- HTKBooks,
- 苏统华, 哈尔滨工业大学人工智能研究室, 2006年10月30日,
- Howard Hung-Ju Chou, Intelligence Information Retrieval Lab., NCKU, Taiwan(R.O.C.).
- HTK 3.4
- Cygwin NT-5.1 1.5.25
- Step 6 - Creating Flat Start Monophones
- Step 7 - Fixing the Slience Models
- Step 8 - Realigning the Training Data
Now the tutorial teaching us to model the model in Fig. 3.9 in Subsection 3.2.2 in HTK Books.
Physical meaning,....
Create a 3 state model for "sp", so we just only one non-emtting state for "sp" model.
How to do that?
- Use text editor to...
- Use HHEd
The content of "sil" is
=================================================================
~h "sil"
< BEGINHMM >
< NUMSTATES > 5
< STATE > 2
<> 39
-9.389361e-001 -1.287944e+000 8.473723e-002 -4.411200e+000 5.332393e-001 1.843251e-001 2.939802e+000 -2.362492e+000 3.039350e-001 5.898609e-003 -3.105349e+000 -1.462931e+000 5.539479e+001 -2.752953e-002 -2.782337e-002 5.648132e-003 4.534409e-002 1.876847e-002 2.492056e-002 1.361921e-002 -1.723138e-002 1.886967e-002 3.497830e-002 1.276191e-002 2.784961e-002 -3.208526e-002 3.180009e-004 1.971325e-003 -3.830043e-003 -1.048350e-002 -1.810746e-003 -1.773861e-003 -9.375007e-004 3.254613e-004 8.180511e-004 3.765909e-003 1.624564e-003 -3.620259e-004 3.278390e-003
< VARIANCE > 39
4.180664e+001 3.271134e+001 3.581472e+001 6.693031e+001 3.528064e+001 5.052157e+001 2.934049e+001 3.423428e+001 3.710680e+001 3.691701e+001 3.710829e+001 2.969890e+001 6.507053e+001 1.337987e+000 1.072017e+000 1.307518e+000 1.887581e+000 1.697909e+000 1.890241e+000 1.785829e+000 2.181937e+000 1.875866e+000 1.797650e+000 1.730149e+000 1.642454e+000 9.926388e-001 1.710149e-001 1.462792e-001 1.848170e-001 2.626473e-001 2.605441e-001 2.970572e-001 3.222261e-001 3.782587e-001 3.125882e-001 3.063583e-001 2.895371e-001 2.911398e-001 1.187189e-001
< GCONST > 1.071964e+002
< STATE > 3
< MEAN > 39
-1.991913e+000 -4.775551e-002 2.959489e+000 2.209434e+000 2.078557e+000 5.562240e+000 5.464221e+000 -4.776323e+000 1.673594e+000 2.683963e+000 -4.633354e+000 -9.166243e-001 4.628856e+001 -1.207492e-001 -8.760695e-002 -7.070365e-002 7.516075e-002 -4.011013e-003 3.128541e-002 8.115381e-002 -3.286631e-002 1.295639e-001 1.558424e-001 5.380721e-002 1.054287e-001 -1.449030e-001 1.667164e-002 2.022874e-002 1.105829e-003 -2.183086e-002 -7.496935e-003 -4.172942e-002 -3.657551e-002 1.193289e-002 -1.476659e-002 -2.710904e-002 1.349834e-002 9.330045e-004 2.211097e-002
< VARIANCE > 39
5.752877e+000 5.706749e+000 9.791572e+000 1.276698e+001 1.414043e+001 1.682921e+001 1.643664e+001 1.884838e+001 1.942560e+001 2.041147e+001 1.927709e+001 1.510888e+001 1.051241e+001 2.168639e-001 3.732721e-001 6.485465e-001 8.246439e-001 9.308486e-001 1.138545e+000 1.447520e+000 1.688959e+000 1.681041e+000 1.680561e+000 1.580671e+000 1.330634e+000 9.859556e-002 3.477598e-002 6.478215e-002 1.191088e-001 1.600942e-001 1.801341e-001 2.153407e-001 2.852951e-001 3.301157e-001 3.403606e-001 3.369383e-001 3.197604e-001 2.676942e-001 1.323066e-002
< STATE > 4
< MEAN > 39
