This paper describes a hidden Markov model (HMM) based speech synthesis system developed for the Blizzard Challenge 2012. In the Blizzard Challenge 2012, we focused on a design of contexts for using audio books as training data and modeling of silence between sentences for synthesizing paragraphs. It is well known that contextual factors affect speech. We use extended contexts for using audio books to construct appropriate model parameter tying structures. In addition, duration models of silence between sentences are created to synthesize more natural speech because connections between sentences are important for synthesizing paragraphs. Subjective evaluation results show that the system synthesized the high intelligible speech.