Environmental Sounds Recognition Based On Spectrogram

The ecological environment is closely related to our life,and animal sounds of ecological environment are viewed as research object,which contain a large amount of rich information.By animal sound recognition,we can make some understanding and analysis on their life habits and distributions to effectively monitor and protect them.However,due to the existence of different environments and noises,the existing method is difficult to ensure the recognition accuracy of animal sound in low signal-to-noise ratio(SNR)condition.To address these problems,this paper proposes a double feature,which consists of projection feature and local binary pattern variance(LBPV)feature,combined with random forests(RF)for animal sound recognition.What's more,short-time spectral estimation algorithm is used to enhance sound signal.The main work includes the followings:(1)Sound enhancement process.The traditional noise estimation algorithms usually require prior assumption that the background noise is steady,so it is not useful for the real non-stationary environment noises.This paper proposes a sound enhancement algorithm by combining the classical short-time spectrum estimation algorithm with a dynamic noise power spectrum estimation algorithm based on two-way searching,to realize the noise reduction of sound signal.(2)Double features extraction.By observing and analyzing the spectrograms of animal sounds,it can be found that these spectrograms have uniqueness and distinguishability,which are suitable for sound recognition.This paper extracts projection feature and local binary pattern variance(LBPV)feature from spectrogram to generate the double feature.Projection feature is set as the first layer of double feature,which is a global feature,got by eigenvalue decomposition and projection on the entire spectrogram matrix.The second layer is LBPV feature,which captures local features of image,combining local binary pattern(LBP)feature with contrast feature effectively.The global feature and local feature,which are complementary,can effectively improve not only recognition rate but also anti-noise performance.(3)Recognition and classification.In view of the advantages of ensemble classifier in effectively improving the recognition accuracy and avoiding the overfitting of data,an ensemble classifier model based on decision tree—Random Forest is used for training and recognizing double feature.And the final recognition result of each test sample depends on the majority voting of many decision trees.RF has good performance as well as fast speed.In the experiments,this paper classifies 40 kinds of common animal sounds under different SNRs with rain noise,traffic noise,and wind noise.As the experimental results show,the proposed method consisting of short-time spectrum estimation,double feature,and RF,can recognize a wide range of animall sounds and still remains a recognition rate over 80%even under OdB SNR.