Most of current pitch prediction methods for mandarin TTS try to get pitch contours from the contextual information with a group of weights assigning. Without a good method in prosody concatenation constraint, the predicted pitch contours are not always stable because of the incomplete accordance between prosody information and text information. The paper presents a new mandarin pitch prediction method with mutual prosodic constraint between syllables. The idea of this mutual constraint is first inspired by lots of observations on corpus, but then it has been strictly proved with performance comparison and feature contribution analysis of CART-Based prosodic parameter prediction. Based on this, a reasonable definition of prosody concatenation cost is presented to measure the naturalness of pitch contours between two adjacent syllables. By minimizing this cost, the model can generate fluent pitch contours, which has been proved to be able to make the TTS system more natural than traditional systems. Keywords: Speech synthesis, TTS, Mandarin, prosody model, pitch generation, mutual constraint.