Modeling the Rational Basis of Musical Expression
Gerhard Widmer · Computer Music Journal · 1995
Expressive interpretation and performance of written music is of central interest to music research. As a manifestation of human musical competence, it is a challenge to musicologists and psychologists alike, and as a specifically phenomenon, expressive performance is a permanent topic of debate and music-theory analysis. Not surpris ingly, this interest has also inspired research into formal models that try to identify the regularities and mechanisms underlying musical expression. Music theorists have tried to relate expression patterns to musical structure. In most cases, the sizes and shapes of expressive gestures are assumed to be more or less closely linked to the phrase or grouping structure, at various levels, of the piece being performed. Some formal computational models have also been proposed. For instance, Sundberg, Askenfelt, and Fryden (1983) have presented a system of rules that express some simple hypotheses about local expression patterns The general approach was to derive hypotheses by intuition, and then test them by experimentation (analysis by synthesis). Other authors take a very different approach, relating the subject to extra-musical phenomena, for instance by hypothesizing analogies between expression and general motion in physical space (e.g., Todd 1990). Common to all these models is the assumption that there is some rational basis to expressive interpretation-in other words, contrary to public opinion, expression is not an intangible, inexplicable, artistic phenomenon that comes out of thin air. Instead, it can in part be traced to structural features of the music performed, and to the performer's awareness of these structures. The objective of the research described here is to investigate this hypothesis more thoroughly, with the help of artificial intelligence (AI). Our starting point is the following questions (typical AI questions, if you will): Is there general musical knowledge that enables one to understand and explain expressive variations; and, can this knowledge be made explicit and modeled in a computer program? The research methodology employed in this project is similar to the philosophy followed in Widmer (1992), where a program was developed that could learn to harmonize melodies. First, we will hypothesize what common, general musical knowledge might be relevant to expressive performance and might, at least in part, explain the phenomenon. This knowledge will then be encoded in an explicit formal model, at the appropriate level of abstraction; care must be taken to ensure that the model is as plausible as possible, both musically and psychologically. And finally, the adequacy of this model will be tested empirically by incorporating it into a learning program that uses the knowledge to learn general performance rules from actual performances by human musicians. In the project described in Widmer (1992), this methodology turned out to be quite successful. A simple, abstract model was constructed that roughly describes how listeners perceive harmonized melodies. It was then shown that this model enabled a computer program to learn to harmonize new melodies more effectively. In the case of expressive performance, we start from the hypothesis that one function of expression is to give the listener cues as to the intended structural interpretation of a piece. Given that this hypothesis is correct, certain aspects of expressive performance should become explainable once we have a model of musical structure understanding. Computer Music Journal, 19:2, pp. 76-96, Summer 1995 ? 1995 Massachusetts Institute of Technology.