On the Use of Absolute Threshold of Hearing-based Loss for Full-band Speech Enhancement

In this paper, we investigate the use of a perceptually motivated loss function for training single-channel full-band speech enhancement models. Specifically, we modify the conventional squared error loss function by incorporating the use of a frequency-importance based weighting scheme utilizing absolute threshold of hearing (ATH). We placed more emphasis on the perceptually relevant frequency bins of the speech spectrogram by applying larger weights to train the speech enhancement model targeting for a higher perceptual quality. We compare the models trained using both the conventional loss and the loss utilizing the proposed ATH-based weighting scheme on the VCTK and 4th DNS challenge datasets. The results demonstrate that the proposed loss using ATH-based weighting scheme achieves better performance than the conventional loss in terms of multiple objective speech quality metrics.