Multiple index combination for Japanese spoken term detection with optimum index selection based on OOV-region classifier

In this paper, a novel index combination method for spoken term detection is proposed. In our method, outputs from four different recognizers (word, syllable, word-syllable, and fragment recognizer) are combined into one confusion network. A novel index-selection method for the multiple index-combination method is then used to suppress the increase of the index size. Two methods are proposed to reduce index size: (1) arc selection and (2) unit selection, both of which are based on an OOV-region classifier score. Experimental results with 39 hours of Japanese lecture recordings showed that the index-selection method achieved a 22% reduction of index size of the best confusion network while maintaining its high accuracy. Compared with the best phoneme-based index from a single recognizer, the proposed method achieved a 25.0% and 14.8% relative error reduction for IV and OOV queries without increasing the index size.