[hts-users:01851] Re: HTS voice built with Festvox 2.1 - HHed failure during duration model clustering

[Esther Klabbers <klabbers@xxxxxxxxxxxx>]

I've tried a similar thing and when I try to run the voice in Festival  
I get a bus error.

Here is my "script":

Copied nitech_us_slt_arctic_hts to hts_demo_cmu_arctic_slt and  
changed .scm scripts in festvox directory accordingly.
Then replaced files in hts directory with the ones obtained from  
training HTS-demo_cmu_arctic_slt.
Ran:
x2x +sf < lf0.win2 > lf0_dyn.win
x2x +sf < lf0.win3 > lf0_acc.win
x2x +sf < mgc.win2 > mcep_dyn.win
x2x +sf < mgc.win3 > mcep_acc.win

swab +f mgc.pdf > mcep.pdf (copy to p_mcep.pdf as there seems to be no  
difference between the two, and the scm file uses only p_mcep).
swab +f duration.pdf > dur.pdf
swab +f lf0.pdf > lf0_new.pdf

cp tree-mgc.inf trees-mcep.inf
cp tree-lf0.inf trees-lf0.inf
cp tree-dur.inf trees-dur.inf
cp label.feats feat.list

I looked at the new mcep file using dmp +i and it seems very similar  
to the old mcep file that was there. Any other hints here?

Thanks, Esther

On Dec 12, 2008, at 1:09 AM, Daniel Tihelka wrote:

> Hallo again,
>
> I just tried to use HTS to build the SLT voice (from file
> HTS-demo_CMU-ARCTIC-SLT.tar.bz2) - make works perfectly without any  
> visible
> problems, synthetic examples in ..../gen/qst001/ver1/hts_engine  
> sound as I
> would expect, files in ..../voices/qst001/ver1/ were generated.
>
> One note: option -w is not supported in sox 14.1.0 (and probably  
> even in some
> little bit earlier version), use -2 instead, when called from  
> Training.pl.
>
>
> Now the question is, how to "map" the files generated into Festival  
> voice. I
> tried the following mapping, but without success - the voice was  
> loaded
> correctly, but SayText("Hallo, this is the first try") command  
> consumed about
> 1.5GB memory and failed with message WALLOC: failed to malloc  
> 671088648 bytes.
>
> I just copied the files in lib/voices/us/voice_name/festvox (in  
> Festival) from
> the SLT voice package cmu_us_slt_arctic-0.95-release.tar.bz2, for  
> the files
> in lib/voices/us/voice_name/hts I have used the following mapping  
> (in order
> required by Festival: --> provided by HTS's Training.pl):
>
> - duration.pdf		--> voices/qst001/ver1/dur.pdf
>  simply renamed
>
> - feat.list			--> voices/qst001/ver1/label.feats
>  simply renamed
>
> - lf0_dyn.win		--> voices/qst001/ver1/lf0.win2
>  3 floats need to be converted to binary form, e.g. 'x2x' from SPTK
>  (I have read it somewhere, but I cannot find it anymore ...)
>
> - lf0_acc.win		--> voices/qst001/ver1/lf0.win3
>  3 floats need to be converted to binary form, e.g. 'x2x' from SPTK
>
> - lf0.pdf			--> voices/qst001/ver1/lf0.pdf
>  used without changes
>
> - mcep_dyn.win	--> voices/qst001/ver1/mgc.win2
>  3 floats need to be converted to binary form, e.g. 'x2x' from SPTK
>
> - mcep_acc.win	--> voices/qst001/ver1/mgc.win3
>  3 floats need to be converted to binary form, e.g. 'x2x' from SPTK
>
> - mcep.pdf		--> voices/qst001/ver1/mgc.pdf
>  simply renamed
>
> - p_mcep.pdf		--> voices/qst001/ver1/????
>  does not exist in HTS, but required by Festival (in case that  
> 'mgc.pdf' file
> would be use for this, what about 'mcep.pdf' then?)
>  I just linked it to mgc.pdf for the test ...
>
> - trees-dur.inf		--> voices/qst001/ver1/tree-dur.inf
>  simply renamed
>
> - trees-lf0.inf		--> voices/qst001/ver1/tree-lf0.inf
>  simply renamed
>
> - trees-mcep.inf		--> voices/qst001/ver1/tree-mgc.inf
>  simply renamed
>
>
> As I said, it did not work. Unfortunately.
>
> So I would like to ask you for some additional hints.  HTS training  
> works
> perfectly, but how to convert the result into Festival? How the  
> voice package
> cmu_us_slt_arctic-0.95-release.tar.bz2 has been built?
>
>
> Thank you very much. Best regards,
> Dan
>
>
> On Wednesday 03 of December 2008, Esther Klabbers wrote:
> > I wish there was more information available on how to use files
> > generated in HTS-demo in Festival. I have been able to run HTS-demo
> > for the SLT voice but when I compare the output .win, .inf and .pdf
> > files they are different than the ones distributed in the
> > nitech_cmu_slt_arctic_hts/hts directory.
> > - For one the standard HTS-demo script produces mgc.pdf instead of
> > mcep.pdf. Are these the same but with a different name? Or are they
> > different. If they are different how can you change the scheme files
> > for Festival to work with the different format.
> > - The .win files are ascii in HTS-demo but something else in the
> > festival voice. How do you convert these?
> >
> > Thanks, Esther Klabbers

Esther Klabbers
Assistant Professor,
Center for Spoken Language Understanding (CSLU),
Division of Biomedical Computer Science (BMCS)
Oregon Health & Science University (OHSU)

20000 NW Walker Road / Beaverton, OR 97006
Office: +1-503-748-3005 / Fax: +1-503-748-1306
http://www.cslu.ogi.edu/people/klabbers