-2.982345e+000 -1.252340e+000 1.087486e+000 7.909203e-001 1.536108e+000 3.573169e+000 5.625374e+000 -3.234990e+000 2.314626e+000 3.188504e+000 -9.258319e-001 1.509047e+000 4.699720e+001 -7.613304e-003 5.702919e-003 -6.563795e-003 -4.346590e-003 -7.446251e-003 -8.997340e-003 -3.822424e-003 -2.726374e-003 -3.682886e-003 -1.174716e-003 1.001520e-002 1.304566e-002 -2.283418e-003 -2.802775e-004 1.980037e-003 1.587337e-003 -6.755204e-004 2.919145e-003 1.646213e-003 -1.079046e-004 1.305768e-003 2.884402e-004 -2.650670e-003 -2.699222e-003 -4.054980e-003 3.949025e-003
< VARIANCE > 39
5.313723e+000 4.299637e+000 5.806711e+000 7.572632e+000 1.195562e+001 1.127259e+001 1.345822e+001 1.842092e+001 1.902783e+001 1.841946e+001 1.679353e+001 1.275744e+001 2.541775e+000 1.125962e-001 2.241242e-001 3.554686e-001 4.804470e-001 7.102868e-001 8.679712e-001 1.053879e+000 1.259253e+000 1.247817e+000 1.199414e+000 1.138910e+000 9.791774e-001 7.236452e-002 2.274701e-002 4.417740e-002 7.068438e-002 9.644291e-002 1.455498e-001 1.809241e-001 2.171511e-001 2.593471e-001 2.625059e-001 2.464305e-001 2.333392e-001 2.013770e-001 1.429966e-002
< GCONST > 6.495581e+001
< TRANSP > 5
0.000000e+000 1.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000
0.000000e+000 9.399074e-001 6.009261e-002 0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 8.703428e-001 1.296572e-001 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 9.800954e-001 1.990458e-002
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000
< ENDHMM >
=================================================================
We copy the red statements to be model of "sp", but remember that only 3 state exist in model "sp" so we have to modify to 2.
=======================================================================
~h "sp"
< BEGINHMM >
< NUMSTATES > 3
< STATE > 2
< MEAN > 39
-9.389361e-001 -1.287944e+000 8.473723e-002 -4.411200e+000 5.332393e-001 1.843251e-001 2.939802e+000 -2.362492e+000 3.039350e-001 5.898609e-003 -3.105349e+000 -1.462931e+000 5.539479e+001 -2.752953e-002 -2.782337e-002 5.648132e-003 4.534409e-002 1.876847e-002 2.492056e-002 1.361921e-002 -1.723138e-002 1.886967e-002 3.497830e-002 1.276191e-002 2.784961e-002 -3.208526e-002 3.180009e-004 1.971325e-003 -3.830043e-003 -1.048350e-002 -1.810746e-003 -1.773861e-003 -9.375007e-004 3.254613e-004 8.180511e-004 3.765909e-003 1.624564e-003 -3.620259e-004 3.278390e-003
< VARIANCE > 39
4.180664e+001 3.271134e+001 3.581472e+001 6.693031e+001 3.528064e+001 5.052157e+001 2.934049e+001 3.423428e+001 3.710680e+001 3.691701e+001 3.710829e+001 2.969890e+001 6.507053e+001 1.337987e+000 1.072017e+000 1.307518e+000 1.887581e+000 1.697909e+000 1.890241e+000 1.785829e+000 2.181937e+000 1.875866e+000 1.797650e+000 1.730149e+000 1.642454e+000 9.926388e-001 1.710149e-001 1.462792e-001 1.848170e-001 2.626473e-001 2.605441e-001 2.970572e-001 3.222261e-001 3.782587e-001 3.125882e-001 3.063583e-001 2.895371e-001 2.911398e-001 1.187189e-001
< GCONST > 1.071964e+002
< TRANSP > 3
0.000000e+000 1.000000e+000 0.000000e+000
0.000000e+000 8.703428e-001 1.296572e-001
0.000000e+000 0.000000e+000 0.000000e+000
< ENDHMM >
=======================================================================
Then use HHEd to modefy hmm4/macros and hmm4/hmmdefs accroding the edit commands in sil.hed.
