I have used char alphabet (a-z) to train the model and the lm works well. But if I used subword alphabet, the lm does not work at all.
The subword looks like below
'ar'
'or_'
'all_'
'by_'
'have'
'g_'
'an_'
'ri'
'him_'
'my_'
'this'
'in'
'k_'
'bo'
'whic'
'ro'
'si'
'po'
'de'
'were'
'un'
'said'
'fi'
Is there any way to use ds_ctcdecoder by subword?