Algorithmic prediction of music complexity judgements
Sebastian Streich, Perfecto Herrera · 2006
In the research presented here, we evaluate the accuracy of algorithmic music complexity predictions. Several algorithms are considered that compute complexity estimates solely based on the music audio signal. The different algorithms focus each one on a different musical facet (e.g. rhythm, instrumentation, harmonies). In order to test the capacities of these algorithms (individually and jointly) we match their output against subjective complexity ratings collected through a listening test and look for significant correlations. Method In the listening test subjects were asked to rate up to 30 music excerpts that were randomly picked from a collection of 82 tracks. Rating consisted on judging the complexity of the music and also indicating their familiarity with and their liking of each excerpt. Additionally, in a separate part of the survey, data about the subjects’ usual music listening habits were collected. An automatic analysis of musical features was used to build predictive models of the excerpts’ complexity. Results In order to evaluate the models we have used a subset of 43 excerpts, those that have been consistently and coherently rated by most of our 16 listeners. With this set, we found some significant correlations between the averaged subjects’ ratings and the algorithmic predictions. In particular, a simple rhythmic complexity model and a model based on the loudness fluctuations by themselves were already significant at the p<0.05 level. Conclusions