An exploration of musical style from human and connectionist perspectives
Giuseppe Buzzanca, Mario Baroni · Archivio istituzionale della ricerca (Alma Mater Studiorum Università di Bologna) · 2006
present work describes two different approaches to musical style recognition. The first includes a behavioral experiment in which human listeners (divided into musi- cians and nonmusicians) are asked to recognize musical style. The main purpose of this approach is to explore the cognitive processes involved in musical style categorization and the influence of prior musical knowledge in style per- ception. The second approach simulates stylistic recogni- tion through a back propagation neural network model, considering recognition as a supervised learning task. The learner (i.e. the model) is given examples of music in the specified style, together with non-examples, thus learning to distinguish between examples and non-examples. How- ever, since musical style recognition is different from tradi- tional supervised learning tasks (due to the freeform na- ture of the input and the depth of structure), considerable care is required in the design of the model. We have chosen back propagation neural networks for our approach for a number of reasons: they have an excellent track record in complex recognition tasks and are capable of inducing the hidden features of a domain. Our model accounts for vari- ous features that are quite likely to be important in the process of musical style recognition in humans: these fea- tures can roughly be described as the accounting for musi- cal input of varying length in an even manner; modeling the hierarchical nature of musical recognition; capturing the importance of directional flow in music. Our model was trained on a corpus comprising the same compositions as the behavioral experiment (i.e. arias by Legrenzi, by sev- eral other coeval composers Rossi, Stradella, Gabrielli, A. Scarlatti and by the computer program LEGRE). The model reaches a fairly high classification accuracy (vary- ing between 66.7 and 100%, depending on the composers considered). The study of the approaches shows similari- ties (the ability to correctly recognize a style depends on the global exposure to that style both in human listeners and the neural model) but also puzzling discrepancies (for example in the case Legrenzi vs LEGRE, where the accu- racy of the model � 100% � is well beyond that of the hu- mans) which the present work tries to assess.