DEVELOPMENT OF BAND-SPLIT RNN AND HYBRID TRANSFORMER DEMUCSFOR MUSIC SOURCE SEPARATION

• Main Author: Kalang Al Qalyubi, Ken
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85018
• Institution: Institut Teknologi Bandung

In recent years, models have been developed in the field of music source separation (MSS). The current state-of-the-art models are Hybrid Transformer Demucs (HT Demucs) and Band-Split RNN (BSRNN). Recent research shows that the pre- trained HT Demucs model can separate six sources (drums, bass, guitar, piano, vocals, and others), tested using the MoisesDB dataset, but scores relatively low on the guitar, piano, and other sources compared to bass, drums, and vocals sources, measured by the utterance-level Signal-to-Distortion (uSDR) metric. However, no research has yet demonstrated the performance of the BSRNN model in separating these six sources. This thesis aims to investigate the performance of the BSRNN and HT Demucs models in separating six sources. For this purpose, BSRNN and HT Demucs models were developed for six-source separation using the MoisesDB dataset. These two models were then evaluated and analyzed to determine the best model for six-source separation. Experimental results show that the HT Demucs model excels in separating all sources compared to the BSRNN model, measured on the uSDR and cSDR metrics with averages of 6.26 dB and 5.88 dB respectively for the HT Demucs model, while the BSRNN model achieved scores of 5.52 dB and 5.38 dB. Additionally, the trained HT Demucs model outperformed the pre-trained HT Demucs model on the piano and other sources by differences of 1 dB and 0.3 dB respectively.