Gaskell and Marslen-Wilson (1996) use data from cross-modal priming to show that word recognition involves phonological inference: listeners more readily recognize a word that is changed from its canonical form if that change is conditioned by a phonological process. Subsequent research has questioned whether word recognition does in fact involve phonological inference, based on evidence that perceptual compensation for assimilation can involve universal, rather than language-specific mechanisms (Gow 2003) and on evidence that changes are accepted even outside of the context in which they are phonologically conditioned (Wheeldon and Waksler 2004). We present new evidence for phonological inference based on a cross-modal priming study with Korean listeners, which shows that they accept an obstruent-tonasal change (e.g. [sok] as [so!] ‘inner part’) only when that change occurs in the context of a following nasal (e.g. [so!ma!m] ‘innermost feelings’), the environment in which it occurs in Korean phonology. We propose that listeners perform a phonological viability check, in which lexical hypotheses are submitted to the phonological grammar to determine if they are viable given the phonological context. We further suggest that a probabilistic viability check can explain apparent differences in robustness of the inference effect across different types of phonological process.