High performance Chinese Spoken Term Detection based on term expansion

This paper mainly focuses on improving the performance of Chinese Spoken Term Detection (S TD) systems using words as searching units. These systems are designed to find instances of particular phrases (called Terms) in voices. Terms are usually segmented into word sequences and searched with voices´ recognition results. Mismatches between recognition results and word-segmentation might affect their performance. To solve this problem, two algorithms are designed to expand the searching spaces. Th e exp anded algorithms improve systems´ performance while lead to a side-effect for its efficiency. To speed up the retrieval tasks, the Finite State Automation (FSA) is used. A token-passing algorithm is the n developed for fast search. Experiments have shown that the proposed term expansion method could effectively improve the STD system´s performance. And using FSA with token-passing algorithm to search could effectively improve searching efficiency.