[hts-users:02245] Re: HERest: 7031 Transition mat all zero

["Lee Siu Wa" <yswlee@xxxxxxxxxxxxxx>]

Hi,

Thanks for your suggestions.

I have tried using '-u mvwdmv' option, the estimation completed for this 
HERest, without error.  However, when I proceed to the next HERest using '-u 
mvwdmv' option again, error occurs:

ERROR [+7390]  StepAlpha: Alpha prune failed eq(5) < sq(6) at time 169
FATAL ERROR - Terminating program /usr/local/HTS-2.1/bin/HERest


Why and how can I resolve this problem?
Is my case different from the one listed in 
http://hts.sp.nitech.ac.jp/hts-users/spool/2009/msg00103.html ERROR [+7390] 
StepAlpha: Alpha prune failed eq(21) < sq(22) at time 403?


Yes, I finished the estimation of monophone models (HInit, HRest, and HERest 
for spec. and dur.), and also triphone models (HERest before decision tree) 
successfully, but when I checked the transition matrices in monophone 
spectrum models given by the 1st HERest ('-u tmvwdmv' option) after HRest 
and HHEd, the transition matrices change from

TRANSP> 7
0.000000e+000 1.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000   <- result after HRest and HHEd
0.000000e+000 6.843743e-001 3.156257e-001 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 8.400140e-001 1.599861e-001 0.000000e+000 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 8.802072e-001 1.197928e-001 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 7.780808e-001 
2.219192e-001 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
6.994264e-001 3.005736e-001
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000
<ENDHMM>

to

<TRANSP> 7
0.000000e+000 1.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000  <- result after 1st HERest
0.000000e+000 0.000000e+000 1.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 1.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 1.000000e+000 
0.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
1.000000e+000 0.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 1.000000e+000
0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 0.000000e+000 
0.000000e+000 0.000000e+000
<ENDHMM>

In general, should we use '-u mvwdmv' option, instead of '-u tmvwdmv'?

Thank you very much.


Best Regards,
Lee Siu Wa



----- Original Message ----- 
From: "Keiichiro Oura" <uratec@xxxxxxxxxxxxxxx>
To: <hts-users@xxxxxxxxxxxxxxx>
Sent: Monday, October 19, 2009 10:22 AM
Subject: [hts-users:02244] Re: HERest: 7031 Transition mat all zero


> Hi,
>
> Could you try '-u mvwdmv' opstion, instead of '-u tmvwdmv' option?
> Did you finnish estimation of monophone model (HInit, HRest, and HERest
> for spec. and dur.) successfully?
>
> Regards,
> Keiichiro Oura
>
>
> Lee Siu Wa wrote:
> > Hi,
> >
> >     I try to re-estimate the spectrum models and duration models using
> > HTS HERest after the decision tree-based clustering.  I use the
> > following command in the 1st and 2nd re-estimation.
> >
> > /usr/local/HTS-2.1/bin/HERest -A -D -T 1 >
> > logs/durCDERest_after_tree3.log -C lists/trn_dur.cfg -H
> > lsp_for_duration_modeling/TRI/hmms/HMM1.2/hmmdefs -N
> > dur/TRI/hmms/HMM1.2/hmmdefs -M lsp_for_duration_modeling/TRI/hmms/HMM1.3
> > -R dur/TRI/hmms/HMM1.3 -m 1 -u tmvwdmv -w 3 -t 1500 100 5000 -S
> > scp/aligned.scp -I mlfs/train_tri_nosp.mlf lists/real_tri_nosp.list
> > lists/real_tri_nosp.list
> >
> > and
> >
> > /usr/local/HTS-2.1/bin/HERest -A -D -T 1 >
> > logs/durCDERest_after_tree4.log -C lists/trn_dur.cfg -H
> > lsp_for_duration_modeling/TRI/hmms/HMM1.3/hmmdefs -N
> > dur/TRI/hmms/HMM1.3/hmmdefs -M lsp_for_duration_modeling/TRI/hmms/HMM1.4
> > -R dur/TRI/hmms/HMM1.4 -m 1 -u tmvwdmv -w 3 -t 1500 100 5000 -S
> > scp/aligned.scp -I mlfs/train_tri_nosp.mlf lists/real_tri_nosp.list
> > lists/real_tri_nosp.list
> >
> > The first re-estimation completed.  However, I entercount the following
> > error in the 2nd re-estimation.
> >
> >   ERROR [+7031]  PutTransMat: Row 6 of transition mat is all zero
> >  FATAL ERROR - Terminating program /usr/local/HTS-2.1/bin/HERest
> >
> > The memory of the machine is relatively sufficient, about 8G.  I will be
> > very thankful if someone could help.
> >
> >
> > Best Regards,
> > Yvonne, Lee Siu Wa