Generating polyphonic musical content with deep generative models

In this thesis, the generation of symbolic polyphonic musical content using deep generative networks has been implemented within the knowledge domain of Western Classical Tonal Music. The production of polyphonic musical content has been examined based on two sub-problems: the generation of monophonic melodies and the generation of monophonic accompaniment. In the study, which uses symbolic sequential data representation, efforts have been made to overcome the problem of preserving long-term relational structure, one of the most significant technical challenges in the field of artificial intelligence-driven melody generation. Consequently, the methods (MTHA-LSTM) and (GAMA-DL) have been developed. MTHA-LSTM is a two-layered hierarchical deep learning method that employs multi-headed attention networks (MTHA) in conjunction with long short-term memory (LSTM) networks. GAMA-DL refers to the general framework of a deep learning method that operates in two stages: high-level musical measure generation and low-level note generation. The developed methods have been tested through quantitative measurement methods and user studies. The obtained results demonstrate the effectiveness of the proposed methods in producing longer and higher-quality musical content by overcoming the long-term dependency problem.