Computer-Aided Orchestration Based on Probabilistic Instruments Models and Genetic Exploration

In this paper we introduce a tool aimed at assisting composers in orchestration tasks. Thanks to this tool, composers can specify a target sound and replicate it with a given orchestra. We discuss the problems raised by the realization of such a tool, concerning instrumental sound description and combinatorial optimization. Then we describe the solution adopted. We propose a machine learning method based on generative probabilistic modeling to represent and generalize instrument timbre possibilities from sample databases. This model allows to deduce the timbre of any mixture of instrument sounds. In a second part, we show that search of sound mixtures that match a given target is a combinatorial optimization problem that can be addressed with multicriteria genetic algorithms.