This paper addresses questions concerning the mental representation of language through the results of a study investigating the effects of various training methods for altering language performance. Specifically, we question whether a single phonological form accompanies an item in the lexicon (the standard assumption of linguistic theory) or whether each item bears two distinct phonological representations in the grammar - one underlying perception, the other driving production. We test these alternatives by examining whether subjects' performance on perception and production tasks is differentially affected by the training method used to teach phonological contrasts that are absent from learners' native system but distinctive in a foreign language. Speakers whose native language is American English were instructed in producing the phonemic quantity contrasts of Japanese by working with either a Level HI tape recorder, a native speaker of Japanese, or a computer that provided real-time visual comparisons between learners' productions and native Japanese models. Analysis of subjects' performance on a battery of pre- and post-training tests reveals that these various training methods may differentially tune production or perception acuity in a manner more consistent with the theory that lexical items carry two distinct phonological forms.