hi master:
I'm newer to sphinx4, but I have a clear target using sphinx4. my target is to achieve confidence score to just one sentence.I don't need the recognizer now. so I see the document of sphinx4 and do something like following:
1. record a wav file(just one)
2. using sphinxtrain to create the sentence's model.
during this operation, some error occured. but I think that error can be igore.the error is :
---------------->
WARNING: "mk_mdef_gen.c", line 455: Out of vocabulary words in transcript will be mapped to SIL!
WARNING: "main.c", line 365: 1 seno's not generated because heap ran out
WARNING: "mod_inv.c", line 257: n_top 8 > n_density 1. n_top <- 1
WARNING: "accum.c", line 626: The following seno never occur in the input data
WARNING: "main.c", line 514: dest cb 45 has > 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 46 has > 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 47 has > 1 corresponding source cb
ERROR: "backward.c", line 431: final state not reached
ERROR: "baum_welch.c", line 331: 10015 ignored
<------------------
these warning and error are from several log files.
3. then I using the result from sphinxtrain buiding a model jar file
4. using the confidence demo from sphinx4 src with some changes.
it start up well,but when I speak something, the error occured in console. then I locate the position and find that the actual parameter token is null when sphinx call the Lattice.collapseWordPath method.
can anybody help me?
1. does the target be achieved using sphinx4 or not?
2. if it can, how to do it ? does my practice right?
thanks a lot!
(a poor english, sorry!)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
However, I hope that the realization of the function is: an English sentence, the user listen to the voice, and then read this sentence. I will give points to the user's pronunciation. And then the user listen to the next one, and then read the next sentence, and then score......
If I provided Sphinx4 with more advanced language, Sphinx4 needs to do two things:
1. To identify which sentence the user reading (as is likely to identify other sentence)
2. For the user pronunciation scoring
So the score may be not accurate!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Let me give you an example:
For example, I have a sound file, it's text is: "Listen to the teacher carefully". I will make the sound files sphinx4 model, then use the confidence demo of sphinx4 load this model. And then run the confidence demo, I said "Listen to the teacher carefully" voice to micophone, I would like to output:
listen (1)
to (0.8)
the (1)
teacher (0.3)
carefully (1)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
My question is
1. I generated model， but there have been some mistakes, would not affect the validity of model
---------------->
WARNING: "mk_mdef_gen.c", line 455: Out of vocabulary words in transcript will be mapped to SIL!
WARNING: "main.c", line 365: 1 seno's not generated because heap ran out
WARNING: "mod_inv.c", line 257: n_top 8> n_density 1. N_top <- 1
WARNING: "accum.c", line 626: The following seno never occur in the input data
WARNING: "main.c", line 514: dest cb 45 has> 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 46 has> 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 47 has> 1 corresponding source cb
ERROR: "backward.c", line 431: final state not reached
ERROR: "baum_welch.c", line 331: 10015 ignored
<------------------
2. In my confidence demo running when NullPointException, this anomaly appears in the Lattice.collapseWordPath methods. Is because the token is null. What should I solve this problem?
I have something to do to leave, thank you for your reply, I have some time to come back
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
> but there have been some mistakes, would not affect the validity of model
They are not critical
> 2. In my confidence demo running when NullPointException, this anomaly appears in the Lattice.collapseWordPath methods. Is because the token is null. What should I solve this problem?
Paste the stacktrace from the exception, provide a data to reproduce it.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
NullPointerException problem seems solved, but another problem emerged: no matter how I read, Result result = recognizer.recognize (); the result is always null.
Here are some documents I created:
I set up keepAllTokens = true, but it reported the following error!
