I'm trying to build my own acoustic model with Sphinx Train.
I followed the tinydoc.txt, but the baum-welch iteration failed :
[root@localhost time]# ./scripts_pl/02.ci_schmm/slave_convg.pl 1
MODULE: 02 Training Context Independent models
Cleaning up directories: accumulator...logs...models...
Flat initialize
Baum welch starting for iteration: 1 (1 of 1)
Using 6 files: 0% 50% 100% Finished
Current Overall Likelihood Per Frame = -565035857.267189
Baum welch starting for iteration: 2 (1 of 1)
Using 6 files: 0% ERROR: 48 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s1 ignored
ERROR: utt> 1 s2 215 0 54 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s2 ignored
ERROR: utt> 2 s3 153 0 36 7 ERROR: "backward.c", line 409: final state not reached
50% ERROR: ERROR: "baum_welch.c", line 300: s3 ignored
ERROR: utt> 3 s4 146 0 36 7 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s4 ignored
ERROR: utt> 4 s5 202 0 48 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s5 ignored
ERROR: utt> 5 s6 253 0 42 6 ERROR: "backward.c", line 409: final state not reached
100% ERROR: ERROR: "baum_welch.c", line 300: s6 ignored
Finished
Baum welch ran successfully for only 0 frames! Aborting..
But I'm not sure about my .dict
I don't use the script make_dict. It works only for US english and I want to make it in french, do you use this script ?
cheers,
Florent
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-04-18
Florent -- you are modeling words, not phones, and therefore your .dict file appears to be correct (as far as I understand word-modeling). The make-dict script would produce a dictionary file in terms of English phones, which is not what you need.
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The module 2 works fine !
I red the documentation and they said that if you're training a little acoustic model (like mine) it is not necessary to build the decision tree and I only have to follow through the CI training.
So, (see tinydoc.txt) I have to run the module 1,2,3,4 and not the 5, 6 ? Ok but when I run the module 7, It said that it cannot find the time.6000.mdef and it is correct because no module buid this file !!!!
Did I forget something ??
Thanks,
Florent
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
So I only have to run is the module 2, so I don't need to run 3...7 ?
Ok, but some times I have this error message during the module 2 :
ERROR: utt> 2 s3 153 0 36 7 ERROR: "backward.c", line 409: final state not reached
It's strange :
For example, I have in my .transcript :
<s> <sil> CREER SPHERE <sil> </s> (s3)
<s> <sil> CREER SPHERE <sil> </s> (s4)
and it works, but when I have this file :
<s> <sil> CREER SPHERE <sil> </s> (s3)
<s> <sil> CREER SPHERE <sil> </s> (s4)
<s> <sil> CREER SPHERE <sil> </s> (s5)
I have this problem ! I trained myself the 3 files and I don't understand.
I red on another message that it happens when there is no silence on the beginning, and on the end of the speech file, is that correct ?
In the SphinxTrain Documentation, they said that word models have some limitations, which are discusses in the non-technical version of this manual. But where is this non-tech manual ?
Thanks,
Florent
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I had some similar problems when I tried to use the SphinxTrainer under Windows...
Do you work under Linux or Windows?
greetz,
Sebastian
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2005-04-21
Yes, it's correct that you need only to train CI models when training word, and not phone models. CI training is done by SphinxTrain module 02. Therefore, you should now be finished!
cheers,
jerry
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I'm trying to build my own acoustic model with Sphinx Train.
I followed the tinydoc.txt, but the baum-welch iteration failed :
[root@localhost time]# ./scripts_pl/02.ci_schmm/slave_convg.pl 1
MODULE: 02 Training Context Independent models
Cleaning up directories: accumulator...logs...models...
