Modeling style from examples

Modeling style from examples

by François Pachet

Abstract: Style is what makes an author (writer, composer, designer, etc.) recognizable. Can machines understand, model, predict and generate artefacts in a given style? In the Flow-Machines project, we address these questions for music and text. I will show several generative models of style able to capture characteristic traits of musical sequences, as well as polyphonies, both in the symbolic (score) and numeric (audio) domains. These models are based on the combination of combinatorial (so-called Markov constraints) and statistical perspectives on music. I will use various reorchestrations of “Ode to Joy”, the European anthem, to illustrate these various approaches. Along the way these increasingly sophisticated models in turn, raise more fundamental questions about nature of music and the mysterious mechanisms of taste and subjective appreciation.