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

[hts-users:03764] Re: Add new feature stream to HTS demo and GV


Many thanks Oura!

Then how can I "control" the stretching effect of GV on the parameters? Or, when we should use GV (I mean for all features or just the mgc and lf0?) Is it easy to apply the GV on only selected streams?   


One more interesting experiment, I tried to use the GV normally. Instead of adding new stream for new feature, I put the new feature in place of the mgc feature (I removed the mgc totally and used the new feature and lf0). Then the demo worked fine! No negatives and the synthesized file based on the new feature was fine.

What do you think?!

Best Regards
Sobh
     

> Date: Fri, 10 May 2013 00:44:01 +0900
> From: uratec@xxxxxxxxxxxxxxx
> Subject: [hts-users:03763] Re: Add new feature stream to HTS demo and GV
> To: hts-users@xxxxxxxxxxxxxxx
> CC: uratec@xxxxxxxxxxxx
>
> Hi,
>
> It is not error.
> Generated parameters are sometimes over-stretched by GV algorithm.
>
> Regards,
> Keiichiro Oura
>
>
>
>
> 2013/5/9 Ibrahim Sobh <im_sobh@xxxxxxxxxxx>:
> > Dear
> > Regarding HTS demo, I've added a new feature and made the necessary changes
> > to the files to make the demo work using the new feature in a new added
> > stream. The demo ran fine, there were no errors and after the HMGenS run,
> > the generated files include the best estimations of the mgc and lf0 as usual
> > and the new added feature files too.
> >
> > However the values of the new feature were not correct! For example,
> > sometimes it is negative (when lf0 is not zero!) where it is trained on only
> > positive values. (The synthesized file based on the new feature was not
> > correct)
> >
> > When I disabled the Global Variance (GV) Step, then the demo worked fine! No
> > negatives and the synthesized file based on the new feature was fine.
> >
> > What was wrong? I need to run the HTS demo for all the features at once
> > without replacement.
> >
> > Many thanks Sobh
>

References
[hts-users:03692] Synthesized speech wave is silence for 1mix, 2mix and stc, NAGARAJ ADIGA
[hts-users:03693] Re: Synthesized speech wave is silence for 1mix, 2mix and stc, Keiichiro Oura
[hts-users:03694] Re: Synthesized speech wave is silence for 1mix, 2mix and stc, NAGARAJ ADIGA
[hts-users:03695] HMGenS add new formant stream, Ibrahim Sobh
[hts-users:03696] HMGenS add formant stream, Ibrahim Sobh
[hts-users:03703] Add new feature stream to HTS demo and HMGenS, Ibrahim Sobh
[hts-users:03704] Re: Add new feature stream to HTS demo and HMGenS, 那兴宇
[hts-users:03705] Re: Add new feature stream to HTS demo and HMGenS, Ibrahim Sobh
[hts-users:03706] Re: Add new feature stream to HTS demo and HMGenS, Ibrahim Sobh
[hts-users:03761] Add new feature stream to HTS demo and GV, Ibrahim Sobh
[hts-users:03763] Re: Add new feature stream to HTS demo and GV, Keiichiro Oura