Naming latencies were measured for single- and multiple-reading (pronunciation) kanji words with two frequency levels in Experiment 1. Results showed that multiple-reading kanji are named much slower than are single-reading controls, the effect of multiple reading being greater than that of frequency. There was found to be measurable competition between two readings of a multiple-reading kanji. In Experiment 2, it was demonstrated that there are multiple-reading kanji which are named as fast as single-reading kanji, thereby showing that no significant competition is induced in them. In light of the theoretical perspective in this study, the conflicting results reported by previous studies are discussed.