CTC HLG decoding high deletion, tune word insertion penalty

[ronggong]

Hi,

My CTC HLG decoding without nbest or whole lattice rescoring gives a quite high deletion. I would like to tune it to balance it with the insertion. Do we have a parameter like word insertion penalty for the lattice?

Thanks.

[csukuangfj]

Does decoding with H have the same issue?

[ronggong]

H decoding has an WER worse than HLG. INS DEL H: 2.30 5.06 HLG: 1.94 5.51

HLG is with lm score = 0.1. Larger lm score gives larger DEL. Basically, HLG increased DEL.

[csukuangfj]

Could you tune https://github.com/k2-fsa/icefall/blob/af735eb75bf55e6b1e41602105a6f939aedbaf5c/egs/librispeech/ASR/conformer_ctc3/decode.py#L245

[ronggong]

lm score = 0.1 HLG score INS DEL 0.2 1.84 4.69 0.4 1.81 4.76 0.6 1.77 4.86 0.8 1.73 5.07 1.0 1.94 5.51

Use small HLG score gives better DEL. Why this happens? Can we make INS and DEL more balance? Maybe by tuning the penalty?