A Deep Learning Loss Function based on Auditory Power Compression for Speech Enhancement

TL;DR

This paper introduces a novel loss function based on auditory power compression for speech enhancement, improving correlation with speech intelligibility metrics and enhancing training effectiveness.

Contribution

The paper proposes a new auditory-inspired loss function that better aligns with speech intelligibility measures and improves speech enhancement training outcomes.

Findings

Higher correlation with objective speech intelligibility scores.

Improved training effectiveness over existing loss functions.

Better overall speech enhancement performance.

Abstract

Deep learning technology has been widely applied to speech enhancement. While testing the effectiveness of various network structures, researchers are also exploring the improvement of the loss function used in network training. Although the existing methods have considered the auditory characteristics of speech or the reasonable expression of signal-to-noise ratio, the correlation with the auditory evaluation score and the applicability of the calculation for gradient optimization still need to be improved. In this paper, a signal-to-noise ratio loss function based on auditory power compression is proposed. The experimental results show that the overall correlation between the proposed function and the indexes of objective speech intelligibility, which is better than other loss functions. For the same speech enhancement model, the training effect of this method is also better than other comparison methods.