Enhancing Creativity and Validation in Explanatory Deep Learning-Based Symbolic Music Generation: A Hybrid Approach With LSTM and Genetic Algorithms

This research proposes an explanatory deep learning-based music generation approach, where the output of a deep learning model is validated through a set of predefined musical rules, with a refinement process applied when inaccuracies are detected. The study focuses on gamelan, a traditional form of Indonesian music. A Long Short-Term Memory (LSTM) network is used to generate musical compositions, while a modified Genetic Algorithm (GA), omitting the selection and crossover operators, performs validation and, when necessary, refinement via mutation. The LSTM network produces initial compositions, and the GA module ensures compliance with musical rules, enhancing both explainability and creativity. The model successfully generates new bars and lines with notation sequences not found in the original dataset, indicating creative variation. Whether produced directly by the LSTM or refined through GA, the generated output demonstrates the system’s ability to innovate while preserving core musical characteristics. Furthermore, the GA-based validation allows the generated music to be interpreted in terms of the underlying rule constraints. The evaluation using the Pearson’s Correlation Coefficient T-test provides supporting evidence that the proposed automatic music generation (AMG) model is capable of learning and generating gamelan music effectively. The LSTM component, functioning based on its ability to creatively generate note sequences, and the GA component, tasked with validation and refinement, have both proven to collaborate effectively and fulfill their respective roles. These findings support the effectiveness of the proposed model in fostering creative exploration of new tonal patterns aligned with the target genre.