Automatic Orchestration for Automatic Composition

The automatic orchestration problem is that of assigning instruments or sounds to the notes of an unorchestrated score. This is related to, but distinct from, problems of automatic expressive interpretation. A simple algorithm is described that successfully orchestrates scores based on analysis of one musical structure – the “Z-chain.” One way for an autonomous system to compose is to initially write parts for different instruments or sounds. A second strategy (one closer to some human practice) is to write a “flat” score first and then orchestrate it. The latter poses a more general problem, since a system capable of orchestrating its own compositions could also be capable of orchestrating other input, from Beethoven sonatas to Tin Pan Alley tunes to Xenakis. The goal of being able to orchestrate “known” music makes the system more suceptible to analysis and evaluation, and therefore to development. Orchestration is a problem of music generation through analysis, occupying a place on the continuum from expressive interpretation (often framed as timing, articulation and dynamics) through to arrangement and variation-making. This class of problems is different from (and somewhat more constrained than) the goal of generating music from scratch. Since a successful orchestration will be reflective of, or synchronized with, things that are happening in the music, an analytical level is essential. The flip side of the orchestration problem is that if we treat orchestration as a kind of sonified analysis, then not only can analysis help us make orchestrations, but careful listening and re-analysis of orchestrations generated through computational analysis can help us evaluate and refine our analytic methods. Copyright c © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. In this paper we define a narrow version of the orchestration problem that deals with assigning notes in the score to instrumental sound samples. A more complete version of the orchestration problem includes articulations and dynamics, while an extended version could include varying and adding to the score (“arranging”). A more constrained version of the orchestration problem would take into consideration the limitations and affordances of physical instrumental performance. Strategies for Expressive Interpretation and Automatic Orchestration There has not been much previous research in the area of automatic orchestration in the sense of autonomously assigning sounds to a score.1 However, there has been research on expressive interpretation. The goal there is generally either to imitate human performance, to impart a mood as described by an adjective, or both. Like automatic orchestration, automatic interpretation starts with a flat score and modifies it through analysis, adding new layers of information. New layers could take the form of instructions for modifying dynamic levels or tempo, addition of accents, or assignment to instrumental sounds. Approaches to expressive interpretation are generally based on determining sets of rules for assigning dynamic and tempo curves, articulations, and accents to notes. These rules are developed from music theories (Bisesi, Parncutt, and Friberg 2011), inferred through listening consultations with professional musicians (Friberg, Bresin, and Sundberg 2006), or derived through machine-learning analysis of human performances (Widmer 2003). Rules are often based on Research on “automatic orchestration” at IRCAM focusses on analysis of timbre and instrumental capabilities for assisted orchestration, not the automatic assignment of sounds to a score (Carpentier et al. 2012). Musical Metacreation: Papers from the 2012 AIIDE Workshop AAAI Technical Report WS-12-16