Search

International audience

In this paper we first introduce a set of functions to predict the timbre features of an instrument sound combination, given the features of the individual components in the mixture. We then compare, for different classes of sound combinations, the estimated values of the timbre features to real measurements and show the accuracy of our predictors. In the second part of the paper, we present original musical applications of feature prediction in the field of computer-aided orchestration. These examples all come from real-life compositional situations, and were all produced with Orchidee, an innovative framework for computer-aided orchestration recently designed and developed at IRCAM, Paris.

In this paper a computational approach of musical orchestration is presented. We consider orchestration as the search of relevant sound combinations within large instruments sample databases. The working environment is Orchidee an evolutionary orchestration algorithm that allows a constrained multiobjective search towards a target timbre, in which several perceptual dimensions are jointly optimized. Up until now, Orchidee was bounded to “time-blind” features, by the use of averaged descriptors over the whole spectrum. We introduce a new instrumental model based on Gaussian Mixture Models (GMM) which allows to represent the complete spectro-temporal structure. We then present the results of the integration of our model and improvement that it brings to the existing system.

cote interne IRCAM: Carpentier07a; None / None; National audience; In this paper we introduce an hybrid evolutionary algorithm for computer-aided orchestration. Our current approach to orchestration consists in replicating a target sound with a set of instruments sound samples. We show how the orchestration problem can be viewed as a multi-ob jective 0/1 knapsack problem, with additional constraints and a case-specific criteria formulation. Our search method hybridizes genetic search and local search, for both of which we define ad-hoc genetic and neighborhood operators. A simple modelling of sound combinations is used to create two new mutation operators for genetic search, while a preliminary lustering procedure allows for the computation of sound mixtures neighborhoods for the local search phase. We lso show in which way user interaction might be introduced in the orchestration procedure itself, and how to lead the search according to the users choices.

We present a work related to the description of instrumental sound in the context of computer aided orchestration. We define the orchestration problem as the search of instrument sound combinations that sound close to a given target. In the presented approach, instrument sounds are described by a set of acoustic features that are extracted from large sound sample databases. The features are then modeled by a probabilistic distribution. Finally the models are combined to approximate the description of a combination of instrument sounds. We discuss the requirement induced by the orchestration problem on the choice of the acoustic features and detail the chosen set of features. In a second part we describe the method used to learn the probabilistic model of different instruments and playing techniques and show that it allows to learn the models from few samples and to deduce the models of some playing techniques, pitches or dynamics that are not available in the training set. Finally we evaluate the proposed...