Post-dialogue confidence scoring for unsupervised statistical language model training

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