A Memory-Based Approach to Meter Induction

Meter induction has been an important topic in the computational modeling of music cognition for quite some time now. In this paper, an attempt is made to model how listeners arrive at a metrical interpretation of a fragment of music. A number of existing models are based on the Gestalt principles of perception, ‘simplicity’ or ease of encoding being a key aspect. An alternative to this approach are models based on the notion of ‘likelihood’, so-called memory-based models. We adapt and evaluate a number of memory-based models for parsing metrical structure. More specifically, we will use the models covered by the DataOriented Parsing (DOP) framework. This framework defines a large class of probabilistic grammars that take sub-trees from an annotated corpus to form a general Probabilistic Tree Grammar. The models are tested on the National Anthems collection, yielding encouraging results.