Vol. 29 Issue 1 Reviews
David Cope: Virtual Bach

Compact disc, 2003, Centaur Records CRC 2619; available from Centaur Records, Inc., 136 St. Joseph Street, Baton Rouge, Louisiana 70802, USA; telephone (+1) 225-336-4877; fax (+1) 225-336-9678; electronic mail info@centaurrecords.com; Web www.centaurrecords.com

Reviewed by Eliot Handelman
Montreal, Quebec, Canada

CD CoverWe seem to accept the achievements of machine composing at an alarmingly uncritical pace. In an article in the EE Times, Colin Johnson writes that “computers can compose a new Bach cantata, but cannot compose anything novel, because their algorithms merely encapsulate a particular style of music” (eetimes.com/at/news/OET20021230S0015). The reference can only be to David Cope, since no one else—as far as I’m aware—has produced work to substantiate the first and last of those claims. Since the mid-1980s, Experiments in Musical Intelligence, Mr. Cope’s program, has been generating Bach-like inventions, Chopin-like nocturnes, Mozart-like symphonies or, as on this CD, Bach-like instrumental concerti and suites, at an unprecedented level of technical accomplishment. There are, in fact, no other contenders for machine-composed music anywhere approaching this level, in consequence of which Mr. Cope’s work has excited enormous admiration and speculation. An article in New Scientist calls it a “requiem for the soul”, and theorists like Leonard Meyer, Fred Lerdahl, Ray Kurzweil, and Douglas Hostadter have expressed interest, approbation, and amazement (arts.ucsc.edu/faculty/cope/biography_page_2.htm).

I seem therefore to be somewhat in the minority by remaining in wonderment as to how any of this stuff is actually generated. The theory is that a few carefully-chosen pieces are automatically analyzed, separated into parts, broken down by commonality, and recombined into new works which, thanks to the preservation of “signature”—phrases and such that recur in a given style—come out sounding as though ghosted by the composer whose works have been sampled. Mr. Cope has written four books that detail the operations, and these have been praised for their lucidity and scope (see, for example, Michael Theodore’s review of The Algorithmic Composer). Yet I find the descriptions, theories, and algorithms offered in these books mystifying. I don’t see how they could not fail to produce an incoherent jumble of bits and pieces of destroyed music. Nor do I see how they can account for the consistently coherent pieces exhibited as the product of a program.

Figure 1

Figure 2Figure 1 shows a few measures near the beginning of the first movement of the Suite for Solo Violoncello (my transcription: I’ve assumed that the performance on the CD uses Baroque “a,” a semitone below A440). Note in the last measure how the “f” descends to “e” before rising to the top note of this section, “g.” Figure 2 shows a parallel passage toward the end of the first section, beginning identically but varied so that the “f” ascends directly to “g.” It’s as though the first phrase holds back from completing the upward scale so that a greater sense of resolution can be created upon its return. Small details of this kind add powerfully to the music’s coherence, and many more could be mentioned. My question is how EMI knows how to do this sort of thing.

In order to try to enlighten myself I studied SARA, the Lisp program that accompanies Virtual Music (Cambridge, Massachusetts: The MIT Press, 2001), which then seemed to be the most extensive and revealing program Mr. Cope has made public. Buried within it are functions like “schenker-plot,” “translate-urlinie,” “layer-analysis,” whose names promise potentially interesting theoretical techniques to handle the kinds of problems that EMI handles so well—ensuring that the music expands its highest register by degrees, creating long-range purposive movement upwards, getting a motive in one phrase to be reflected and developed in another at an appropriate scale degree, etc. In the first movement of the cello suite, the first part rises step by step to “a;” only in the second part does it rise beyond, first to “b” and then to “c.” This ensures a sense of upward development necessary for the perceived continuity of the line. But how does the program know how to do this? This is just where some sort of Schenkerian manipulation would seem to be requisite. But the fancy-sounding function names seem to lead down a garden path, returning a list of descending numbers that would seem to code linear movement, but turn out to code for fragments to be selected in a database that is compiled so as to discourage inspection. There is no other reason to compile raw data with a Lisp compiler. As far as I can tell, no layer-analysis, or indeed musical analysis of any kind, takes place in SARA. If EMI does use a kind of Schenkerian analysis, the equivalent details in SARA are hidden from prying eyes.

