Recognition and Classification of Philippine Indigenous Musical Instruments Using Convolutional Neural Networks with Timbral and Pitch Characterization

This research employed music information retrieval to develop a system for recognizing and classifying Philippine Indigenous Musical Instruments using Custom CNN and VGG-16 architectures. Audio samples were collected from the Katunog database, UPCE archives, and a live recording. The VGG-16 model consistently achieved higher accuracy and F1-scores than the Custom CNN for both individual and family instrument classification, regardless of whether Mel Spectrogram or MFCC data were used. Mel Spectrogram performed better on both CNN architectures. Additionally, most models exhibited higher performance metrics at a k-value of 10 for k-fold cross-validation. The PIMIs were then characterized in terms of timbre and pitch to explore their uniqueness. When using Self-Organizing Maps, MFCC and Mel Spectrogram data showed strong clustering, explaining the high performance of CNN models in recognition and classification tasks.