Abstract
Recently, there have been many efforts in developing confidence measures for speech recognition output. Usually, these measures are applied to the final result of the decoder. However, using these measures early in the search process can guide the search to more promising paths. In this paper we use a confidence metric, the Average Best_Base_Phoneme Rank, to dynamically tune the contribution of the language model score. The advantage of this metric is that it uses information already available during the decoder search. The performance of this guided search approach was tested in two experiments for the ATIS and WSJ data sets and results show significant reductions in recognition error rates.