Using Spectro-Temporal Features for Environmental Sounds Recognition

The paper presents the task of recognizing environmental sounds for audio surveillance and security applications. A various characteristics have been proposed for audio classification, including the popular Mel-frequency cepstral coefficients (MFCCs) which give a description of the audio spectral shape. However, it exist some temporal-domain features. These last have been developed to characterize the audio signals. Here, we make an empirical feature analysis for environmental sounds classification and propose to use the log-Gabor-filters algorithm to obtain effective time-frequency characteristics. The LogGabor filters-based method utilizes time-frequency decomposition for feature extraction, resulting in a flexible and physically interpretable set of features. The Log-Gabor filters-based feature is adopted to supplement the MFCC features to yield higher classification accuracy for environmental sounds. Extensive experiments are performed to prove the effectiveness of these joint features for environmental sound recognition. Besides, we provide empirical results showing that our method is robust for audio surveillance Applications.