Industrial Noisy Speech Enhancement Using Joint Time-Frequency Loss Function Based on U-Net

Single-channel speech enhancement research in complex industrial production environments is limited. Current methods, whether based on attention mechanisms or generative adversarial networks, primarily focus on learning speech characteristics in the time domain, neglecting the frequency spectrum. Additionally, existing frequency-domain algorithms lack accuracy in spectral and phase matching of noisy speech, rendering them unsatisfactory for industrial noise environments. To address this issue, this paper proposes the TFU-Net model, a time-frequency joint loss function algorithm based on deep learning and U-Net. It incorporates a combined loss function of Least Absolute Error (LAE) and Mean Square Error (MSE) for speech enhancement in industrial noise environments. Experimental results demonstrate that the frequency-domain loss function, when combined with the time-domain loss function, yields better speech enhancement under industrial environmental noise.