Estimation confidence of the output hypothesis of a speech recognizer can be used in many practical applications of speech recognition technology. In this paper, we propose to estimate the confidence of a hypothesized word directly as its posterior probability. There are tow methods to calculate posterior probabilities, one of which is based on word graphs and the other is based on N-best lists. We present experiment results on database provided by China National 863. And we also use confidence measures in unsupervised speaker adaptation.