In this paper, a decoding method incorporating word-level confidence measures for improved speech recognition is presented. At first, we focus on the estimation of confidence measures from the word graph and evaluate them in word graph rescoring (2nd-pass in 2-pass search system). Next, we propose the lexical tree search (1st-pass in 2-pass search system) incorporating the word-level confidence measures and an iterative decoding based on the confidence measures, resulting in the reconstrucion of the word graph. The experimental results showed that this method achieved a slight improvement at word accuracy.