Page  00000001 AVA An Experimental, Grammar/Case-based Composition System to Variate Music Automatically Through the Generation of Scheme Series Wolfgang Chico-T6pfer Darmstadt University of Technology Wilhelminenstr.7, D-64283 Darmstadt, Germany email: Abstract. AVA1 - a database-orientated composition system centered on variation as the basic composition principle - fills a gap left by other current composition systems: By offering a variety of alternatively selectable implementations of modern music structure models, it allows testing and comparing each of them on the same musical input. Therefore, AVA allows empirical investigation into the adequacy of music structure models. Among the currently implemented ones are the GTTM reduction model, the IR model, an event model, and SPEAC. The applied musical data base is based on Mozart's piano variations. AVA, implemented in SALIERI, supports the addition of new models through its architectural design focussing on modularity and extendability. 1 Introduction Automated music composition already occupied well-known composers from the 18th century. Back then, it became popular to design musical dice games. Such a musical dice game is basically an early musical data base made of a collection of measure-sized segments that are categorized according to their musical function and intervallic properties. An n-bar-piece could be generated by retrieving and combining n data base segments that keep a fixed functional and intervallic structure despite their seemingly random generation process through subsequent dice tossing. Composers learned from designing such games, making use of this knowledge in their own works. However, this knowledge can only be found implicitly in the works of such composers. Developing a database-orientated composition system that combines segments which were not devised for recombination in a rule-based way, allows making explicit what constitutes musically meaningful composition. David Cope's EMI [Cope; 1991] and SARA [Cope; 1996] are examples of such systems. They reveal certain music models whose application led to some impressive results as far as the generation of music arguably following the style of a certain composer is concerned. Nevertheless, they do not provide empirical means to investigate the adequacy of the applied models. As to the overall output of these systems, the stylistically unsatisfactory compositions turn out to be dominating [Cope; 1997]. In contrast to these and other similar systems, AVA recombines by using a variety of implementations of modern music structure models such as the GTTM reduction model [Lerdahl, Jackendoff; 1983], the IR model [Narmour; 1990], the event model inspired by [Fitch, Leach; 1995], a specific music structure model that views music as a series of events, and David Cope's functional model SPEAC [Cope; 1991]. AVA provides means for the selection of different models which may be tested and compared with each other on the same musical input. 1Aleatoric Variation Algorithm

Page  00000002 2 Process Stages and Architectural Design Basically, AVA coilsists of the compolleilts showil withill the cross (Figure 1). Reductional Analysis SCEM User I ~ ~ ~ -4 THEMEll 4-4K17l~l3) 44 PlHRASE~i Analysis Synthesis Figure 1: AVA's modules and their application based on a theme including a description of the theme's reductional structure passed by the user The user starts AVA by ertterirtg a theme as well as a reductiortal theme descriptiort amortg other parameters to specify the variatiort to be gerterated. Figure 1 shows with Mozarts K.265's A part, art example of ETNA, the elaboratiort-orierttated tree rtotatiort for the AVA user [Chico-T6pfer; 1997a] to support her or him ill firtdirtg the most adequate theme reductiort. The bottom lirte shows a represerttatiort of the melody of K.265's A part as a complex: sequence. This is basically a tagged SALIERI sequence [Hoos et~cd.; 1998a] to express phrasal artd orrtamerttal aspects which cart also be expressed ill Extended GUIDO [Hoos; 1998b]. Thert VAR, the mail module, takes corttrol, operatirtg ill two phases: It first artalyzes the theme (horizorttal phase ill Figure 1) artd thert gerterates - based ort the reductiortal theme descriptiort artd the currertt musical data base - a variatiort as specified by the user (vertical phase irt Figure 1). For the artalysis, VAR. makes use of PHRASE, AVA's lirtear2 phrase structure artalyzer. For the syrtthesis, VAR. makes use of CBR, AVA's case-based reasortirtg, to select artd recombirte the segmertts from the musical data base. These segmertts are really musical schemes ie. abstract represerttatiorts of music expressed irt a scheme lartguage for the irtterpretatiort of music (SLIM, see [Chico-Topfer; 1997a]) givert irt cortverttiortal music rtotatiort. Thus, precisely speakirtg, it is more adequate to call AVA's musical data base a musical schemes base (MsB). Its structure is basically made of a collectiort of phrases