École Nationale Supérieure de
Techniques Avancées

Abstract

Musical texture is a high level feature. This work exposes a semantic, syntactic and object-oriented structure modeling pianistic texture of the classical repertoire and aimed at automatizing its analysis. A dataset of 1164 annotated textures was built in order to train supervised machine learning models. Furthermore, 62 musical descriptors were implemented to describe high-level phenomena in symbolic music. Tested linear models are found to be globally more efficient for the prediction of 15 distincts textural elements. The analysis of these results and the statistical study of the corpus underline the parts to be improved , but finally issue findings which are musically consistent.