An artificial neural network model for analysis and synthesis of pianists’ performance styles
Roberto Bresin · The Journal of the Acoustical Society of America · 1999
A model for synthesis and analysis of professional pianists’ performance styles by means of artificial neural networks (ANNs) is presented. The style of pianists was synthesized using Bruno Repp’s measurements [J. Acoust. Soc. Am. 92, 2546–2568 (1992)] of Schumann’s Trumerei as performed by 24 famous pianists. It will be shown that, by using significant score parameters, ANN can learn and reproduce the style of a pianist at a local level. The behavior of the trained ANN and the performance characteristics of the pianists were analyzed by running the trained ANN in a special way. All ANN input neurones were locked to fixed values, except one or two neurones whose input values were allowed to change in time. For example, if one of these neurones is related to the duration of the current note, the output of the ANN gives a curve showing expressive deviations produced by the ANN, when the duration of the note is varying. In this way it is possible to extract behavior rules of performing action of the ANN, and compare them with the performance style of pianists as well as of rule-based performance systems. The model is developed in the Java language and is available on the Internet.