Learning and Recalling Melodies

Abstract

Using melodic recall paradigm data, we describe an algorithmic approach to assessing melodic learning across multiple attempts. In a first simulation experiment, we reason for using similarity measures to assess melodic recall performance over previously utilized accuracy-based measures. In Experiment 2, with up to six attempts per melody, 31 participants sang back 28 melodies (length 15–48 notes) presented either as a piano sound or a vocal audio excerpt from real pop songs. Our analysis aimed to predict the similarity between the target melody and participants’ sung recalls across successive attempts. Similarity was measured with different algorithmic measures reflecting various structural (e.g., tonality, intervallic) aspects of melodies and overall similarity. However, previous melodic recall research mentioned, but did not model, that the length of the sung recalls tends to increase across attempts, alongside overall performance. Consequently, we modeled how the attempt length changes alongside similarity to meet this omission in the literature. In a mediation analysis, we find that a target melody’s length, but not other melodic features, is the main predictor of similarity via the attempt length. We conclude that sheer length constraints appear to be the main factor when learning melodies long enough to require several attempts to recall. Analytical features of melodic structure may be more important for shorter melodies, or with stimulus sets that are structurally more diverse than those found in the sample of pop songs used in this study.