Current Research in Computer-Generated Music

1 JECT OVERVIEWS Current Research in Computer-Generated Music The six overviews that follow reflect varied ongoing research. Reporting from such diverse locales as Singapore, Europe, and the US, the authors explore the spheres of computer-aided composition, synthesis of musical scores, computer simulation, and composing by musical analog. Algorithms for Musical Composition: A Question of Granularity Stephen W. Smoliar Institute of Systems Science, National University of Singapore Neng Mui Keng Terrace, Kent Ridge, Singapore 051 I omputer programs that “com- the most frustrating applications of com- mathematical models as alternatives to pose” music have been around puter composition in the “Illiac Suite.” Markov processes. The spectral densi- almost as long as the computers The appeal lay in the belief that knowl- ties of a wide variety of physical quanti- thatc run them. Initial approaches se- edge of past events could serve as a ties tend to vary as the inverse of the lected notes for a composition using a valuable predictor for subsequent frequency: l(f. Voss and Clarke demon- random process with constraints to elim- events. The frustration lay in the dis- strated such a correlation using the au- inate or modify undesirable choices. The covery that there was either too much dio signal of Bach’s first Brandenburg computer composition “Illiac Suite for or too little of this knowledge. If the Concerto and then detected the same String Quartet,” by Lejaren HiIler and number of past events considered was correlation in signals from classical music Leonard Isaacson, demonstrates the too small, the result sounded as random radio stations.’ Voss and Clarke con- range of these approaches. as if the notes had simply been selected ducted similar studies of the music of Experiments with Markov processes from an arbitrary set of weighted prob- other composers, including Scott Joplin were probably the most appealing and abilities without any prior knowledge. and Milton Babbitt. Observing this dis- If the number was too large, what was tribution in many different sources of “predicted” was nothing more than a music. they reasoned that they could replication of the original data from synthesize new music by a random pro- Audio examples which the probabilities were computed. cess based on the same distribution. There seemed to be no middle ground Figure 1 shows an example. Examples of the music dis- between the absolute predictability of Eric Iverson has explored another cussed in these project overviews are available on CD and audiocas- duplication and the absolute unpredict- approach.* He applies principles of sette. Please see the order form ability of random noise. autocatalysis, a chemical process that on page 9. The quest for an appropriate middle enables the reproduction of molecular ground led to an investigation of new strings, to the synthesis of “strings” of 54 001%Ylh2/91/0700-005J$01.00@ 1991 IEEE COMPUTER music-either pitches or intervals. Fig- ure 2 shows an example of his results. The variety in approaches to algorithmic composition is impressive, but the results are disappointing. Do these results arise from a critical commonality in these many experiments? I believe the commonality is the assumption that musical composition requires decisions about the placement of notes. This assumption is natural enough. It has its roots in a pedagogic tradition in which the composition student is concerned with manipulating notes on music paper. However, music need not be notated to be composed. A composer can compose while playing an instru- ment. This is generally known as improvisation. Anyone who has heard good jazz improvisation can appreciate how unlikely it is that musicians who impro- vise read from some notation they write in their head. Thus, if we want a founda- Figure 1. Music synthesized by a random process based on a L’fdistribution. tion for algorithmic composition, the (Reproduced with permission.*) materials provided by the pedagogy of musical composition may be far less valuable than those of the actual prac- with the Dice Composer was to make resolved: These examples are models. tice of making music. In this article I sure that the terminals chosen for any In musical composition, such models argue that making music is concerned given measure were interchangeable. include specific compositions or excerpts with a higher level of granularity than Thus, by raising the level of granularity from compositions. David Cope is a good that of the notes on music paper. to entire measures, Mozart could use example of a working composer who chance techniques to produce accept- has managed to put a model-based ap- Raising the level of granularity. Mu- able music in the classical genre in a proach into practice. (See Cope’s arti- sical composition is not necessarily a manner never duplicated by any algo- cle beginning on page 22 of this issue.) matter of putting notes together in the