[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02253] Re: Error in train transitions matrix


Technically, state alignment using HSMM is not implemented.
Your .hed file converted HSMMs to HMMs.
Therefore, '-R dur.mmf', '-N durdir', and 'durlist' are not required for HERest.

Keiichiro Oura

Xingchi Xian wrote:
Hi all,

I want use the HSMM I've trained to do state alignment, thus I use Simon's recommendation to deal with this issue. I do this as follows.

1.Untie the shared transition stuctrue and retie it in terms of each phone, my .hed file is:
TR 2
UT {*.transP}
TI T_a {(*-a+*).transP}
TI T_ai {(*-ai+*).transP}
TI T_an {(*-an+*).transP}
TI T_ang {(*-ang+*).transP}
TI T_ao {(*-ao+*).transP}
AT 2 2 0.9 {*.transP}
AT 2 3 0.1 {*.transP}
AT 3 3 0.9 {*.transP}
AT 3 4 0.1 {*.transP}
AT 4 4 0.9 {*.transP}
AT 4 5 0.1 {*.transP}
AT 5 5 0.9 {*.transP}
AT 5 6 0.1 {*.transP}
AT 6 6 0.9 {*.transP}
AT 6 7 0.1 {*.transP}

There is no problem at this stage. The model file is stored in untiedt.mmf
2. Retrain the transition matrix using the commond below.

/usr/local/HTS-2.1/bin/HERest -A -B -C xxx/configs/trn.cnf -D -T 1 -S xxx/data/scp/train.scp -I xxx/data/labels/full.mlf -m 1 -u t -w 3 -t 1500 100 5000 -H xxx/models/qst001/ver1/cmp/untiedt.mmf -N xxx/models/qst001/ver1/dur/untied.mmf -M xxx/models/qst001/ver1/cmp -R xxx/models/qst001/ver1/dur -C xxx/configs/nvf.cnf -s xxx/stats/qst001/ver1/cmp.stats.untiedt -w 0.0 xxx/data/lists/full.list xxx/data/lists/full.list However, it shows that Processing Data: xxx_0001.cmp; Label xxx_0001.lab
Retrying Beta pass at 1600.0
Retrying Beta pass at 1700.0
Retrying Beta pass at 1800.0
Retrying Beta pass at 1900.0
Retrying Beta pass at 2000.0
Retrying Beta pass at 2100.0
Retrying Beta pass at 2200.0
Retrying Beta pass at 2300.0
 WARNING [-7324]  StepBack: Bad data or over pruning
ERROR [+7390]  StepAlpha: Alpha prune failed sq(43) > qHi(42) at time 1109

What's wrong with my manipulation? (Machine: 64bit Linux, 16GB mem)

Thanks for help!

Xian Xingchi

[hts-users:02254] Re: Error in train transitions matrix, Xingchi Xian
[hts-users:02250] Error in train transitions matrix, Xingchi Xian