[hts-users:02194] Re: Format and Generation rule of Full context labels in HTS-demo(NIT-ATR503)

[Jonas Beskow <beskow@xxxxxxxxxxxxx>]

Hi Seung-Uk,
the key to understanding the somewhat bizarre-looking full context label 
format (took me a while!) - with different delimiters between every 
other feature - is the questions file. The format allows you to write 
the grep-like expressions (questions) that pick out all labels matching 
a particular contextual feature. (Example: a question matching all 
labels where next-phone is an "s", can be written as "*+s=*" if 
next-phone is the only feature surronuded by delimiters "+" and "="). 
This means that you're free to design the full-context labels any way 
you see fit (if the format used in the demos doesn't suit your needs) as 
long as the search questions are written accordingly. And make sure not 
to use any delimiters that can legally appear in the features themselves!

good luck
- Jonas

Seung-Uk Lee skrev:
> Hello,
>
>  
>
> I have tried the HTS-demo and now I am analyzing it to apply the same 
> or similar process to Korean Speech Synthesis.
>
> I chose Japanese version of the demo because of close similarity 
> between Japanese and Korean.
>
> However, I could not find how the full context labels given created. 
> There is no explanation about the format or generation rule at all.
>
>  
>
> I found the English version (CMU_ARCTIC) has a document named /an 
> example of context-dependent label format for HMM-based speech 
> synthesis in English./ This format looks not the same as Japanese one.
>
>  
>
> Could anyone give some hint or doc about it?
>
>  
>
> Thanks in advance.
>
>  
>
>  
>
> Regards,
>
>  
>
> Seung-Uk Lee.
>
>  
>
> Voice Interface Lab., KAIST,
>
> Daejeon, Korea.
>
> elvanore@xxxxxxxxxxxxxxxxxx 
> <mail_write_form.jsp?Receive=elvanore@xxxxxxxxxxxxxxxxxx>
>
> Phone@Lab: +82-42-350-5556
>
> Mobile: +82-16-643-1415
>
>  
>
>