Supposedly, the pieces on Virtual Bach are completely determined by works composed by J. S. Bach. Mr. Cope states “The Experiments in Musical Intelligence Bach Suite for solo violoncello (2000) follows the forms of Bach’s six cello suites quite closely” (artsjournal.info/e-zine/musical_notes/cope.html). I fail to see this: for instance I can’t locate anything in the Bach Suites that suggests a recapitulation within a first section of the type shown above, much less one that cadences, midway through the piece, on the tonic. I suspect that a database-specific derivation can’t account for all details of these pieces. This is because of the presence of anomalous passages impossible to derive from the modeled composer’s works. In one such anomaly, a theme in a Chopin-like nocturne—included in the Appendix of Computers and Musical Style (The Computer Music and Digital Audio Series, Volume 6, Madison, Wisconsin: A-R Editions, 1991)—is repeated up an octave, a procedure not found anywhere in the Chopin Nocturnes, and which, it struck me, ran against the fatalistic interplay of free development escaping from within a static (and never transposed) thematic framework, which does seem to me to be a primary affective characteristic of Chopin’s style. But then how does the program know when to perform that spurious transposition? When I questioned Mr. Cope about it he admitted that he needed to supplement the system with rules, despite having denied this in an exchange held through the online discussion list of the Society for Music Theory (“… EMI are data-dependent programs, not rules-based programs …” Message-ID: Pine.NEB.4.44.0207242019340.17600-100000@panix1.panix.com).

But, what kind of rules are they, and how particular are they to a given composition? Even allowing a distinction between rules and a prerogative to edit EMI’s output, I still feel mystified: to perform thematic transposition by rule, must not the program have some model of what a theme is? Consider the problem of getting a program to produce a sequence of the kind in Figure 1. Supposedly, this is based on an underlying model, but I don’t see how this can be done unless the program is able to analyze and label passages in the modeled music as “sequences.” It does seem to me in fact that compositions of EMI’s order of complexity would require an automatic labeling system, surely including boxes labeled “theme 1,” or “parallel,” with slots indicating which parts of such objects are essential for its recognition, and so on. How else to derive the transition network from the models that will determine a composition’s formal unfolding? I haven’t found any information about these matters in Mr. Cope’s writings. It may be supercilious to add that SARA’s coding style is bereft of data abstraction, which would not encourage a belief that EMI uses a much different approach, leaving the whole question of automatic labeling moot. Some other approach would have to account for the way in which the program successfully deals with musical units at the level of theme and motive.

How well do these pieces resemble the composers they are claimed to model? In my opinion, they model them rather poorly. The more I’ve listened or looked the more they strike me as generally superficial imitations which import stylistic and formal techniques alien to, and always simpler than, the style in question. I mentioned in this context the anomalous ritornello in the cello suite, or the octave transposition in the Nocturne. For all I know, EMI’s music might have been written by hand so as to evoke the idea of a computer program that models music, in which the subtle details are stylistically wrong. It might be worthwhile doing some independent analysis to compare, for example, the harmonic content of an EMI-written piece with those which may have served as models. I would not guess that very strong correlation would be found.

Do these experiments help to resolve the problem of style? I hinted at my feeling about this in my remarks about Chopin: a style, for me at least, includes a series of views related to music’s affective and narrative outlook. The style of Beethoven, for instance, is in part about obstacles, ascendancy, freedom, a way of controlling how and when and why a turn of phrase is to sound hopeful or resigned. Mr. Cope’s theorizations dispense with these kinds of meta-level considerations, and yet, I can’t see how to construct a sensible piece of tonal music without at least keeping tabs on an overall sense of something like expressive purpose. I’m willing to accept that I’m wrong in this, and I look forward, in fact, to genuine evidence through which I can disabuse myself of this belief, which greatly complicates for me the whole problem of machine composition.

