<![CDATA[Introduction: Balinese Selonding gamelan is an endangered sacred repertoire, and automatic recognition of its musical pieces can support documentation and preservation. Method: This study investigates the automatic classification of Selonding gamelan music using a Convolutional Neural Network (CNN). The dataset consists of 10 traditional Selonding compositions. Recordings were segmented into fixed 15-second excerpts, converted to WAV, normalized, and transformed into time–frequency features using two approaches: Mel-Frequency Cepstral Coefficients (MFCC) and Constant-Q Transform (CQT). A CNN-based classifier was trained and evaluated using 5-fold cross-validation for each feature representation. Results: The MFCC-based model achieved stable high performance, with mean accuracy of 94.67% (±2.11%), mean precision of 94.97% (±1.90%), mean recall of 94.67% (±2.11%), and mean F1-score of 94.63% (±2.12%) across folds. In contrast, the CQT-based model performed notably worse, reaching only 58.04% mean accuracy and 53.28% mean F1-score, with large variance across folds. These results indicate that MFCC features capture the discriminative timbral characteristics of Selonding more effectively than CQT under the current experimental setting. Conclusion: Overall, the findings show that a CNN trained on MFCC features can reliably distinguish Selonding compositions using only short (15-second) audio segments, despite limited data. This suggests that deep learning is a feasible strategy for indexing, retrieval, and long-term preservation of Balinese gamelan repertoires.]]>
Copyrights © 2025
