Skip to main content
EURASIP Journal on Audio, Speech, and Music Processing

Table 3 Testing speed comparisons of UP-NNLMs and NNLMs for different hidden layers

From: Empirically combining unnormalized NNLM and back-off N-gram for fast N-best rescoring in speech recognition

Model

Speed × 103 (words/second)

 

10

50

100

200

300

400

RNNLMa

0.78

0.23

0.11

0.056

0.041

0.029

+Class layer

70.87

28.5

13.94

7.39

4.21

3.10

UP-RNNLM

400.01

197.75

80.30

26.93

11.95

7.20

FNNLMa

2.72

1.63

0.35

0.291

0.214

0.14

+Class layer

124.89

66.83

36.02

16.16

9.55

5.75

UP-FNNLM

678.48

209.16

83.95

25.33

17.77

11.63

fast-UP-FNNLM

746.43

557.18

406.47

291.35

240.38

201.75

  1. aThe implementations of RNNLM and FNNLM are based on the open source toolkits, CSLM and RNNLM. The matrix and vector operations in CSLM toolkit are optimized via MKL Library, so that the evaluation of FNNLM is faster than RNNLM with the same size of hidden layer.
Back to article page