Content-based Audio Retrieval Research

With rapid development of network technology and multimedia technology, alarge number of multimedia information can be obtained from the Internet. Audio isan important part of the multimedia. As rapid increase in audio data, how toautomatically manage these data has become a prominent issue. Especially to thewide variety of music data, people demand a rapid and efficient method for themanagement of their classification (according to different styles or singers, etc.). Thisrequires an efficient automatic classification technology to collate audio data, and itcan serve audio search or related analysis. In this dissertation, we present a study oncontent-based music classification.Currently most research of music classification focus on music feature extractionand classification. Music features include time-domain features, such as short-termenergy, short-term zero rates, frequency-domain features, such as bandwidth,spectrum center, and Mel-frequency cepstral coefficients (MFCC) based on theexperience of hearing. Many features of music can be used for music classification.And classification algorithms can use the existing classification model of a largenumber of efficient algorithms, for example, Gaussian Mixture Model (GMM),Neural Networks (NN), Support Vector Machine (SVM), Hidden Markov Model(HMM).According to the above study, in the dissertation we use hidden Markov Modelbased on Mel-frequency cepstral coefficients. In music feature extraction aspect, weconsider perceptual characteristics and Mel-frequency cepstral coefficients as afeature vector; in music classification aspect, we use hidden Markov model as theclassifier for music clustering and classification. Clustering uses a supervised learning.Classification divides the test samples into categories according to the largestlikelihood value. We identify the results of the samples of the same audio using thevoting method, thus determine the category of the audio. Classification accuracy hasbeen further improved. According to the above methods, we did some simulationexperiment, and analyzed the experimental results. Audio data in this paper dividedinto five categories. We compared the performance of four different classifications,and tested the interference model. The results show HMM classification performance has a certain advantage and a strong anti-interference. Audio classification technologyis an important supplementary means for the audio retrieval and other audioprocessing. Through content-based music classification, it brings the convenience tomusic retrieval or related analysis. Therefore, in the study field of content-basedmusic retrieval, the content-based music classification is a very important andmeaningful work.