I am trying to create CI model.
During first iteration bw generate no error, but fires warrning like
WARNING: "accum.c", line 579: The following seno never occur in the input data
12 13 14 45 46 47
Where as 12 13 14 belongs to "ao" (vowel) and 45 46 47 belongs to "y" (consonent) phone which occurs number of times in the corpus
(If I am wrong in the interpretation of above warning, please correct me)
And on next Iteration
For some utt, it generate error like
141 ERROR: "backward.c", line 409: final state not reached
What is the reason? Why this error didn't fired by first bw command?
Your suggestion valueable help for me?
Thanking you
TusharP
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
seems as we have the same problems.
When the first bw iteration finishes, it raises the same warning and in the next step, almost every file generates the "final state not reached" error.
Maybe the bw application deletes senones that doesn't occur in the first part of iteration, but they are needed in the second part.
Perhaps it works when the whole iteration is done in one part? I will try it and then post the result.
Greetings,
Jan Pawellek
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-04-11
With respect to the first warning from accum.c, you did the right thing by looking in the .mdef file to identify the phones in question. You have told us that these phones occur in your training data, but have you verified that those words are in the training dictionary and that they contain AO and Y? For more advice on this warning see http://fife.speech.cs.cmu.edu/speech/sphinxman/logfiles.html#100 .
The second warning, "final state not reached' is also mentioned at the URL above. In my experience, this usually occurs when the transcription does not match the utterance -- it means that bw was unable to align the utterance's feature frames with the sequence of HMMs specified by the utterance's transcription. On the first iiterations, these may just occur for no apparent reason, but they will go away after a few iterations. But if your transcriptions contain errors, they will persist. When this occurs, the utterance is simply ignored for that iteration only. (bw does not delete senones.)
Note that (as pointed out in the above URL), if you are doing the bw computation in parts, then some parts may not contain every phone, and the "seno never occur" warnings will happen for that reason, but when the parts are combined in the "norm" step, the "missing" phones will be seen. Check the norm log to be sure.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello Everybody
I am trying to create CI model.
During first iteration bw generate no error, but fires warrning like
WARNING: "accum.c", line 579: The following seno never occur in the input data
12 13 14 45 46 47
Where as 12 13 14 belongs to "ao" (vowel) and 45 46 47 belongs to "y" (consonent) phone which occurs number of times in the corpus
(If I am wrong in the interpretation of above warning, please correct me)
And on next Iteration
For some utt, it generate error like
141 ERROR: "backward.c", line 409: final state not reached
What is the reason? Why this error didn't fired by first bw command?
Your suggestion valueable help for me?
Thanking you
TusharP
Hello,
seems as we have the same problems.
When the first bw iteration finishes, it raises the same warning and in the next step, almost every file generates the "final state not reached" error.
Maybe the bw application deletes senones that doesn't occur in the first part of iteration, but they are needed in the second part.
Perhaps it works when the whole iteration is done in one part? I will try it and then post the result.
Greetings,
Jan Pawellek
With respect to the first warning from accum.c, you did the right thing by looking in the .mdef file to identify the phones in question. You have told us that these phones occur in your training data, but have you verified that those words are in the training dictionary and that they contain AO and Y? For more advice on this warning see http://fife.speech.cs.cmu.edu/speech/sphinxman/logfiles.html#100 .
The second warning, "final state not reached' is also mentioned at the URL above. In my experience, this usually occurs when the transcription does not match the utterance -- it means that bw was unable to align the utterance's feature frames with the sequence of HMMs specified by the utterance's transcription. On the first iiterations, these may just occur for no apparent reason, but they will go away after a few iterations. But if your transcriptions contain errors, they will persist. When this occurs, the utterance is simply ignored for that iteration only. (bw does not delete senones.)
Note that (as pointed out in the above URL), if you are doing the bw computation in parts, then some parts may not contain every phone, and the "seno never occur" warnings will happen for that reason, but when the parts are combined in the "norm" step, the "missing" phones will be seen. Check the norm log to be sure.
cheers,
jerry