Acoustic Model Discussions

Flat
CMLLR adaptation: problems with HERest and FORCECXTEXP
User: giampila22
Date: 3/8/2013 6:55 am
Views: 6764
Rating: 2

Hello everybody!!

I am trying to implemente CMLLR adaptation. When I run HERest, I got an errore due to the absence of a particualr triphone in my HMM. I am using a configuration file in which I set:

ALLOWXWRDEXP=F
FORCECXTEXP=T

should not this start looking for monophone in case there is not the relative triphone instead of reporting an error?

HVite does it.. does it not work in the same way for HERest?

If not, does this mean that I should have a model for any triphone present in the adaptation data?

I will be very very grateful for any kind of help you can give me.. I am really getting crazy! :

Thank you very much in advance!!

--- (Edited on 3/8/2013 6:55 am [GMT-0600] by Visitor) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: nsh
Date: 3/8/2013 7:45 am
Views: 193
Rating: 3

In order to let others understand your problem please provide the exact error message from HERest

Remember, the more information you provide the faster you get the answer.

 

--- (Edited on 3/8/2013 16:45 [GMT+0300] by nsh) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: giampila22
Date: 3/8/2013 12:52 pm
Views: 175
Rating: 3

Thank you for your answer:)

This is the error I am getting:

ERROR [+7321]  CreateInsts: Unknown label i-n+o
 FATAL ERROR - Terminating program /home/giampila/Desktop/htk-3.4.1/HTKTools/HERest

I am using different data sets for training and adaptating and, apparentely, some of the triphones in the adaptation sets are not in the training set.

If I run HERest using phones instead of triphones, it run, but the perfomarnce are really low.

--- (Edited on 3/8/2013 12:52 pm [GMT-0600] by Visitor) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: nsh
Date: 3/8/2013 1:54 pm
Views: 208
Rating: 4

It's probably better to create a full tied list with the dictionary with HHEd AU command. That way you will be able to find models for all possible variants from your dictionary and it will be better than monophone fallback.

 

> FORCECXTEXP=T

> should not this start looking for monophone in case there is not the relative triphone instead of reporting an error?

 

 

To fallback to monophones you should use the FORCECXTEXP=F

See in the sources

 

iv)  Linking of models to network nodes.

Model names are determined from the phone name and the surrounding context names.

a)  Construct CD name and see if model exists.

b)  Construct CI name and see if model exists.

If ALLOWCXTEXP==FALSE (a) is skipped and if FORCECXTEXP==TRUE (b) is skipped. 

When no matching model is found and error is generated.

 

--- (Edited on 3/8/2013 22:54 [GMT+0300] by nsh) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: giampila22
Date: 3/11/2013 7:11 am
Views: 167
Rating: 4

You mean that I should train the model using fulllist as a reference instead of the tied one?

 

--- (Edited on 3/11/2013 7:11 am [GMT-0500] by Visitor) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: nsh
Date: 3/11/2013 1:41 pm
Views: 181
Rating: 2

Sorry for misunderstanding, I never said that.

I said you need to create a new tied list with all possible triphones from the words from the dictionary, not the one you are using. Please read documentation about AU command.

--- (Edited on 3/11/2013 21:41 [GMT+0300] by nsh) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: giampila22
Date: 3/11/2013 1:58 pm
Views: 176
Rating: 3

Yes, I tried. But still I get the same error and I can not understand why.

The AU command should synthesis new models from the decision trees for the unseen labels.

So.. why do I still get the same error? If HHEd does not find the model, the AU command should compensate, right? Why does this not happen?


Thank you very much and sorry about the previous comment: I should have read better the documentation about the AU command.

--- (Edited on 3/11/2013 1:58 pm [GMT-0500] by Visitor) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: nsh
Date: 3/11/2013 3:58 pm
Views: 172
Rating: 3

> So.. why do I still get the same error? If HHEd does not find the model, the AU command should compensate, right? Why does this not happen?

Who knows. If you provide your files it would be easier to understand the problem.
Anyway, the second way is to set FORCECXTEXP to F, could you try that?

--- (Edited on 3/11/2013 23:58 [GMT+0300] by nsh) ---

Re: CMLLR adaptation: problems with HERest and FORCECXTEXP
User: giampila22
Date: 3/12/2013 11:25 am
Views: 2086
Rating: 4

Thanks a lot! It seems it works now!

I probably made some stupid mistakes I was not able to see..

--- (Edited on 3/12/2013 11:25 am [GMT-0500] by Visitor) ---

PreviousNext