This paper introduces the Toshiba Mandarin Text-to-Speech (TTS) system submitted to the Mandarin benchmark of the Blizzard Challenge 2009. The basic framework keeps unchanged with the system in 2008 and we modify the system in several aspects: automatically find bad units in the database when preparing the speech corpus, add a G2P procedure after the text analysis to increase accuracy of the predicted pinyin for heteronyms, introduce prosody layer information into the prosody modeling and modify the fusion methods to fuse units in frequency domain. The subjective evaluation results show that these modifications improve the performance.