The Research On Audio Signal Classification

Advances in the multimedia and internet bring more and more multimedia information. How to analyze, store and classify the huge amount of data efficiently, especially for those audio data is an imperative problem. As one of research hotspots in multimedia data process, audio classification has been applied in the field of audio retrieval. There are two key problems of content-based audio classification, which are how to extract more robust features from audio data and how to implement the audio classifier.This thesis, which is based on the expounding the development status of content-based audio classification nowadays, emphasized the research on audio analysis and abstract, classifier design and test. Then two classifiers were designed based on GMM and HMM, which can classify the silence, speech, music and speech with background sound.The main research contents and results of this thesis can be concluded as follows:(1) Audio classification is always based on the audio features. And the selection of the audio feature must be represented important classification features. So analyses and extraction of audio features are the base and key of the audio classification. Based on the analyses of the audio features in the time domain, frequency domain and acoustics, the audio features are extracted at frame-level and clip-level, including Short Time Energy, Zero-crossing Rate, Sub-band Spectrum Energy, Spectral Centroid, Bandwidth, MFCC and Fundamental Frequency.(2) The difficulty of the audio classifier is how to design the classifier. By analyzing the typical methods of classification, this dissertation designed two classifiers based on GMM and HMM, which make the classification of the silence, speech, music and speech with background sound come true.(3) The test results show that the feature selected are effective and the classification accuracy is good.