Acoustic properties of words vary based on idiolects of different speakers, yet listeners appear to understand this varying speech input effortlessly. How do listeners adapt to acoustic variation across speakers? Previous research has shown that listeners can implicitly shift phonetic category boundaries in a speaker-specific fashion during a two-step paradigm involving lexical decision on critical words with a word-medial ambiguous phoneme, followed by a phonetic categorization task of that ambiguous phoneme (see Kraljic and Samuel, 2005). Other studies report that this boundary shift does not generalize to other speakers (Kraljic et al., 2008). However, for a given speaker, it does generalize to different positions within a word, for instance, from word-medial to word-initial (Jesse and McQueen, 2011). Importantly, these generalization tests used identical ambiguous phonemes in each position, rather than context-appropriate tokens, allowing the possibility of token-specific learning rather than true generalization. The current research extends these findings by providing distinct, position-appropriate tokens of the ambiguous phoneme during training and test items, and shows that the generalization to new word position is retained. Thus, generalization within a speaker is not token-specific, but instead exhibits some level of abstraction.