when i try to execute livedecode after train my data in Sphinx 3 for LINUX I recieve the following error:
FATAL_ERROR: "kbcore.c", line 243: Mdef #senones(8172) != mgau #senones(846)
Training:
I have train 92 words dictionary, just for probe the training.
I recive an error in bw loops and two words are ignored (perhaps I have to record them again)
I recive an error that repeats (22Mb of html file) that say
ERROR: "gauden.c", line 1418: var (mgau=0, feat=0, density=3,component=38)<0
.
.
.
ERROR: "gauden.c", line 1418: var (mgau=845, feat=0, density=3,component=38)<0
but the training process finishes all Modules (but module 09)
What happen? I think the error is in training, but I don't know what do I have to do.
Thanks a lot.
Javi
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Look at the model definition file. You will see the number of senone is 8172.
While if you look at the mean and variance files. (using command printp provided by SphinxTrain). Then I will guess you will get 846.
Therefore decode will give you a system error because means and mdef have to match in order to allow correct mapping of phoneme to its HMM parameters
It usually doesn't turn out this way and I have no concrete ideas of why. We probably need further probing. From your error message, it only say that some variance was smaller than zero. I will guess that might be caused by 1) too few training data, 2) some tweakings you have done.
One way to debug this problem is,
1, First tell us you have too few training data. There is no need to go further if you only have less than 1 hour of speech and the material is imbalance. Any trainer will probably fail in such a case.
2, At stage 2,4,7, try to check whether the size of mdef and the size of the means are the same. They have to be same. If that is not true, that means procedure in between might already have.
We probably need further rounds to understand why. Tell us what you observe. Also use the Support tab to send us further materials you found it useful for debugging.
Arthur
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-06-06
Thank you for your requiest
I have found the error: I was using path/model_parameter/time.alltriphones.mdef with path/model_architecture/time.cd.4000/variance (etc)
With time.4000.mdef there is no problem.
Greetings
Javi
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello all and thank you in advance,
when i try to execute livedecode after train my data in Sphinx 3 for LINUX I recieve the following error:
FATAL_ERROR: "kbcore.c", line 243: Mdef #senones(8172) != mgau #senones(846)
Training:
I have train 92 words dictionary, just for probe the training.
I recive an error in bw loops and two words are ignored (perhaps I have to record them again)
I recive an error that repeats (22Mb of html file) that say
ERROR: "gauden.c", line 1418: var (mgau=0, feat=0, density=3,component=38)<0
.
.
.
ERROR: "gauden.c", line 1418: var (mgau=845, feat=0, density=3,component=38)<0
but the training process finishes all Modules (but module 09)
What happen? I think the error is in training, but I don't know what do I have to do.
Thanks a lot.
Javi
Look at the model definition file. You will see the number of senone is 8172.
While if you look at the mean and variance files. (using command printp provided by SphinxTrain). Then I will guess you will get 846.
Therefore decode will give you a system error because means and mdef have to match in order to allow correct mapping of phoneme to its HMM parameters
It usually doesn't turn out this way and I have no concrete ideas of why. We probably need further probing. From your error message, it only say that some variance was smaller than zero. I will guess that might be caused by 1) too few training data, 2) some tweakings you have done.
One way to debug this problem is,
1, First tell us you have too few training data. There is no need to go further if you only have less than 1 hour of speech and the material is imbalance. Any trainer will probably fail in such a case.
2, At stage 2,4,7, try to check whether the size of mdef and the size of the means are the same. They have to be same. If that is not true, that means procedure in between might already have.
We probably need further rounds to understand why. Tell us what you observe. Also use the Support tab to send us further materials you found it useful for debugging.
Arthur
Thank you for your requiest
I have found the error: I was using path/model_parameter/time.alltriphones.mdef with path/model_architecture/time.cd.4000/variance (etc)
With time.4000.mdef there is no problem.
Greetings
Javi