Hi,
i have build an acoustic model with SphinxTrain for the aurora corpus and want to use it now in Sphinx4. I have read the S4 FAQ part and the doc "How to use Models from SphinxTrain in Sphinx 4" and ahve done all the decribed steps and build an JAR file for the model. But the model dont work with my modified WavFile demo. I changed the config.xml for the WavFile demo and included the new model... When i modify the FeatureExtraction for the new S3 Model from:
and leave out the last two steps CMN and FE the decoder brings up an error:
Exception in thread "main" java.lang.ClassCastException
at edu.cmu.sphinx.linguist.acoustic.tiedstate.MixtureComponent.getScore(MixtureComponent.java:136)
at edu.cmu.sphinx.linguist.acoustic.tiedstate.GaussianMixture.calculateScore(GaussianMixture.java:152)
at edu.cmu.sphinx.linguist.acoustic.tiedstate.GaussianMixture.getScore(GaussianMixture.java:85)
....
can anyone say something about that ???
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2006-03-23
Thanks, the decoder works now ;-)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2006-03-19
I tryed this patch always and it generates the same model file link my modyfied Sphinx4 ant build file... The Problem is not the conversion of the model files but the feature extracion definition in the config.xml file itself. If i leave out the last two lines how desribed in the FAQ the decoder crashes because the last step is necesarry ???
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2006-03-19
Chris -- Indeed those "last two lines" are necessary for normal Sphinx-4 decoder operation.
That FAQ, "Does the Sphinx-4 front end generate the same features as the SphinxTrain wave2feat program?" deals with the academic question of the relation between the cepstrum features generated by wave2feat (the feature computation program in SphinxTrain) and those computed in the S4 FE. If you use the S4 FE as suggested there, and then use some other component* to write that cepstrum data to a file, it'll be the same as that computed by wave2feat.
*I don't know whether such a data-output component exists in S4.
Left unsaid is the further information that neither SphinxTrain nor Sphinx-N train or decode using only those basic cepstral data. Subsequent processing in both does (1) cepstral-mean normalization and (2) generation of 1st- and 2nd-differences, which all go into the feature-set that's used for training and decoding.
In any case, it should not have misled you to believe that you should remove those 2 lines in order to use your new acoustic model in Sphinx-4.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
i have build an acoustic model with SphinxTrain for the aurora corpus and want to use it now in Sphinx4. I have read the S4 FAQ part and the doc "How to use Models from SphinxTrain in Sphinx 4" and ahve done all the decribed steps and build an JAR file for the model. But the model dont work with my modified WavFile demo. I changed the config.xml for the WavFile demo and included the new model... When i modify the FeatureExtraction for the new S3 Model from:
<component name="mfcFrontEnd" type="edu.cmu.sphinx.frontend.FrontEnd"> <propertylist name="pipeline"> <item>streamDataSource</item> <item>premphasizer</item> <item>windower</item> <item>fft</item> <item>melFilterBank</item> <item>dct</item> <item>batchCMN</item> <item>featureExtraction</item> </propertylist> </component>and leave out the last two steps CMN and FE the decoder brings up an error:
Exception in thread "main" java.lang.ClassCastException
at edu.cmu.sphinx.linguist.acoustic.tiedstate.MixtureComponent.getScore(MixtureComponent.java:136)
at edu.cmu.sphinx.linguist.acoustic.tiedstate.GaussianMixture.calculateScore(GaussianMixture.java:152)
at edu.cmu.sphinx.linguist.acoustic.tiedstate.GaussianMixture.getScore(GaussianMixture.java:85)
....
can anyone say something about that ???
Thanks, the decoder works now ;-)
The page you've seen could be too difficult to follow. Have you tried Jerry Wolf's script?
http://sourceforge.net/tracker/index.php?func=detail&aid=1174946&group_id=1904&atid=301904
Arthur
Hi Athur,
the link above dont work... can you write it again...
Chris
It's http://sourceforge.net/tracker/index.php?func=detail&aid=1174946&group_id=1904&atid=301904
If that doesn't work, follow the "Patches" link at the top of the CMU Sphinx SpurceForge page. It's the top one listed there.
cheers,
jerry
It's http://sourceforge.net/tracker/index.php?func=detail&;aid=1174946&group_id=1904&atid=301904
good luck
I tryed this patch always and it generates the same model file link my modyfied Sphinx4 ant build file... The Problem is not the conversion of the model files but the feature extracion definition in the config.xml file itself. If i leave out the last two lines how desribed in the FAQ the decoder crashes because the last step is necesarry ???
Chris -- Indeed those "last two lines" are necessary for normal Sphinx-4 decoder operation.
That FAQ, "Does the Sphinx-4 front end generate the same features as the SphinxTrain wave2feat program?" deals with the academic question of the relation between the cepstrum features generated by wave2feat (the feature computation program in SphinxTrain) and those computed in the S4 FE. If you use the S4 FE as suggested there, and then use some other component* to write that cepstrum data to a file, it'll be the same as that computed by wave2feat.
*I don't know whether such a data-output component exists in S4.
Left unsaid is the further information that neither SphinxTrain nor Sphinx-N train or decode using only those basic cepstral data. Subsequent processing in both does (1) cepstral-mean normalization and (2) generation of 1st- and 2nd-differences, which all go into the feature-set that's used for training and decoding.
In any case, it should not have misled you to believe that you should remove those 2 lines in order to use your new acoustic model in Sphinx-4.
cheers,
jerry
What Jerry said is correct, my question here is why you would want to remove the two lines?
They are necessary to generate CMN and create delta coeff.
Arthur