This paper describes the development of PKU mandarin speech synthesis system for Blizzard Challenge 2009, which is built in the framework of corpus-based unit concatenation synthesis. The system employs a trainable VTR model named HTM to label the VTR trajectories in corpus and predict the target VTR features. In addition, a CART based prosody model is built to predict the prosody parameters of the target units. In corpus building, the speech waveform in the corpus is converted to parametric representation by STRAIGHT algorithm. And in voice building, the speech waveform is constructed from the connected STRAIGHT parameters of selected units.