Hi everybody,

I would like to evaluate speaker adaptation on a digits corpus in German. I have trained using HTK, a monophone system on another german corpus which work quite well in term of accuracy on the digits corpus. But I have problems in making the adaptation process (MLLR) to run on this corpus ...

So I have multiple questions:

- In the doc and examples, most of the examples for speaker
adaptation given are with triphones models ... I suppose that adapting context independent monophone acoustic models is also possible ?

- If yes, here are the steps that have been followed (following htk doc):

  1- training of a 64 mixture monophone system (43 german phones - 3 states per phone - 64 mixture per states)
         HERest.exe -B -D -T 4 -L new-seg -S mfcc-list-all.scp -s
statphones -H .\hmm01\hmmdefs -M hmm02\ listphones.txt

  2- Tree clustering using following command:
        HHEd.exe -B -T 4 -H .\hmm02a\hmmdefs -M classes regtree.hed listphones.txt
        where regtree.hed is:
        LS "statsphone"
        RC 32 "rtree" {sil.state[2-4].mix}
        in the class directory I have a file named global
        ~b ``global''
        <MMFIDMASK> *
        <PARAMETERS> MIXBASE
        <NUMCLASSES> 1
        <CLASS> 1 {*.state[2-4].mix[1-64]}

        This is creating 3 different files which seems quite ok :
rtree.base; rtree.rtree and hmmdefs in the classes directory

HHEd
 43/43 Models Loaded [5 states max, 64 mixes max]

LS statsphone
 Loading state occupation stats
  Stats loaded for 43 models
  Mean Occupation Count = 59334.495429

RC 32 rtree
 Building regression tree with 32 terminals
Creating regression class tree with ident rtree.tree and baseclass
rtree.base
 Models
  sil
  1 models in itemlist
          sil.state[2].stream[1]
          sil.state[3].stream[1]
          sil.state[4].stream[1]
Splitting Node 3, score 1.007575e+008
Splitting Node 5, score 6.008558e+007
Splitting Node 6, score 3.361824e+007
Splitting Node 2, score 2.519530e+007
Splitting Node 4, score 2.258807e+007
Splitting Node 7, score 1.867613e+007
Splitting Node 8, score 1.807741e+007
Splitting Node 9, score 1.189350e+007
Splitting Node 13, score 1.053742e+007
Splitting Node 14, score 9.306629e+006
Splitting Node 12, score 8.674780e+006
Splitting Node 16, score 7.772738e+006
Splitting Node 17, score 7.701269e+006
Splitting Node 20, score 7.331825e+006
Splitting Node 15, score 7.193971e+006
Splitting Node 19, score 5.830185e+006
Splitting Node 30, score 4.630366e+006
Splitting Node 18, score 4.581620e+006
Splitting Node 22, score 4.436380e+006
Splitting Node 25, score 4.378486e+006
Splitting Node 29, score 3.968522e+006
Splitting Node 27, score 3.646834e+006
Splitting Node 23, score 3.525451e+006
Splitting Node 32, score 3.492046e+006
Splitting Node 26, score 3.205845e+006
Splitting Node 37, score 3.171115e+006
Splitting Node 35, score 2.855839e+006
Splitting Node 24, score 2.766869e+006
Splitting Node 33, score 2.633669e+006
Splitting Node 38, score 2.627900e+006

Saving new HMM files ...
Edit Complete

    3- Adaptation
    HERest.exe -L adapt-lab -C config.global -S 001.scp -u a -H
.\hmm02a\hmmdefs -J .\classes -K xform mllr list-phones.txt
  ERROR [+999]  Components missing from Base Class list (1548 8256)
  ERROR [+999]  BaseClass check failed
 FATAL ERROR - Terminating program HERest.exe

I have checked the following error and this seems to be due to the fact that mixture components were missing in the model; although after having checked it seems that the model is containing the right number  of phones with 64 mixtures each
    (running a command like cat .\hmm02a\hmmdefs | grep "MIXTURE" | wc -> gives a result of 8256 (43phones*3states*64 mixtures as expected)

I have to say that I don't understand what I'm doing wrong here ... Any help would be useful here.

Thanks for your help.

Best Regards.

--- (Edited on 8/13/2009 2:38 am [GMT-0500] by frenchfrog) ---