It has been observed in the literature that stem-final coronal obstruents of nouns in Korean are generally in variation with [s] in the prevocalic position: for example, /pathh-ɨl/ [pathh-ɨl] ~ [pasɨl] ‘field, accusative’. In addition, nouns with final noncoronal stops have variants ending in lenis stops: for example, /iph-e/ [iph-e] ~ [ipe] ‘leaf, locative’. Recent survey and experimental studies reveal that a wide set of coronal obstruents [s, ch, th, c, t] may occur as variants. Moreover, there is an order of preference among them: in general, s >> ch, th >> c, t. This paper first shows that the observed relative preference among variants is matched by the distribution of lexical final obstruents in noun stems. Building on Albright’s Paradigm Learning Model, I provide a unified account for the occurrence of most variants and their relative preference by proposing stochastic rules deriving the paradigmatically related forms of noun stems. In addition, a wug-test is carried out to investigate the productivity of such rules. Results suggest that these rules are productive, providing the evidence for their cognitive presence.