--------------------------------------------------------------------------------------------------------------------
$ HHEd -H ./hmms/hmm4/macros -H ./hmms/hmm4/hmmdefs -M ./hmms/hmm5 sil.hed ./lists/monophones1
--------------------------------------------------------------------------------------------------------------------
The new hmmdefs, will be added the following statements,
======================================================================================
~s "silst"
< MEAN > 39
-9.389361e-01 -1.287944e+00 8.473723e-02 -4.411200e+00 5.332393e-01 1.843251e-01 2.939802e+00 -2.362492e+00 3.039350e-01 5.898609e-03 -3.105349e+00 -1.462931e+00 5.539479e+01 -2.752953e-02 -2.782337e-02 5.648132e-03 4.534409e-02 1.876847e-02 2.492056e-02 1.361921e-02 -1.723138e-02 1.886967e-02 3.497830e-02 1.276191e-02 2.784961e-02 -3.208526e-02 3.180009e-04 1.971325e-03 -3.830043e-03 -1.048350e-02 -1.810746e-03 -1.773861e-03 -9.375007e-04 3.254613e-04 8.180511e-04 3.765909e-03 1.624564e-03 -3.620259e-04 3.278390e-03
4.180664e+01 3.271134e+01 3.581472e+01 6.693031e+01 3.528064e+01 5.052157e+01 2.934049e+01 3.423428e+01 3.710680e+01 3.691701e+01 3.710829e+01 2.969890e+01 6.507053e+01 1.337987e+00 1.072017e+00 1.307518e+00 1.887581e+00 1.697909e+00 1.890241e+00 1.785829e+00 2.181937e+00 1.875866e+00 1.797650e+00 1.730149e+00 1.642454e+00 9.926388e-01 1.710149e-01 1.462792e-01 1.848170e-01 2.626473e-01 2.605441e-01 2.970572e-01 3.222261e-01 3.782587e-01 3.125882e-01 3.063583e-01 2.895371e-01 2.911398e-01 1.187189e-01
< GCONST > 1.071964e+02
======================================================================================
And original ~h "sil" and ~h "sp" become to be like following,
====================================================
~h "sp"
~s "silst"
0.000000e+00 7.000000e-01 3.000000e-01
0.000000e+00 8.703428e-01 1.296572e-01
0.000000e+00 0.000000e+00 0.000000e+00
~h "sil"
< BEGINHMM >
< NUMSTATES > 5
< STATE > 2
< MEAN > 39
-9.389361e-01 -1.287944e+00 8.473723e-02 -4.411200e+00 5.332393e-01 1.843251e-01 2.939802e+00 -2.362492e+00 3.039350e-01 5.898609e-03 -3.105349e+00 -1.462931e+00 5.539479e+01 -2.752953e-02 -2.782337e-02 5.648132e-03 4.534409e-02 1.876847e-02 2.492056e-02 1.361921e-02 -1.723138e-02 1.886967e-02 3.497830e-02 1.276191e-02 2.784961e-02 -3.208526e-02 3.180009e-04 1.971325e-03 -3.830043e-03 -1.048350e-02 -1.810746e-03 -1.773861e-03 -9.375007e-04 3.254613e-04 8.180511e-04 3.765909e-03 1.624564e-03 -3.620259e-04 3.278390e-03
< VARIANCE > 39
4.180664e+01 3.271134e+01 3.581472e+01 6.693031e+01 3.528064e+01 5.052157e+01 2.934049e+01 3.423428e+01 3.710680e+01 3.691701e+01 3.710829e+01 2.969890e+01 6.507053e+01 1.337987e+00 1.072017e+00 1.307518e+00 1.887581e+00 1.697909e+00 1.890241e+00 1.785829e+00 2.181937e+00 1.875866e+00 1.797650e+00 1.730149e+00 1.642454e+00 9.926388e-01 1.710149e-01 1.462792e-01 1.848170e-01 2.626473e-01 2.605441e-01 2.970572e-01 3.222261e-01 3.782587e-01 3.125882e-01 3.063583e-01 2.895371e-01 2.911398e-01 1.187189e-01
< GCONST > 1.071964e+02
~s "silst"
< STATE > 4
< MEAN > 39
