This paper describes the development of the NICT/ATR speech synthesizer for the Blizzard Challenge 2008 and discuss the official results. The submitted system is based on the hidden Markov model speech synthesis technology and utilizes an improved excitation approach based on residual modeling, in order to remove artifacts related to the parametric way in which speech is synthesized. Although development time was limited, the results show that the system in question achieves good performance in terms of naturalness and intelligibility.