deepGTTM-III: Multi-task Learning with Grouping and Metrical Structures

Abstract

This paper describes an analyzer that simultaneously learns grouping and metrical structures on the basis of the generative theory of tonal music (GTTM) by using a deep learning technique. GTTM is composed of four modules that are in series. GTTM has a feedback loop in which the former module uses the result of the latter module. However, as each module has been independent in previous GTTM analyzers, they did not form a feedback loop. For example, deepGTTM-I and deepGTTM-II independently learn grouping and metrical structures by using a deep learning technique. In light of this, we present deepGTTM-III, which is a new analyzer that includes the concept of feedback that enables simultaneous learning of grouping and metrical structures by integrating both deepGTTM-I and deepGTTM-II networks. The experimental results revealed that deepGTTM-III outperformed deepGTTM-I and had similar performance to deepGTTM-II.