Résumé:
The primary goal of this thesis is to conduct extensive experimental comparisons among Sound Event Detection systems using a combination of DCASE 2016 and DCASE 2017 datasets. We have carried out two sets of experiments. First, We have examined 3 different types of features extracted from short time frames of each audio recording (Mel frequency cepstral coefficients MFCCs, Log Mel-band Energy and a combination of MFCCs with AMFCCs and AAMFCCs) and 4 classification paradigms while varying their parameters (Support Vector Machine, Convolutional Neural Network, Adaboost and Random Forest). We have supported our analysis and discussion with numerous statistical tests to analyze and compare the effect of the above mentioned features and classifiers on the detection performance. Our experimental findings indicate the effectiveness of the ensemble-based classifiers (Random Forest and Adaboost) along with MFCCs and the Support Vector Machine classifier along with MFCCs + AMFCCs + AAMFCCs. More specifically, for both classifiers, we have obtained an overall class
based detection accuracy of 83% using 1 second segment based evaluation technique. Second, we have investigated the effect of the number of features on the generalization performance of the Random Forest classifier through invoking a feature selection approach, namely Minimum Redundancy-Maximum Relevance. However due to the lack of data in this field of research, this technique has caused a slight degradation in the performance of the system as the selected features were not sufficient for learning an effective model.
Keywords: Sound Event Detection, Feature Engineering, Feature Selection, Machine Learning.
