Modeling Moods in Violin Performances

In this paper we present a method to model and compare expressivity for different Moods in violin performances. Models are based on analysis of audio and bowing control gestures of real performances and they predict expressive scores from non expressive ones. Audio and control data is captured by means of a violin pickup and a 3D motion tracking system and aligned with the performed score. We make use of machine learning techniques in order to extract expressivity rules from score-performance deviations. The induced rules conform a generative model that can transform an inexpressive score into an expressive one. The paper is structured as follows: First, the procedure of performance data acquisition is introduced, followed by the automatic performance-score alignment method. Then the process of model induction is described, and we conclude with an evaluation based on listening test by using a sample based concatenative synthesizer.