Acoustic Model Discussions

Flat
Problem in Adapation using Hvite-HEadapt
User: mvoxforge
Date: 1/11/2013 8:57 am
Views: 6854
Rating: 8

Hi I have followed htk 3.2 tutorial example steps to creat a speech recognition system on wsjcam0. my system has +50K tied states and 8 comp GMMs and I used word-internal triphones.

After I have done adaptation and test on another corpus and i got unexpected results, I have checked the accuracy of my process the only odd thing was in the equivalent step to: http://www.voxforge.org/home/dev/acousticmodels/linux/adapt/htkjulius/adapt/step-4 step my adaptPhones.mlf had some biphones instead of phones for instance when i have a sentence with word 'looked' with phon transcription 'l uh k t sp' according to beep dictionary  I see the following form in my adaptphones.mlf file:

sp
l
t-uh
k
t
sp

 

I dont know why i have biphones in my list could you please guide me.

Cheers.

--- (Edited on 1/11/2013 8:57 am [GMT-0600] by mvoxforge) ---

Re: Problem in Adapation using Hvite-HEadapt
User: TonyR
Date: 1/11/2013 9:24 am
Views: 187
Rating: 9

Are you sure that your main model has the triphones you expect and not monophones and biphones?

 

Tony

 

--- (Edited on 11-January-2013 3:24 pm [GMT+0000] by TonyR) ---

Re: Problem in Adapation using Hvite-HEadapt
User: Visitor
Date: 1/11/2013 9:37 am
Views: 243
Rating: 10

Thanks Tony, Yes I have followed the instruction in HTK for creating word internal triphones.

--- (Edited on 1/11/2013 9:37 am [GMT-0600] by Visitor) ---

Re: Problem in Adapation using Hvite-HEadapt
User: TonyR
Date: 1/11/2013 9:56 am
Views: 208
Rating: 7

That wasn't the question I asked - the question I asked was whether you had a model that contained the word internal triphones you expected.   To check this you convert the MMF to ASCII and then have a look.

You are calling:

HVite -A -D -T 1 -l '*' -o SWT -b SENT-END -C config -H macros -H hmmdefs -i adaptPhones.mlf -m -t 250.0 150.0 1000.0 -y lab -a -I adaptWords.mlf -S adapt.scp dict tiedlist 

The -m flag says output model alignment - so if the ouptut models are monophones and biphones then chances are these are the models you have in your MMF.   Of course, you can call the models what you like in the MMF if they are mapped via the tiedlist to get you the complete range of context dependent phones, so it may be doing what you want, but you have to check.

--- (Edited on 11-January-2013 3:56 pm [GMT+0000] by TonyR) ---

Re: Problem in Adapation using Hvite-HEadapt
User: Visitor
Date: 1/11/2013 12:15 pm
Views: 189
Rating: 7

I think I understand your question now. I looked at my hmmdefs and I could see hmm definition for monophones, biphones and triphones.

I mean examples of bellow exists in my hmmdef :

~h "a"

 

...

~h "a-b"

...

~h "a+b"

...

~h "a-b+c"

...

--- (Edited on 1/11/2013 12:15 pm [GMT-0600] by Visitor) ---

Re: Problem in Adapation using Hvite-HEadapt
User: TonyR
Date: 1/11/2013 8:45 pm
Views: 176
Rating: 9

Okay.  What is the content of your file "config"?   Does it have FORCECXTEXP  = TRUE?

I haven't seen your problem before but if you have all monophones and don't force context expansion then maybe HVite uses these models instead of the full word internal models.

 

Tony

-- 

Dr Tony Robinson
CEO Cantab Research Ltd
Phone +44(0)1223 794497

--- (Edited on 12-January-2013 2:45 am [GMT+0000] by TonyR) ---

Re: Problem in Adapation using Hvite-HEadapt
User: Visitor
Date: 1/12/2013 4:36 pm
Views: 187
Rating: 9

Thanks you very much for your reply.

The fist time I called hvite my config file didn't have FORCECXTEXP=T but in the last stage when I called hvite I had it.

if we consider the adaptation process as form below:

hvite--> hhed--> headapt--> hvite--> hresults

in the first 3 stage my config file was:

TARGETKIND = MFCC_D_A_C_K_Z_0
TARGETRATE = 100000.0
SAVECOMPRESSED = T
SAVEWITHCRC = T
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 20
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = F

and in stage 4 when I called hvite my config file was complete:

FORCECXTEXP = T
ALLOWXWRDEXP = F
TARGETKIND = MFCC_D_A_C_K_Z_0
TARGETRATE = 100000.0
SAVECOMPRESSED = T
SAVEWITHCRC = T
WINDOWSIZE = 250000.0
USEHAMMING = T
PREEMCOEF = 0.97
NUMCHANS = 20
CEPLIFTER = 22
NUMCEPS = 12
ENORMALISE = F

 

 

 

 

--- (Edited on 1/12/2013 4:36 pm [GMT-0600] by Visitor) ---

Re: Problem in Adapation using Hvite-HEadapt
User: mvoxforge
Date: 1/12/2013 5:16 pm
Views: 200
Rating: 10

I changed all my conf file so they include FORCECXTEXP=T

I run a few experiments, and my adaptphon.mlf file has changed. now I can see my word internal triphones but they dont look alright.

For instance for word 'looked: l uh k t sp'

my triphone looks like :


l+ao
l-uh+k
uh-k+k
k-t
sp

and for word 'lands: 'l ae n d z sp'

my triphone looks like :

l+ae
l-ae+m
ae-n+d
n-d+s
dh-z
sp

 

 

--- (Edited on 1/12/2013 5:16 pm [GMT-0600] by mvoxforge) ---

Re: Problem in Adapation using Hvite-HEadapt
User: TonyR
Date: 1/13/2013 2:46 am
Views: 212
Rating: 11

Why don't these look right to you?    You have the right central phone in all cases and you have word internal triphones, so they look rigtht to me.   For example, you'll see that l+ao aliases to l+uk in your tiedlist.

Also I note that you are using HEadapt, so you must be using a very old version of HTK.   You might want to upgrade to version 3.4.1.

 

Tony

 

--- (Edited on 13-January-2013 8:46 am [GMT+0000] by TonyR) ---

Re: Problem in Adapation using Hvite-HEadapt
User: Visitor
Date: 1/13/2013 5:46 am
Views: 2485
Rating: 10

Thank you. I have checked my tiedlist and what you said was true.

Below are the reasons I didnt chose HTK3.4 and used HTK 3.2.1 instead.

1) I thought it is a more stable version and there might be bugs in 3.4. 

2) I also had lots of attempts to use HDecoder in 3.4 and I was not successful so I decided to permanently switch to 3.2.

3) My goal is to compare MLLR and MAP adaptation to other adaptation techniques. So I thought as long as I can do these tasks in 3.2 I dont need to use 3.4.

Thank you very much for your very useful comments.

M.N

 

--- (Edited on 1/13/2013 5:46 am [GMT-0600] by Visitor) ---

PreviousNext