This study examined the role of stimulus variability in learning non-native phonetic contrasts (suprasegmentals) for word identification by adults, considering whether all learners benefit from high-variability training. We trained native English-speaking adults to use Mandarin lexical tones to identify 18 English pseudowords. Subjects were randomly assigned to two experimental conditions: a multi-talker group in which learners trained on stimuli produced by four talkers, and a single-talker group, in which each learner trained on only one of the four talkers. Before training, all subjects were tested on their ability to identify these pitch patterns in a non-lexical context. Subjects with high pitch identification ability learned more successfully than those with lower pitch identification ability. Further, multi-talker training was beneficial only for learners with high pitch identification ability, whereas learners with low pitch identification ability benefited more from single-talker training.