Toward Better Machine Understanding of Rhythm: Rhythmicness as a Regression Problem

Abstract

The concept of musicality refers to the ambiguously defined qualities of melodiousness, harmoniousness, and rhythmicity. Prior computational research has investigated musicality solely as a classification task discriminating music from speech or other environmental sounds, however, research in psychoacoustics indicates that people perceive musicality as a continuous rather than discrete feature audio. To address this, we propose a novel task of assessing musicality, and the absence thereof, as a continuous feature of sound. We present novel datasets and a regressive approach for assessing the extent to which a given piece of audio contains a structured rhythm. We test a variety of convolutional models, including a state-of-the-art model from the related task of beat-tracking, to try and predict the extent to which a given audio sample differs from a metrically regular pattern of note onsets. We make contributions toward determining which models are suited for this task, and establish model performance baselines for future research on this task.

Keywords: CNN, Musicality, Rhythm, Transformers, Deep learning

DOI: 10.54941/ahfe1007220