Error identification for large vocabulary speech recognition

This paper proposes two methods for identifying recognition error. The first method is a two-level schema - given the recognition hypothesis of an utterance, an utterance classifier (UC) is first applied to decide if the hypothesis is error-free or erroneous; followed by a word classifier (WC) which is applied to each word hypothesis in the erroneous utterance to decide if the word hypothesis is a misrecognition. The second method is a one-level schema in which a word classifier is applied directly to all word hypotheses to detect word recognition errors. We compare the two methods at both word and utterance levels. Experimental results show that the two methods are comparable in terms of word error detection. However, the two-level schema is very effective in filtering out error-free utterance hypotheses, which offers a key advantage to economize on word error detection.