Dynamic tuning of language model score in speech recognition using a confidence measure

Speech recognition errors limit the capability of language models to predict subsequent words correctly. An effective way to enhance the functions of the language model is by using confidence measures. Most of current efforts for developing confidence measures for speech recognition focus on applying these measures to the final recognition result. However, using these measures early in the search process may guide the search to more promising paths. In this work we propose to use a word-based acoustic confidence metric estimated from word posterior probability to dynamically tune the contribution of the language model score. The performance of this approach was tested on a conversational telephone speech corpus and results show significant reductions in recognition error rates.