面向新闻视频内容分析的音频分层分类算法  

Hierarchical audio classification algorithm for news video content analysis

作　　者：冀中[1] 苏育挺[1] 宋星光[1] 安欣[1] 

机构地区：[1]天津大学电子信息工程学院,天津300072

出　　处：《计算机应用研究》2009年第5期1673-1675,共3页Application Research of Computers

基　　金：国家"863"计划资助项目(2006AA01Z407);天津市应用基础及前沿技术研究计划重点资助项目(07JCZDJC05800)

摘　　要：提出了一种规则和隐马尔可夫模型相结合的音频分层分类算法，首先利用规则将新闻节目中的音频分为静音、语音和音乐三类，然后采用隐马尔可夫模型进一步将语音和音乐细分为男主持人语音、女主持人语音、交替报道、独白语音、现场语音和音乐六类。实验结果表明，男主持人语音、女主持人语音以及音乐的分类效果最好，查准率和查全率均可达90％以上；交替报道的分类性能最差，查准率为57．5％，查全率为79．3％；其他类别的分类性能居中，在70％～90％左右。与同类算法相比，该算法分类性能较高。This paper proposed hierarchical audio classification algorithm, which first classified the news audio stream into silence, speech and music with rule-based classifier, and then employed hidden Markov models to categorize the speech and music to male-anchor speech, female-anchor speech, alternate speech, monologue speech, live report and music. The experiment results show that the classification works best in male-anchor speech, female-anchor speech and music, in which precision and reall can both reach more than 90%. The classification performs worst in alternate speech with precision of 57.5% and with recall of 79.3%. The performance of classification in other types is at the average level with precision and recall ranging from 70% to 90%. Compared with the other representative algorithm, this method works well with relatively high precision.

关 键 词：音频分类 内容分析 隐马尔可夫模型 新闻视频 视频检索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]