The creative side of Maximum Entropy: a scientific paper provides a new model of capturing and generating music melodies

Music can be looked as a complex network of interacting components, whose knowledge allows to generate new sequences from original phrases. A new paper, out in Nature Scientific Reports, explores the statistics of melodies, presenting an innovative Maximum Entropy model that can be used for generating sequences that mimic some aspects of the musical style of a given corpus.

Titled Maximum entropy models capture melodic styles and co-authored by ISI Foundation Research Leader Vittorio Loreto and ISI Associated Researcher Francesca Tria, with Jason Sakellariou and François Pachet of Sony CSL/Sorbonne Universités, the research doesn't use the Markov chains (the most common strategy in automatic music generation), instead providing a k-nearest neighbour model with pairwise interactions only.

Results show that long-range musical phrases don't need to be explicitly enforced using high-order interactions, but can instead emerge from multiple, competing, pairwise interactions. The new model, based on Maximum Entropy principle, outperforms both fixed-order and variable-order Markov models in generating new melodies that emulate the style of a given musical corpus, providing sensible alternative realizations that borrow from the original phrases without simply plagiarizing them.

Maximum entropy models capture melodic styles, Jason Sakellariou, Francesca Tria, Vittorio Loreto and François Pachet, Scientific Reports, 23 august 2017