Content-based Audio Classification And Its Application In Multimedia Retrieval

With the development of digital and internet technologies, the multimedia informationexperience a explosive growth. Among numbers of multimedia data types, audio has its ownspecial structure and rich semantics, and different types of audio has a different meaning ofevent. Audio classification occupy a very important role in the content-based audio retrieval,is the basis of audio structure, and achieved the structure of audio stream to a certain extent.Audio classification is directly related to the in-depth analysis of the extent and degree ofaccuracy, as well as the extraction of semantic audio content. This paper focus oncontent-based audio stream classification and retrieval for exploratory study, and alsoexplored its application value content-based multimedia retrieval.The main work as follows:1. Study the results of multimedia retrieve techniques at home and abroad especially thetheory and techniques of audio content analysis and content-based audio classificationand retrieval.2. Analyze basic principles and methods of audio content, audio segmentation, audioclassification and retrieval, system framework of content-based audio classification andretrieval, focused on description of the audio feature analysis and extraction techniques.Details the Mel cepstral coefficients (MFCC) and its extraction methods.3. Against the features of threshold selecting limitation and unsuitability for shortwindow audio change detection of BIC change points algorithm, design and achieve theentropy trend-based audio feature transition points detection algorithm, the test provedthat the algorithm has good performance: the average miss detection rate is only2.8%.4. Adopt method based on statistical learning theory, considering limit ofraining data and classification efficiency factor, design and achieve GMM-based andHMM-based classifier, the experiments show good performance: the accuracy rateare86.36%and86.59%respectively.5. Design and achieve a complete content-based audio stream classificationsystem module and user query interface, innovative achieve positioning and classification inthe long-term audio stream, which is the primary first step of content-based audio events retrieval. The system integrate automatic audio segmentation and audio classificationalgorithm module, and use the results of audio classification to amend before thesegmenting points in order to eliminate false alarm points. And test result shows the falsealarm rate was sharply reduced by35.33%, but due to the introduction of automaticsegmentation techniques, the classification accuracy rate has a certain degree of decline,because each voice its own characteristics was weakened in the adjacent of different voices.