Training studies of the past decade demonstrated that laboratory training can improve the ability of adults to perceive non-native speech contrasts. Those studies used various training methods, which differed in several variables, such as target contrast, task, stimulus material, stimulus sequence, and amount of training. Recently, the effects of these variables have been systematically examined in a series of studies [e.g., Lively et al., J. Acoust. Soc. Am. 96, 2076--2087 (1993); Magnuson et al., J. Acoust. Soc. Am. 97, 3417(A) (1995); Bradlow et al., Proc. ICPhS, 562--565 (1995)]. When Japanese speakers were trained on /r/--/l/ minimal pairs using an identification task with natural tokens, the effects were generalized to novel talkers and novel words. The following findings were also obtained: First, the training using a single talker sometimes failed, suggesting the importance of high variability in the stimulus set. Second, the training in perception resulted in improvement in production, implying a close link between speech perception and speech production. Third, trainees retained their perception and production ability as much as 3 to 6 months after the conclusion of the training, proving that the laboratory training produced long-term modifications in both domains. Implications for the optimization of this training method will be discussed in the context of theories of perceptual learning and development of phonological categories.