| Popis: |
Background The notes in a melody are perceived as differently important even when they are performed without any emphasis or expression. This is a crucial aspect for forming the perception of the musical structure such as the metrical grid (Large and Jones, 1999). Several models have been suggested that predict perceived accents from musical structure (e.g. Thomassen, 1984; Bisesi et al., in submission). These models often use a top-down approach starting from theoretical principles with rather limited data sets. Thus, a general model trained on a large database is still lacking. Aims We aim to formulate a model that predicts perceived accents from the score using a set of context-dependent local principles derived e.g. from pitch curve, note duration, and metrical position, and to use this model to investigate which principles are being used by listeners to determine perceived accents. Methods Our approach is data-driven instead of starting from a specific theory. We computed several features from the score representation, expressing different hypotheses about the origin of perceptual accents (see also Müllensiefen et al., 2009). A representative set of 60 melodies (30 vocal and 30 instrumental) was selected from three different styles (baroque, romantic, and post-tonal). They were recorded on a Yamaha Disklavier without any performance variations. Thirty amateur musicians listened to all melodies and rated each note according to their perceived importance. Music analysis of a selection of answers was carried out in order to formulate hypotheses about the underlying principles used by listeners in their decisions. These hypotheses, as well as other principles gathered from the literature, constituted our ground-truth. Most of them were local principles derived from neighbouring notes’ context, and were related to pitch contour, duration, and metrical position. These features were extracted for each note from the score representation of each melody. Using different machine learning methods, the average perceptual data was predicted from this set of features. Results The resulting data is currently analysed so we lack robust results. However, preliminary analyses indicate that (1) listeners used different strategies, (2) the task was relatively hard, in particular for the complex melodies, (3) a meaningful average can be extracted across the listeners, corresponding to the principles we have assumed, and (4) the model is able to predict the average with adequate accuracy. Further music analyses will address differences between different music styles, participants’ backgrounds, and individual strategies. Conclusions Our model will provide further insight in how listeners do perceive accents in melodies featuring different degrees of structural complexity and musical styles. |
