Computational Approach of Musical Orchestration - Constrained Multiobjective Optimization in Large Sound Sample Databases


Abstract

Among all techniques of musical composition, orchestration has never gone further than an empirical activity. Practicing and teaching orchestration — the art of blending instrument timbres together — involve hard-to-formalize knowledge and experience that computer music and composition systems have for years stayed away from. The state-of-the-art orchestration tools search for sound combinations within instrument sample databases that best match a target timbre defined by the composer. To this end, those methods use either decomposition or matching pursuit algorithms, and therefore circumvent to the combinatorial problem of orchestration. We propose in this thesis an original approach for the discovery of relevant sound combinations, in which we explicitly address combinatorial issues and tackle orchestration in its inner complexity. Initial considerations of the problem in a multiobjective knapsack framework shows that non-linearity and non-additivity of objective functions require a wider theoretic approach. We suggest a generic and easily extendible formalization of orchestration as a constrained multiobjective search towards a target timbre, in which several perceptual dimensions are jointly optimized. We first validate our approach on a small-size problems test set, with a rather simple, spectral-based timbre description. We show that theoretic solutions of the optimization problem correspond to perceptually relevant orchestration proposals. We then introduce a time-efficient evolutionnary orchestration algorithm allowing the discovery of optimal solutions. By estimating acoustic properties of sound mixtures, our method suggests orchestration proposals in relation with perceptual criteria and favors the exploration of somehow nonintuitive sound mixtures. From there, the search may be pursued in specific directions. To enhance the control of symbolic features in orchestrations proposals, we define a formal framework for global constraints specification and introduce an innovative repair metaheuristic. Thanks to this method the search is led towards regions fulfulling a set of musical-related requirements. We finally present a composer-friendly computer-aided orchestration prototype, in which timbre space exploration is encouraged by a multiple viewpoints representation and an interactive mecanism for guessing the composer’s listening preferences. We end this thesis with relevant application examples in real musical works.