Comparison of classification results obtained by using cyclostationary features, MFCC, proposed algorithm and development of an environmental sound classification system

This paper presents a method for automatic classification of environmental sounds into three major categories, namely bird sounds, human speech and noise. Three various algorithms are used to classify the sounds individually and their efficiencies are compared. The classification algorithms are based on cyclostationary features, Mel Frequency Cepstral Coefficients (MFCC) and our proposed algorithm uses the autocorrelation properties, Dynamic Energy Ratio (DER), Zero Crossing Rate (ZCR) of the sound signals for classification. Initially a large dictionary of sounds of three categories is created. Various features of the dictionary are studied and the distinct features for each category are identified and the system is trained accordingly to classify the test sample into its respective category by fixing individual thresholds. The criteria for fixing the thresholds for each category have been discussed in the paper.