[hts-users:01585] Re: GV causes clipping

This problem may indeed come from the dynamic range of the generated speech samples.

I had something similar, and noticed that, at least under Unix, when you cast an out-of-(short) range double to a short, it systematically gets -32768 as a value.

I did fix this by normalizing the sample values before casting them to shorts.

Hope that helps,

Geoffrey

2008/7/30 Jun Xu <xujun00@xxxxxxxxxxxxxxxxxxxxx>

Hi all

I also encountered this problem using LF0 with MGC.

When GV is enabled, the amplitude is enhanced and sometimes it overflows. I don't know if hts_engine is using "double" as its internal data type. Maybe checking and decreasing the total amplitude before converting to "short" will do help?

----------------------------------------------------------------------

Best wishes,

Jun Xu

Room 3-522, FIT Building,

Tsinghua University,

Beijing, China, 100084

From: xavigonzalvo@xxxxxxxxx [mailto:xavigonzalvo@xxxxxxxxx] On Behalf Of Xavi Gonzalvo
Sent: Wednesday, July 30, 2008 7:17 PM
To: hts-users@xxxxxxxxxxxxxxx
Subject: [hts-users:01583] Re: GV causes clipping

Hi,
It happened to us with mcep.
Decreasing the weight of GV for spectrum alleviates the problem, but it doesn't solve it as it can appear in other files.
Regards

2008/7/30 Nicholas Volk <nvolk@xxxxxxxxxx>

Thanks for the reply. Prettu much what I expected but just wnated to check
I wasn't missing something obvious...
I'm using hts_engine. Changing the factor from 1.0 to, say, 0.5 solves the
problem.

br,
Nicholas

> Hi,
>
> Nicholas Volk wrote:
>
>> I'm applied global variance to a 22kHz male voice. There's some
>> occasional
>> clipping when using gv. Without GV the results are fine. (Without gv the
>> min and max of the waveform are around +-20000, not far from SLT demo
>> voice.)
>> Have you had this clipping problem with any voices?
>
> This is because the use of GV enhances dynamic ranges of speech
> parameter trajectories. It sometimes over-enhancement and causes
> overflow.
>
> BTW, which are you using, hts_engine or HMGenS?
>
> 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
> ______________________________________________________________________
>
>