In this paper, a word-length-dependent confidence measure for large vocabulary Chinese keyword spotting is proposed to deal with the problem caused by the significant difference in keyword length, i.e. false alarms are likely to occur for shorter keywords, while false rejections are more likely to occur for longer keywords. The proposed confidence measure is based on not only the acoustic scores for the component sub-syllabic units of the keywords, but a set of word-length-dependent parameters trained with the minimum classification error criteria.
展开▼