Inferring Efficient Hierarchical Taxonomies for MIR Tasks: Application to Musical Instruments

A number of approaches for automatic audio classification are based on hierarchical taxonomies since it is acknowledged that improved performance can be thereby obtained. In this paper, we propose a new strategy to automatically acquire hierarchical taxonomies, using machine learning methods, which are expected to maximize the performance of subsequent classification. It is shown that the optimal hierarchical taxonomy of musical instruments (in the sense of inter-class distances) does not follow the traditional and more intuitive instrument classification into instrument families.