[hts-users:04400] Re: synthesis with STRAIGHT

[Rasmus Dall <R.Dall@xxxxxxxxxxxx>]

Hi,

> The corpus we are training on is from 3 different speakers and we have
> LOWERF0 and UPPERF0 set to 110 and 280 respectively, same as in the SLT
> demo, for all speakers, for both STRAIGHT and hts-engine.

Whethe ror not this is actually the issue, you should see a quality  
improvement by tuning these parameters. This is the input to the F0  
extractor which will assume anything outwith this range is wrong. Try  
and analyse your speakers individually by extracting F0 from a large  
number of utterances using a wide F0 range, then narrow it down until  
it seems right.

> Are there any
> other parameters to change, in particular for voiced/unvoiced?

I'm not entirely sure if using straight. But if not using straight and  
you use SPTK as is the default then in the SPTK/bin folder you should  
be able to find the pitch.c file where you can see the options, -T for  
RAPT and -t for SWIPE.

These options should be passed on line 117 in the makefile of the data  
folder in HTS-demo.

- Rasmus

Quoting Erica Cooper <ecooper@xxxxxxxxxxxxxxx> on Mon, 28 Mar 2016  
13:18:45 -0400:

> Thanks, all, for the helpful replies!
>
> The hts-engine sample was trained using hts-2.2.  The STRAIGHT samples were
> trained using hts-2.3 and STRAIGHT V40.  When I say we are using the same
> data for both, I just mean the same corpus -- the STRAIGHT training
> includes the 'bap' features.
>
> The corpus we are training on is from 3 different speakers and we have
> LOWERF0 and UPPERF0 set to 110 and 280 respectively, same as in the SLT
> demo, for all speakers, for both STRAIGHT and hts-engine.  It sounds like
> most likely these ought to be changed, which I will try.  Are there any
> other parameters to change, in particular for voiced/unvoiced?  I checked
> both data/Makefile and scripts/Config.pm and didn't see anything that
> looked relevant.
>
> Best,
> Erica
>
>
> On Mon, Mar 28, 2016 at 9:32 AM, Blaise Potard <bpotard@xxxxxxxxx> wrote:
>
> > Hello,
> >
> > I am not entirely sure which version of the HTS demo or STRAIGHT you are
> > using, as I don't think the demo normally sounds like this. Regardless,
> > when you say you use the exact same data for hts_engine and STRAIGHT
> > synthesis, you mean you are not using mixed excitation at all?
> >
> > In any case, 1mix / 2mix / stc will produce different parameters from what
> > hts-engine is generating, so if you want to have a fair comparison, you are
> > probably better off dumping the filter / excitation feature coefficients
> > from hts_engine using the -om / -of parameters, and do the synthesis from
> > the generated coefficients using STRAIGHT.
> >
> > If you still have problems with the synthesis using the parameters
> > generated by HTS-engine, then probably you are using a bad version of
> > STRAIGHT.
> >
> > If you don't have problems with the synthesis, then it is likely something
> > wrong happened during model training, or, maybe as Rasmus mentioned, during
> > the feature extraction.
> >
> > Regards,
> > Blaise
> >
> > 2016-03-25 14:21 GMT+00:00 Erica Cooper <ecooper@xxxxxxxxxxxxxxx>:
> >
> > > Hi,
> > >
> > > We've started using STRAIGHT for synthesis, and we've found that for our
> > > data, it sounds worse than synthesis with hts-engine, despite the STRAIGHT
> > > SLT demo voice sounding very nice.  We are using the exact same data with
> > > both STRAIGHT and hts-engine synthesis, but the STRAIGHT-synthesized
> > > utterances sound 'hoarse.'  1mix, 2mix, and stc are all not so good.  I was
> > > wondering whether there is any advice for which parameters might be changed
> > > to solve this.
> > >
> > > original hts-engine voice:
> > > http://www.cs.columbia.edu/~ecooper/audio/eng_alice01.wav
> > > STRAIGHT 1mix:
> > > http://www.cs.columbia.edu/~ecooper/audio/1mix_alice01.wav
> > > STRAIGHT 2mix:
> > > http://www.cs.columbia.edu/~ecooper/audio/2mix_alice01.wav
> > > STRAIGHT stc:   http://www.cs.columbia.edu/~ecooper/audio/stc_alice01.wav
> > >
> > > Thanks,
> > > Erica



--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.