Use DNN to estimate the phase information of the speech in STFT filed to improve the speech quality and intelligibility.
"Phase-Aware Speech Enhancement Based on Deep Neural Networks", Naijun Zheng and Xiao-Lei Zhang
DNN_sim.zip is the Matlab code
RUN.m line 6: please change the order as follows: isseq = [1,1,1]; % control the three stages: [mix, get featrue, train and test DNN]
The demo package contains the wav demo generated by the DNN models and BLSTM models
========================================================================================= Example 1. using TIMIT corpurs. A. Factory1 noise, SNR=0dB, female speakers (“She had your dark suit in greasy wash water all year.“)
B. Bucanner1 noise, SNR=-3dB, female speakers (“She had your dark suit in greasy wash water all year.“)
C. Babble noise, SNR=0 dB, male speakers (“Don’t ask me to carry an oily rag like that.“)
D. Babble noise, SNR=-5 dB, male speakers (“But even mother’s loving attitude will not always prevent misbehavior.“)
E. Destoryoper noise (unknown) , SNR=-3 dB, both genders (“She had your dark suit in greasy wash water all year.“)
============================================================================================== Example 2. using CHiME-4 corpurs (simulated data). % the training and test data informaiton of our experiment can be referred to [training and test information] [tr05_simu]
A. CHiME-4, using mask approximation (MA) loss function
B. CHiME-4, SNR=-3 dB, using mask approximation +masked signal approximation (MA+mSA) loss function
133: “The bond market and the dollar also tumbled”, “BUS”,
616: “We\’ve got enormous problems and we have to learn to live within our resources”, “PED”,
843: “This, some analysts say, may mean that the London market reaction will be confined to a brief adjustment at the opening.”, “CAF”,
892: “Maintaining those levels ,COMMA that’s going to be the challenge .PERIOD \”DOUBLE-QUOTE”, “STR”,
================================================================================================== Example 3. using CHiME-4 corpurs using BLSTMs.