Lattice Segmentation Based Confusion Network Generation Method

Confusion network (CN), a compact representation of lattice, has been arousing more and more attention in the area of speech recognition. In the paper, a novel lattice segmentation based CN generation method is proposed for remarkably reducing time complexity. Confidence based lattice segmentation algorithm is designed to cut lattices. The experimental results on a Chinese continuous speech database show that the significant reduction in generation time is achieved but almost no decline in the quality of CN compared to word-clustering algorithm (WCA) without arc-pruning operation. In the case of consuming same time, recognition performance of the proposed method is beyond that of WCA.