Skip to main content

Table 3 WER of the MC MDT experiments on test set 8-14 (Far-talking microphone)

From: Multi-candidate missing data imputation for robust speech recognition

 

WER (%)

 

8. Clean

9. Car

10. Babble

11. Restau

12. Street

13. Airport

14. Train

Average

1. No MDT

21.35

30.51

47.71

53.26

55.52

46.95

55.41

44.39 ± 0.5

2. BE PROSPECT MDT

15.93

23.02

36.39

41.29

42.14

34.65

41.32

33.54 ± 0.5

3. GB GD MC MDT

15.93

23.63

37.15

42.27

41.43

35.29

40.91

33.80 ± 0.5

4. GB MAP MC MDT

15.51

23.58

37.17

42.13

41.99

35.00

41.55

33.84 ± 0.5

5. GB GD MC MDT + CGS

15.86

23.96

37.43

42.49

41.58

35.15

40.83

33.90 ± 0.5

6. GD CLBR

19.39

27.93

40.54

45.75

48.14

38.88

45.45

38.01 ± 0.5

7. MAP CLBR

17.86

27.54

43.96

48.10

48.10

39.73

46.98

38.90 ± 0.5

8. SB MC MDT

15.88

23.01

37.23

41.93

41.39

34.99

41.71

33.73 ± 0.5

9. Spectral MC MDT

15.77

24.42

38.74

44.57

42.42

36.17

42.54

34.95 ± 0.5

10. Cepstral MC MDT

16.05

23.04

36.57

42.37

41.74

35.07

41.84

33.62 ± 0.5

  1. GB, Gaussian-based; SB, State-based; CGS, CG selection; CLBR, cluster-based reconstruction; BE, Backend.