We present a musical analyzer for a generative theory of tonal music (GTTM) that enables us to output the results obtained from analysis that are the similar to those obtained by musicologists on the basis of deep learning by learning the analysis results obtained by musicologists.

Directly learning the relationship between an input score and output analysis result is impossible. Therefore, we first constructed a deep belief network (DBN) with latent musical knowledge that could output whether each GTTM rule was applicable or not on each note transition by learning the relationship between the scores and positions of applied grouping preference rules with deep learning. After learning all the grouping preference rules, the network underwent supervised fine-tuning by back propagation using the labeled datasets of local grouping boundaries.