which are regarded as cases. However, urtlike cortverttiortal CBR systems eg. [Rossell; 1996] a sirtgle case is rtot usually takert as the urtique source. Rather, parts of several cases are used to gerterate a solutiort. AVA's CBR is opert to the user irt a rtumber of ways because its behaviour cart be set by so-called user switches. Such a switch is a procedure to activate or deactivate a certairt part of the CBR behaviour. Two user switch types are supported: ort the orte hartd, there are user switches to select the applicatiort of a certairt music structure model. Each music structure model is hereby cortrected with a corresportdirtg search method that is applied irt the CBR module. Thus to switch to the evertt model eventsearch~on must be called; to switch to the JR model, Ut_search~on is rteeded. Ort the other hartd, there are user switches to set gerteral search properties: * swit ch~discriminat ion (DIS): iff ort, ortly schemes from MSB phrases of equal type are selectable with respect to the theme phrase to be variated; e.g. if the theme phrase is art A phrase, ortly schemes from A MSB phrases are allowed. * switchlaithfulness (PUR): iii ort, schemes are selected from as few MSB phrases as possible; otherwise MSB phrases are left asap, thereby ertsurirtg a maximum usage of all the MSB phrases 2in number of compared measures

Page  00000003 * switchfunctionality (FCT) with switch values 1. s: the SPEAC rules [Cope; 1991] are followed 2. e: the musical functions of the theme and their sequence are exactly followed; a selected scheme must have the same function as the theme at the given point 3. f: only the SPEAC categories of the musical functions of the theme are followed 4. 0: only the most similar function of each scheme is selected with respect to the theme. If there is no similarity, an arbitrary function of the scheme in question is chosen. This allows the application of schemes with seeming functional incompatibility. Note that the last switch sets the functional model e.g. SPEAC which means that actually two music structure models are combined e.g. the event model and SPEAC. The last step in the synthesis phase consists of a call of SCHEME, AVA's scheme series interpreter, to get a corresponding SALIERI/GUIDO segment from the scheme recombination. This can be processed by SALIERI into a corresponding MIDI file if required. 3 Experiments To give an idea of AVA's capabilities I carried out a series of experiments using the event model. This resulted in an evaluation of 48 compositions generated by AVA and summarized in three tables [Chico-T6pfer; 1997a] one of which is reproduced here (see [Chico-T6pfer; 1997b] to listen to the corresponding MIDI files). Each experiment was given marks on a scale from 0 (very low) to 4 points (very high). The experimental frame was defined by a small MSB and the use of the event model. Despite AVA's greater flexibility the variation type was restricted to melody variation. This follows quite naturally from the applied MSB being exclusively made of schemes taken from melody variations. Theme and MSB material was K.573, 265 and 398. To use a form compliant with AVA's underlying phrase model, a fragment of the theme of K.398 was taken. The generated variations were evaluated through intuitive listening with respect to the following three categories: * structural quality (SQ): the most important evaluation category consisting of criteria that judge the musically meaningful placement of schemes and adequate schemes choice; one criterion is the correct ending of a phrase by a cadence. Another is functional correctness depending on the sequence of functions and its musical logic; e.g. a sequence V-IV or v-IV (as in exp39814, exp39815) is musically illogical. Furthermore, single scheme pitches are evaluated according to their fitting the function in use. Correct voice-leading is only checked by listening. * recombination degree (COM): identifies the proportion of segments that stem from a single MSB phrase to those which are combinations made of schemes from several MSB phrases; a lower recombination degree indicates fewer combinations. * recognizability (COG): characterizes how much of a variation is actually recognizable as a theme derivation. This involves the recognition of the melodic and/or harmonic course. The applied event model proved to generate musically and structurally meaningful variations despite the small MSB. Half of all the variations were given high SQ marks, only one variation was juged to be of low SQ. The event model was also confirmed in its phrase-structural aspects. For instance, there is no variation that simply interrupts a phrase (note that K.398 variations are based on a theme fragment). The power of the event model below the phrase-structural level, however, could not be verified clearly. There were both convincing and unconvincing scheme recombinations. This ought to be examined more closely in further and larger scaled experiments.

