A multi-feature approach for musical instrument classification using machine learning

This study examines the performance of a collection of spectral audio features, including RMS Energy, Zero Crossing Rate (ZCR), and Spectral Centroid, for musical instrument classification by using the Random Forest and XGBoost classifiers. These machine learning algorithms demonstrate enhanced precision in complex classification scenarios and improve the ability to discriminate among highly correlated instrument classes. Machine learning approaches were employed in this study due to being explainable, computationally efficient, and suitable when deep learning is not feasible under the constraints of hardware or data. As part of the experimental setup, the audio features were obtained from the Philharmonia dataset, which includes 20 instrument classes. Seven different configurations were evaluated, including each feature set individually, as well as their pairwise and triplet combinations. The highest performance in terms of accuracy was obtained when all attributes were utilized: 0.91 with Random Forest and 0.93 with XGBoost. These machine learning algorithms were particularly well adapted to distinguish acoustic differences in music. Confusion matrix analysis indicated that both models worked best for instruments with clear acoustic characteristics, such as guitar and banjo. The findings suggested that the combination of multiple complementary features improves the classification performance of musical instruments.