Hi
I’ve been trying to build a regression tree without using decision tree for speaker adaptation with SI model.
So, I changed Config.pm file (‘dec’ -> ‘reg’).
%tknd = ('adp' => 'reg', # tree kind (reg or dec)
And I wanted to reduce # of regression tree, so I changed Config.pm file($nClass = 32 -> 2)
$nClass = 2; # number of regression classes
The problem is that the output files (mapmmf/re_clustered_all.mmf, synthesized wave files) using 2 regression classes were same as those using 32 regression classes.
Is there more option that should be changed to fix this problem?
And another problem is that no new files in xforms or anywhere else. Is it normal situation?
I’ve been trying to see Adaptation Transform from xforms, but I couldn’t see anything about it.
I’m using HTS 2.2.
Log file about this below:
======================================================================================
Start building regression-class trees for adaptation at Thu Mar 8 03:26:09 KST 2012
======================================================================================
HHEd -A -C configs/trn.cnf -D -T 1 -p -i -C configs/dec_cmp.cnf -H models/qst001/ver1/cmp/re_clustered.mmf -M models/qst001/ver1/cmp/regTrees edfiles/qst001/ver1/cmp/reg.hed data/lists/full.list
HTK Configuration Parameters[9]
Module/Tool Parameter Value
# SHRINKOCCTHRESH Vector 5 500.0 100.0 100.0 100.0 100.0
# MINDUR 5
# MAXSTDDEVCOEF 10
# APPLYDURVARFLOOR TRUE
# DURVARFLOORPERCENTILE 1.000000
# VFLOORSCALESTR Vector 5 0.01 0.01 0.01 0.01 0.01
# NATURALWRITEORDER TRUE
# NATURALREADORDER TRUE
# APPLYVFLOOR TRUE
HHEd
108964/108964 Models Loaded [7 states max, 2 mixes max]
// load stats file
LS stats/qst001/ver1/cmp.stats.re-clustered
Loading state occupation stats
Stats loaded for 108964 models
Mean Occupation Count = 12.104720
// construct regression class tree
RC 2 reg
Building regression tree with 2 terminals (5 streams)
Creating regression class tree with ident reg.tree and baseclass reg.base
Splitting Node 1, score 1.000000e+10
(Stream splitting)
Splitting Node 3, score 1.000000e+10
(Stream splitting)
Splitting Node 5, score 1.000000e+10
(Stream splitting)
…
stream 1: #terminals=2
stream 2: #terminals=2
stream 3: #terminals=2
stream 4: #terminals=2
stream 5: #terminals=2
Edit Complete
================================================================================
Start speaker adaptation (speaker independent) at Thu Mar 8 03:46:35 KST 2012
================================================================================
Iteration 1 of adaptation transform reestimation
HERest -A -C configs/trn.cnf -D -T 1 -S data/scp/adapt.scp -I data/labels/full.mlf -m 1 -u ada -w 3 -t 1500 100 5000 -h */phj* -H models/qst001/ver1/cmp/re_clustered_all.mmf -N models/qst001/ver1/dur/re_clustered_all.mmf -C configs/adp.cnf -K models/qst001/ver1/cmp/xforms SI+reg_feat1 -H models/qst001/ver1/cmp/regTrees/reg.base -H models/qst001/ver1/cmp/regTrees/reg.tree -Z models/qst001/ver1/dur/xforms SI+reg_feat1 -N models/qst001/ver1/dur/regTrees/reg.base -N models/qst001/ver1/dur/regTrees/reg.tree models/qst001/ver1/cmp/tiedlist models/qst001/ver1/dur/tiedlist
HTK Configuration Parameters[26]
Module/Tool Parameter Value