Exception in thread "main" java.lang.NullPointerException
at edu.cmu.sphinx.result.Lattice.collapseWordPath (Lattice.java: 218)
at edu.cmu.sphinx.result.Lattice.collapseWordToken (Lattice.java: 194)
at edu.cmu.sphinx.result.Lattice. <init> (Lattice.java: 146)
at edu.cmu.sphinx.result.MAPConfidenceScorer.score (MAPConfidenceScorer.java: 70)
at edu.cmu.sphinx.demo.test.Test.main (Test.java: 67)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Confidence example, I set up the keepAllTokens = true, Confidence can run well. I wonder whether the model that I have problems? Wav file is requested to comply with the rules? Wav files of my bit rate is 128kbps
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I also tested several times, I found: Sphinx4 the demo to work. However, if the custom model, then the process will always be the following error:
Exception in thread "main" java.lang.NullPointerException
at edu.cmu.sphinx.result.Lattice.collapseWordPath (Lattice.java: 218)
at edu.cmu.sphinx.result.Lattice.collapseWordToken (Lattice.java: 194)
at edu.cmu.sphinx.result.Lattice. <init> (Lattice.java: 146)
at edu.cmu.sphinx.result.MAPConfidenceScorer.score (MAPConfidenceScorer.java: 70)
at edu.cmu.sphinx.demo.test.Test.main (Test.java: 67)
I think it was the issue of model, my model is produced in accordance with the http://www.bakuzen.com/?p=16. How to solve the ah? Very anxious
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I've just committed a fix to your issue in sphinx4 trunk, now it will not give you NPE, but it will not give you result as well.
The reason for that is that you are using not enough variants in your language model and your acoustic model is trained only on 1 sentence. So neighther your acoustic model nor language model aren't representative enough.
The acoustic model even for a single speaker should contain at least 50 variants of each phone.
The language model to get confidence score should contain all variants of the badly spoken sentences with all mistakes, not just the correct sentence.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
hi master:
I'm newer to sphinx4, but I have a clear target using sphinx4. my target is to achieve confidence score to just one sentence.I don't need the recognizer now. so I see the document of sphinx4 and do something like following:
1. record a wav file(just one)
2. using sphinxtrain to create the sentence's model.
during this operation, some error occured. but I think that error can be igore.the error is :
---------------->
WARNING: "mk_mdef_gen.c", line 455: Out of vocabulary words in transcript will be mapped to SIL!
WARNING: "main.c", line 365: 1 seno's not generated because heap ran out
WARNING: "mod_inv.c", line 257: n_top 8 > n_density 1. n_top <- 1
WARNING: "accum.c", line 626: The following seno never occur in the input data
WARNING: "main.c", line 514: dest cb 45 has > 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 46 has > 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 47 has > 1 corresponding source cb
ERROR: "backward.c", line 431: final state not reached
ERROR: "baum_welch.c", line 331: 10015 ignored
<------------------
these warning and error are from several log files.
3. then I using the result from sphinxtrain buiding a model jar file
4. using the confidence demo from sphinx4 src with some changes.
it start up well,but when I speak something, the error occured in console. then I locate the position and find that the actual parameter token is null when sphinx call the Lattice.collapseWordPath method.
can anybody help me?
1. does the target be achieved using sphinx4 or not?
2. if it can, how to do it ? does my practice right?
thanks a lot!
(a poor english, sorry!)
However, I hope that the realization of the function is: an English sentence, the user listen to the voice, and then read this sentence. I will give points to the user's pronunciation. And then the user listen to the next one, and then read the next sentence, and then score......
If I provided Sphinx4 with more advanced language, Sphinx4 needs to do two things:
1. To identify which sentence the user reading (as is likely to identify other sentence)
2. For the user pronunciation scoring
So the score may be not accurate!
I suggest you to upload your files to public resource and give us a link, I don't understand neither what you have done nor what do you want to do.
Probably you could write your questions in your native language. I suspect google translate will do better job.
sorry, chinese is here:
我想使用sphinx4实现对一个英文句子打分。
我不想使用sphinx4来识别用户说的是哪句话，我只是想知道用户说的每一个单词是否正确以及正确的程度。所以我有一个疑问：sphinx4是否能够做到这点呢？如果能够做到这点，我应该怎么做呢？
非常感谢！
Ok, you are on the right path, and what exact problem do you have now?
Let me give you an example:
For example, I have a sound file, it's text is: "Listen to the teacher carefully". I will make the sound files sphinx4 model, then use the confidence demo of sphinx4 load this model. And then run the confidence demo, I said "Listen to the teacher carefully" voice to micophone, I would like to output:
listen (1)
to (0.8)
the (1)
teacher (0.3)
carefully (1)
I got this. What is your problem doing that?
My question is
1. I generated model， but there have been some mistakes, would not affect the validity of model
---------------->
WARNING: "mk_mdef_gen.c", line 455: Out of vocabulary words in transcript will be mapped to SIL!
