Many Spoken Term Detection (STD) systems use query expansion to return an increased number of keyword candidates and make posterior probability a confidence feature to reject false alarms. However, some keyword candid...Many Spoken Term Detection (STD) systems use query expansion to return an increased number of keyword candidates and make posterior probability a confidence feature to reject false alarms. However, some keyword candidates hold high posterior probability although these are not recognized correctly. We investigate the Word Activation Force (WAF) model that compatibly encodes syntactical and semantic information into sparse coding directed networks. A high-level confidence feature Keyword Activation Force (KAF) based on WAF is proposed. KAF can be used for detecting false alarms by considering information about the neighbors to provide a more reliable and accurate keyword affinity. Compared with the baseline system, a relative reduction of 30.94% in average error rate could be achieved when KAF is combined with the posterior probability and the language model score.展开更多
基金supported by National Natural Science Foundation of China under Grants No.61005004,No.61175011,No.61171193 the Next-Generation Broadband Wireless Mobile Communications Network Technology Key Project under Grant No.2011ZX03002-005-01+1 种基金 the 111 Project under Grant No.B08004 Scientific Research Foundation for the Returned Overseas Chinese Scholars,State Education Ministry
文摘Many Spoken Term Detection (STD) systems use query expansion to return an increased number of keyword candidates and make posterior probability a confidence feature to reject false alarms. However, some keyword candidates hold high posterior probability although these are not recognized correctly. We investigate the Word Activation Force (WAF) model that compatibly encodes syntactical and semantic information into sparse coding directed networks. A high-level confidence feature Keyword Activation Force (KAF) based on WAF is proposed. KAF can be used for detecting false alarms by considering information about the neighbors to provide a more reliable and accurate keyword affinity. Compared with the baseline system, a relative reduction of 30.94% in average error rate could be achieved when KAF is combined with the posterior probability and the language model score.