-2.982345e+00 -1.252340e+00 1.087486e+00 7.909203e-01 1.536108e+00 3.573169e+00 5.625374e+00 -3.234990e+00 2.314626e+00 3.188504e+00 -9.258319e-01 1.509047e+00 4.699720e+01 -7.613304e-03 5.702919e-03 -6.563795e-03 -4.346590e-03 -7.446251e-03 -8.997340e-03 -3.822424e-03 -2.726374e-03 -3.682886e-03 -1.174716e-03 1.001520e-02 1.304566e-02 -2.283418e-03 -2.802775e-04 1.980037e-03 1.587337e-03 -6.755204e-04 2.919145e-03 1.646213e-03 -1.079046e-04 1.305768e-03 2.884402e-04 -2.650670e-03 -2.699222e-03 -4.054980e-03 3.949025e-03
< VARIANCE > 39
5.313723e+00 4.299637e+00 5.806711e+00 7.572632e+00 1.195562e+01 1.127259e+01 1.345822e+01 1.842092e+01 1.902783e+01 1.841946e+01 1.679353e+01 1.275744e+01 2.541775e+00 1.125962e-01 2.241242e-01 3.554686e-01 4.804470e-01 7.102868e-01 8.679712e-01 1.053879e+00 1.259253e+00 1.247817e+00 1.199414e+00 1.138910e+00 9.791774e-01 7.236452e-02 2.274701e-02 4.417740e-02 7.068438e-02 9.644291e-02 1.455498e-01 1.809241e-01 2.171511e-01 2.593471e-01 2.625059e-01 2.464305e-01 2.333392e-01 2.013770e-01 1.429966e-02
< GCONST > 6.495583e+01
< TRANSP > 5
0.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00
0.000000e+00 7.519259e-01 4.807409e-02 2.000000e-01 0.000000e+00
0.000000e+00 0.000000e+00 8.703428e-01 1.296572e-01 0.000000e+00
0.000000e+00 2.000000e-01 0.000000e+00 7.840764e-01 1.592367e-02
0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00
< ENDHMM >
====================================================
Because the commands in sil.hed,
============================
AT 2 4 0.2 { sil. transp }
AT 4 2 0.2 { sil. transp }
AT 1 3 0.3 { sp. transp }
TI slist { sil.state[3], sp.state[2] }
============================
AT i j prob itemList(t) in page 256 in HTKBook 3.4.
The probability will be rescaled so that summation of p is equal to 1.0.
For example in ~h "sp",
We modify it by AT 1 3 0.3 { sp.transp }, so the sp.transp
==================================
< TRANSP > 3
0.000000e+000 1.000000e+000 0.000000e+000
0.000000e+000 8.703428e-001 1.296572e-001
0.000000e+000 0.000000e+000 0.000000e+000
==================================
to be rescaled to
==================================
< TRANSP > 3
0.000000e+00 7.000000e-01 3.000000e-01
0.000000e+00 8.703428e-01 1.296572e-01
0.000000e+00 0.000000e+00 0.000000e+00
==================================
TI means Tie itemlist to be macroname.
-------------------------------------
$ Ti macroname itemlist
-------------------------------------
To know more about TI command, please refer to Chapter 10.3 in HTK Book. (PS: the section 10.3 and section 10.4 should be exchanged.)
To know more about HHed, please refer to Chapter 10 in HTK Book.
Continue...
2 comments:
hye... Howard
i hope you can help me.. Actually im stuck at step 6, which is about http://jrgemini.blogspot.com/2009/02/htk-chapter-3-section-2-step-6.html
you had state that the problem of: ERROR [+6510] LOpen: Unable to open label file .\data\train\feature\S0001.lab
FATAL ERROR - Terminating program HERest..
is that mean that we have to create .lab files for all the content inside phones0.mlf. my phones0.mlf content is about 200, so is it i have to create 200 .lab files???
i hope you can help me..
Post a Comment