PENGEMBANGAN APLIKASI PEMISAHAN AUDIO ALAT MUSIK TRADISIONAL INDONESIA DARI REKAMAN LAGU MENGGUNAKAN U-NET

• Main Author: Zuriel, Enrico
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/80726
• Institution: Institut Teknologi Bandung

Music has been in human civilization for ages and has been a way to communicate and convey expressions. In Indonesia, all tribes have their own musical culture. Unfortunately, recent technological advancements have not been utilized optimally to introduce those musical cultures to young students and foreigners in Indonesia. Therefore, technological advancements for preserving and introducing musical culture, including its instruments, is important. One of the technologies that have not been researched as often is audio separation technology for traditional musical instrument sources. Said technology can help musicians incorporate traditional Indonesian sounds to their music and help academics understand musical instruments through recordings without the disturbance of other musical instruments in the recordings. Research regarding musical instrument source separation has been done since 2018, with U-Net being one of the more popular model given its promising Source-to-Distortion ratio when tested. Therefore, this research tests the use of U-Net on traditional musical instruments recordings and integrates it onto a simple web application. The resulting application shall be a proof-of-concept of the feasibility of the technology’s benefits. The developed application is able to separate traditional musical instrument sources from its original recording. The U-Net Architecture uses Short-Time Fourier Transform (STFT) values to determine the underlying musical instrument source and separate it. This research shall determine the optimal optimizer, weight decay, and learning rate parameters for the U-Net model. The model shall then be tested and integrated to the application’s interface. This research uses digitally-generated ensemble recordings of Angklung and Javanese Gamelan, with the Angklung’s audio being the audio that is separated from its recording. Based on comparative experiments of various optimizer, weight decay, and learning rate values, it is concluded that the the most appropriate optimizer is Adam with a weight decay of 10-5 and learning rate of 10-4 which gives an average SDR score of 7.26 dB. The integrated application is also able to properly handle common use cases and can compete with other audio separation tools on the Internet in terms of processing speed.