An Iterative Post-processing Approach for Speech Enhancement

Abstract

Speech enhancement has been widely used in speech recognition, multimedia systems and hearing aids etc. In this study, we explore a new post-processing strategy for speech enhancement. The main goal of proposed post-processing method is to reduce speech distortion and improve speech quality and intelligibility after enhancement. First, a masking-based speech enhancement system based on deep neural network is implemented. Then, an iterative global variance equalization post-processing is proposed to adopt on estimated masks. We evaluate the intelligibility and quality of enhanced speech and observe that the proposed post-processing method achieves higher speech intelligibility and less speech distortion at low signal-to-noise ratios (SNRs) comparing to the baseline system without post-processing or previous post-processing methods. The experiments under unseen noises also show that the proposed post-processing strategy can improve the model generalization at multiple noise types.