Monolingual speakers of Japanese were trained to identify English /r/ and /l/ using Logan et al.'s [J. Acoust. Soc. Am. 89, 874-886 (1991)] high-variability training procedure. Subjects' performance improved from the pretest to the post-test and during the 3 weeks of training. Performance during training varied as a function of talker and phonetic environment. Generalization accuracy to new words depended on the voice of the talker producing the /r/-/l/ contrast: Subjects were significantly more accurate when new words were produced by a familiar talker than when new words were produced by an unfamiliar talker. This difference could not be attributed to differences in intelligibility of the stimuli. Three and six months after the conclusion of training, subjects returned to the laboratory and were given the post-test and tests of generalization again. Performance was surprisingly good on each test after 3 months without any further training: Accuracy decreased only 2% from the post-test given at the end of training to the post-test given 3 months later. Similarly, no significant decrease in accuracy was observed for the tests of generalization. After 6 months without training, subjects' accuracy was still 4.5% above pretest levels. Performance on the tests of generalization did not decrease and significant differences were still observed between talkers. The present results suggest that the high-variability training paradigm encourages a long-term modification of listeners' phonetic perception. Changes in perception are brought about by shifts in selective attention to the acoustic cues that signal phonetic contrasts. These modifications in attention appear to be retrained over time, despite the fact that listeners are not exposed to the /r/-/l/ contrast in their native language environment.