Cunhang Fan(Institute of Automation, Chinese Academy of Sciences), Jianhua Tao(National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences), Bin Liu(CASIA), Jiangyan Yi(Institute of Automation, Chinese Academy of Sciences) and Zhengqi Wen(CASIA)
Monaural speech dereverberation is a very challenging task because no spatial cues can be used. When the additive noises exist, this task becomes more challenging. In this paper, we propose a joint training method for simultaneous speech denoising and dereverberation using deep embedding representations. Firstly, at the denoising stage, the deep clustering (DC) network is used to extract noise-free deep embedding representations from the anechoic speech and residual reverberation signals. These deep embedding representations are represent the inferred spectral masking patterns of the desired signals so that they could discriminate the anechoic speech and the reverberant signals very well. Secondly, at the dereverberation stage, we utilize another supervised neural network to estimate the mask of anechoic speech from these deep embedding representations. Finally, the joint training algorithm is used to train the speech denoising and dereverberation network. Our experiments are conducted on the TIMIT dataset. Experimental results show that the proposed method outperforms the WPE and BLSTM baselines. Especially in the low SNR (-5 dB) condition, our proposed method produces a relative improvement of 7.8% for PESQ compared with BLSTM method and relative reductions of 16.3% and 19.3% for CD and LLR measures.