deepGTTM-I&II: Local Boundary and Metrical Structure Analyzer Based on Deep Learning Technique

Abstract

This paper describes an analyzer for detecting local grouping boundaries and generating metrical structures of music pieces based on a generative theory of tonal music (GTTM). Although systems for automatically detecting local grouping boundaries and generating metrical structures, such as the full automatic time-span tree analyzer, have been proposed, musicologists have to correct the boundaries or strong beat positions due to numerous errors. In light of this, we use a deep learning technique for detecting local boundaries and generating metrical structures of music pieces based on a GTTM. Because we only have 300 pieces of music with the local grouping boundaries and metrical structures analyzed by musicologist, directly learning the relationship between the scores and metrical structures is difficult due to the lack of training data. To solve this problem, we propose a multi-task learning analyzer called deepGTM-I&II based on the above deep learning technique to learn the relationship between scores and metrical structures in the following three steps. First, we conduct unsupervised pre-training of a network using 15,000 pieces of music in a non-labeled dataset. After pre-training, the network involves supervised fine-tuning by back propagation from output to input layers using a half-labeled dataset, which consists of 15,000 pieces of music labeled with an automatic analyzer that we previously constructed. Finally, the network involves supervised fine-tuning using a labeled dataset. The experimental results indicate that deepGTTM-I&II outperformed previous analyzers for a GTTM in terms of the F-measure for generating metrical structures.