Abstract
A future broadcasting system will broadcast along with regular programming a wide variety of program-related information called metadata. Our speech recognition technology is applied to extract information that corresponds to highlights from the commentary of a football game. The proposed method uses additional phoneme models of excitedly uttered words which may be important key words for highlight extraction. The proposed method obtained 14% of key word error reduction, which improved extraction accuracy of scenes making a shoot by 18% in precision and 17% in recall.
