Objective and Subjective Analysis of Audio Source Separation using GAN

The study centers on audio separation and conducts a thorough quality assessment through objective and subjective analysis. A key area of research in source separation is Blind Source Separation (BSS), which aims to estimate source signals from mixed signals with minimal information about the sources or the mixing process. Researchers have primarily utilized various signal processing techniques to address the source separation challenge. Recently, deep learning methods, especially those involving sound and image recognition, have shown significant effectiveness and have been applied in previous studies in this field. These networks employ various mathematical functions to aggregate multiple inputs, resulting in separated outputs. The model in this study was trained using audio samples from the Audio Source Separation Dataset, with training conducted through Generative Adversarial Networks (GANs). The generated estimates were evaluated using various objective analysis methods designed for assessing audio signals. Experimental evaluations demonstrated the effectiveness of the approach through metrics such as Signal-to-Distortion Ratio (SDR), Signal-to-Interference Ratio (SIR), Signal-to-Artifacts Ratio (SAR), and Image-to-Signal Ratio (ISR).