Previous studies have shown listeners' ability to remember fine phonetic details (e.g., Mullennix et al., 1989), providing support for the episodic view of speech perception. The imitation paradigm (Goldinger, 1998, Shockley et al., 2004), in which subjects' speech is compared before and after they are exposed to target speech (= study phase), has shown that subjects shift their production in the direction of the target, indicating not only episodic memory in speech perception but also the close tie between speech perception and production. The purpose of the current study is to determine whether and how this imitation effect is generalized to new stimuli, as well as to investigate the locus of the effect. In the study phase, subjects listened to a word containing items with initial /p/ and /t/ that had extended VOT. In the pre- and post-study phases, subjects produced words from lists including 1) the words in the study list, 2) the segments /p/ and /t/ in new words, and 3) the segment /k/, which did not occur during study. Initial results replicated Shockley et al. (2004), showing the imitation effect of extended VOT: subjects produced longer VOT for post-study than pre-study. Furthermore, the results indicated the trend that the modeled feature for one segment (/p/) was generalized to a new segment (/k/). The specific source of this imitation effect is still unknown. That is, when a subject shifts production of a particular sound in a particular word, it is uncertain whether the shift has been caused by change in the speaker's register (e.g., 'speak more carefully'), or by change in lexical/phonetic representations. If the shift in production is indeed due to some change in lexical/phonetic representation, what would be the size of linguistic unit influenced by the effect? If the imitation effect is truly due to episodic memory, only the manipulated variable (in this case, VOT) should be affected. On the other hand, if we observe changes in other variables (e.g., the following vowel, entire word), the imitation effect is more likely to be due to some more global aspect of speech production. The implications of the data for exemplar versus abstract views of speech perception will be discussed.