Humans are adept at understanding speech despite the fact that our natural listening environment is often filled with interference. An example of this capacity is phoneme restoration, in which part of a word is completely replaced by noise, yet listeners report hearing the whole word. The neurological basis for this unconscious fill-in phenomenon is unknown, despite being a fundamental characteristic of human hearing. Here, using direct cortical recordings in humans, we demonstrate that missing speech is restored at the acoustic-phonetic level in bilateral auditory cortex, in real-time. This restoration is preceded by specific neural activity patterns in a separate language area, left frontal cortex, which predicts the word that participants later report hearing. These results demonstrate that during speech perception, missing acoustic content is synthesized online from the integration of incoming sensory cues and the internal neural dynamics that bias word-level expectation and prediction.