Sound Events Recognition Based On Orthogonal Matching Pursuit In Low SNR

The aim of Sound events recognition is recognizing real sound in the audio data,which plays an important role in audio forensics,location-tracking and sound source identification.However,the background noise is complex and ever-changing in real environment,which seriously affect the recognition of sound event.To address the problem,the paper consider the animal sound event of natural environment as research object and propose a sound event recognition method based on Orthogonal Matching Pursuit(OMP)and Deep Belief Network(DBN)in low signal-to-noise ratio(SNR).The main work includes the followings:(1)Optimized OMP sparse decomposition.In order to solve the problem that the high computation complexity in searching the best atom of OMP sparse decomposition,Particle Swarm Optimization(PSO)is proposed to optimized the searching process,which reserved the main body of sound signal and realized OMP sparse decomposition rapidly.(2)Second self-adaption reconstruction of optimized OMP.Concerning the problem that the traditional noise estimation algorithm require the statistics knowledge of noise in advance,a second self-adaption reconstruction method based on optimized OMP is used.Firstly,sound signal is reconstructed self-adaption by optimized OMP to reserve the main body of sound signal.Then,the short-time spectrum estimation algorithm is employed to strengthen the residue signal after the first reconstruction and compensated the first reconstructed sound signal.At last,combining the two-stage reconstructed sound signal to reduce the influence of nonstationary noise and improve the precision of reconstructed sound signal.(3)Composite time-frequency feature extraction.According to observing and analyzing the time-frequency distribution of sound signal,A composite time-frequency feature OOMP is extracted from the reconstructed sound signal,which is composed of the time feature Pitch,frequency feature MFCC and time-frequency feature optimized OMP.The OOMP feature can retain the time and frequency information,describe and representation sound signal better,and improve the recognition performance effectively and have a better anti-noise performance.(4)Deep Belief Network classification.Due to the pure recognition precision and data overfitting,DBN is employed to classify and recognize the OOMP feature to achieve a high recognition performance.DBN use greedy layer-wise unsupervised learning algorithm to extraction and abstract the input data form bottom to up.Then,back propagation is supervised used to fine tune all weights to get a strong distinguish ability model to classify and recognize sound events effectively.In the experiments,this paper classifies 40 kinds of common sound events under different SNRs with wind noise,thunderstorm noise,and lakeside in rain noise.As the experimental results show,the proposed method not only can effectively recognize a wide range of sound events but also suit for recognizing sound evens in low SNR circumstances.