This study investigated the role and time-course of phonology in accessing meaning during silent reading. In Experiment 1, the homophone effect was replicated in a semantic categorization task in French. When deciding whether a stimulus belonged to a semantic category (FOOD), subjects made more errors to homophones (MEET) than to orthographic controls (MELT). In Experiment 2, event-related brain potentials (ERPs) were used to study the online development of this effect. If access to meaning was mediated by phonology, smaller N400 components should be obtained to homophones than to orthographic controls. The ERP data exhibited a full-blown N400 component to homophones that did not differ from the N400 to controls. No differences between homophones and controls were found before the N400. After the N400, however, homophones differed from controls, with ERPs to homophones being similar to those of correct category exemplars. The results suggest that the final selection of a word's meaning does not depend on its phonological form. This result is incompatible with a strong phonological view according to which the only way to meaning is via a word's phonology.