The Comparison of Different Feature Extraction Methods in Musical Instrument Classification

In this paper, we analyze four different methods for audio feature extraction and compare their efficiency in the context of musical instrument classification. We study spectrograms, Mel spectrograms, Linear-Frequency Cepstral Coefficients (LFCCs) and Mel-Frequency Cepstral Coefficients (MFCCs) in combination with three different Deep Learning architectures: VGG-16, ResNet-34 and a custom CNN. We investigate the behavior of our models in two different classification scenarios to determine a possible correlation between the number of classes and the efficiency of each method. For this purpose, we took samples from the London Philharmonic Orchestra dataset and ran the experiment for three and fifteen classes of musical instruments belonging to three different instrument families: Woodwinds, Strings and Brass.