Skip to main content

Advertisement

Table 3 Recognition performance of second-pass re-scoring

From: A hybrid input-type recurrent neural network for LVCSR language modeling

Tasks LM 4gr RNNLMs
   s-H W PM H-F H-R1 H-R2
  #Word 30 K 35 K 0 K 35 K 30 K 25 K
  #PM 21 K 0 K 25 K 25 K 5 K 10 K
BN PER (%) 19.60 19.11 19.12 18.73 18.84 18.81
  WER (%) 22.92 22.22 22.23 22.05 22.16 22.13
VT PER (%) 19.02 19.02 19.26 18.89 18.79 18.41
  WER (%) 21.75 21.85 22.17 21.69 21.57 21.27
AVG. PER (%) 19.31 19.07 19.19 18.81 18.82 18.61
  WER (%) 22.34 22.04 22.20 21.87 21.87 21.70