Automatic Classification of Leading Interactions in a String Quartet
Floriane Dardard, Giorgio Gnecco, Donald Glowinski · ACM Transactions on Interactive Intelligent Systems · 2016
The aim of the present work is to analyze automatically the leading interactions between the musicians of a string quartet, using machine-learning techniques applied to nonverbal features of the musicians’ behavior, which are detected through the help of a motion-capture system. We represent these interactions by a graph of “influence” of the musicians, which displays the relations “is following” and “is not following” with weighted directed arcs. The goal of the machine-learning problem investigated is to assign weights to these arcs in an optimal way. Since only a subset of the available training examples are labeled, a semisupervised support vector machine is used, which is based on a linear kernel to limit its model complexity. Specific potential applications within the field of human-computer interaction are also discussed, such as e-learning, networked music performance, and social active listening.