Lower SNR Environment Sound Recognition

The aim of sound recognition in acoustic scene is trying to recognize the real events hidden in the audio data.Sound recognition plays an important role in many applications,such as audio forensics,environmental sounds recognition,bio-acoustic monitoring,auditory scene analysis(ASA),environmental security surveillant,real-time military focus detection,location-tracking and sound source identification,patient monitoring,abnormal event detection and troubleshooting,submitting the key information of early maintenance and so on.Because of different environments,background noise which concurrently exist with environmental sounds are different and appear as a non-stationary signal.Therefore,how to effectively recognize environmental sound in different acoustic scenes,especially under low signal-to-noise ratio(SNR)condition,is still a challenging task.In this paper,the main research of this paper includes the following four aspects:1)The low SNR environment sound framework.In low SNR acoustic scenes,the background noise can interfere environment sound recognition rates.Hence,how to find an effective framework to eliminate this interference is the key to this topic.The proposed framework,use the training sample mixed by environment sounds and background noise to train the classifier.2)Endpoint detection and SNR estimation:the test signal is decomposed into intrinsic mode functions(IMFs)using empirical mode decomposition.Then,we detect end points of the 2nd to 6th intrinsic mode functions,and give the final decision by voting.The result of end point detection is for SNR estimation of testing environment sounds.3)Feature extraction methods:for variability,nonstationary and unstructured of signals this paper proposed a feature extraction method combined gray level co-occurrence matrix(GLCM)with higher order singular value decomposition(HOSVD).First,environment sound signals are transformed into a frequency spectrum using fast Fourier transform(FFT),and then use GLCM as the texture features of spectrum.Finally,a plurality of gray co-occurrence matrix is decomposed using HOSVD to get eigenvalues for classification.4)Muti-random forest classifier and sub-random forest classifier:this paper proposed a muti-random forest(M-RF)classifier to solve the problem of of SNR estimation inaccurately.Muti-random forest train random forests using multiple SNR mixed training samples.Then,those forests are reformed as a new forests.For sub-band power distribution(SPD)feature,we proposed a sub-random forest classifier for environment sound recognition.The feature vector is separated into several sub feature vector to training the random forest.Then,those forests are reformed as a new forests,which is the same as M-RF.In the experiment,forty kinds of three categories environment sounds including birds,mammals and insects are utilized to conduct research and contrast experiments.And the experimental results show that the proposed method still remains an average recognition rate over 70%even in-5dB acoustic scene.