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

[hts-users:02269] Re: More questions to HSMM

Dear Heiga and 燕鹏举
I am rechecking the formal hts mails. When I read your discussion about HSMM, I am confused by your modification of HERest.c:main(). List below:
Oops, I found a bug in HERest.c:main().  Please replace
         if (up_durLoaded && (updateMode&UPMODE_UPDATE) && uFlags_dur) {
         if ((updateMode&UPMODE_UPDATE) && uFlags_dur) {
I don't get the meaning of removing 'up_durLoaded'. As I understand, if '-n' is not turned on, 'up_durLoaded' will be False. In HTS decument, '-n' is defined as a "dir to find duration model definitions". But in the macro definitions 'up_durLoaded' is explained as /*Set semi-Markov training */ option.
When '-n' isn't turned on, the duration model won't be loaded. Does that means the first time of calling HERest? So the models is still HMM and will soon be transformed into HSMM? Then why the unmodified code worked fine in other scenario except Dr. Yan's?
Maybe my question is silly but pleeeeease tell me about. Thanks a lot!

Na Xingyu
Ph. D Student
Recearch Center of Digital Communication Technology
Beijing Institute of Technology
Beijing, China

发件人: Heiga ZEN (Byung Ha CHUN)
发送时间: 2009-09-02  19:10:35
收件人: hts-users
主题: [hts-users:02207] Re: More questions to HSMM
燕鹏举 wrote:
> The problem of HERest -g still exists that I get always N(0,1) for duration
> pdfs.
> I did use -u tmvwdmv. What I did is just (based on HERest parameters in step
> $ERST0) removing the -N and -R parameters and adding a -g parameter. See the
> following command:
> HERest -A -C .../trn.cnf -D -T 1 -S .../train.scp -I .../mono.mlf -m 1 -u
> tmvwdmv -w 3 -t 1500 100 5000 -H .../cmp/monophone.mmf -M .../cmp -g
> .../dur/monophone.mmf .../mono.list
> Is there anything incorrect or I missed?
Oops, I found a bug in HERest.c:main().  Please replace
         if (up_durLoaded && (updateMode&UPMODE_UPDATE) && uFlags_dur) {
         if ((updateMode&UPMODE_UPDATE) && uFlags_dur) {
for now.  I think this fix should work, but I haven't tested it yet.  Please
report whether this fix solves your problem.  I hope Oura-kun also tests this
and includes this fix in the future release.
Best regards,
Heiga ZEn (Byung Ha CHUN)
Heiga ZEN (Byung Ha CHUN)
Speech Technology Group
Cambridge Research Lab
Toshiba Research Europe
phone: +44 1223 436975
This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email 

[hts-users:02270] Re: More questions to HSMM, 燕鹏举
[hts-users:02198] TRANSP of monophone reestimation, 燕鹏举
[hts-users:02200] Re: TRANSP of monophone reestimation, Heiga ZEN (Byung Ha CHUN)
[hts-users:02201] Re: More questions to HSMM, 燕鹏举
[hts-users:02202] Re: More questions to HSMM, Heiga ZEN (Byung Ha CHUN)
[hts-users:02206] Re: More questions to HSMM, 燕鹏举