search

k2-fsa / icefall

https://k2-fsa.github.io/icefall/
Apache License 2.0
928 stars 295 forks source link

the LG result is weird #839

Open didadida-r opened 1 year ago

didadida-r commented 1 year ago

Hi, I am testing the LG result in K2, but i found the result is bad and something cases are unusual. The result can be divided into two case. Empty and Incomplete, like #403 .

Env setting: zipformer, fast_beam_search_nbest_LG, and OOV word is not very much. lastest K2 and icefall, logadd by default, like #277.

a. Empty the whole decode result is empty, and in some cases, this can be alleviated when degrading the lm scale. unlike #277, the OOV is not the leading factor of this phenomenon. e.g. src = ['这'， ‘个’， ‘多’， ‘少’， ‘钱’， ‘啊’] lm-scale-0.1-hyp = ['这'， ‘个’， ‘多’， ‘钱’， ‘啊’] lm-scale-0.15-hyp = [] lm-scale-0.2-hyp = [] ...

b. Incomplete the result result is incomplete. This can be found in the beginning, end and middle of a uttrance. Also, if the avg duration of the testset audio & text is longer, the Incomplete case will much more common, resulting much more del error.

iggygeek commented 1 year ago

I have the same problem using my own ARPA LM , although I validated with ./local/validate_bpe_lexicon.py