The Perception of Accents in Pop Music Melodies

Abstract

We examine several theoretical and empirical approaches to melodic accent perception and propose a heuristic classification system of formalized accent rules. To evaluate the validity of the accent rules a listening experiment was carried out. 29 participants had to rate every note of 15 pop music melodies presented as audio excerpts and as monophonic MIDI renditions for their perceived accent strength on a rating scale. The ratings were compared to accent predictions from 38 formalized, mainly binary accent rules. Two statistical procedures (logistic regression, and regression trees) were subsequently used in a data mining approach to determine a model consisting of an optimally weighted combination of smaller rule subset to predict the accents votes of the participants. Model evaluation on a set of unseen melodies indicates a very good predictive performance of both statistical models for the participants' votes obtained for the MIDI renditions. The two models derived for the audio data perform less well but still at an acceptable level. An analysis of the model components shows that Gestalt rules covering several different aspects of a monophonic melody are of importance for human accent perception. Among the aspects covered by both models are pitch interval structure, pitch contour, note duration, metrical position, as well as the position of a note within a phrase. In contrast, both audio models incorporate mainly rules relating to metre and syncopations. Potential applications of the presented accent models in automatic music analysis as well as options for future research following this computational approach are discussed.

Acknowledgements

We thank the 29 participants of the experiment, Steffen Just for entering the experimental data and Marcus Pearce for revising several drafts of this paper. Daniel Müllensiefen is supported by EPSRC grant EP/D038855/1.

Notes

¹The term logistic regression is also commonly used to refer to this type of model emphasizing the type of transform (logit) applied to the response variable. However, we like to refer to this model as a binomial model emphasizing the type of error distribution of the response variable that we assume, leaving some leeway to explore other options of transformation (e.g. probit transformation) of the response variable.

²In fact, a rank correlation between the Pearson values and the AUC values comparing all accent rules and the proportion variable for the data of subset 1 yields a correlation value of 0.85.

³The parameter estimation was done using an approximation to the maximum likelihood criterion as obtained by the iteratively weighted least squares (IWLS) procedure (e.g. Venables & Ripley, 2002, p. 185). We used the functions glm, stepAIC, and update as implemented in the statistical software R.