Estimation of time-varying parameters in rule systems for music performance

In studies of music performance, rule systems have been used to model the expressive deviations introduced by musicians. In this work we present a methodology for the estimation of the rule parameters to reproduce a given human performance as closely as possible. To achieve best fit, a least square algorithm was used. The estimation can be carried out on different time scales to improve the fit when the performer used different expressive strategies during the piece. The method has been tested on different pieces, recorded both in a controlled experiment, and also in ecological settings. Results confirm this methodology, and allow an objective comparison of parameter sets already studied in literature. The model can also be used to compare different rule systems, both from a numerical point of view, and also by listening the synthesized performances obtained using the estimated values.