The high-variability training paradigm (multiple words, phonetic contexts, and talkers) has been successful for perceptual learning of tone contrasts. Here, it is extended to training native English listeners on Tokyo Japanese pitch-accent contrasts. Participants had no previous experience with lexically contrastive pitch patterns. They learned to identify three pitch patterns in disyllabic Japanese words: 1st-syllable accented, 2nd-syllable accented, and unaccented. Immediate feedback was provided to the training group but not to the control group. The results showed an effect of training on pitch-pattern identification accuracy that was also generalized to new words spoken in new contexts. In contrast, the control group improved only on the 1st-syllable accented pattern. Error analysis suggested that the unaccented pattern is the most difficult to identify. The results are discussed in terms of native language bias and individual bias affecting second language learning in the prosodic domain.