Accuracies in Algorithmic Predictors of Musical Emotion

Abstract

Music information retrieval (MIR) is a growing area of study that aims to algorithmically analyse musical features from audio recordings. Curiously, despite increasing use of MIR algorithms, few studies have examined their accuracy. Here, we evaluate the accuracy of two functions within MIRToolbox—a software library widely used in the field of music cognition for automated musical analyses. We focus on two key musical features—modality (specific note groupings that contribute to the emotional aspect of music) and attack rate (a global measure of timing information). To compare algorithmic estimates of modality and timing against known information about these features, we used “ground truth” data from a widely, historically important set of pieces. Specifically, we analysed (a) modality and (b) attack rate in four complete performances of Frederic Chopin’s Préludes (Chopin, 1839). Timing analyses revealed accurate predictions for slow pieces, but reduced accuracy for fast pieces. Mode analyses were generally accurate, but consistency varied between performers (79% to 88%). For each performer, the algorithm incorrectly predicted mode in at least three, and at most five, excerpts out of 24. This preliminary exploration of popular MIR algorithms offers valuable insights that shed light on the limitations of widely-used tools. Addressing these limitations will lay the foundation for future research endeavours that aim to delve deeper into the application of these algorithms across a wide range of musical works and genres.