论文部分内容阅读
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.
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 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.