This study aims to develop a predictive model for classifying emotions in instrumental music excerpts using machine learning techniques. The methodology involves experiments with three distinct datasets, including a newly created dataset for this research. The models employed include Random Forest, Multi-Layer Perceptron, and Convolutional Neural Network architectures. The experimental results, conducted with a single dataset for testing and validation, were generally positive. However, when attempting to generalize the models using different datasets, a considerable reduction in generalization capability was observed. Despite this, the study presents promising results, indicating that increasing the number of training data can significantly improve the models’ generalization capability.