The fast generation method based on lattice segmentation for high-quality confusion network

Aimed at the problem that the existing confusion network generating methods cannot keep a tradeoff between the network generation speed and the quality of confusion network, the paper investigates two major lattice segmentation methods with the purpose of using them to reduce the impacts of segmentation to the quality of confusion networks, and based on this, presents a high-quality method for fast generating confusion networks based on lattice segmentation. The method segments the large-scale lattice from automatic speech recognition (ASR) into sequences of smaller sub-lattices and then generates the confusion networks from these sub-lattices, thus remarkably decreasing the computation scale and increasing the network generating speed. The balance between the generation speed and the network quality is controlled by the segmentation number. The experimental results show that the proposed method can significantly improve the speed of confusion network generation while hold almost the same quality compared with the traditional word-clustering method without lattice segmentation. At the same speed, the proposed method can obtain a lower tonal syllable error rate than the word-clustering method with lattice pruning.