change in lexical status also altered perception of stimuli in the category boundary region, the best exemplars of the category remained relatively fixed. Thus contextual variables that have comparable effects in the boundary region can in some cases be dissociated with respect to their effects on a category's best exemplars. The second set of studies, which used a speeded categorization task, examined how differences in perceived category goodness affect the speed of phonetic categorization. The results showed that although categorization time was relatively slow for poor exemplars in the voiced-voiceless boundary region, poor exemplars with VOT values longer than the best exemplars (and thus far from the category boundary) were identified as quickly as the best exemplars themselves. This suggests that categorization time depends more on an exemplar's location in perceptual space vis-à-vis a boundary with a competing category than on its perceived category goodness. Taken together, the new findings underscore the importance of considering both internal category structure and phonetic category boundaries when developing models of phonetic categorization.