Skip to main content

Table 7 Average WER [%] investigating the effect of iteration numbers of bMMI and f-bMMI discriminative training with SAT on the REVERB challenge dev set using eight-channel data

From: Effectiveness of dereverberation, feature transformation, discriminative training methods, and system combination approach for various reverberant environments

 

MFCC

PLP

 

Number of iterations

    
 

1

2

3

4

1

2

3

4

bMMI

        

SIMDATA

8.70

8.41

8.18

7.63

9.02

8.64

8.47

8.06

REALDATA

29.21

28.34

28.16

28.71

29.74

29.26

28.91

28.45

f-bMMI

        

SIMDATA

8.07

7.56

7.30

7.01

8.47

7.93

7.57

7.32

REALDATA

27.70

27.29

27.16

26.94

29.36

27.86

27.15

26.78