Vowel devoicing happens in Japanese when the high vowel is between voiceless consonants. The aim of this study is to investigate the lexical representation of vowel devoicing. A long-term repetition-priming experiment was conducted. Participants shadowed words containing either a devoiced or a voiced vowel in three priming paradigms, and their shadow responses were analyzed. It was found that participants produced the phonologically appropriate allophone most of the time based on the consonantal environments. Shadowing latencies for the voiced stimuli were faster than for the devoiced stimuli in the environment where the vowel should be voiced; while, no significant RT difference was observed between the two forms in the environment where vowel devoicing was expected. In addition, a priming effect between the devoiced and voiced stimuli emerged only in the devoicing environment. The results suggest that since vowel devoicing is very common in spoken Japanese, the devoiced form may be stored in the lexicon. The results also suggest a link between the two forms in the lexicon and a direct access between an input and a lexical representation without going through intermediate levels that usually cost extra processes. (Contains 6 tables, 1 figure and 1 note.)