[hts-users:02189] Re: Problem about unstable synthesized voices(using Speaker dependenttraining with STRAIGHT demo)

[Kei Hashimoto <bonanza@xxxxxxxxxxxxxxx>]

Dear all,

The default setting of STRAIGHT version V40_007c was changed from before
version.
So, in data/Makefile.in, we need to modify the frame length for the
spectral parameter extraction from 1024 to 2048.

- $(MGCEP) -a $(FREQWARP) $${GAMMAOPT} -m $(MGCORDER) -l 1024 -j 0 -f
0.0 -q 3 > mgc/$${base}.mgc;
+ $(MGCEP) -a $(FREQWARP) $${GAMMAOPT} -m $(MGCORDER) -l 2048 -j 0 -f
0.0 -q 3 > mgc/$${base}.mgc;

After changing the frame length, I got good voices.

Best Regards,
Kei Hashimoto


> Dear Evan,
>  
> I have tried several versions. But I can't get any stable results. 
> Especailly, in my case, V40_007c didn't work at all.. How could you make 
> it work?
>  
> Have a nice day!
>  
> Best Regards,
> Yuan-Fu
>
> 2009/8/11 evan huang <vinsanity1101@xxxxxxxxx 
> <mailto:vinsanity1101@xxxxxxxxx>>
>
>     Hello everyone,
>
>     I also encounter this problem. In fact, I cannot get reliable result
>     of the straight demo for English.
>     The STRAIGHT version is V40_007c. What version is preferred for the
>     STRAIGHT demo?
>     Thanks for any comments in advance.
>
>     Best Regards,
>
>     On Sun, Aug 9, 2009 at 8:15 AM, <yfliao@xxxxxxxxxxx
>     <mailto:yfliao@xxxxxxxxxxx>> wrote:
>      >
>      > Dear All,
>      >
>      > I got an unstable problem about the synthesized voice and would
>     like to know
>      > how to debug it.
>      >
>      > Basically, I adopt the scripts/settings from "Speaker dependent
>     training
>      > with STRAIGHT demo" (for English voice), and changed the data to
>     Mandarin
>      > speech (in fact Mandarin data from Blizzard Challenge 2009).
>      >
>      > But it seems that the synthesized voices are not very stable. I
>     mean that
>      > for some utterances, I got good voices. But for some utterances,
>     I got
>      > unstable voices (pop, white noise or even silence).
>      >
>      > I had also put the same data to "Speaker dependent training demo"
>     (without
>      > STRAIGHT) and got no problems. So I don't know where is the
>     problem (from
>      > STRAIGHT?) and how to solve it.
>      >
>      > I appreciate very much your help and Have a nice day!
>      >
>      > Best Regards,
>      > Yuan-Fu
>      >
>      > PS: I used HTS 2.1 and STRAIGHT version V40_006b (no error during
>     training
>      > procedure)
>      >
>      >
>      >

--
-------------------------------
Nagoya Institute of Technology
Tokuda and Lee lab.
Kei Hashimoto
bonanza@xxxxxxxxxxxxxxx
-------------------------------