Enhanced Feature Extraction Approaches for Detection of Sound Events

This work focusses on using digital signal processing techniques to analyze and extract audio features and use them to predict the type of event that might have taken place in an audio signal using supervised machine learning approaches. The performance of five classification approaches using different feature subsets were analysed. Feature subsets include frequencies of the segmental features, frequencies of the supra-segmental features and combination of both. This gives an insight about the relative importance of the feature subsets and also the need for extracting new features from existing features. Features were extracted after the audio signal was filtered using a lowpass Butterworth filter with a cutoff frequency of 1500 Hz; it was inferred that the including features of the difference signals improved the performance of the learning algorithms. The work also includes tuning the parameters of the classification approaches to improve the performance. The observations and inferences of the experimental results can potentially be used for designing robust surveillance systems for rare event detection.