U-Net Based Speech Denoising Model for Kiosk Environments

In the context of kiosk systems, denoising is a critical preprocessing step that enhances speech recognition reliability. It is especially relevant for applications like chatbots, as it effectively removes ambient noise and sharpens the speech signal. The neural network model presented in this paper is specifically tailored for this purpose. It incorporates domain-specific noise data and employs TFC-TDF blocks to effectively denoise targeted speech, as depicted in the provided figure. Our experimental results, encompassing various metrics, indicate a significant performance boost attributable to the TDF layer. A comparative analysis of TFC and TFC-TDF models underscores that the observed enhancement in performance stems not merely from an increased parameter count due to the TDF layer but from the inherent structural benefits of the TDF design.