Whether humans spontaneously sound out words in their mind during silent reading is a matter of debate. Some models of reading postulate that skilled readers access the meaning directly from print but others involve print-to-sound transcoding mechanisms. Here, we provide evidence that silent reading activates the sound form of words before accessing their meaning by comparing event-related potentials induced by highly expected words and their homophones. We found that expected words and words that sound the same but have a different orthography (homophones and pseudohomophones) reduce scalp activity to the same extent within 300 ms of presentation compared with unexpected words. This shows that phonological access during silent reading, which is critical for literacy acquisition, remains active in adulthood.