Confidence Measure Based Unsupervised Speaker Adaptation

摘要

Unsupervised adaptation is the most convenient mode for the user of a speech recognition system. However the performance of unsupervised adaptation is worse than that of the supervised mode because of the recognition errors. This paper introduces a kind of word-lattice based ocnfidence measure to evalute the reliability of the recognition result and discard the uncertain parts fro mthe adaptation speech. Experiemtns demonstate that the confidence can improve the performance of unsupervised adaptation considerably.