# BANDWIDTH IntVec 3 40 40 40 IntVec 1 1 IntVec 1 1 IntVec 1 1 IntVec 3 5 5 5
# BLOCKSIZE IntVec 3 40 40 40 IntVec 1 1 IntVec 1 1 IntVec 1 1 IntVec 3 5 5 5
# DURSPLITTHRESH Vector 5 100.0 100.0 100.0 100.0 100.0
# SPLITTHRESH Vector 5 500.0 100.0 100.0 100.0 100.0
# KEEPXFORMDISTINCT TRUE
# SMAPSIGMA 1
# USESMAP TRUE
# MLLRDIAGCOV FALSE
# DURUSEBIAS TRUE
# DURTRANSKIND CMLLR
# USEBIAS TRUE
# TRANSKIND CMLLR
# DURREGTREE /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/dur/regTrees/reg.tree
# DURBASECLASS /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/dur/regTrees/reg.base
# REGTREE /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/cmp/regTrees/reg.tree
# BASECLASS /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/cmp/regTrees/reg.base
# DURADAPTKIND TREE
# ADAPTKIND TREE
# MINDUR 5
# MAXSTDDEVCOEF 10
# APPLYDURVARFLOOR TRUE
# DURVARFLOORPERCENTILE 1.000000
# VFLOORSCALESTR Vector 5 0.01 0.01 0.01 0.01 0.01
# NATURALWRITEORDER TRUE
# NATURALREADORDER TRUE
# APPLYVFLOOR TRUE
HMM:
HERest ML Updating: XForms
System is SHARED
110824 Logical/87000 Physical Models Loaded, VecSize=138
3 MMF input files
Duration model:
HERest ML Updating: XForms
System is SHARED
110824 Logical/3798 Physical Models Loaded, #States=5
3 MMF input files
Pruning-On[1500.0 100.0 5000.0]
…
========================================================================
Start speaker adaptation (SI+MLLR+MAP) at Thu Mar 8 03:56:20 KST 2012
========================================================================
HERest -A -B -C configs/trn.cnf -D -T 1 -S data/scp/adapt.scp -I data/labels/full.mlf -m 1 -u pmvwdpmv -w 3 -t 1500 100 5000 -h */phj* -H models/qst001/ver1/cmp/re_clustered_all.mmf -N models/qst001/ver1/dur/re_clustered_all.mmf -M models/qst001/ver1/cmp/mapmmf -R models/qst001/ver1/dur/mapmmf -C configs/map.cnf -H models/qst001/ver1/cmp/regTrees/reg.base –H models/qst001/ver1/cmp/regTrees/reg.tree -J models/qst001/ver1/cmp/xforms SI+reg_feat6 -E models/qst001/ver1/cmp/xforms SI+reg_feat6 -a -N models/qst001/ver1/dur/regTrees/reg.base -N models/qst001/ver1/dur/regTrees/reg.tree -Y models/qst001/ver1/dur/xforms SI+reg_feat6 -W models/qst001/ver1/dur/xforms SI+reg_feat6 -b models/qst001/ver1/cmp/tiedlist models/qst001/ver1/dur/tiedlist
HTK Configuration Parameters[13]
Module/Tool Parameter Value
# HMAP TRACE 2
# MIXWEIGHTFLOOR 3
# MINEGS 1
# DURMAPTAU 50
# MAPTAU 50
# MINDUR 5
# MAXSTDDEVCOEF 10
# APPLYDURVARFLOOR TRUE
# DURVARFLOORPERCENTILE 1.000000
# VFLOORSCALESTR Vector 5 0.01 0.01 0.01 0.01 0.01
# NATURALWRITEORDER TRUE
# NATURALREADORDER TRUE
# APPLYVFLOOR TRUE
HMM:
HERest MAP Updating: Means Variances MixWeights
System is SHARED
110824 Logical/87000 Physical Models Loaded, VecSize=138
3 MMF input files
Duration model:
HERest MAP Updating: Means Variances
System is SHARED
110824 Logical/3798 Physical Models Loaded, #States=5
3 MMF input files
Pruning-On[1500.0 100.0 5000.0]
…
Saving hmm's to dir /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/cmp/mapmmf
Saving duration models to dir /home/htk/Database/Heamin/HTS/HTS_AD_K_STRAIGHT_2_2_m/models/qst001/ver1/dur/mapmmf
Reestimation complete - average log prob per frame = 9.669560e+01
- total frames seen = 6.099200e+04