An Audio Classification Algorithm For News Video Retrieval

Content-based audio signal analysis is one of the most important parts in multimedia processing, which needs to discriminate different types and deal with them in different methods. Audio classification plays an important role in it and is a base work of audio content analysis and audio structuring. It is widely applied in content-based audio/video retrieval and other multimedia application systems.Based on the development of past research, this paper deeply explores the audio semantic content of the news program, analyses and defines audio structure units on different hierarchies. According to the characters of news audio structure and content, six audio types in news program are defined: silence, pure-music, anchorman/anchorwoman pure speech, live report and alternated speech. This paper solves the following problems: audio feature analysis and extraction, audio classification and speech tracking clustering based on divergence shape distance.First, discriminating features among different audio types are researched on frame and clip level respectively. According to the six audio types, the paper proposes five clip features, such as silence ratio, zero crossing rate standard deviation, frequency centroid standard deviation, pitch standard deviation and Mel Frequency Cepstrum Coefficients as the input of the classifier. The performance is analyzed through experiments meanwhile.Secondly, according to pattern recognition theory, an audio classification framework is proposed. HMM classifier is analyzed deeply, and then a new audio hierarchical classification algorithm based on rules and HMM is proposed. The experiments show its high performance.At last, in order to understand the semantics content of different speakers, the paper employs an unsupervised approach to track and cluster the anchorman speech based on divergence shape distance.