Paper Info
Title
A context aware-based deep neural network approach for simultaneous speech denoising and dereverberation
Abstract
Generally, the recorded speech signal is corrupted by both room reverberation and background noise leading to a reduced speech quality and intelligibility. In order to deal with the distortions caused by the joint effect of noise and reverberation, we propose a context-aware-based deep neural network (DNN) approach for simultaneous speech denoising and dereverberation. The proposed system consists of two stages such as denoising stage and the dereverberation stage. In the denoising stage, the additive noise is suppressed by estimating a phase-sensitive mask using DNN. Then, the noise-free reverberant speech is processed through the dereverberation stage. In the dereverberation stage, a reverberation-time-aware DNN-based model is used to perform dereverberation by adopting two reverberation time-dependent parameters such as frameshift size and acoustic context size to get the benefits of the characteristics of the superposition and frame-wise temporal correlations in different reverberation circumstances. Finally, we integrate both the modules and employ the integrated module for joint training using a multi objective loss function to further optimize both the denoising and dereverberation stages. Experimental results show that the proposed approach has shown significant performance improvements over prevalent benchmark dereverberation algorithms on IEEE corpus, REVERB challenge, and TIMIT corpus datasets under several reverberation circumstances.
Year
DOI
Venue
2022
10.1007/s00521-022-06968-1
Neural Computing and Applications
Keywords
DocType
Volume
Speech dereverberation, Phase sensitive mask, Deep learning, Two-stage model, Speech denoising, Joint training
Journal
34
Issue
ISSN
Citations 
12
0941-0643
0
PageRank 
References 
Authors
0.34
22
2
Name
Order
Citations
PageRank
Sidheswar Routray100.34
Qirong Mao226134.29