Skip to main content

Table 4 WER [%] on the REVERB challenge dev set using eight-channel data and MFCC features

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

   

SIMDATA

REALDATA

   

Room 1

Room 2

Room 3

Avg

Room 1

Avg

 

Feature

Type

Near

Far

Near

Far

Near

Far

 

Near

Far

 

CSP+BF+derev.

MFCC

ML

10.79

12.19

11.02

16.71

11.47

20.43

13.77

40.36

42.83

41.60

+NLMS

  

11.11

12.27

11.81

17.40

12.34

21.46

14.40

38.37

40.74

39.56

GMM

+LDA+MLLT

ML

8.38

10.30

9.91

14.94

10.19

17.28

11.83

34.06

37.18

35.62

 

+basis fMLLR

 

7.74

9.22

8.80

13.33

9.05

15.28

10.57

27.39

30.14

28.77

  

bMMI

6.64

8.21

7.25

11.39

7.10

11.50

8.68

24.89

27.96

26.43

  

f-bMMI

6.19

7.40

7.39

10.13

6.58

10.24

7.99

22.58

26.25

24.42

  

f-bMMI c

6.39

7.33

7.44

9.86

6.70

10.44

8.03

22.71

27.41

25.06

 

+SAT

ML

7.25

9.32

8.70

12.79

8.33

13.80

10.03

28.88

32.88

30.88

  

bMMI

5.24

7.10

6.56

9.93

5.98

10.98

7.63

26.58

30.83

28.71

  

f-bMMI

5.01

6.76

5.96

9.07

5.84

9.40

7.01

24.27

29.60

26.94

  

f-bMMI c

5.16

6.93

6.11

9.49

5.96

9.67

7.22

24.27

29.73

27.00

SGMM

 

ML

5.65

7.62

7.47

10.97

7.00

11.45

8.36

25.27

30.35

27.81

  

bMMI

4.57

6.05

6.19

9.27

6.01

9.89

7.00

24.70

30.01

27.36

  

bMMI c

4.72

6.10

6.09

9.56

6.18

10.01

7.11

24.39

30.01

27.20

DNN

 

CE

6.49

7.45

7.84

11.44

7.25

11.97

8.74

25.27

29.32

27.30

  

bMMI

5.56

6.27

6.24

9.29

5.71

10.44

7.25

23.27

28.84

26.06

  

bMMI c

5.26

6.05

6.21

9.10

5.61

10.06

7.05

22.65

28.50

25.58

  1. In addition to the proposed dereverberation method, BF with direction of arrival estimation by CSP analysis and NLMS adaptive filters were used. Subscript letter “c” represents the proposed “complementary” system. Italicized data were the best systems in each condition