| Resumen |
The task of music genre recognition (MGR) has significant applications in the music industry, including copyright control and music recommendations. Traditional feature engineering approaches in classical machine learning (ML) have utilized spectral features like Mel Frequency Cepstral Coefficients (MFCC), alongside time and frequency domain features such as zero-crossing rate, spectral centroid, and spectral rolloff. This study investigates the impact of harmonic, tonal, and rhythmic features, specifically tonnetz, chroma, and tempo, on the performance of ML models. We implemented Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and XGBoost, both in their traditional forms and with the integration of active learning, to classify genres in the GTZAN dataset. Our results demonstrate that active learning significantly improves model accuracy, with the highest accuracy achieved by the active SVM model at 80.3% using a combination of tonnetz and chroma features. This study underscores the importance of tonal and rhythmic features, particularly tonnetz, in optimizing MGR models. Future work includes expanding the dataset using the developer API from a music streaming platform and exploring alternative feature representations such as auditory filter banks. |
