Audio Segmentation And Keywords Spotting System For Sports Game

Speech recognition and keywords spotting increasingly develop in recent years. Compared with continuous speech recognition, keywords spotting technology has more flexibility and applications. The keywords spotting for TV sports games discussed in this paper is such an application. As we know, the sports game audio is very complicated, which means a good pre-processing module (audio segmentation and classification) for keywords spotting system is very important. In this dissertation, the key research contents are as follows:Segmentation and classification of TV sports game audio:the composition of sports game audio is very complicated, including anchor's speech, music, cheering and so on. Here three audio segmentation methods are discussed:phoneme decoder; the combination of BIC and GMM; the combination of BIC and phoneme decoder. We do experiments on different matches respectively using this three different methods, and we can see from the results that the method which combines BIC and GMM has a higher recall rate and precision than two other methods. Therefore this method is suggested to use for sports game audio segmentation.Keywords spotting:some improvements are made on our speech lab's keywords spotting baseline system. The baseline system, which is designed for broadcast news speech, works poorly for sports game audio. To solve this problem, language model adaptation and acoustic model adaptation are adapted with the baseline system. We mainly do four experiments:MAP acoustic model adaptation to different matches with language model unchanged; language model adaptation to different matches with acoustic model unchanged; language model and acoustic model adaptations with the mixture of adaptation data and both the language model and the acoustic model adaptations with corresponding adaptation data of each match types. The result shows that the last method can improve baseline system better than the other three methods.