Skip to main content

Table 4 Mean-squared error (MSE) between the log-magnitude spectra of reference (clean) and DNN enhancement

From: Speech dereverberation for enhancement and recognition using dynamic features constrained deep neural networks and feature adaptation

Processing

Static

Delta

Acceleration

DS

226.21

10.76

2.06

DS+DNN5

91.81

6.53

1.31

DS+DNN7

91.34

6.28

1.22

DS+DNN7LS

87.70

5.99

1.18

DS+DNN8

88.77

5.99

1.20

  1. The enhanced spectra are taken from the DNN’s output layer without going through the resynthesis to avoid the effect of imperfect phase. The MSE is averaged over all the six simulated test conditions of eval set