An Instrument Timbre Model for Computer Aided Orchestration

In this paper we propose a generative probabilistic model for instrument timbre dedicated to computer aided orchestration. We define the orchestration problem as the search of instruments sounds combinations that sound close to a given target. A system that addresses this problem must know a large variety of instruments sounds in order to be able to explore the timbre space of an orchestra. The proposed method is based on gaussian mixture modeling of signal descriptors and on a division of the learning problems that allows to learn many different instrument sounds with few training data, and to deduce the models of sounds that are not in the training set but that are known to be possible.