I'm currently working on a speech recognition project. This project require to recognize just single word instead of whole sentence. Currently the WER is just too high. How do I programatically reduce the WER? Can I check the tokens and scores and filter out result with low scores or too many tokens?
Thanks.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
A lot of this depends upon your task. If you have a relatively small vocabulary where the user might say any of the words with equal probability, you might consider using a trivial JSGF grammar that is a simple word list. If you have some idea of the probabilities of the words, you could also try adding weights to the words.
To tune the engine, there are a number of parameters you can set. For JSGF grammars, you might try setting the following parameters in the linguist:
The bunch of these together should tell calls to getBestResultNoFiller to return "<unk>" if Sphinx-4 thinks the utterance is out of grammar. You can read more about this in the FAQ:
Hi,
I'm currently working on a speech recognition project. This project require to recognize just single word instead of whole sentence. Currently the WER is just too high. How do I programatically reduce the WER? Can I check the tokens and scores and filter out result with low scores or too many tokens?
Thanks.
Hi Weiheng:
A lot of this depends upon your task. If you have a relatively small vocabulary where the user might say any of the words with equal probability, you might consider using a trivial JSGF grammar that is a simple word list. If you have some idea of the probabilities of the words, you could also try adding weights to the words.
To tune the engine, there are a number of parameters you can set. For JSGF grammars, you might try setting the following parameters in the linguist:
<property name="addOutOfGrammarBranch" value="true"/> <property name="outOfGrammarProbability" value="1E-60"/> <property name="phoneInsertionProbability" value="1E-10"/> <property name="phoneLoopAcousticModel" value="wsj"/>The bunch of these together should tell calls to getBestResultNoFiller to return "<unk>" if Sphinx-4 thinks the utterance is out of grammar. You can read more about this in the FAQ:
http://cmusphinx.sourceforge.net/sphinx4/doc/Sphinx4-faq.html#out_of_grammar
Note, however, that if you have individual words that can be easily confused (e.g., "on"/"off" "cat"/"fat"), you may still end up with poor results.
Hope this helps,
Will
PS - You might also take a look at the demo/sphinx/confidence/ demo.