Skip to main content

Table 7 False rejection rates for OOV words when using Kaldi-base ASR with N-gram plus RNN re-scoring LM

From: Dynamic out-of-vocabulary word registration to language model for speech recognition

  Position Freq. 2 mora 3 mora 4 mora 5 mora
ONE 1 .988 .696 .5
ALL .253 .151 .071 .022
500r Random 500 .747 .395 .214 .109
1kr Random 1000 .639 .326 .232 .152
2kr Random 2000 .602 .302 .214 .130
5kr Random 5000 .482 .267 .214 .130
500c POS 500 .687 .244 .232 .044
1kc POS 1000 .518 .256 .196 .043
2kc POS 2000 .410 .289 .214 .065
5kc POS 5000 .373 .221 .179 .022