Robust Eco-environmental Sounds Recognition Based On Multi-band Spectral Subtraction Method

With the development of our society, the ecological environment which is related to our life is getting increased attention. In the ecological environment, a wide variety of creatures produce different sounds, these sounds contain a wealth of information and are closely related with their living environment and life habits, as well as human activities. Therefore, the study of ecological sounds recognition has an important significance to protect eco-environment. However, there are varieties of unpredictable noise in the actual ecological environment. Thus it will be more valuable to study the task of recognizing eco-environmental sounds in a noise situation.The main work includes the followings:1) Feature extraction and feature selection:First, extract a variety of different ecological environment sound features; second, using each feature as the training sample respectively and building classification model with the same classifier; Last, considering classification ability and anti-noise ability, we compare the results of different sound features, then reserve features which have better recognition rate.2) Dimensionality reduction:We get distinctive feature vector with higher dimensions after feature extraction and feature selection. In order to reduce data processing complexity and preserve data inherent geometric properties at the same time, in this paper, we make use of ISOMAP which based on multi-scale transform for dimensionality reduction.3) Spectral subtraction noise reduction:In real situation of ecological environment sounds recognition, we would encounter various types of noise situation. In order to reduce the noise interference, we propose a sound enhancement method for the non-stationary noisy environment by combining adaptive averaging periodogram algorithm with multi-band spectral subtraction algorithm, to realize the noise reduction of ecological environment sounds.4) Two-layer sound recognition algorithm:Considering the superiority of SVM in solving the pattern recognition problem of small, nonlinear structure and high dimension samples, in this paper, we use SVM as the first layer classifier; then taking advantage of the excellent clustering ability of GMM, we use it as the second layer classifier to adjust the result of previous layer and get the final result. They constitute a SVM_GMM two layer classifier.In the paper we recognize ecological sound under low SNR noise environment, meanwhile, provide a reference for better research on other sounds.