WARNING: "main.c", line 365: 1 seno's not generated because heap ran out
WARNING: "mod_inv.c", line 257: n_top 8> n_density 1. N_top <- 1
WARNING: "accum.c", line 626: The following seno never occur in the input data
WARNING: "main.c", line 514: dest cb 45 has> 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 46 has> 1 corresponding source cb
WARNING: "main.c", line 514: dest cb 47 has> 1 corresponding source cb
ERROR: "backward.c", line 431: final state not reached
ERROR: "baum_welch.c", line 331: 10015 ignored
<------------------
2. In my confidence demo running when NullPointException, this anomaly appears in the Lattice.collapseWordPath methods. Is because the token is null. What should I solve this problem?
I have something to do to leave, thank you for your reply, I have some time to come back
> but there have been some mistakes, would not affect the validity of model
They are not critical
> 2. In my confidence demo running when NullPointException, this anomaly appears in the Lattice.collapseWordPath methods. Is because the token is null. What should I solve this problem?
Paste the stacktrace from the exception, provide a data to reproduce it.
NullPointerException problem seems solved, but another problem emerged: no matter how I read, Result result = recognizer.recognize (); the result is always null.
Here are some documents I created:
test.lm : http://www.box.net/shared/4h02mp9cq8
config.xml : http://www.box.net/shared/e7xzp08hf0
model jar : http://www.box.net/shared/0nh7ihn0mp
The class is same as the Confidence.java.
Make sure you have
<property name="keepAllTokens" value="true"/>
in config file
I set up keepAllTokens = true, but it reported the following error!
Exception in thread "main" java.lang.NullPointerException
at edu.cmu.sphinx.result.Lattice.collapseWordPath (Lattice.java: 218)
at edu.cmu.sphinx.result.Lattice.collapseWordToken (Lattice.java: 194)
at edu.cmu.sphinx.result.Lattice. <init> (Lattice.java: 146)
at edu.cmu.sphinx.result.MAPConfidenceScorer.score (MAPConfidenceScorer.java: 70)
at edu.cmu.sphinx.demo.test.Test.main (Test.java: 67)
In svn trunk everything should be fine
Confidence example, I set up the keepAllTokens = true, Confidence can run well. I wonder whether the model that I have problems? Wav file is requested to comply with the rules? Wav files of my bit rate is 128kbps
I also tested several times, I found: Sphinx4 the demo to work. However, if the custom model, then the process will always be the following error:
Exception in thread "main" java.lang.NullPointerException
at edu.cmu.sphinx.result.Lattice.collapseWordPath (Lattice.java: 218)
at edu.cmu.sphinx.result.Lattice.collapseWordToken (Lattice.java: 194)
at edu.cmu.sphinx.result.Lattice. <init> (Lattice.java: 146)
at edu.cmu.sphinx.result.MAPConfidenceScorer.score (MAPConfidenceScorer.java: 70)
at edu.cmu.sphinx.demo.test.Test.main (Test.java: 67)
I think it was the issue of model, my model is produced in accordance with the http://www.bakuzen.com/?p=16. How to solve the ah? Very anxious
Upload a sample so I can test it.
The following are the three important files to download Address:
test.lm: http://www.box.net/shared/4h02mp9cq8
config.xml: http://www.box.net/shared/e7xzp08hf0
model jar: http://www.box.net/shared/0nh7ihn0mp
Class I use is Confidence class, in addition to names, there is no change in
Give me the complete example ready to run that reproduces the problem with sources and wave file you test on. Pack everything into a single archive.
The link below for all resources, you can click test.bat run the app
http://www.box.net/shared/ybmsv657ke
I've just committed a fix to your issue in sphinx4 trunk, now it will not give you NPE, but it will not give you result as well.
The reason for that is that you are using not enough variants in your language model and your acoustic model is trained only on 1 sentence. So neighther your acoustic model nor language model aren't representative enough.
The acoustic model even for a single speaker should contain at least 50 variants of each phone.
The language model to get confidence score should contain all variants of the badly spoken sentences with all mistakes, not just the correct sentence.
Do you mean is: the current version of sphinx4 can not meet my demand. Is it?
Did I ever wrote even something similar? Sphinx4 is the package that you need.
You just need to use it with more advanced language and acoustic models, not the ones you need.
If Sphinx4 can not meet my request, do you know which libary can meet my request?