Research On Content-based Audio Information Classification And Retrieval Technology

Rapid advances in the multimedia technology and computer processing capacity make people be faced with large digital "Information Ocean". It's more and more important for people to retrieval these information quickly and effectively. Therefore, multimedia information retrieval has been rapidly developed from the 1990s and become one of the most important areas in the information retrieval. And in the beginning most works were focused on content-based image retrieval and content-based video retrieval. However, researches on audio retrieval are relative lagging, just because abundant semantic information contained in audio data is always ignored, and audio data are non-structuring. As more and more audio data appear, content based audio retrieval has been one of research hotspots in multimedia retrieval.This dissertation, which is based on the summarization of former research findings, deals with several problems of content-based audio retrieval. Researches are emphasized on audio feature analysis, classifier design and speech information retrieval.The main research contents and results of this dissertation can be concluded as follows:(1) Research on audio feature classificationIt's always based on subjective or objective audio features for audio classification, and the selection of audio features must be represented important classification features in time domain and frequency domain. Hence, analyses and extractions of audio features are the base of audio classification. How to extract features and keep them independent mutually is the important problems to be settled, which can reduce information redundancy.In this paper, the Independent Component Analysis(ICA) method is introduced into audio feature analysis. The method can extract pivotal and high dimension independent features and improve feature separability. Furthermore, the Support Vector Machine(SVM) is used with its approved classified performance to classify audio data features. A hybrid model is presented to audio features classification, which is combined with ICA and SVM.(2) Design and of audio classifier In content-based audio retrieval, classification objects are continuous audio data. There is another problem to design a classifier, which can represent time statistic characteristics with certain separation abilities.Based on former research in speech recognition, the General Model(GM), which is extended from Hidden Markov Model(HMM), is applied into audio classification and retrieval. The hybrid SVM/GM approach is presented and used in audio classification and retrieval.(3) Research on broadcast news audio retrievalSpeech is one of important kinds of audio data, for example, there is abundant speech information in broadcast news or tapes of symposiums. Because of speech characters, such as intuitive, physical, and convenience, and so on, it's worth to discuss how to directly utilize speech to retrieval audio information in multimedia effectively.Audio retrieval of broadcast news is discussed in this dissertation, mainly in audio classification, aidio retrieval and speech recognition. Finally, a content-based audio retrieval prototype system is designed.