We are interested in developing an automatic emphasis correction system, which converts any unemphasized word in an utterance into emphasized. Analyzing how prosody changes from unaccented to accented is crucial for the task. While previous works on prosody reconstruction only model the prosody contour itself instead of the increment, we propose a framework to study the prosody increment induced by pitch accents from real speech in a statistically rigorous manner. This framework also infers the degree of emphasis of each word to account for the additional prosody variations due to metalinguistic factors. According to the analysis results, this framework provides a lot of useful insights into the prosody increment, which are consistent with many existing studies on pitch accent and emphasis.