Perceived accent in melodies: Computational, musicological, and perceptual issues

In our presentation we introduce a project being carried out on the modelling of the perceived accents in melodies. The main goal was to make a model that can predict the perceptual data from the score using a set of context-dependent local principles derived for example from the pitch curve, note durations, and metrical position. The general approach was data-driven rather that starting from a specific theory. It means that we computed features from the score representation that express several alternative hypotheses about the origin of the perceptual accents. A representative set of 60 melodies were selected from three different genres (baroque, romantic, and post-tonal) each with vocal and instrumental melodies. They were recorded on a Yamaha Disklavier without any performance variations. Twenty-one amateur musicians listened to all melodies and rated each note according to the perceived importance. A rather large set of different features were extracted from the score representation of all melodies. Using different machine learning methods, the average perceptual data was predicted from this feature set. We are currently analysing the data and we will discuss the results from several computational, musicological, and perceptual points of view, for example: What is the relationship between the musical style and the perceived accents in melodies? What is the relationship between performance, musical style, and perceived accent? From a perceptual point of view, do the cognition, education, age, culture, the familiarity with a particular repertoire (as listener and as performer), affect the perception of the accents in melodies? Is it possible to make an average across subjects when there are different strategies for marking the accents? And finally, how can computation, perceptual, and musicological approaches be combined in order to study the perceived accents in melodies.