Performance Comparison for EMD Based Classification of Unstructured Acoustic Environments Using GMM and k-NN Classifiers

Abstract

The environment around us is very rich in acoustic information, with its scope extending beyond mere speech signals. Perception of acoustic information plays a significant role in allowing human beings to comprehend sounds that they hear in their surroundings. This paper focuses on the development of feature extraction and accurate classification of variety of acoustic sounds in unstructured environments, where adverse effects such as noise, distortion are likely to dominate. This work attempts to classify ten different unstructured real-world acoustic environments using empirical mode decomposition (EMD) which considers inherent non-stationarity of acoustic signals by decomposing the signal into intrinsic mode functions (IMFs). These IMFs are used for feature extraction. This work suggests utility of composite feature set for classification and proposes an optimized, robust, best-suitable feature set for classification of the diverse acoustic environments. For the classification task, Gaussian mixture model (GMM) and k-nearest neighbor (k-NN) classifiers are used. Utilization of this optimized best-suitable feature set yields the maximum classification accuracy of 100 % with GMM classifier and an average accuracy of 95 % for k-NN (k = 1, 3, 5). Lastly, this study presents the use and comparison of various performance metrics to evaluate classification techniques used.