Page  188 ï~~UNDERSTANDING MUSIC WITH Al Otto E. Laske New England Computer Arts Association, Inc. Needham, MA 02192, USA Copyright Â~ Otto E. Laske 1991 Abstract An overview of the book UNDERSTANDING MUSIC WITH Al, co-edited by Mira Balaban, Kemal Ebcioglu, and Otto Laske at the AAAI Press, Menlo Park, CA, USA, 1991. Keywords: Al and Music, Cognitive Musicology 1. Definition of Al and Music For the purposes of this book preview, let me adopt the definition of Al and Music as a discipline whose goal it is "to build a composer, or perhaps a listener" (Minsky, Foreword). This definition implies two main points, viz., (1) that the central topic of such research is musical activity, thus music as an action, not just as an understanding, and that (2) the main tool for carrying out the research is the computer program. While the practicing musician, to whom this preview is addressed, is most likely going to be comfortable with this orientation, nevertheless the question will arise: what is such research good for in light of the purpose of making music, or teaching music to others. To answer this question in more detail, let me first familiarize you with the contents and organization of the book in question. 2. Contents and Scope of the Book UNDERSTANDING MUSIC WITH Al (The AAAI Press, in conjunction with The M.I.T. Press, 1991) is a collection of research papers presented at 1988 and 1989 workshops on Al and Music, both in the U.S. and in Europe. The editors of the volume wanted to present a collection of readings in Al and Music that would represent the cutting edge of present-day research in this young field. The book opens with a Foreword in the form of a conversation between Marvin Minsky and Otto Laske about music and Al, and a Preface by the editors. These introductory texts are followed by groupings of papers centered around the following seven topics: 1. Two Views on Cognitive Musicology 2. General Aspects of Modeling Musical Activities 3. Composition 4. Analysis 5. Performance 6. Perception 7. Learning and Tutoring. The book addresses itself to four different academic communities: A. the professional musician, especially the musician working with computers B. the professional music technologist and designer of music systems C. the professional Al researcher D. the professional cognitive scientist and cognitive psychologist. In order to reach these different audiences, technical jargon has been kept to a minimum, and actual programming code is nearly absent. The book is available near the end of 1991. In this outline, I am primarily speaking to the first two constituencies. ICMC 188

Page  189 ï~~3. Why Bother? A practicing musician or teacher of music is not ordinarily interested in "building a composer, or perhaps a listener," using computer programs. Rather, he is interested in developing and using tools that will aid him in solving everyday musical problems. Moreover, the universe of discourse in Al often seems arcane and not always straightforward to musicians, being derived as it is from the discussion of problems centering on knowledge rather than action, on mental representations rather than on sound objects giving rise to mental representations. So, why should a musician bother to acquire and read the book I am previewing here, especially since the book does not contain a tutorial on Al? (As an AI tutorial embedded within a cognitive science perspective, I recommend: S tillings,1987). For a composer, performer, theorist, historian, teacher or tool designer what could be of interest in this collection of readings? The main thing of interest to a practicing musician might be that, in this book, the popular assumption that one already knows what music is (thus being oriented to past music) is suspended in favor of an empirical inquiry into specific musical tasks, and into the capabilities underlying the performance of such tasks. This viewpoint opens up a perspective toward "possible musics," or musics in progress that is a familiar notion to professional musicians. While the book also comprises analyses of past musics (especially in the section on music analysis), the emphasis of most of the chapters is on the problem of how one might produce music of a certain kind, and what capabilities, rules, and control structures it might take to do so. As well, by perusing papers on processes involved in making music, the reader will naturally absorb a lot of vocabulary that will stand him in good stead when delving into the Al and Music literature more deeply. Lastly, the reader might be able to apply constructs and schemata learned from this book in his own musical work, or the design and implementation of a particular music system. 4. The Individual Sections 4.1 Two Views on Cognitive Musicology AI and Music might be viewed as an approach to treating musicology as a part of cognitive science, since it is the thinking and problem-solving processes in music that are in the center of attention. 0. Laske and P. Kugel address two different aspects of such an approach. 0. Laske conceives of cognitive musicology as an action science, meaning a science that "creates communities of inquiry in communities of social practice," and for this reason sees the discipline as an instrument of social change, not just a theoretical endeavor. By contrast, P. Kugel is more concerned with the difference between problem-solving and creativity. Making a distinction between ordinary (result-yielding) and limiting computations (whose definitive result can not be ascertained), Kugel suggests that musical competence is not entirely computable, but nevertheless should be turned into a 'technique' as much as possible. 4.2 Modeling of Musical Activity Modeling musical activities by computer is difficult because musical objects, in a way comparable to poetic forms, have no pre-assigned connotation, and thus lack a semantic dimension that is independent of music's intended usage, history, or producing agent. For this reason, no direct semantics can be provided; there are as many possible conceptiOns of the nature of a 'musical object' or 'musical primitive' as there are belief systems. The papers in this section address issues of modeling time, hierarchy, and types in music. S. Smoliar's view is that the primitives of a music system ought to be the mental states of some associated musical agent. By contrast, the other contributors to the section (B. Bel, E.B. Blevis et al., M. Balaban, and F. Courtot) assume that musical primitives are ICMC 189

