This paper details a speech synthesis system developed at NICT for the Blizzard Challenge 2010. The system depends on an HMM-based speech synthesis technique that possesses two distinctive features: HMM training under global-variance constraint on the parameter trajectory and trainable mixed excitation for source-filter vocoding. For this year’s entry, we added some modifications to the system we developed for last year's Challenge. The major improvement is on the scheme for the training of the unvoiced filter that is a component of our mixed excitation model. Despite the fact that our excitation modelling has room for further improvement, the official results show that the system achieves reasonable performance for all assessment categories.