Style Manipulation as a Creative Device

Creative artifacts are often obtained by combining ideas, patterns, ways of doing – in a word, styles – to new configurations or situations. We describe a new approach to content authoring tools based on this idea. The approach is based on letting users explicitly manipulate style, as a computational object. In such a context, one of the key technical problem is to generate sequences that imitate a style, while satisfying arbitrary constraints, coming from the domain of study.
We describe a novel sequence generation approach based on an old statistical tool: Markov chains. We show that it is possible to explore the complete set of sequences that a Markov model can generate, using combinatorial optimization techniques. We show that the addition of even simple constraints bias the initial Markov model in interesting ways, which are not fully understood. We describe some recent results in constrained Markov generation who pave the way for novel and exciting style manipulation applications, in music composition, improvisation and literary text writing, and give a few examples developed within the Flow Machines project.