Spoken Term Detection by Result Integration of Plural Subwords using Confidence Measure

摘要

Spoken term detection (STD) systems should not restrict query terms because they are likely to be out-of-vocabulary (OOV) of a general speech recognizer such as technical terms or personal names. To detect OOV terms, we have proposed a method that uses subword such as monophone, triphone and newly proposed subwords such as 1/2phone and 1/3phone. The method enabled to detect OOV terms and the proposed subwords worked well compared with triphone. We could improve the STD performance by integrating plural detection results obtained from several subwords. Each utterance in spoken documents has plural scores obtained from the plural subwords. The integrated score is computed by giving a weight to the score of each subword and summing up the weighted scores. We note a confidence measure that is obtained from a subword based speech recognizer, when determining the weights. In case of integrating the plural scores obtained from several subwords for a certain section, the section that has a high confidence measure for a query is more reliable than other sections that have lower confidence measure. Therefore, the STD performance is improved by giving a higher weight to the subword that has a high confidence measure and low weights to the other subwords. The proposed method automatically determines the weights dynamically for each candidate section according to the confidence measures of the subword models in the candidate section. Through evaluation experiments using a lecture corpus, we could confirm that the proposed method improved the performance of a conventional linear integration method.