EMI apparently has a public and a private side. The public side includes programs which, like SARA, are always conceptually simplistic, and which, as a contributor to rec.music.compose recently noticed, don’t at all seem related to the problem of creating style imitations, in which observation I concur. The extensiveness of the private side, which would have to account for the sorts of details I’ve indicated, might suggest, to a skeptic, the possibility that the E in EMI stands for Emperor. I, for one, need much better information as to how these pieces actually were written before I’m willing to accept that they were, in some non-trivial sense, computer-composed. They may have been. But I see no reason to accept this on the basis of the available evidence.

It’s true that Mr. Cope may be caught in the bind of having created an astonishing and unique program whose secrets he doesn’t wish to divulge, for whatever reasons. In that case I would like to suggest that he might, at least, make EMI accessible to his students, or to attendees of his summer workshops, so that the rest of us can compare results and have something to think about.

Claims about what Artificial Intelligence (AI) and computers could do were grossly exaggerated in the 1970s and 1980s, so that programs like RACTER could pass as legitimate expressions of autonomous computational writing without much suspicion, but that time is disappearing: everyone now knows how hard basic AI problems (like spam-detection) really are. And a new generation of Music Information Retrievers is showing how hard it is merely to find beats and modified themes. Computer-oriented music theorists haven’t yet succeeded at dependable automatic harmony analysis; no one has succeeded in writing a program that produces even moderately-successful melodies, and there still isn’t a good counterpoint-generating program around. Our society is both losing its naiveté about computation and, perhaps, becoming a little impatient with the unstoppable proliferation of consumable objects rather than valid, productive, meanings. Inexorably, all music-listeners are becoming composers; we’re beginning to return to a time when the distinction between producer and consumer is merely qualitative. This is why the entire interest in Virtual Bach is in how it is produced: there’s nothing in my feeling about the current state of the world that suggests a need for the sort of music it produces. As a technical achievement, if that’s what it is, I find these works good enough to strain my credulity. I would call upon Mr. Cope to supply (or at least promote) the missing evidence that all is as he claims.

Invited Response from David Cope:

Thank you for the opportunity to respond to this review. I will try to make my comments as brief as possible here, because I wish to spend my time more productively than responding to such off-the-wall allegations.

Mr. Handelman's review is inaccurate in many regards. For example, the SARA program does not accompany Virtual Music, but rather accompanies my book Experiments in Musical Intelligence (Madison, Wisconsin: A-R Editions, 1996). Mr. Handelman also states that SARA is the "the most extensive and revealing program Mr. Cope has made public." Clearly, to anyone who has read my books, Alice (the program that accompanies my book The Algorithmic Composer) is far more extensive than SARA.

Mr. Handelman states: "I find the descriptions, theories, and algorithms offered in these books mystifying" and that he can't "see how they can account for the consistently coherent pieces exhibited as the product of a program" and doesn't "see how they could not fail to produce an incoherent jumble of bits and pieces of destroyed music." Yet, in Virtual Music I give a measure-by-measure account of how a Mozart sonata movement was composed using exactly the processes which Mr. Handelman characterizes as mystifying (see particularly pp. 141-175). Mr. Handelman also states (later in his review) that "there still isn't a good counterpoint-generating program around." I disagree again. Bill Schottstaedt's 1989 "Automatic Counterpoint" in Current Directions in Computer Music Research (Max Mathews and John Pierce, eds., Cambridge, Massachusetts: The MIT Press, pp. 199-214) is, in my opinion, an excellent counterpoint program which has been around for 15 years. David Jones's counterpoint program (discussed in his article "Composer's Assistant for Atonal Counterpoint" in Computer Music Journal 24/4, Winter 2000, pp. 33-43), falls in the same class of excellence. I have made my own Gradus counterpoint program available through my Web site in relation to an article in CMJ 28/3 (Summer 2004, pp. 12-27).

