Internet Talent Show Audio Classification Research

With the rapid development of communication, network and computer techniques, many multimedia data, such as images, video, audio, etc, has become a main information media form in the information processing fields. These data accounts for about 70% of the data transmitted on the internet. Among them, audio media is more and more widely used in all kinds of domains. It can say that the audio media is the most important form of all media except visual media. Therefore, facing the mass of audio data, how can divide them into specific category for people to choose and use is become one of the important topic at present. On the basis of predecessors' achievements, this paper focuses on the Internet talent show audio classification.The work of this paper mainly includes the following several aspects:(1)Initial processing of the original audio signals of talent show on the Internet, including the digit transformation of audio signal, pre-emphasis, windowing and audio signal framing, etc;(2)According to the characteristics of audio features and the extraction method, the features of the audio data after the initial processing are extracted;(3)By analyzing the feature selection algorithm based on Adaboost, this paper put forward a new feature selection algorithm to select audio features;(4)Using multilevel classification method, training Gaussian mixture model with the selected feature set at each stage;(5)Testing the classifiers with the test data.Experiments show that, the classified accuracy with the best feature set selected by the suggested feature selection algorithm is better than the classified accuracy without the suggested feature selection algorithm.In addition, the best feature set selected by the new selection algorithm could avoid the feature repeating which may appear by using the feature selection algorithm based on Adaboost, and could avoid the divergence because of the improper sample selection.