Learning to distinguish nonnative words that differ in a critical phonetic feature can be difficult. Speech training studies typically employ methods that explicitly direct the learner's attention to the relevant nonnative feature to be learned. However, studies on vision have demonstrated that perceptual learning may occur implicitly, by exposing learners to stimulus features, even if they are irrelevant to the task, and it has recently been suggested that this task-irrelevant perceptual learning framework also applies to speech. In this study, subjects took part in a seven-day training regimen to learn to distinguish one of two nonnative features, namely, voice onset time or lexical tone, using explicit training methods consistent with most speech training studies. Critically, half of the subjects were exposed to stimuli that varied not only in the relevant feature, but in the irrelevant feature as well. The results showed that subjects who were trained with stimuli that varied in the relevant feature and held the irrelevant feature constant achieved the best learning outcomes. Varying both features hindered learning and generalization to new stimuli.