Musical Meta-Creation: The Discourse of Process

Musical meta-creation is an artistic and technical field in which the results (i.e. music) often cannot be empirically tested. The processes by which music is generated, however, afford a rational level of discourse. This paper suggests ways in which we might orient our discussion and our system-building in order to explore musical possibility and use computational modeling to grow into a new kind of creative musical practice. Evaluating Musical Meta-Creation: Process and Discourse The development of musically creative systems is both an artistic and a technical endeavor. Designers of computational music generating systems may be artists using computation as a private means of creating musical works. Other artists may consider the meta-artwork to be the piece they’re presenting, itself deserving of aesthetic appraisal. Musicologists, cognitive scientists, and mathematicians may develop music generating systems in order to test hypotheses. Artists, scientists, or artist-scientists may want to model a corpus, model a composer’s musical mind, develop a new style, explore the limits of musical listening, or map an open space of musical possibility. Musical meta-creation is a technical, mathematical field, often practiced by scientists, and yet it is invested with an irreducibly informal, subjective criterion at the ouput level. Music can’t be formally, scientifically evaluated.1 But the professional requirement of scientists, as well as the influence of science, currently culturally ascendant, on humanists and artists, brings the desire for empirical validation methods for meta-creative systems. The hypothesis and method of testing depends on the goal of the experiment – and the general problem of create (good) (great) (groovy) (new) music may have no available formal test at all. For the problem of modeling a corpus, however, the “indistinguishability test” is common (Pearce and WigCopyright c 2013, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Measuring effects of music with respect to human test subjects is one workaround which may be satisfactory for some purposes, but nonetheless is not equivalent to a direct evaluation of music. gins 2001), in which listeners are enlisted to decide whether output is indistinguishable from a target corpus.2 Supposing a system passes the indistinguishability test, however, it’s not clear what this (by itself) reveals, since we don’t know just how and at what level the test subjects listen. For instance, (Tillmann and Bigand 1996) found that chopping up works of Bach or Mozart into short segments and reversing the order of the segments has little effect on expressiveness and coherence for non-musicians. And given that many generative systems rely on corpus recombination, it’s possible that a positive indistinguishability test could be interpreted as a confirmation (or elaboration) of these findings – but presented as a generative model of music. And it’s not just recombinative systems that are subject to this danger. For instance, it’s easy to use genetic methods to relatively quickly “evolve” a sentence of Shakespeare – given that the sentence is available as a model! (Dawkins 1996) This is not to say that corpus-bound, evolutionary or recombinative musical systems are no achievement – they may in fact hold extraordinary lessons for us. But those lessons may be subtle. If we want to know just what has been modeled, a careful look at the generative process is necessary. As in Borges’ “Pierre Menard”, the attempt is made to recreate existing work in a new way, at a new historical moment. And as with the few paragraphs of “the Quixote” that Menard managed to recreate, the interest in musical re-creation lies precisely in the means by which it may occur. And it remains that many of us engaged in exploratory artistic and mathematical work don’t have goals that can be verified through indistinguishability. It doesn’t serve us to have an empirical standard in “testing” output. Music, generally speaking, is “evaluated” by critics and audiences, within cultural, social, and political environments, and often with results betrayed by history.3 These It’s also been noted that despite a superficial similarity, this is not a Turing test, e.g. because it doesn’t deal with semantics, general knowledge of the world, etc. (Ariza 2009) (Slonimsky 2000) is an entertaining catalogue of critical assaults on great composers including Beethoven, Chopin, Stravinsky, Tchaikovsky, and Wagner. One might ask “where does music function if not within cultural, social, and political environments” (as one reviewer did). Aside from the fact that cultural and social environments change and therefore evaulation within one or Musical Metacreation: Papers from the 2013 AIIDE Workshop (WS-13-22)