To search, Click below search items. |
| All Published Papers Search Service |
|
|
|
Title |
A Comparative Study of Generative LSTM Models for Multi-Instrumental Music Composition |
|
Author |
Ko Ko Aung, Yasushi Nakabayashi, Ryuji Shioya, and Masato Masuda |
| Citation |
Vol. 25 No. 4 pp. 11-26 |
|
Abstract |
The field of music generation using deep learning has primarily focused on Western instruments supported by the standard MIDI system, limiting research attention on traditional instruments from non-Western cultures. This study addresses this gap by introducing a novel approach to data acquisition and model training for traditional instruments, using Burmese traditional instruments as a case study. By employing sound-font technology, we indirectly convert audio data into MIDI-like symbolic representations, enabling compatibility with standard deep learning workflows.We then develop and evaluate three generative LSTM architectures ? Variational LSTM, Conditional LSTM, and Hierarchical LSTM ? to assess their performance in generating music for these instruments. Comparative evaluation focuses on both objective performance metrics and the adaptability of each architecture to the specific characteristics of traditional music data.This paper contributes to expanding the scope of generative music research, demonstrating how modern deep learning approaches can be adapted to preserve and revitalize musical traditions. The findings highlight the advantages and limitations of each LSTM architecture, offering practical guidance for future researchers working with underrepresented musical forms and non-Western instrumental datasets. |
|
Keywords |
Music Generative Model, LSTM model, Sound font system , Burma Traditional Instrument |
|
URL |
