A comparative study of blind source separation methods

Blind source separation is a popular research topic used for decomposing mixed signals, particularly in the field of music. In addition to exploring machine learning-based approaches, this study aims to examine the performance of classical algorithms in separating audio signal sources. The evaluation of different genres is a significant aspect of this study as the performance of the methods may vary across various musical genres and different audio components. This consideration provides a novel perspective and contributes to a comprehensive analysis of the algorithms. Using the MusDB-HQ dataset, we conducted experimental studies comparing classical algorithms, including FastICA, NMF, and DUET, with implemented architectures such as Hybrid Demucs, Spleeter, Open Unmix, and Wave-U-Net. The audio components were assessed based on several factors, including time, genre, and signal-to-distortion ratio (SDR) scores, after artificially mixing the signals. The results demonstrated the superior performance of machine learning models over classical methods. Specifically, Wave-U-Net achieved the highest SDR scores for drums, other, and mixture components (2.041, -2.087, and 0.941, respectively), while Spleeter showed the highest SDR scores for vocals and bass components (3.145 and 0.066, respectively). Additionally, this study highlights the influence of different genres on algorithm performance, providing valuable insights for music production and related applications. Overall, this study contributes to the existing knowledge in the field of audio source separation by comparing classical algorithms and machine learning models, considering genre variations, and evaluating performance across different audio components. The findings have implications for the development of improved algorithms and their application in various musical genres.