Page  00000004 FCT PHR DIS COG COM SQ NAME s T T 3 1 4 exp3982 s I T 2 3 3 exp3981 s 1 1 3A 2 exp3983 s 1 1 3A 1 2 exp3984 ~ 1 1 3 1 3 exp3988 ~ 1 T 2 3 2 exp3986 ~ 1 1 2 3 2 exp3987 ~ T 1 3 2 3 exp3985 f T T 3 1 3 exp3989 f I T 2 3 2 exp39810 f 1 1 3 3 2 exp39811 f T 1 3 2 3 exp39812 I T T 3 2 2 exp39813 I I T 2 3 3 exp39814 1 1 1 2 2 2 exp39815 I T I 3A 2 2 exp39816 Table 1: AVA'5 variations of a theme fragment of KC.398, their switch values and evaluation. 4 Future Prospects Due to space corttrairtts this paper cart ortly give a brief impressiort of AVA artd its poterttial. Regardirtg the future, various aspects such as other music structure models or the formalizatiort of evaluatiort criteria have yet to be explored. Moreover, future versiorts of AVA are plartred to support art ilicreasilig flexibility ill the choice of music structure models. Irt additiort, art automated documerttatiort ort the origirts of the applied schemes is plartred ill order to obtairt a history that explairts the elemertts that make each gerterated compositiort. References [Chico-T pfer; 1997a] Wolfgang Chico-T pfer; AVA: Emn experimentelles, grammatik- und fallbasiertes Kompositionssystem in SALIERI zur automatischen Variation von Musik durch Generierung von Schemafolgen; Diploma Thesis, October, 1997; TUD Computer Science Department of Automata Theory and Formal Languages (Germany). [Chico-T pfer; 1997b] Wolfgang Chico-T pfer; AVA: A grammar/case-based composition system to variate music for musicological studies in Salieri. http / / /SALJERJ/ava/ava.html [Cope; 1991] David Cope; Computers and musical style; The Computer Music and Digital Audio Series, Volume 6.A-R Editions; Madison, Wisconsin 1991. [Cope; 1996] David Cope; Experiments in musical intelligence; The Computer Music and Digital Audio Series, Volume 12; A-R Editions; Madison, Wisconsin 1996. [Cope; 1997] David Cope; The Composer's Underscoring Environment: CUE; Computer Music Journal, 2 1:3, pp.20-37, Fall 1997. [Fitch, Leach; 1995] John Fitch, James Leach. Nature, Music and Algorithmic Composition. Computer Music Journal, 19:2, pp.23-33, Summer 1995, Bath UK. [Hoos et~czl.; 1998a] Holger H. Hoos, Jiirgen Kilian, Kai Renz, and Thomas Helbich. SALIERI - a General, Interactive Computer Music System; Proc. ICMC-98. [Hoos et~czl.; 1998b] Holger H. Hoos, Keith A. Hamel, Kai Renz, and Jiirgen Kilian; The GUIDO Notation Format - A Novel Approach for Adequately Representing Score-Level Music; Proc. ICMC-98. [Lerdahl, Jackendoff; 1983] Fred Lerdahl, Ray Jackendoff. A Generative Theory of Tonal Music. MIT Press Series on Cognitive Theory and Mental Representation. J.Bresnan, L.Gleitman, S.J.K~eyser editors. MIT Press, 1983. [Narmour; 1990] Eugene Narmour; The Analysis and Cognition of Basic Melodic Structures: the Implication-Realization model; University of Chicago Press. [Rossell; 1996] Josep Lluis A. Rossell; Saxex: un sistema de Raonament Basat en Casos per a l'expressi6, musical; Universitat Pompeu Fabre - Institut Universitari de l'Audiovisual; Barcelona, Desembre 1996.