This study investigates the specificity of predictive coding in spoken word comprehension using event-related potentials (ERPs). We measured word-evoked ERPs in Catalan speakers listening to semantically constraining sentences produced in their native regional accent (experiment 1) or in a non-native accent (experiment 2). Semantically anomalous words produced long-lasting negative shift (N400) starting as early as 250 ms, thus reflecting phonological as well as semantic mismatch. Semantically expected but phonologically unexpected (non-native forms embedded in a native context) produced only an early (∼250 ms) negative difference. In contrast, this phonological expectancy effect failed for native albeit phonologically unexpected target words embedded in a non-native context. These results suggest phonologically precise expectations when operating over native input, whilst phonologically less specified expectations in a non-native context. Our findings shed light on contextual influence during word recognition, suggesting that word form prediction based on context is sensitive and adaptive to phonological variability.