AUDIO DENOISING USING U-NET ARCHITECTURE

Abstract

Audio denoising is a pivotal task in audio signal processing. This paper presents a machine learning approach using a U-Net architecture to denoise musical audio signals affected by four distinct noise types: white noise, urban noise, reverberation, and noise cancellation artifacts. The model was evaluated on datasets derived from IRMAS and UrbanSound8K. Objective and subjective evaluation metrics were used, which show the model's effectiveness in filtering white and urban noise. However, performance on reverberation and noise cancellation artifacts is limited, indicating areas for future architectural and methodological improvements.