Flat initialize
Baum welch starting for iteration: 1 (1 of 1)
Using 6 files: 0% 50% 100% Finished
Current Overall Likelihood Per Frame = -565035857.267189
Baum welch starting for iteration: 2 (1 of 1)
Using 6 files: 0% ERROR: 48 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s1 ignored
ERROR: utt> 1 s2 215 0 54 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s2 ignored
ERROR: utt> 2 s3 153 0 36 7 ERROR: "backward.c", line 409: final state not reached
50% ERROR: ERROR: "baum_welch.c", line 300: s3 ignored
ERROR: utt> 3 s4 146 0 36 7 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s4 ignored
ERROR: utt> 4 s5 202 0 48 6 ERROR: "backward.c", line 409: final state not reached
ERROR: ERROR: "baum_welch.c", line 300: s5 ignored
ERROR: utt> 5 s6 253 0 42 6 ERROR: "backward.c", line 409: final state not reached
100% ERROR: ERROR: "baum_welch.c", line 300: s6 ignored
Finished
Baum welch ran successfully for only 0 frames! Aborting..
I don't understand why it failed.
That's my .dict :
CREER CREER
SPHERE SPHERE
my .filler
<s> SIL
</s> SIL
<sil> SIL
/SMACK/ +SMACK+
my .phone
SIL
CREER
SPHERE
+SMACK+
my .transcription
<s> <sil> /SMACK/ <sil> CREER SPHERE <sil> </s> (s1)
<s> <sil> /SMACK/ <sil> CREER <sil> SPHERE <sil> </s> (s2)
<s> <sil> CREER SPHERE <sil> </s> (s3)
<s> <sil> CREER SPHERE <sil> </s> (s4)
<s> <sil> /SMACK/ <sil> CREER SPHERE <sil> </s> (s5)
<s> <sil> CREER <sil> SPHERE <sil> </s> (s6)
I have 6 little wav files
I have 5-state hmms, 500 senones...
Any ideas ???
Thanks !
Florent
I work under Linux.
But I'm not sure about my .dict
I don't use the script make_dict. It works only for US english and I want to make it in french, do you use this script ?
cheers,
Florent
Florent -- you are modeling words, not phones, and therefore your .dict file appears to be correct (as far as I understand word-modeling). The make-dict script would produce a dictionary file in terms of English phones, which is not what you need.
cheers,
jerry
I have finally solved this problem !
The module 2 works fine !
I red the documentation and they said that if you're training a little acoustic model (like mine) it is not necessary to build the decision tree and I only have to follow through the CI training.
So, (see tinydoc.txt) I have to run the module 1,2,3,4 and not the 5, 6 ? Ok but when I run the module 7, It said that it cannot find the time.6000.mdef and it is correct because no module buid this file !!!!
Did I forget something ??
Thanks,
Florent
So I only have to run is the module 2, so I don't need to run 3...7 ?
Ok, but some times I have this error message during the module 2 :
ERROR: utt> 2 s3 153 0 36 7 ERROR: "backward.c", line 409: final state not reached
It's strange :
For example, I have in my .transcript :
<s> <sil> CREER SPHERE <sil> </s> (s3)
<s> <sil> CREER SPHERE <sil> </s> (s4)
and it works, but when I have this file :
<s> <sil> CREER SPHERE <sil> </s> (s3)
<s> <sil> CREER SPHERE <sil> </s> (s4)
<s> <sil> CREER SPHERE <sil> </s> (s5)
I have this problem ! I trained myself the 3 files and I don't understand.
I red on another message that it happens when there is no silence on the beginning, and on the end of the speech file, is that correct ?
In the SphinxTrain Documentation, they said that word models have some limitations, which are discusses in the non-technical version of this manual. But where is this non-tech manual ?
Thanks,
Florent
I had some similar problems when I tried to use the SphinxTrainer under Windows...
Do you work under Linux or Windows?
greetz,
Sebastian
Yes, it's correct that you need only to train CI models when training word, and not phone models. CI training is done by SphinxTrain module 02. Therefore, you should now be finished!
cheers,
jerry