Feature Discovery for Sequential Prediction of Monophonic Music

Learning a model for sequential prediction of symbolic music remains an open challenge. An important special case is the prediction of pitch sequences based on a corpus of monophonic music. We contribute to this line of research in two respects: (1) Our models improve the stateof-the-art performance. (2) Our method affords learning interpretable models by discovering an explicit set of relevant features. We discover features using the PULSE learning framework, which repetitively suggests new candidate features using a generative operation and selects features while optimizing the underlying model. Defining a domain-specific generative operation allows to combine multiple music-theoretically motivated features in a unified model and to control their interaction on a fine-grained level. We evaluate our models on a set of benchmark corpora of monophonic chorales and folk songs, outperforming previous work. Finally, we discuss the characteristics of the discovered features from a musicological perspective, giving concrete examples.