This page is part of the SALIERI webpages.
by Wolfgang Chico Töpfer
Automated music composition is not so new as one might at first think. Various examples can already be found in the 18th century. Back then, it became popular among composers to design so-called Musikalische Würfelspiele. Such a musical dice game is basically an early musical data base made of a collection of measure-sized segments which were categorized according to their musical function and intervallic properties. An n-bar-piece could be generated by retrieving and concatenating n data base segments that keep a fixed functional and intervallic structure despite their seemingly accidental generation process through subsequent dice tossing. Composers learned themselves 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 concatenates segments which were not devised for reconcatenation in a rule- based way, allows making explicit what constitutes musically sensible composition.
David Cope's EMI (or SARA, 1997) may already explain a special sub category of musically sensible composition by the application of certain music models, yielding some impressive results as far as the generation of music - arguably following the style of a certain composer - is concerned. Nevertheless, they fail in giving an empirical proof on the adequacy of the applied models. As to the overall output of these systems, it remains rather unclear how the relation between the stylistically valid and invalid output really is.
Neither does MELONET - a composition system that generates melodic variations, arguably in the style of the chorale variations of J.Pachelbel (University of Karlsruhe, Germany) - offer a proof due to its opaque, connectionistic approach.
In contrast to these and other similar systems, AVA (short for Aleatoric Variation Algorithm) - a database-orientated composition system - fills this gap by offering a variety of implementations of modern music structure models that may be applied alternatively: the GTTM reduction model by Fred Lerdahl and Ray Jackendoff, the Implication-Realization model by Eugene Narmour, the Event model inspired by John Fitch and James Leach (Bath, UK) and David Cope's functional model SPEAC. Each of these models is controlled by switches so that different models may be tested and compared with each other on the same input, a theme that must obey a pre-defined theme phrase structure.
Therefore, AVA allows empirical proofs on the adequacy of music structure models. For instance, an experimental series that resulted in the generation of 48 different tonal variations showed that the SPEAC model ensured a musically sounder functional structure than those variations that do not follow any functional restraints.
Furthermore, different criteria have been developed for the empirical evaluation of the appliable models. The main criterion is called structural quality (SQ) and includes all aspects of the musically sensible placement of database segments. For example, one aspect is that every phrase ought to be finished by a cadential part. Other aspects involved are functional correctness and proper voice-leading.
Moreover, AVA follows an object-orientated design implemented in SALIERI that allows the addition of other music structure models without greater efforts. Future extensions of AVA will also include a full documentation on the origins of the segments that are applied in a variation, permitting a precise analysis of each generated variation.