[hts-users:04526] adaptation failed in HTS-demo_CMU-ARCTIC-ADAPT

[Zhen Wei <zhwei@xxxxxxxxxxxxx>]

Hi all!

These days I use HTS-demo_CMU-ARCTIC-ADAPT but vctk database to do the hmm adaptation. the labels and cmp (25d-mgc+delta+delta delta, 5d-bap+delta+delta delta, 1d-lf0+delta+delta delta,1d-u/v, 94d in total) of all speakers are prepared outside this demo by myself, so the f0 range is uniform for male and uniform for female.

The procedure with all the switches on except FALGN MCDGV MKUNG (useGV=0) can go through successfully, but the results are not expected ones. The results of SI and SAT should be the average voice, but my results sound that you can specify the man voice and women voice in one sentence, which means that they seem to be hard to merge with each other. Meanwhile, the results of SI+dec_feat3 and SAT+dec_feat3 should be the adaptive results, but my results sound have no difference with SI and SAT, which means that the adaptation training failed.

Some configuration is listed as follows:
$usestraight = '1';

@SET = ('cmp','dur');
if(!$usestraight){
    @cmp = ('mgc','lf0');
}else{
    @cmp = ('mgc','lf0','bap');

%mdlf = ('mgc' => '1.0', # tree size control param. for MDL
         'lf0' => '1.0',
         'bap' => '1.0',
         'dur' => '1.0');

%mocc = ('mgc' => '10.0', # minimum occupancy counts
         'lf0' => '10.0',
         'bap' => '10.0',
         'dur' => ' 5.0');

%gam = ('mgc' => '000', # stats load threshold
         'lf0' => '000',
         'bap' => '000',
         'dur' => '000');

%msdi = ('mgc' => '0', # msd information
         'lf0' => '1',
         'bap' => '0',
         'dur' => '0');

%ordr = ('mgc' => '25', # feature order
         'lf0' => '1',
         'bap' => '5',
         'dur' => '5');

# Speaker adaptation Setting ============
$spkrPat = "\"*/*_*\""; # speaker name pattern

# regression classes
%dect = ('mgc' => '500.0', # occupancy thresholds for regression classes (dec)
         'lf0' => '100.0', # set thresholds in less than adpt and satt
         'bap' => '100.0',
         'dur' => '100.0');

$nClass = 32; # number of regression classes (reg)

# transforms
%nblk = ('mgc' => '3', # number of blocks for transforms
         'lf0' => '1',
         'bap' => '3',
         'dur' => '1');

%band = ('mgc' => '25', # band width for transforms
         'lf0' => '1',
         'bap' => '5',
         'dur' => '0');

$bias{'cmp'} = 'TRUE'; # use bias term for MLLRMEAN/CMLLR
$bias{'dur'} = 'TRUE';
$tran = 'feat'; # transformation kind (mean -> MLLRMEAN, cov -> MLLRCOV, or feat -> CMLLR)

# adaptation
%adpt = ('mgc' => '500.0', # occupancy thresholds for adaptation
         'lf0' => '100.0',
         'bap' => '100.0',
         'dur' => '100.0');

$tknd{'adp'} = 'dec'; # tree kind (dec -> decision tree or reg -> regression tree (k-means))
$dcov = 'FALSE'; # use diagonal covariance transform for MLLRMEAN
$usemaplr = 'TRUE'; # use MAPLR adaptation for MLLRMEAN/CMLLR
$usevblr = 'FALSE'; # use VBLR adaptation for MLLRMEAN
$sprior = 'TRUE'; # use structural prior for MAPLR/VBLR with regression class tree
$priorscale = 1.0; # hyper-parameter for SMAPLR adaptation
$nAdapt = 3; # number of iterations to reestimate adaptation xforms
$addMAP = 1; # apply additional MAP estimation after MLLR adaptation
$maptau{'cmp'} = 50.0; # hyper-parameters for MAP adaptation
$maptau{'dur'} = 50.0;

# speaker adaptive training
%satt = ('mgc' => '10000.0', # occupancy thresholds for adaptive training
         'lf0' => '2000.0',
         'bap' => '2000.0',
         'dur' => '5000.0');

$tknd{'sat'} = 'dec'; # tree kind (dec -> decision tree or reg -> regression tree (k-means))
$nSAT = 3; # number of SAT iterations

I'm really sorry to bother you! Can anyone help me with these problems? I don't know why these happened cause I just changed the data and modify some path according to the data.

Best regards,
zhen wei