Native speakers of Mandarin who learn English as a second language (L2) frequently misidentify unreleased tokens of /t/ and /d/ in the final position of English words. The purpose of this study was to evaluate two training methods that might be used to increase Mandarin adults’ accuracy in identifying such stops. Subjects were assigned to receive training using a two-alternative forced-choice procedure or a categorical same/different discrimination procedure. Small but significant increases in the percentage of correct identifications of /t/ and /d/ tokens were obtained for both groups of native Mandarin subjects, although their gains did not differ significantly. These gains were still evident for both groups 2 months after completing the training. The effect of training for both groups generalized to words that were not used in training. Contrary to the hypothesis that identification training promotes more robust long-term memory representations than same/different training, the magnitude of generalization observed for the two groups did not differ significantly. However, subjects seemed to maintain the effects of the same/different training better than the effects of the identification training. Taken together, the results obtained here challenge the view that identification training is superior to same/different training as a means for training novel phonetic contrasts.