rithmic technique concerned with se- right way. Long before there was jazz, lecting individual notes. An alternative agenda. When we talk people were making music by working If raising the level of granularity in- informally about musical composition, with a higher level of granularity. creases the power of random processes we rarely talk about individual notes. Mozart’s Dice Composer, for example, applied to musical composition, per- Composition is a matter ofworking with generated (to use the terminology of haps we are thinking about the wrong “musical ideas.” We use this phrase in- computational linguistics) random sen- things when we focus our attention on tuitively, without pinning down just what tences by selecting productions from a the selection of individual notes. Work it denotes. Nevertheless, following the simple context-free grammar. Each of in artificial intelligence shows that such lead of model-based reasoning, we can the terminal symbols in this casewas an low-level decisions may actually be sub- assume that musical ideas include our entire measure of keyboard music. Each ordinate to a model-based control struc- memories of past listening experiences. measure of the score, in turn, corre- ture. The underlying premise of such To some extent all composers pick up spondedto 11productions, eachof which control is that the knowledge sources materials from past experiencesand find filled in that measure with one of those used to make decisions include actual new ways to assemble them. terminals. The bulk of Mozart’s work examples of how problems have been What do such intuitions tell us about Figure 2. Music synthesized by applying principles of autocatalysis to the selection of pitches and intervals. (Reproduced with permission. *) July 1991 55 PROJECT OVERVIEWS algorithmic composition? The primary music powerful enough to compel us to look where the ordinary is practiced to lesson is that our agenda for algorith- sit still and listen - even one of Bach’s find a foundation for studies of the ex- mic composition is off the mark. If we two-part inventions or Domenico Scar- traordinary. n are determined to seek out algorithmic latti’s many sonatas - is probably too rules, then our search should be direct- sophisticated for the analysis required ed by two questions: in algorithmic composition. Finding much past evidence of the ordinary in References l How do we identify units of mate- these extraordinary pieces may be diffi- rial of the appropriate granularity? cult simply because the ordinary rarely 1. R.F. Voss and J. Clarke, “‘l/f noise’ in l Given a collection of those units, survives the eroding forces of history. Music: Music From l/fNoise,“J. Acous- how do we properly assemble them? However, the ordinary is very much tical Sot. America, Vol. 63, No. 1, Jan. with us in the present. We tend to ig- 1978, pp. 258-263. What made Mozart’s Dice Composer a nore the routine work of composers 2. E. Iverson and R. Hartley, “Metaboliz- clever piece of work was his recognition obliged to provide themes and jingles ing Music,” Proc. Int’l Computer Music that he could think about simple dances for commercial applications. These com- ConF, Computer Music Assoc., San Fran- (following the style of the dance move- posers must create on demand, proba- cisco, 1990, pp. 298-301. ments of Bach’s keyboard suites) in bly to the extent that none of their tasks terms of individual measures. He could receive extensive cognitive effort. Here Stephen W. Smoliar is investigating several then ask how any one measure could be is where we may find artisans with a aspects of music cognition and perception in varied so that it would always “fit” in clear understanding of the units of ma- addition to his artificial intelligence research the context of preceding and succeed- terial with which they work. We should at the National University of Singapore. ing measures. He gave these variations enough flexibility that any alternative for the first measure could be followed by any alternative for the second measure, and so on for the duration of the entire composition. Scoresynth: A System for the Perhaps Mozart’s exercise was a par- ody of the rather routine approach to Synthesis of Music Scores Based on composition he observed in others around him. For Mozart, the model be- Petri Nets and a Music Algebra came music only after he suitably per- turbed it, but he could not address questions of perturbation until he had a model Goffredo Haus and Albert0 Sametti in place. The minuet movements that Laboratorio di Informatica Musicale, Dipartimento di Scienze dell’Informazione, Mozart wrote for his more memorable Universitri degli Studi di Milano, via Moretto da Brescia, 9, I-20133 Milano, Italy orchestral and chamber music could not readily conform to this dice model, since the model had no good way of imple- menting material that returns in a simi- usic structures can be de- traditional and nontraditional music lar, but not identical, form.

Load more