This paper proposes a State-Correlated Duration model for HMM-based speech synthesis system. It uses an improved forward-backward algorithm to estimate the state-duration transition probability between the neighboring states. In the synthesis part, we determine the state duration taking account of the state-duration transition probability. Experiment results show that the speech we synthesized using the new duration model has a higher quality. Keywords: State-Correlated Duration model, improved forward-backward algorithm, state duration, state-duration transition probability