A method of developing ensemble classifiers for recognizing audio data of various nature

The work developed a method of building ensemble classifiers for recognizing audio data of various nature. The method is a tentative date and requires the following steps. In the first step, the input datasets are selected, which are transformed and divided into training and test samples, respectively. RAVDESS datasets are chosen for the task of audio emotion recognition, music genre recognition is performed on the GTZAN dataset. In the second step, the following seven classifiers were created and investigated as elements of ensemble classifiers: K Nearest Neighbors, Support Vector Machine, Random Forest, XGBoost, Multilayer Perceptron, Convolutional Neural Network and Long Short-term Memory. In the process of training elementary classifiers, the corresponding hyperparameters were adjusted using the Grid Search approach. In the third step, elementary classifiers were combined using the stacking ensemble method with such types of aggregation as soft voting, hard voting, soft voting using the GOMPERTZ function. All possible ensemble combinations starting with three elementary classifiers for recognizing audio emotions and music genres were tested. Therefore, the total number of ensembles studied in the work was 297. The research results for the problem of audio emotion classification showed that the accuracy of recognition according to the Accuracy metric of the best ensemble classifier is 8.1% higher than that of the best elementary classifier in its composition, which is based on on MLP, and according to the F1 metric, this indicator is 8% higher. For the task of recognizing music genres, the corresponding indicators are higher by 5.6 % and 5.2 %, respectively.