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

[hts-users:02247] Re: HERest: 7031 Transition mat all zero


Hi,

Do you use MAXSTDDEVCOEF in configfile?
I use MAXSTDDEVCOEF = 10 starting from decision tree (i.e. HHEd, HERest onwards), but not in monophone HRest and HERest.

Do you use '-g' option in HRest for initialization of state duration model?
Yes, I use '-g' option to get an initial set of duration model, but without setting MAXSTDDEVCOEF.

If you want to use HSMM, the parameters in <TRANSP> should be
<TRANSP> 7
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
0 0 0 0 0 0 0
Yes, I need the state duration models, so I use HSMM. If I also want to use HMM to obtain standard spectrum models, can I obtain HSMM state duration model and HMM spectrum model from one HERest execution? or I need to train the model separately?

Thank you for your help.

Best Regards,
Lee Siu Wa



----- Original Message ----- From: "Keiichiro Oura" <uratec@xxxxxxxxxxxxxxx>
To: <hts-users@xxxxxxxxxxxxxxx>
Sent: Monday, October 19, 2009 1:08 PM
Subject: [hts-users:02246] Re: HERest: 7031 Transition mat all zero


Hi,

Do you use MAXSTDDEVCOEF in configfile?

Do you use '-g' option in HRest for initialization of state duration model?
If you want to use HSMM, the parameters in <TRANSP> should be
<TRANSP> 7
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 0 0
0 0 0 0 0 1 0
0 0 0 0 0 0 1
0 0 0 0 0 0 0
On the other hand, if you want to use HMM, state duration model should not be used.

Regards,
Keiichiro Oura


Lee Siu Wa wrote:
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









Follow-Ups
[hts-users:02248] Re: HERest: 7031 Transition mat all zero, Keiichiro Oura
References
[hts-users:02241] HERest: 7031 Transition mat all zero, Lee Siu Wa
[hts-users:02244] Re: HERest: 7031 Transition mat all zero, Keiichiro Oura
[hts-users:02245] Re: HERest: 7031 Transition mat all zero, Lee Siu Wa
[hts-users:02246] Re: HERest: 7031 Transition mat all zero, Keiichiro Oura