0:01
hello this is daniel povey and today
0:03
we're going to ask him we trained libre
0:05
speech model using call these scripts
0:07
what is the next step what can we do now
0:09
to improve its word error rate
0:15
well
0:17
so
0:18
when you ask that question i'm going to
0:20
assume that you trained
0:22
like to the very end of the run.sh so
0:25
you have the like the chain system
0:28
so i mean already that's that's a pretty
0:30
good system
0:32
uh
0:33
but if you want to improve the water
0:35
rate further
0:38
i think the main thing you can do is to
0:40
use a better language model so
0:44
uh
0:45
like the the default uh decoding in
0:48
keldi is i think with the foreground
0:50
language model that script should be
0:51
testing with the foreground that that's
0:53
as good as you can get from
0:56
uh an engram language model you know a
0:58
graph-based decoding but you can improve
1:00
that by rescoring with an iron and a
1:02
lamb
1:03
there are some scripts in there to
1:05
rescore with an rnn so this is a chaldea
1:09
based iron nlm it's not uh one of those
1:12
pie torch-based transformers or
1:13
something so i mean it's a pretty basic
1:15
rnlm these days people can do better
1:19
and we do have some scripts somewhere in
1:20
chaldea
1:22
that you can run a pie torch based rnlm
1:24
but
1:25
i think i would recommend to use the
1:27
kaldi one for now simply because there's
1:29
fewer things that can go wrong
1:34
will we do rescoring with this new rnn
1:37
lm yeah you'll do lattice rescoring uh
1:40
we we don't normally do decoding in the
1:42
first pass with the iron and the lamp
1:45
so you generate you decode the entire
1:47
utterance and then you restore the
1:48
lattice
1:50
okay thank you
1:53
okay bye