Skip to main content

Table 9 WER [%] on the REVERB challenge eva set

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

  

Near

Far

Near

Far

Near

Far

 

Near

Far

 

1 ch

Kaldi baseline

13.23

14.13

15.54

29.69

20.06

37.44

21.68

50.62

45.98

48.30

 

derev.

12.50

13.43

14.61

24.71

17.09

32.62

19.16

44.75

43.32

44.04

 

GMM (f-bMMI)

7.27

8.17

8.82

14.11

10.54

18.76

11.28

28.65

29.54

29.10

 

GMM (SAT,f-bMMI)

6.44

7.22

7.57

13.97

9.52

18.44

10.53

28.87

29.78

29.33

 

SGMM (SAT, bMMI)

5.81

6.54

7.22

13.84

8.70

18.17

10.05

27.75

28.36

28.06

 

DNN (SAT, bMMI)

5.90

6.84

7.35

12.57

9.40

16.55

9.77

25.97

25.69

25.83

 

ROVER 5)

5.30

5.61

6.30

11.16

7.76

14.95

8.51

23.79

23.60

23.70

8 ch

CSP+BF+derev.

10.94

11.69

10.98

16.33

12.79

21.39

14.02

34.33

36.93

35.63

 

+NLMS

10.94

12.32

11.38

17.59

13.46

22.96

14.78

35.32

35.28

35.30

 

GMM (f-bMMI)

6.57

6.93

6.80

9.93

7.47

12.76

8.41

20.22

23.19

21.71

 

GMM (SAT, f-bMMI)

6.17

6.64

6.51

10.13

7.40

13.15

8.33

20.63

23.67

22.15

 

SGMM (SAT, bMMI)

5.86

6.44

6.29

9.23

6.96

12.83

7.94

20.66

23.50

22.08

 

DNN (SAT, bMMI)

5.64

6.18

6.16

9.29

7.08

12.40

7.79

19.35

22.28

20.82

 

ROVER 5)

4.96

5.62

5.58

8.18

5.73

10.47

6.76

16.90

20.29

18.60

 

ROVER 6)

5.00

5.56

5.38

8.15

5.73

10.70

6.75

17.47

20.36

18.93

  1. All systems except ROVER are single systems. MFCC feature was used for single system, and MFCC and PLP features were used for ROVER 5). Italicized data were the best systems in each condition