Post-dialogue recognition confidence scoring for improving statistical language models using untranscribed dialogue data

This paper presents a new recognition confidence scoring method, which is used for selecting speech recognition results of untranscribed user utterances. The selected recognition results can be used for training statistical language models of speech recognizers in spoken dialogue systems. The method uses features that can only be obtained after the dialogue session, in addition to such features as the acoustic scores of recognition results. Experimental results showed that the proposed confidence scoring improves correct/incorrect classification of recognition results and that using the language models obtained through our approach results in higher recognition accuracy in speech recognition than those achieved by conventional methods.