A growing body of recent work on the phonetics–phonology interface argues that many phonological patterns refer to psychoacoustic similarity – perceived similarity between sounds based on detailed acoustic information. In particular, two corresponding elements in phonology (e.g. inputs and outputs) are required to be as psychoacoustically similar as possible (Steriade 2001a, b, 2003; Fleischhacker 2005; Kawahara 2006; Zuraw 2007). Using a corpus of Japanese imperfect puns, this paper lends further support to this claim. Our corpus-based study shows that when Japanese speakers compose puns, they require two corresponding consonants to be as similar as possible, and the measure of similarity rests on psychoacoustic information. The result supports the hypothesis that speakers possess a rich knowledge of psychoacoustic similarity and deploy that knowledge in shaping verbal art patterns.