I recommend you to use latest version

As for 3.2.1 it expects you both -k and -K when -K should be the name of the transform and -k argument should have special format:

-k UID <user_id>

note that it has two strings separated by space as argument.

--- (Edited on 6/16/2009 10:41 am [GMT-0500] by nsh) ---

Also you need either rtree or this patch applied

http://htk.eng.cam.ac.uk/pipermail/htk-developers/2003-October/000508.html

--- (Edited on 6/16/2009 12:28 pm [GMT-0500] by nsh) ---

Hi, I have 2 issues. One with 3.2.1 and the other with 3.4

3.2.1

1.@nsh: even after I apply the patch, I get this message

-------------------------------------------------------------------------------

>>/home/user/htk/bin.linux/HVite -S adapt.scp -H hmmdefs -H macros -K trans.tmf -k UID unknown -j 1 -i results -w wdnet dict tiedlist
  ERROR [+7421]  InitialiseRegClasses: MMF does not contain any identifier;
Use HHEd to generate one

-----------------------------------------------------------------------------

i think it is with the -k tag. i just gave a random name "unknown" for my UID. What am I supposed to do with the -k ?Also,  -K tag is just for storing the output transform right?

3.4

2. I am okay with using the 3.4 version also, but there were no complete tutorials on adaptation. The one in the HTKbook gives me errors(in step 13).

--------------------------------------------------------------------------

>>/home/user/Desktop/tmp/bin/HERest -C config -C config.global -S adapt.scp -I adaptPhones.mlf -H hmm15/macros -u a -H hmm15/hmmdefs -z -K xforms mllr1 -J classes -h ’*/%%%%%%_*.mfc’ -K trans.tmf dict tiedlist
  ERROR [+2319]  HERest: output TMF file expected
 FATAL ERROR - Terminating program /home/user/Desktop/tmp/bin/HERest
-----------------------------------------------------------------------

So, if there is a good tutorial you could suggest for speaker adaptation using the 3.4 version, I am fine with using that also.

Anything is fine as long as I am able to do unsupervised adaptation!

Thanks a lot!

Cheers!

--- (Edited on 6/17/2009 1:40 am [GMT-0500] by Visitor) ---

> 3.2.1

The patch is actually to HEAdapt, minor bit for HVite is required.

Otherwise as I wrote you still need regression class tree in model definition, it doesn't create one itself. You need to follow the initial step of supervised adaptation to create a tree in the model and then you can continue with unsupervised.

> 3.4 

I think it's easier and you can always help to improve HTK book :)

ERROR [+2319]  HERest: output TMF file expected

It waits for the argument for -z.

 -z s    Save all xforms to TMF file s        TMF

--- (Edited on 6/17/2009 11:04 pm [GMT-0500] by nsh) ---

Hi,

thanks for that. i decided to get the 3.2.1 working first since it seems that it might be the first one to work :)

>>You need to follow the initial step of supervised adaptation to create a tree in the model

I did this and stored the new models with regression tree info. in hmm17/

when I then attempted to use HVite for incremental adaptation,

--------------------------------------------------------------------

>>/home/user/htk/bin.linux/HVite -S adapt.scp -H hmm17/hmmdefs -H hmm17/macros -K trans.tmf -k UID unknown -j 1 -i results dict tiedlist -C config
Pruning-On[20000000000.0 20000000000.0 40000000001.0]
  ERROR [+6510]  LOpen: Unable to open label file ../test/1.lab

FATAL ERROR - Terminating program /home/user/htk/bin.linux/HVite

---------------------------------------------------------------------

why is it expecting a label file here(i do not have transcriptions for my adaptation data)?

and if I include my word net file, i get this message

-------------------------------------------------------------------

/home/user/htk/bin.linux/HVite -S adapt.scp -H hmm17/hmmdefs -H hmm17/macros -K trans.tmf -k UID unknown -w wdnet -j 1 dict tiedlist -C config
Pruning-On[20000000000.0 20000000000.0 40000000001.0]
  ERROR [+7332]  CreateInsts: Cannot have Tee models at start or end of transcription
 FATAL ERROR - Terminating program /home/user/htk/bin.linux/HVite
-------------------------------------------------------------------

also

>>minor bit for HVite is required.

is there a patch for HVite also? pl. direct me to that if there is one

pl. suggest possible solutions to the label file problem..

Thanks again!

--- (Edited on 6/18/2009 12:38 am [GMT-0500] by Visitor) ---

you need wdnet. about tee models, your dict should have no sp

--- (Edited on 6/18/2009 3:46 am [GMT-0500] by nsh) ---

Hi,

ya.. i could figure that out on my own(could have been more patient in posting).. I could generate the tmf file also. But im now trying to find out how to use them to transform the model.

Can HERest work without transcriptions? Or what tool must I use for that?

An example to use a tmf file to transform mmf files would indeed be useful.

Cheer

--- (Edited on 6/18/2009 3:59 am [GMT-0500] by Visitor) ---

ok.. Ive managed to do it finally :) though the results are very bad which I hope to look into soon.

So, this is the approach I took.

1. Use the SI models that I downloaded

2. Create new models with reg tree information

3. Use HVite with its -K, -k and -j options to create the *.tmf file

4. then to recognise, use HVite and the -J option to load the transform along with the MMF files

Both the adaptation and test data contain digit utterances and there are a very few files(8 to adapt and 8 to test) to adapt to my voice(8 files with 5 to 6 digit utterances with pauses in between words)

Supervised adaptation accuracy= 81.4%

unsupervised adaptation accuracy= 44.19%(too low i guess)

I know the results are data specific but is this amount of disparity common?(given that I have used very less data to adapt)

Also, I attempted to do adaptation with phone level transcripts created by HVite and got  62.79% accuracy. Very strangely this rudimentary method with no confidence measurements seems to win out on the other unsupervised method(atleast in this case)

questions:

1. Are the steps I followed correct for unsupervised adaptation?

2. Any comments on the amount of adaptation data and the numbers I anaged to churn out?

Thanks a lot for your time! Really appreciate that

Cheer

--- (Edited on 6/18/2009 6:56 am [GMT-0500] by Visitor) ---

So, can anyone tell me if the method i took for unsupervised adaptation on the right lines??

 this is the approach I took.

1. Use the SI models that I downloaded from voxforge

2. Create new models with reg tree information using HHEd

3. Use HVite with its -K, -k and -j options to create the *.tmf transform file

4. then to recognise, use HVite and the -J *.tmf option to load the transform along with the -H hmmdefs -H macros option for MMF files

Pl. let me know if this approach is fine.

thanks a lot.

cheer

--- (Edited on 6/23/2009 11:57 pm [GMT-0500] by Visitor) ---

Y, it look correct

--- (Edited on 6/24/2009 2:11 am [GMT-0500] by nsh) ---