A study on deep neural network based speech enhancement by combining outlier-robust loss
Published in The Journal of the Acoustical Society of Korea, 44(1), 2025
Mask-based speech enhancement is usually trained with a mean squared error term in the frequency domain. MSE is sensitive to outliers, so a handful of badly-estimated bins can dominate the gradient.
This work replaces that term with two outlier-robust alternatives, Huber loss and Berhu loss, combined with SI-SNR.
Results. On VCTK, both SI-SNR + Huber and SI-SNR + Berhu improve all three metrics – SDR, PESQ and STOI – over the conventional SI-SNR + MSE objective.
Recommended citation: Bongsu Jung and Wooil Kim. (2025). "A study on deep neural network based speech enhancement by combining outlier-robust loss." The Journal of the Acoustical Society of Korea, 44(1), pp. 66–73. doi:10.7776/ASK.2025.44.1.066
Download Paper | Download Bibtex