We used event-related potentials (ERPs) and acoustic analyses to investigate the processing of prosodic pitch accents as a function of their position in a sentence. Accents in sentence-medial positions were characterized by a higher fundamental frequency (F0) and an increased duration. They elicited two different negative ERP components around 400 ms, depending on the predictability of the accent. When the accent was predictable, the negativity was fronto-laterally distributed and identified as the previously known Expectancy Negativity. Unpredictable accents elicited a more broadly distributed N400 with a central maximum, reflecting difficulties in semantic processing. For sentence-initial pitch accents, words had a higher F0 but of the same duration as sentence-initial words without pitch accents. These pitch accents elicited a P200 but no negativity in a 400 ms time window. The P200 was modulated by the onset latency of the F0 peak rather than its magnitude. We discuss the possibility of a delayed processing of sentence-initial accents when the actual occurrence of an F0 peak can be identified by comparing the F0 of the sentence-initial word to a reduced F0 of a word occurring later in the sentence.