If I am training for Hebrew language is it necessary to create linguistic question file
How can I create the linguistic question file to classify the vowels and consonant
If I create this file where should I place it for the run_all.pl script
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If I am training for Hebrew language is it necessary to create linguistic question file
How can I create the linguistic question file to classify the vowels and consonant
If I create this file where should I place it for the run_all.pl script
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2006-03-19
The short answer is, I believe, no.
The "linguistic questions" for the generation of the pruning tree are generated automatically and independent of the language or phone set.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
There are two way to generate decision tree. One way is to do it as SphinxTrain/script_pl/slave.treebuilder.pl does, that is to generate the questions automatically.
Another way (which I think is actually more prominent) is to use a pre-defined question set. This could be done by (not tested)
1, not running
2, Point $CFG_QUESTION_SET in ./etc/sphinx_train.cfg to you question set file.
Then, buildtree.pl will be able to call your cusotomized question set into your decision-tree building.
Now back to your question, it sounds like you want to build a question set for Hebrew. What people usually consider is to ask questions on
a, phones,
b, phone class.
Try to start with this first. You should be able to get something reasonable out of your training.
Arthur
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
thank Arthur
Some thing strange
When I direct the configuration file to my question file I make mistake so the tree did not build at all because sphinxTrainer did not locate the file when I run performance test on the resulting model-without any trees the WER decreases to 19.2 it was 23.1 with the original question i.e. without changing any thing.
When I fix the configuration file and direct it to my question file the trainer build the tree but the WER increase once again
Question : is it possible to have better performance even though there is no trees at all ?.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If I am training for Hebrew language is it necessary to create linguistic question file
How can I create the linguistic question file to classify the vowels and consonant
If I create this file where should I place it for the run_all.pl script
YES,YES On the performance test using the same audio used in training
If I am training for Hebrew language is it necessary to create linguistic question file
How can I create the linguistic question file to classify the vowels and consonant
If I create this file where should I place it for the run_all.pl script
The short answer is, I believe, no.
The "linguistic questions" for the generation of the pruning tree are generated automatically and independent of the language or phone set.
cheers,
jerry
Hi Nihadisra,
There are two way to generate decision tree. One way is to do it as SphinxTrain/script_pl/slave.treebuilder.pl does, that is to generate the questions automatically.
Another way (which I think is actually more prominent) is to use a pre-defined question set. This could be done by (not tested)
1, not running
if (system("perl $scriptdir/make_questions.pl -cfg \"$cfg_file\"")) {
$return_value = 1;
exit ($return_value);
}
in the script $scriptdir/slave.treebuilder.pl
2, Point $CFG_QUESTION_SET in ./etc/sphinx_train.cfg to you question set file.
Then, buildtree.pl will be able to call your cusotomized question set into your decision-tree building.
Now back to your question, it sounds like you want to build a question set for Hebrew. What people usually consider is to ask questions on
a, phones,
b, phone class.
Try to start with this first. You should be able to get something reasonable out of your training.
Arthur
thank Arthur
Some thing strange
When I direct the configuration file to my question file I make mistake so the tree did not build at all because sphinxTrainer did not locate the file when I run performance test on the resulting model-without any trees the WER decreases to 19.2 it was 23.1 with the original question i.e. without changing any thing.
When I fix the configuration file and direct it to my question file the trainer build the tree but the WER increase once again
Question : is it possible to have better performance even though there is no trees at all ?.
When you test with context-independent models, do you get 19.2% WERR?