[hts-users:00011] Re: context-dependent label format

[Heiga ZEN <zen@xxxxxxxxxxxxxxxx>]

Hi, Jan Kmet

On Sat, 31 Jan 2004 17:08:23 +0100 (CET)
Jan Kmet <kmet@xxxxxxxxx> wrote:

> 1)
> First problem is context-dependent label format.
> I read some  article, but there are still some strange things for me...
> 
>  Are there any more detailed document for this format.
>  And are there any tools, which help in label formating?

Sorry, there is no detailed documents about context-dependent label format, 
and there is no tools for helping context-dependent label formating.

To design context-dependent label format for Slovak, 
first you should design the context-dependent label structure according to Slovak's language structure.
For example, we have designed label structure for English, Portuguese, Japanese, and Chinese as follows. 

   English : phoneme -- syllable -- word -- phrase -- utterance
Portuguese : phoneme -- syllable -- word -- phrase -- utterance
  Japanese : phoneme -- mora -- word -- accent-phrase -- phrase -- utterance
   Chinese : half-syllable -- syllable -- word -- phrase -- utterance 

If Slovak has the structure like English, 
I think you can use the same one used in English.  

Secondly, you need to choose which contexts are used in your context-dependent label.
I think most of contexts in English also can be used in Slovak. 

>  for example: I don't know meaning for word "content"
>  in section gpos,

It means "content word".
We use gPOS which is predicted by festival (not manually corrected) instead of POS.
If you have correct POS, you should use POS.   

>   and don't know what is TOBI end tone

It means "TOnes and Break Indices".
Some documents about TOBI are available on following web site.

http://ling.osu.edu/~tobi

Documentation about feature functions in the festival speech synthesis system
will help you understand the context-dependent label format. 

http://www.festvox.org/docs/manual-1.4.3/festival_32.html

> 2)
>  in HTS-Demo there is a shell script utt2lab,
>  Using this I can do some lab file,
>  but it lookse that they are diffrent...
> 
>  Here are my steps:
> 
>   1) FESTIVAL: (utt.save (SynthText "Hello World") "hello.utt")
> 
>   2) using utt2lab on hello.utt , making hello.lab
> 
>   3) using of mofyfied HTS-Demo,
>       only for generation from already trained HMM's
>    ( Training.pl : only section $GEN=1  # Generate Speech )
> 
> 
>   4) error message: Generating speech parameter : hello.lab
>       ERROR [+9999] Generator: Cannot find duration model for
>   'x^x-pau+hh=ax@x_x/A:0 ......  '
> 
>       where is the problem?

This error means that there is no HMMs correspond to 'x^x-pau+hh=ax@x_x/A:0...' in mmf file.
I think this model is not contained in training data.
You have to generate unseen models by using constructed decision trees before synthesis.

Please try following steps:

1) Add unseen models into full_${SPEAKER}_list

   % cat full_${SPEAKER}_all.list hello.lab | sort -u > temp
   % mv temp full_${SPEAKER}_all.list

2) Generate unseen models and synthesize speech
   
   using of modified HTS-demo,
   only for generation of unseen models and speech synthesis from HMMs
   (Training.pl : $MKU=1, $GEN=1)


I think you use the demonstration distributed with older version (1.1b or 1.1?).
We have released new demonstration by using CMU_ARCTIC database with HTS-1.1.1.
Please try it !

Best regards,

Heiga Zen
 
-- 
 ------------------------------------------------
  Heiga Zen
  Department of Computer Science and Engineering
  Graduate School of Engineering
  Nagoya Institute of Technology
  Japan

  e-mail: zen@xxxxxxxxxxxxxxxx
     web: http://kt-lab.ics.nitech.ac.jp/~zen
 ------------------------------------------------