Enhanced Instrument Identification Using CNNs: A Comparative Study with Random Forest Models

This study explores the effectiveness of instrument identification algorithms (IIAs) using Convolutional Neural Networks (CNNs) and Random Forest (RF) models. Utilizing the NSynth dataset from Tensorflow, we began by evaluating the original CNN model, then enhanced its structure and feature selection to improve performance. The updated CNN model was compared to both the unmodified CNN and the RF model. The unmodified CNN achieved a practical validation accuracy of 55% and a training accuracy of 66%, indicating an 11% loss. The RF model demonstrated an average accuracy of 65%, with accuracies ranging from 92% for strings to 45% for reed instruments. Despite an initial inclination to refine the RF model, our research showed that CNNs could achieve higher accuracy with minimal preprocessing adjustments. By switching the input from mel spectrograms to MFCCs, we observed significant improvements. Our enhanced CNN model outperformed both the unmodified CNN and the RF model, highlighting the potential of CNNs in IIA development. However, this approach also resulted in increased training times and processing demands.