Study Of Recognition Technology For Abnormal Sound Based On MFCC

With the development of society, some potential conflicts emerge, which constantly brings to the attention and thinking on abnormal sounds. In order to improve the intelligence of intelligent security system, the discriminant for abnormal sounds is urgent to be introduced into traditional video surveillance.For a long time, the research and development of abnormal sounds identification is slow,lags far behind other sound progress, it is mainly because that people can't find the characteristic parameters which describe the nature of abnormal sounds. Use for reference the human auditory features on the advantages of Listening to contend content,more and more characteristic parameters to imitate the auditory perception of human ear is put forward,which has shown great potential for development in the field of information science. This paper makes a study on extracting sound characteristics based on Mel frequency cepstrum coefficient (MFCC) and its improved feature extraction method about several typical abnormal sounds collected in the parking lot, through the support vector machine (SVM) to complete the recognition of all kinds of abnormal sounds.The following is the main contents of this paper:Firstly, the way to preprocess the abnormal sounds is mainly including pre-emphasis and adding window frame. The purpose is to eliminate numerical magnitude difference between the samples in the process of acquisition, avoid the volume high and low effects on the overall sound quality, and highlight its own characteristics of each sample. To reduce the calculation during the process, firstly detect endpoint, which can confirm the start and end of the abnormal sound.Secondly, in view of the abnormal sound signal after the pre-progress, characteristic parameters are extracted based on Mel frequency cepstrum. In the process of feature extraction, using the human ear on the perception of sound frequency characteristics,constructs a set of triangle filter bank like cochlear auditory mechanism, its role is to get spectrum energy of each frame sound signal from linear frequency domain mapping to Mel frequency domain. Then logarithmic transformation is done to the output of the triangular filter the nonlinear frequency spectrum, at last, by discrete cosine transform in mapping into cepstrum domain, to complete the MFCC feature parameter extraction.Thirdly, in the process of calculating the MFCC, against the Fourier transform of limited time-frequency resolution and the defects of calculation process generating harmonic interference, the wavelet transform is used to make the corresponding improvement, which the features of the extracted parameters more conform in line with the human ear hearing characteristics and better robustness to noise. At the same time, the empirical mode decomposition (EMD) method is introduced in the process of feature extraction, mining more dynamic characteristics, to obtain composite improved MFCC feature extraction method.Fourthly,complete the classification and identification of abnormal sounds. According to extracted features of abnormal sounds, the abnormal sound is modeled and tested though support vector machine (SVM).In the stage of train and test, combine binary classification SVM to achieve multiple classification. According to the parameters effect on the size of abnormal sound model generalization ability, choose a suitable type of kernel function, so as to get the best training model, to determine the category of test sample.