Maximum Entropy Models Capture Melodic Styles

By Jason Sakellariou, Francesca Tria, Vittorio Loreto, and François Pachet

Jason Sakellariou, Francesca Tria, Vittorio Loreto, and François Pachet. “Maximum Entropy Models Capture Melodic Styles”. Scientific Reports, 7, article 9172, 2017.

François Pachet and Pierre Roy reimplemented the model in Python. The repository includes a teaching version that uses Python loops, a faster NumPy version, and examples for melodies, chord sequences, and text.

About the paper

How can a model learn the sound of a melodic style and still produce new phrases? This study uses maximum entropy to model statistical relationships between pairs of musical symbols at different distances in a melody.

Abstract in brief

The authors learn a probability distribution from pairwise statistics in a corpus of melodies and sample new sequences from it. They evaluate how well the samples reproduce the statistics of the original style and measure how much material is copied from the training corpus.

Related research

The study was supported by the Flow Machines project. The Flow Machines synthesis paper and examples place it in the broader research on assisted music creation.