can i update language model with different domain content without recreating graph. as sometimes, it required on client side if they have different content for transcription as per their domain. if possible, then do explain the way
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The language model is encoded in the graph so ideally you should recompile
the graph if the LM changes, but there are other ways to do this, e.g.
lattice rescoring. Some of the recipes show how to do this, search for
"rescore" in the run.sh scripts. There are two ways: using a graph-based
LM and using an ARPA LM; the latter is more efficient if you have an
ARPA-style language model, search for "carpa" in the scripts.
can i update language model with different domain content without
recreating graph. as sometimes, it required on client side if they have
different content for transcription as per their domain. if possible, then
do explain the way
It hasn't been implemented yet in real-time decoding but shouldn't be super
hard to do. WER will depend how different the LMs are. If extremely
different, then it would be better to make the graph.
Dan
Hello,
can i update language model with different domain content without recreating graph. as sometimes, it required on client side if they have different content for transcription as per their domain. if possible, then do explain the way
The language model is encoded in the graph so ideally you should recompile
the graph if the LM changes, but there are other ways to do this, e.g.
lattice rescoring. Some of the recipes show how to do this, search for
"rescore" in the run.sh scripts. There are two ways: using a graph-based
LM and using an ARPA LM; the latter is more efficient if you have an
ARPA-style language model, search for "carpa" in the scripts.
Dan
On Tue, Mar 17, 2015 at 12:23 PM, Ashish Dave badboys4life007@users.sf.net
wrote:
yes i found it using an ARPA LM but does it affect on performance with real time decoding and on WER with nnet2 in compare to graph-based LM
It hasn't been implemented yet in real-time decoding but shouldn't be super
hard to do. WER will depend how different the LMs are. If extremely
different, then it would be better to make the graph.
Dan
On Tue, Mar 17, 2015 at 3:53 PM, Ashish Dave badboys4life007@users.sf.net
wrote: