Grammatical word graph re-generation for spontaneous speech recognition

We propose a novel recognition method for generating an accurate grammatical word-graph allowing grammatical deviations. Our method uses both an n-gram and a grammar-based statistical language model and aligns utterances with the grammar by adding deviation information during the search process. Our experiments confirm that the word-graph obtained by our proposed method is superior to the one obtained by only using the n-gram with the same word-graph density. In addition, our recognition method can search enormous hypotheses more efficiently than the conventional wordgraph based search method.