i've built accoustic model and using grammar.
when i try to using julius live system can recognize word/sentence well.
but, when i try to build language model using htk to make reverse trigram live with julian and using accoutic model, the performance is worse than using grammar.
i try language model using 3136 training sentence and 180 word list (wlist).
when i executing julius there is no error message.
is it true that the performance using language model is worse than using grammar?
here is the step that i already did in language model htk:
- reverse all train corpus
- make unigram and bigram
- make trigram
LBuild -f TEXT -T 1 -c 3 1 -n 3 -l lm_2k\bg1 lm_2k\2k.wmap lm_2k\tg1_1_new.arpa holmes.1\data.0 holmes.1\data.1 holmes.1\data.2 lm_2k\data.0
- mkbingram -nlr bg1_new.arpa -nrl tg1_1_new.arpa julius2.bin
from the step above, is there wrong step or missing step?
--- (Edited on 4/28/2011 2:44 am [GMT-0500] by agnes) ---