This paper describes the NICT speech synthesis system submitted to the Blizzard Challenge 2009: a hidden Markov model (HMM)-based synthesizer constructed by training trajectory HMMs considering global variance. To improve naturalness of the synthesized speech a mixed excitation approach based on closed-loop residual modeling through the training of state-dependent filters is employed. According to the official results the system in question performs well in terms of naturalness and intelligibility although synthesized speech does not sound very similar to the original speaker.