Audio Signal Feature Extraction And Classification Research

In the background of big data era,more and more audio data is stored in the network.The automatic and intelligent classification system for audio is needed and growing up.The research of audio signal classification is one of the most important part to promote the development of such system.Feature extraction,feature set optimization and classifier design are the three most important links in studies of audio signal classification.We can find through studies and analysis of existing literatures that the focus of audio classification research mainly include three aspects:the study of different feature extraction methods,research on feature selection algorithms used to optimize feature set,and classification performance of different classification algorithms.Studies of this thesis focus mainly on these three parts,and main works of this thesis include following parts:1.A total of 89 audio features that including time-domain features,frequencydomain features,cepstrum-domain features and other features were extracted using different audio feature extraction methods to construct the original audio feature set.2.Mainly studied three feature selection algorithms which respectively based on Pearson correlation coefficient,entropy method,Relief algorithm.Proposed an improved method of feature selection based on correlation coefficient.And these algorithms are used to finish the optimization of feature set.Verified the validity and feasibility of these four feature selection algorithms through designed experiments,and compared the advantages and disadvantages of different algorithms.3.Three classifiers include K-nearest neighbor classifier,Decision tree Classifier and BP neural network classifier were designed.Studied the problem of BP neural network that tend to fall into local optimal,and designed an improved BP neural network classifier based on simulated annealing algorithm.And put the original audio feature set into these four audio classifiers,completed the experiments of classification of speech and music,music genre,and musical instruments respectively.Experimental results show that the average classification accuracy rate of improved BP neural network classifier,tradition BP neural network classifier,Decision tree Classifier andK-nearest neighbor classifier were 96.15%,92.86%,93.60%,85.98% respectively.4.Completed the classification experiment with four optimized feature sets and the improved BP neural network classifier.Proved that the classification result obtained by using the feature set corresponding to the improved correlation coefficient feature selection algorithm was best.The average classification accuracy rate obtained from the classification experiments of speech and music,music genre,music instrument were 97.78%,92.69%,98.50% respectively.