EMD based efficient discrimination of real-world environmental sounds using SVM classifier

Sounds from real-world environments are playing a vital role in fundamental and applied research which crosses the limits of different domains. Natural and artificial environmental sounds around us are affluent in terms of acoustic information, with their scope expanding much beyond conventional speech/music signals. The sixth sense of such acoustic information is important for allowing humans to comprehend sounds that they pick up in their atmosphere. This paper is associated with implementation of feature extraction and efficient discrimination of real-world environmental sounds. In the study presented here ten different environmental sounds are discriminated by means of Empirical Mode Decomposition (EMD). EMD takes into account intrinsic non-stationarity associated with acoustic signals by decomposing the original signal into an ensemble of Intrinsic Mode Functions (IMFs). Feature extraction is carried out using these IMFs. This work recommends the use of hybrid feature set for discrimination and suggests an optimized, best suitable feature vector for discrimination of diverse environmental sounds. One-against-all support vector machines (OAA-SVM) classifiers are used for discrimination. This most appropriate feature set yields the maximum discrimination accuracy of 98.75 % with SVM classifier.