Redes neurais artificiais para geração de acordes em melodias musicais

• Ano de defesa: 2022
• Autor(a) principal: Costa, Lucas Francesco Piccioni
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação; UTFPR
• Palavras-chave em Português: Teoria musical; Sistemas tutoriais inteligentes; Desempenho - Avaliação; Redes neurais (Computação); Music theory; Intelligent tutoring systems; Performance - Evaluation; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/31456

Harmony can be defined as the art of combining sets of musical sounds in order to frame a melody. Furthermore, as in any form of art, it takes time, dedication, study, and experience to make coherent musical harmonies possible, a complex task for beginners. A system capable of automating this process is of great value in helping harmony and composition students, both beginners and experienced since they can find new ways to build musical structures. The literature presents several solution alternatives, such as computational methods based on rules, evolutionary algorithms, and applying Artificial Neural Networks. Works that used Bidirectional LSTM Networks from symbolic data of standardized melodies achieved considerable results considering the number of possible classes. This work proposes new approaches to the problem of automatic musical harmonization, aiming at results that are more accurate than those found in the literature and diversified through preexisting harmonies, in addition to presenting a new form of musical representation using images and an innovative architecture system called Intelligent Harmonizer System. Thus, a melody described in symbolic format is processed to serve as input to the system, which consists of a Convolutional Neural Network trained based on data from original melodies and harmonies. In this way, there is harmony at the exit within the proposed objectives. The results were analyzed using statistical instruments, which showed reasonable success rates above what is found in the literature, establishing a new milestone. The resulting system presents advantages both for the field of music, useful for musicians of all levels of knowledge and theoretical study of music, and computer science, contributing to the development of musical structures encodings and the construction of innovative neural network architectures.