Bug: wrong vacab index for WFST decoding?
iou2much opened this issue · 3 comments
iou2much commented
https://github.com/athena-team/athena/blob/master/athena/models/mtl_seq2seq.py#L153
predictions = [self.vocab[prediction] - 1 for prediction in words_prediction]
Seems a bug, not suppose to minus 1 here
Some-random commented
Yeah it probably is a bug to complement the fact our previous vocab file didn't start with 0. Can you make sure removing the minus 1 here will get correct results on AISHELL and make a pr for us?
iou2much commented
Yes. I've already try to remove -1 here, and get correct result. I'll make a pr. thanks