The present study investigated whether the ability to encode the sounds of difficult second-language (L2) contrasts into novel nonnative lexical representations is modulated by the phonological form of the words to be learned. In 3 experiments, German learners of English were trained on word-picture associations with either novel minimal pairs only differing in the difficult /[epsilon]/-/ae/ contrast (Experiments 1 and 2; e.g., "tendek-tandek") or pairs that additionally differed in their second syllables (Experiment 3; e.g., "tenzer-tandek"). Word recognition was assessed by means of a visual-world eye-tracking task. We asked whether learners would be more successful at encoding a distinction between the 2 vowels in the minimal-pair than in the nonminimal-pair items because of the central role of the contrast for accurate word learning with minimal pairs. Results from eye-fixation analyses at test showed that learners recognized /ae/-items faster than /[epsilon]/-items when they were minimal pairs and these pairs had already appeared together on the screen on a number of training trials (Experiment 1 vs. 2). This asymmetry could not be replicated with nonminimal pairs (Experiment 3). In line with previous studies, the asymmetry in Experiment 2 is taken as evidence of lexical separation for /[epsilon]/ and /ae/. Accordingly, we argue that exposure to the minimal-pair stimuli highlighted the challenging distinction by enhancing listeners' attention to the critical sounds and conclude that the encoding of difficult L2 contrasts into the lexicon is more likely when the phonological form of the words emphasizes relevant phonological differences that may otherwise remain unnoticed.