Determination of prosodic phrase break from text is one of the important problems in generating good prosody for Chinese text-to-speech system. In this paper, we propose a statistical approach for detecting prosodic phrase breaks. Part-of-speech sequence information is used as the primary information. The history of the previous breaks is considered as constraint in this work. The probabilities are calculated using CART approach. During the prediction process, the breaks are dynamically determined. Viterbi algorithm is applied to find the best break sequence. We achieved a result of recall of about 89% and precision of about 85 % for our testing data.