Evaluating Feature Selection Algorithms for Machine Learning-Based Musical Instrument Identification in Monophonic Recordings

Abstract

Musical instrument identification (MII) research has been studied as a subfield of the Music Information Retrieval (MIR) field. Conventional MII models are developed based on hierarchical models representing musical instrument families. However, for MII models to be used in the field of music production, they should be developed based on the arrangement-based functions of instruments in musical styles rather than these hierarchical models. This study investigates how the performance of machine learning based classification algorithms for Guitar, Bass guitar and Drum classes changes with different feature selection algorithms, considering a popular music production scenario. To determine the effect of feature statistics on model performance, Minimum Redundancy Maximum Relevance (mRMR), Chi-sqaure (Chi2), ReliefF, Analysis of Variance (ANOVA) and Kruskal Wallis feature selection algorithms were used. In the end, the neural network algorithm with wide hyperparameters (WNN) achieved the best classification accuracy (91.4%) when using the first 20 statistics suggested by the mRMR and ReliefF feature selection algorithms.