Detection Of Audio Events With Scene Dependence

Audio information always plays a critical role in multimedia analysis, retrieval and index. Automatically classification and semantic analysis of key audio effects are the most important steps in content based calculation of audio documents. Most of the previous work is focused on the detection and classification of pure and single audio event. However, much of the multimedia information is consisted of several simultaneous audio events which leads to a complicated audio environment such as in movies and increases the difficulty in audio retrieval and analysis. When processing a complex audio scene where several key audio events occur simultaneously, the existing methods shows obvious limitation. The major difficulties might be the mixture of different audio events, which leads to the complication of probability distribution of sample data.In this paper, some researches and methods are employed to detect key audio events and overlap audio segments in movies and the data from cell phone end. In such complex audio environments, GMM and SVM are employed to detect and classify single audio event and an unsupervised method is used to model overlap audio segments according to a certain semantic audio scene. The emphases are the detection of overlap audio segments and the content based analysis of such audio scene. Main works are as followed:1. Single audio event detection and classificationBased on the detection and classification of single audio event, key audio detection plays an important role in advertisements filter, harmful audio detection, automatically audio classification, and highlight extraction in movies. It has insignificant influence in both practical application and academic researches. Both GMM and SVM are used to model and classify the key audio events in movies. In connection with the data features of complicated audio signal, GMM is employed in further work according to the results of experiments. A preliminary study of the classification of imbalance data set is also carried out:down-sampling.2. Overlap audio segments detection based on the calculation of information entropyPeople are more interested in the sense of an audio scene with complete semantic content when analyzing audio documents. The researches on overlap audio scene detection and classification need to combine the audio elements together, which are presentable for the audio scene information. Because several audio effects might occur simultaneously in a certain scene, the traditional methods for audio events detection are limited and ineffective. A single audio event cannot simply stand for an audio scene with several audio events in it. Aimed at solving such problem, a solution for overlap audio segments detection is proposed:IEC-ED.3. Detection of Audio events with scene dependenceAs for the classification of overlap audio segments, UBM-GMM is applied in the research. Since the audio events are presented in a certain scene, the models for these events are supposed to contain part of the information of the scene. First, a UBM, which contains as much information as in a scene, can be trained with the samples of the scene, and then, for each audio event, a GMM is trained to statistically present the audio event.