Autocorrelation of Pitch-Event Vectors in Meter Finding

摘要

Computational researchers often use autocorrelation techniques to identify the meter of a musical passage, tracking the ebs and flows of loudness or -if using symbolic data- peaks and valleys of note attacks. This paper investigates the relative success of various harmonic and pitch events compared to a note-attack model when identifying musical meter using autocorrelation. This study implements such a process using several different parameters: note attacks, pitch class change, set class probabilities, and scale-degree set probabilities. These outputs are measured against a ground truth derived from each piece's notated time signature. The relative success of each parameter is tracked using F scores. While the study shows that loudness-oriented parameters are overall more successful, the paper discusses how its findings add to our understanding of musical meter and the role played by pitch parameters in metric accents.