Speech perception in noise is dependent on stimulus-driven and knowledge-driven processes. Here we investigate the neural correlates and time course of discrepancies between incoming auditory information (i.e. stimulus-driven processing) and linguistic expectations (knowledge-driven processing) by including 20 normal hearing adults in a MEG study. Participants read 48 rhyming sentence pairs beforehand. In the scanner, they listened to sentences that corresponded exactly to the read sentences except that the last word (presented after 1600 millisecond delay and with 50% intelligibility) was only correct in half of the cases. Otherwise, it was 1) phonologically but not semantically related, 2) semantically but not phonologically related, or 3) neither phonologically nor semantically related to the sentence. Participants indicated by button press whether the last word matched the sentence they had read outside the scanner. Behavioural results showed more errors in condition 1 than in conditions 2 or 3, suggesting that phonological compatibility overrides semantic discrepancy when intelligibility is poor. Event-related field analysis demonstrated larger activity on frontal sites for correct than unrelated words, suggesting that the former were more accurately expected than the latter. An early M170 component was also observed, possibly reflecting expectation violation in the auditory modality. Dipole analysis will reveal whether M170 could be modulated by type of linguistic discrepancy. Distributed-network analysis will further our understanding of the time course and neural correlates of discrepancies between incoming auditory information and linguistic expectations.