Gamelan Melody Generation Using LSTM Networks Controlled by Composition Meter Rules and Special Notes

Abstract

This study proposes a Gamelan melody generation system based on three characteristics, which are the melodic patterns recognition, composition meter rules that control the duration of notes, and the special notes (pitches) selection which represent ambiguous rules in determining the Gamelan musical mode system. Long-Short Term Memory (LSTM) networks were trained using the sequence prediction technique to generate symbolic based Gamelan melodies. The dataset collected from sheet music was converted into ABC notation format, added with codes representing the composition meter and special notes, and restructured into a character-based representation format. The LSTM network training showed good results in the melodic patterns recognition but the networks take less than 10 attempts for the LSTM network to successfully generate one melody. The evaluation was conducted using experts’ judgment. Three generated melodies were sent to experts to be read, hummed and judged. Overall, the evaluation results showed that the generated melodies can comply with the characteristics of the Gamelan melodic patterns, the composition meter and the special notes.