Skip to main content

Advertisement

Table 5 Word error rates (WERs) of Hub5’00-SWB and RT03S-FSH for 100-best/1,000-best rescoring with NNLM and UP-NNLM

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

Model 100-best WER (%) 1,000-best WER (%)
  Hub5’00-SWB RT03S-FSH Hub5’00-SWB RT03S-FSH
1-best (KN3) 17.3 20.2 17.3 20.2
+UP-FNNLM 16.6 (-0.7) 19.0 (-1.2) 16.5 (-0.8) 19.0 (-1.2)
+UP-RNNLM-C400 16.5 (-0.8) 19.0 (-1.2) 16.4 (-0.9) 19.1 (-1.1)
+FNNLM 15.7 (-1.6) 18.4 (-1.8) 15.6 (-1.7) 18.4 (-1.8)
+RNNLM-C400 15.4 (-1.9) 18.2 (-2.0) 15.2 (-2.1) 18.4 (-1.8)
KN5 17.1 (-0.2) 19.5 (-0.7) 16.9 (-0.4) 19.4 (-0.8)
+UP-FNNLM 16.4 (-0.9) 18.6 (-1.6) 16.2 (-1.1) 18.6 (-1.6)
+UP-RNNLM-C400 16.1 (-1.2) 18.5 (-1.7) 16.0 (-1.3) 18.5 (-1.7)
+FNNLM 15.6 (-1.7) 18.3 (-1.9) 15.4 (-1.9) 18.2 (-2.0)
+RNNLM-C400 15.3 (-2.0) 18.1 (-2.1) 15.2 (-2.1) 18.1 (-2.1)
FNNLM 15.9 (-1.4) 18.7 (-1.5) 15.8 (-1.5) 18.7 (-1.5)
RNNLM-C400 15.4 (-1.9) 18.4 (-1.8) 15.5 (-1.8) 18.6 (-1.6)