Page  190 ï~~ultimately sounding objects, however abstractly they may be specified (as, e.g., in Balaban's contribution, where the exact nature of a musical primitive is left open). 4.3 Composition Composition theory is a new topic evolving in Al and Music. While music analysts, music theorists, and musicologists continue to view composition as a process reducible to those of an idealized listener (thereby merging composer, listener, and analyst), composition theory is geared to understanding the control structure of compositional processes as processes sui generis. (This means, in particular, to understand the knowledge base required by such processes, and their relationship to the musical results they produce.) Light is shed on this problem area by C. Ames & M. Domino, R. D. Riecken, S. C. Marsella & C. F. Schmidt, and O. Laske. While Ames et al. discusses a system for generating music in jazz, rock, and ragtime styles, Riecken explores the notion of musical constructs having a specific 'emoting potential.' Marsella et al. investigate the problem reduction approach as a tool for explicating compositional control structures (gleaned from notated works), while O. Laske presents an overview of two systems for knowledge acquisition in music that bypass verbalization, being directly geared to documenting musical action. 4.4 Music Analysis In most cases, the process by which music is analyzed is as much of an unknown as the music under scrutiny. Music analysts typically have a remarkable disinclination toward analyzing their own processes (see, e.g., Dalmonte, R. and M. Baroni 1991). This state of affairs is beginning to change. Contributors to this section, employing synthesis by analysis, test music-analytical concepts as a basis for a computational synthesis of specific types of music, from jazz to classical. K. Ebcioglu builds a sophisticated rule system for modeling four-part chorale harmonization (inventing a new programming language along the way), while J. Maxwell creates an algorithmic model of Bach's French Suite movements, and D. Cope describes experiments geared to producing music based on the notion of augmented transition networks, taken from linguistics. What emerges is the insight that music analysis is a peculiar discovery process that, to a high degree, can be construed as rule-based. 4.5 Performance Aside from the work by Sundberg, real-time performance of music is still a rather neglected research topic. An interesting contribution is made in this section by B. Bel et J. Kippen, both 'ethnomusicologists' researching Indian drumming. They show that a grammar can be used to render the verbally labeled drumming sequences of an expert, and simultaneously employ him for 'debugging' the proposed grammar. In a very different approach, S. Ohteru et al. deal with musical performance based primarily on vision, artificially supplied by a TV camera. 4.6 Perception Perception and Listening diverge in the sense that the latter term is far more complex, implying processes of problem solving, analogical and metaphorical reasoning, story understanding, and emotional mapping (if not others). Therefore, the linkage of perception and listening is not straight-forward. In this section, contributions by C. Linster, B. 0. Miller et al., and P. Desain et al. deal with strictly perceptual problems, such as how a machine might perceive musical rhythm and meter, and how one might quantize the continuous flow of music utilizing traditional and connectionist approaches. ICMC 190

Page  191 ï~~4.7 Learning and Tutoring Even less researched than listening is the topic of musical learning and tutoring, despite much work done in music education. Both M. J. Baker and G. Widmer take on this topic, but in very different ways. Baker suggests an architecture for a system that guides novices through a musical task by way of generating expert accounts of its associated problems, followed by a negotiation dialog with the student. Widmer applies 'explanationbased' learning techniques to the problem of writing two part species counterpoint, enabling his system to infer new rules by way of generalization. 5. Conclusion Clearly, our book opens only a tiny window into the new realm of Al and Music. But equally clearly, musicians of all creeds will find in this volume many fruitful suggestions for their practical and theoretical work. Without an increasingly stronger collaboration between the communities of musical practice and Al research, Al and Music is going to be a still-born infant with no future to speak of. We hope that this book will convince musicians that they are the experts this field is looking for and, equally, the experts this book is about. References Balaban, M. et al. 1991. Understanding Music with Al. Menlo Park, CA: The AAAI Press. Brachman R. J. et al. 1985. Readings in Knowledge Representation. Los Altos, CA: Morgan Kaufmann. R. Dalmonte, M. Baroni (eds.) 1991. Proceedings, 2nd European Conference on Music Analysis. Trento, Italy. M. W. Firebaugh 1988. Artificial Intelligence: A Knowledge-Based Approach. Boston, MA: Boyd & Fraser Publishing Co. Gardner, H. 1985. The Mind's New Science. New York: Basic Books. Stillings, N. A. et al. 1987. Cognitive Science: An Introduction. Cambridge, MA: The M.I.T. Press. (A Bradford Book) ICMC 191