The Research Of Audio Classification Based On The Characteristics Of Source-Sound

Audio classification lets users search audios by one feature or a combination of them. Accurate classification of audio signals not only can improve the efficiency and reduce complexity of the follow analysis, but also be applied to the area of military investigators, speech recognition, musical instruments classification and so on. The accuracy of audio classification depends on two points: the feature for classification and the classifier. Therefore, this paper will rearch around the two aspects, and the main rearch and results are as follows:1. The natural and artificial audio classification based on sample entropy is used. Then all samples are divided into 13 subclasses with multi-feature on the premise of the two major categories. The physical of sample entropy is rearched at first. Then use kmeans to classify the natural and artificial audio after the average and variance of sample entropy is extracted. Simulation experiment show the accuracy can reach 83.846%. Integrate multi-feature to divide all samples into 13 subclasses after the discussion of the diversity of audios, and the accuracy can reach 85% without any prerequisites of type-information.2. Audio classification based on atomic parameters of sparse decomposition using MP is proposed. MP algorithm can decompose signal into a group of linear time-frequency atoms which can describe the essential characteristics of signal pronunciation. After analyzing the realationship of audio signal and the atomic parameters of sparse decomposition, 10 kinds of musical instruments and 5 kinds of brids are classified with the atomic parameters of sparse decomposition using MP. The realation between first n atoms and the classification accuracy is researched in simulation experiment, and the first 6 atoms are selected as the feature which can achieve 17.15% and 10% higher than MFCC to complete the classification task.3. Musical instruments classification based on Gabor atomic parameters of sparse decomposition using MP with less iteration steps is proposed after the regularity of specific instrument signal's Gabor atomic parameters of sparse decomposition is researched. Simulation experiment compares the atomic parameters, accuracy of classification and the running time of 10 kinds of musical instruments between the original algorithm and the improved one. The results show that although the accuracy decrease slightly, the running time is about 1/3 of the original algorithm.4. Audio classification based on SOM network which initial weights improved by genetic algorithm is used. Genetic algorithm can make up the problem of local minimum which caused by the randomness of initial weights. Simulation experiment extract the 12-order MFCC of musical instruments, bird sounds and sound of weather, then classify these sounds with improved SOM network. The results show that the improved SOM network has better performance.