Abstract: As confidence measures can well judge whether an observation matches a model, they can be used to do hypothesis testing and to locate the errors, thus improving accuracy and robustness of speech recognition. In this paper, the theory of confidence measures is introduced in the context of speech recognition, main methods of model construction and performance evaluation are described, and various applications are reported. Our experiments results are also presented, which show that confidence measures are very effective in such aspects as decoding and pruning, speaker adaptation, and utterance verification and rejection.