At a number of other points in Mr. Handelman's review, he makes statements which are simply not correct. For example, he argues that one of my functions "would seem to code linear movement, but turn out to code for fragments to be selected in a database that is compiled so as to discourage inspection." The databases for SARA on the CD-ROM that accompanies my book Experiments in Musical Intelligence come in two sets: one set compiled for speed of loading (when SARA was completed in the early 1990s, speed was still a critical issue for many desktop computers), and a second set of non-compiled databases appearing right next to the compiled databases which can be read from Lisp, or for that matter from any application that can read text files, without problems. Mr. Handelman's use of the phrase "discourage inspection" has me baffled.

Mr. Handelman's comments concerning the computer-composed Bach-style cello suite are quite interesting (e.g., "It's as though the first phrase holds back from completing the upward scale so that a greater sense of resolution can be created upon its return. Small details of this kind add powerfully to the music's coherence, and many more could be mentioned."). In fact, I can find just as many or more stylistically wrong or unmusical turns of phrases in these works (as he himself does later in his review). I have written about such successes and failures extensively in all of my books on Experiments in Musical Intelligence and, in lieu of covering them again here, point the reader to any of these books as reference.

(I should note here in response to Mr. Handelman's comment regarding an Experiments in Musical Intelligence's Chopin-style nocturne that the theme "is repeated up an octave, a procedure not found anywhere in the Chopin Nocturnes." However, there are octave transpositions of musical ideas in Chopin's Nocturnes: see, for example, mm. 51-2 and 53-4 of Nocturne, Op. 48, No. 2, and a more spectacular example in this same nocturne in mm. 119-121 and 123-145. Octave transpositions also appear in Op. 27, No. 1, mm. 29-32 and mm. 37-40, for example. But, I agree that these are rarities, and that most likely the Experiments in Musical Intelligence nocturne's transposition was the result of a variation rule, rather than an inherited characteristic.)

Mr. Handelman's comment that "As far as I can tell, no layer-analysis, or indeed musical analysis of any kind, takes place in SARA" dumbfounds me. The code is riddled with music analysis functions. In fact, my book Experiments in Musical Intelligence, which introduces SARA (an acronym for Simple Analytical Recombinance Algorithm, by the way), contains an entire chapter (chapter 2—26 pages) called "The Analysis Component," which is followed by yet another chapter (chapter 3, "The Pattern-Matching Component," 44 pages). Both of these chapters, and others, contain actual code examples from SARA that clearly involve musical analysis. He later adds: "models that will determine a composition's formal unfolding? I haven't found any information about these matters in Mr. Cope's writings." Again, I point to chapters 4 and 5 of Experiments in Musical Intelligence in which I write extensively on these issues. I also discuss them in The Algorithmic Composer (see particularly, chapter 5, “Structure and Coherence”).

Mr. Handelman claims that I "might, at least, make EMI accessible to his students, or to attendees of his summer workshops, so that the rest of us can compare results and have something to think about." I often discuss, demonstrate, and provide sample code to relevant classes. It is very true that these programs (even my version, for that matter) do not produce perfect specimens each time they are run; however, I have been very clear about this in my public and private discussions of my work. I have also distributed more code over the years relating to my work than anyone else I know.

SARA is modeled on Experiments in Musical Intelligence, with every aspect of its basic principles represented. Contrary to what Mr. Handelman states, SARA, when used properly, can produce some very sophisticated results. For example, many of the musical examples in my book Experiments in Musical Intelligence were composed using SARA. As well, a number of smaller works by Experiments in Musical Intelligence were created by SARA (I consider Experiments in Musical Intelligence to be a general umbrella for all of my computer programs). I will be the first to say, however, that SARA is not a full implementation of Experiments in Musical Intelligence—many of the functions are abbreviated for clarity and demonstration purposes. As for the functions which Mr. Handelman describes, they are explained in the book (Experiments in Musical Intelligence) and I will leave my remarks at that.

