A central issue in speech recognition is how contrastive phonemic information is stored in the mental lexicon. The conventional view assumes that this information is closely related to acoustic properties of speech. Considering that no word is ever pronounced alike twice and that the brain has limited capacities to manage information, an opposing view proposes abstract underspecified representations where not all phonemic features are stored. We examined this proposal using event-related brain potentials, in particular mismatch negativity (MMN), an automatic change detection response in the brain that is sensitive to language-specific phoneme representations. In the current study, vowel pairs were presented to subjects, reversed as standard and deviant. Models not assuming underspecification predict equal MMNs for vowel pairs regardless of the reversal. In contrast, enhanced and earlier MMNs were observed for those conditions where the standard is not phonologically underspecified in the mental representation. This provides the first neurobiological evidence for a featurally underspecified mental lexicon.