Some Investigations on Deep Learning based Single Channel Speech Enhancement
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Speech is the primary means of communication among humans and a natural interface
newlinefor human-machine interaction. However, the presence of ambient noise in real-life
newlinesituations poses challenges for both humans and machines. The objective of speech
newlineenhancement is to eliminate or reduce the impact of background noise on degraded
newlinespeech.
newlineIn this study, we investigated some of the speech enhancement models based on
newlineNon-Negative Matrix Factorization (NMF) and Deep Neural Network (DNN).
newlineWe initiate our research by investigating speech enhancement techniques based on
newlineNMF. We proposed a regularized NMF with an adaptive wiener filter method for speech
newlineenhancement, where prior distributions are used for transformed domain magnitudes of
newlinespeech and noise spectral components to implement an iterative posterior NMF model.
newlineAn adaptive factor and#945; is introduced in the wiener filtering approach to adjust the weights between noise levels and estimated speech based on signal-to-noise level for the gain function, which helps to further enhance the speech quality.In contrast to the statistical and templet-based methods, the combination of NMF (proposed iterative posterior NMF) techniques with statistical techniques (Adaptive Wiener filtering) improved the quality of enhanced speech in the presence of both stationary and nonstationary noises.
newlineHowever, when dealing with a mixture of a large number of sources, it can be challeng-
newlineing to accurately determine the exact number of sources, and do not generalize well to
newlineunseen noisy conditions.
newlineSecond, we proposed a novel framework that enables temporal context aggregation
newlineusing U-Net with shuffle attention-based skip links. The advantage of shuffle attention
newlinemechanism is that the channel and spatial attention are carried out simultaneously for
newlineeach sub-feature in order to prevent potential noises while also highlighting the proper semantic feature areas by combining the same features from all locations. Our test findings demonstrate that the proposed model offers greater enhancement perfor