In his comment "For all I know, EMI's music might have been written by hand so as to evoke the idea of a computer program that models music," Mr. Handelman seems to be suggesting that I doctor the output or have actually composed the 1,000-plus works (including three grand operas, three symphonies, and three piano concerti) that Experiments in Musical Intelligence has produced, while somehow writing four (soon to be five) books, maintaining an exhaustive teaching schedule, and composing many works in my own style. While I appreciate the compliment that this implies, I can assure those reading this that I am not up to this feat. Mr. Handelman adds later on that "EMI apparently has a public and a private side." My writings have revealed all of the basic components of the Experiments in Musical Intelligence. My available code (accompanying books, articles, Web site, etc.) has ample examples of each of these components. My hope has been (and continues to be) that rather than debating whether or not this or that composition had this or that amount of human intervention or why I should or should not reveal some magic bullet, those who wish to know more about my work would study the principles I have presented in my publications, and create their own computer-composing programs. This, it seems to me, is a far more productive way to proceed. I am convinced that there are many individuals out there who are more than capable of producing programs whose music will outclass Experiments in Musical Intelligence.

Mr. Handelman states that "There are, in fact, no other contenders for machine-composed music anywhere approaching this level…" I disagree. For example, the Dominik Hörnel and Wolfram Menzel (1998) article, "Learning Musical Structure and Style Using Neural Networks" in Computer Music Journal 22/4 (Winter 1998), pp. 44-52, includes several extraordinary examples of their machine-composed music of which I am a devout admirer. Likewise, Shlomo Dubnov and Gérard Assayag in a 2002 article, "Universal Prediction Applied to Music Generation with Style" in Proceedings of the Fourth Diderot Mathematical Forum, (Springer-Verlag), provide some quite remarkable output. There are other examples. Mr. Handelman might do well to search the available literature before making such assertions.

Mr. Handelman quotes me in his review as stating that my programs "are data-dependent programs, not rules-based programs." I try my best to make my programs as solely data-dependent as possible. However, it is impossible for any program, in my estimation, to fully achieve this goal. Therefore, I admit absolutely that my programs contain some rules (a simple look at my available code will convince anyone of that).

As every programmer knows, producing publishable code is exhausting and requires an immense amount of time. Code must be carefully documented and programs should be perfected to the point of not requiring user revisions. I have spent years making sure the code that I publish won't fail under any conditions and is thoroughly documented. Such programs (and Experiments in Musical Intelligence is several programs, not just one) also need extensive user manuals. Experiments in Musical Intelligence would require several years of my life to fully put in this condition and I, for one, continue to develop my programs in lieu of such expenditures of time. I release as much code as I possibly can under these circumstances (as my soon-to-be published book will verify—I will be releasing 21 separate self-authored programs in my forthcoming book, Computer Models of Musical Creativity).

Finally, Mr. Handelman concludes his review with the comment "I would call upon Mr. Cope to supply (or at least promote) the missing evidence that all is as he claims." I am tempted to promote my newest book, which will be released soon, Computer Models of Musical Creativity, by stating that it contains the ultimate missing treasure I have withheld in my previous volumes. In many ways, this book does contain details not expressed previously (though these are due to the fact that I continue to revise my programs and these details have surfaced only recently); however, there is no magic bullet in this book either. Any success my work has achieved (and, believe me, the negative reviews of my work far outnumber the positive reviews) has been as a result of continuous hard work over nearly 25 years of research.

I wish to thank Mr. Handelman for contributing to the controversy about Experiments in Musical Intelligence. I suspect that many individuals will, as a result, want to take a closer look at its output (available at www.spectrumpress.com) and read about my work, and for this I am most grateful.


Sincerely,


Dave Cope, Professor of Music


All Experiments in Musical Intelligence scores are now available from www.spectrumpress.com. Information on the third Workshop in Algorithmic Computer Music (WACM) is available from summer.ucsc.edu/wacm. Mr. Cope’s Web site can be found at arts.ucsc.edu/faculty/cope.