Eco-environmental sounds classification with time-frequency features under noise conditions

Eco-environmental sounds depict the sound content of varieties of creatures´ survival and activities in the ecological environment at a time interval. Research on eco-environmental sounds is useful in monitoring of the wildlife and their evolution with time. Due to varieties of noises in the ecological environment, we consider the task of eco-environmental sounds classification under noise conditions. Time-frequency representations have the potential to be powerful features for nonstationary signals. Especially, time-frequency domain features can classify sounds with noise where using frequency-domain features (e.g., MFCCs) fail. Hence, a classification approach using time-frequency features for eco-environmental sounds under noise conditions is presented in this paper. Matching pursuit (MP) algorithm is proposed to extract time-frequency features (MP-based features, for short) of effective signals. Besides statistical features extracted under Choi-Williams distribution (CWD-based features, for short) also perform more effectively than other conventional audio features under noise conditions. Considering the effectiveness of features and robustness of classifier, a classification model using time-frequency features (the combination features of MP-based features and CWD-based features) and support vector machine (MP+CWD-SVM for short) is proposed. Experimentally, CWD+MP-SVM is able to achieve a higher classification rate for eco-environmental sounds under noise conditions. The result shows that time-frequency features and SVM classifier have better noise immunity.