A Method of Constructing Ensemble Classifiers for Recognizing Audio Data of Various Nature

A method of constructing ensemble classifiers for recognizing audio data of various nature was created in this work. For the creation of this method the following things were done: for audio emotion recognition and genre recognition problems two datasets - RAVDESS and GTZAN were chosen, transformed and used as train and test data. 7 classifiers, such as K Nearest Neighbors, Support Vector Machine, Random Forest, XGBoost, Multilayer Perceptron, Convolutional Neural Network and Long Short-term Memory were constructed as the elements of ensemble classifiers that were combined with stacking using soft, hard voting aggregation and soft voting with GOMPERTZ fuzzy ranking. The research results indicate that the ensemble classifier’s accuracy is 8.1 percent higher than that of the best individual classifier within it (MLP) and F1 metric is also 8 percent higher for audio emotion classification problem. For music genre recognition the correspondent numbers are 5.6 percent and 5.2 percent higher.