Effects-of-Skip-Connections-in-CNN-based-Architectures-for-Speech-Enhancement

/**********************************************************************************************************************************/

This document includes some samples and the relative spetrograms processed by NN-based systems.

• Author: Yupeng Shi, Nengheng Zheng, Yuyong Kang, Weicong Rong
• e-mail: [email protected]
• Date: 05/12/2019

/*********************************************************************************************************************************/

Speech enhancement based deep neural networks (e.g., CNN)

This project is a Python implementation to investigate on the effects of skip connections applied in CNN structures.

samples description

clean: the clean speech
seen_noise: 16 noise types used in the training stage.
unseen_noise: 4 noise types excluded in the training stage 
seen_nosiy:  the noisy speech corrupted by the seen noises 
seen_enhanced: 
              CNN0: the denoised speech enhanced by CNN0 
              CNN1: the denoised speech enhanced by CNN1 
              CNN2: the denoised speech enhanced by CNN2 
              wiener: the denoised speech enhanced by a parametric Wiener filtering
unseen_noisy: the noisy speech corrupted by the unseen noises 
unseen_enhanced: 
              CNN0: the denoised speech enhanced by CNN0 
              CNN1: the denoised speech enhanced by CNN1 
              CNN2: the denoised speech enhanced by CNN2 
              wiener: the denoised speech enhanced by a parametric Wiener filtering

some spectrograms

clean speech noise noisy speech Wiener filtering enhanced CNN0 enhanced CNN1 enhanced CNN2 enhanced

future works

The relative codes will be uploaded soon.

References

[1] Y. P. Shi, W. C. Rong, and N. H. Zheng, "Speech enhancement using convolutional neural network with skip connections,"in the 11th international symposium on Chi-nese spoken language processing (ISCSLP), 2018.

[2] N. H. Zheng, Y. P. Shi, W. C. Rong, and Y. Y. Kang, "Effects of Skip Connections in CNN-based Architectures for Speech Enhancement," accpeted by Journal of Signal Processing Systems, 2019.