A classic problem in speech perception is the information listeners use to distinguish words.  This must be solved by every child over development as they determine constellation of acoustic cues used by their language. Classic approaches posit that people learn to perceive individual sounds first and then assemble them to acquire words. We asked the converse question: does learning words help acquire individual sounds, more specifically whether learning highly similar words improves speech perception. Learning similar words may force better encoding of the individual sounds that comprise them as there are fewer cues available to distinguish them.  For example, distinguishing pin and pen requires attention to fine-grained distinctions in the vowel; while pin and lamp can be distinguished by any number of factors. We examined this by teaching participants a set of novel words. One group learned words which differed on numerous dimensions, while a second learned a set in which half were similar. We then compared how well the subjects encoded speech in three tasks. We found that participants learning similar words were more likely to reject mispronounced words (e.g. pin spoken as peen), and showed slower reaction times when recognizing mispronounced words.  Both findings generalized to the dissimilar words learned by this group, and suggest that word representations are more specific after training on similar words.  A separate vowel categorization task also showed that participants learning similar words had more distinctly defined vowel categories suggesting that the benefits of learning similar sounding may enhance speech perception more generally.