Two experiments were carried out to extend Logan et al.'s recent study [J. S. Logan, S. E. Lively, and D. B. Pisoni, J. Acoust. Soc. Am. 89, 874-886 (1991)] on training Japanese listeners to identify English /r/ and /l/. Subjects in experiment 1 were trained in an identification task with multiple talkers who produced English words containing the /r/-/l/ contrast in initial singleton, initial consonant clusters, and intervocalic positions. Moderate, but significant, increases in accuracy and decreases in response latency were observed between pretest and posttest and during training sessions. Subjects also generalized to new words produced by a familiar talker and novel words produced by an unfamiliar talker. In experiment 2, a new group of subjects was trained with tokens from a single talker who produced words containing the /r/-/l/ contrast in five phonetic environments. Although subjects improved during training and showed increases in pretest-posttest performance, they failed to generalize to tokens produced by a new talker. The results of the present experiments suggest that variability plays an important role in perceptual learning and robust category formation. During training, listeners develop talker-specific, context-dependent representations for new phonetic categories by selectively shifting attention toward the contrastive dimensions of the non-native phonetic categories. Phonotactic constraints in the native language, similarity of the new contrast to distinctions in the native language, and the distinctiveness of contrastive cues all appear to mediate category acquisition.