A visual semantic categorization task in English was performed by native English speakers (Experiment 1) and late bilinguals whose first language was Japanese (Experiment 2) or Spanish (Experiment 3). In the critical conditions, the target word was a homophone of a correct category exemplar (e.g., A BODY OF WATER-SEE; cf. SEA) or a word that differed from the correct exemplar by a phonological contrast absent in the bilinguals' first language (e.g., USED FOR COOLING DOWN-FUN; cf. FAN). Homophones elicited more false positive errors and slower processing than spelling controls in all groups. The Japanese-English bilinguals, but not the Spanish-English bilinguals, also displayed "near-homophone" effects (i.e., homophone-like effects from minimal pairs on nonnative contrasts). We conclude that second-language visual word recognition is influenced by first-language phonology, although the effect is conditioned by the first-language orthographic system. Near-homophone effects can occur when the orthographic systems of the late bilingual's two languages are different in type (e.g., alphabetic vs. non-alphabetic), but may be blocked if the languages use the same writing script (e.g., Roman alphabet). (Contains 3 footnotes and 5 tables.)