[hts-users:01663] Re: HMM state tying

[Keiichiro Oura <uratec@xxxxxxxxxxxxxxx>]

Hi, Mirko

Mirko Vladic wrote:
> Hello.
>
> I am an graduate student and I would like to make a HMM Serbian tts 
> system for a student project. I have some knowledge in speech 
> recognition modeling with HTK HMMs.
>
> I used the HTS demo scripts for building an English TTS based on arctict 
> database. I have problems with the file questions*.hed, which is used 
> for EN state phonetic tying. I understand the 1. part of the file which 
> includes questions with phoneme features (wovels, consonants,...). I 
> believe that the second part uses some questions for F0 and duration, is 
> that true ? I already have questions with phoneme features for my 
> language, but none for F0 and duration. Is there a tool which I could 
> use to generate those questions?

In HTS demo scripts, a common question set is used for spectrum, F0 and
duration modeling. Thus, the questions for phoneme features can be used
for F0 modeling and the questions for non-phoneme features can also be
used for spectrum modeling. If you want to build a "Serbian TTS system,"
you should design a "Serbian question set" by hand.

> I looked at many articles wchich feature HMM TTS syntesis, but there is 
> still one thing I don't fully understand. In the demo I saw some label 
> files which are used as input for voice syntethization (Alice*.lab). In 
> the files I have seen the phonetic trancription and also other 
> parameters. Are these parameters also used in HMM TTS?

Yes. The files (Alice*.lab) are used in voice synthesis part.

Regards,
Keiichiro Oura