The Orchestration project is an integral part of Sample Orchestrator.
It hopes to provide composers with a tool to assist orchestration that is integrated in the software environments created at IRCAM to assist composition, such as OpenMusic. By taking advantage of the knowledge acquired through large databases of instrumental sounds—taking into account the breadth of the possibilities of musical instruments—and the progress in the understanding of pitch, we hope to help composers explore the array of sound possibilities available with an orchestra. The problem consists of finding the combinations of instrumental sounds that are the closest to a target defined either from a recorded sound or from a more abstract process (overlapping pitches, for example) via a search by similarity defined by the user.
The major scientific issues lie, from a signal analysis point of view, in the choice of a group of ‘characteristics’ that can be calculated from samples of instrumental sounds, pertinent in terms of their timbre description, and that make it possible to predict—without having to listen or to analyze—the timbre of an overlapping of sounds. The chosen approach consists of incorporating the knowledge gained from the analysis of large sound databases and a set of "instrument models", making it possible to estimate the probability that a certain sound will be played by an instrument or by a group of instruments.
For the time being, this is limited to static harmonic sounds and to periodic temporal vibrations (such as a tremolo or a vibrato) but in time the application will be improved to the point where will work for all types of sounds.
Another issue concerns the management of the combinatorial analysis adapted to the manipulated data, the complexity of the problem makes it impossible to assess the range of possible solutions. The method currently being explored through an interactive environment lets us offer a group of solutions to a composer, in a reasonable time frame, where certain zones of research could be favored over others that are less likely to lead to interesting solutions.
The search for an orchestration therefore becomes an iterative process, during with pertinent solutions and user preferences surface simultaneously. Technically speaking, the interesting solutions are close to a multi-criteria Pareto distribution optimization problem.