Automatic speech recognition (ASR) systems are being integrated into a wider variety of tasks involving human{machine interaction. In evaluating these systems, however, it has become clear that more accurate means must be developed for detecting when portions of the decoded recognition hypotheses are either incorrect or represent out{of{vocabulary utterances. This paper describes the use of confidence measures based on likelihood ratio based optimization procedures for decoding and rescoring word hypotheses in an HMM based speech recognizer. These techniques are applied to spontaneous utterances obtained from a \movie locator" based dialog task.