[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02358] Re: hts voice output problem


Hi,

Have many utterance final p's you have in your training data?
Probably not enough to make them significant.

The release of an utterance is typically in the the last state(s).
If you have the default number of states, you could investige the contests
of your 5th mgc/mcep tree. Scripts for visualising trees have been
discussed in the mailing list in the past.
After a thorough investigation of decision tree questions, you
hopefully understand what causes the issue. Then you might be able to
create questions that catch the phonetic phenomena you want to appear in
the signal.

I have sometimes edited HHEd.c so that certain questions that I'm
interested in do indeed appear in a given place within a given decision
tree, but that is quite hard to do.

br,
  Nicholas


> Hi all,
> Thanks Nickolay.I've built clunits voice with my source database and "p" 
> pronounciation is good but my hts voice is not.I should say that "p"
> pronounciation is just bad in "cvc" and "cvcc" syllables but  not in "cv"
> syllable.I hear "p" in cv but "f" in cvc or cvcc.I just have problem in
> these syllables when i have "a" or "@" or "e" vowels after "p".I have
> correct pronounciation of "p" in "po" or "pu" even in "cvc" and "cvcc"
> syllables.I'm so confused.
> please help me to solve this problem.
>
> Best Regards
> A.A
>
>
> --- On Wed, 1/6/10, Nickolay V. Shmyrev <nshmyrev@xxxxxxxxx> wrote:
>
> From: Nickolay V. Shmyrev <nshmyrev@xxxxxxxxx>
> Subject: Re: [hts-users:02332] hts voice output problem
> To: hts-users@xxxxxxxxxxxxxxx
> Cc: "Angelina Anderson" <angelina_anderson@xxxxxxxxx>
> Date: Wednesday, January 6, 2010, 12:53 AM
>
> Ð? Ð?Ñ?н, 01/01/2010 в 23:51 -0800, Angelina Anderson пиÑ?еÑ?:
>> Hi,
>> I've built hts voice.  The output is good but the quality of some
>> phones aren't good when i have one word or the word is in the end of
>> the sentence. For example "p" is heard like "f" or "ch" is heard "sh".
>> I don't know why?
>>
>
> The parameters for the phones are constructed by averaging the examples
> from your source database. Thus bad pronunciations are caused by
> majority of the examples not following your intentions. Like you might
> have few incorrectly labelled "p" in your source database in the end of
> the utterance and such mistakes affect the trained parameters.
>
> It's hard to debug such issues, basically you need to check all the
> possible sources. It often helps to build a unit selection voice first
> and clean it up. In unit selection the source of the sound is easily
> trackable and you can find the reason of bad pronunciation faster. Once
> your source database is cleaned and all errors are fixed you can rebuild
> hts voice and it will have better quality as well.
>
>
>
>
>
>



References
[hts-users:02357] Re: hts voice output problem, Angelina Anderson