[hts-users:03110] Re: Issue in hts synthesized alice samples
- Subject: [hts-users:03110] Re: Issue in hts synthesized alice samples
- From: "Nicholas Volk" <nvolk@xxxxxxxxxx>
- Date: Sun, 6 Nov 2011 13:19:48 +0200
- Delivered-to: hts-users@xxxxxxxxxxxxxxx
- Importance: Normal
Hacky solution proposals coming up:
1) If your language uses a trilled /r/, you might have issues with
voiced/voiceless boundries. You could try to make the all voiced at least
when surrounded by voiced phones. (Example context: /ara/.)
2) The default number of states (5) is sufficient for a single tap, but it
really can not handle a situation where the tongue hits the alveolar wall,
say, 3 times. There's just too much happening to be descriobed within 5
states. If there are multiple trills in your languge you could try to use
/arra/ for /ara/.
There are pros and cons for both of these tactics.
Solution 2 requires a new voice model to be built to work properly,
but you can test it to some degree with the old model as well.
> I think the problem of /r/ may be complicated for many languages. However,
> you can try to improve the phoneme related context and questions,
> and extracted pitch / spectrum parameters.
> On Sun, Nov 6, 2011 at 12:15 PM, Anil John M <aniljohn80@xxxxxxxxx> wrote:
>> Hi All,
>> I have used HTS demo scripts to build a voice for my language. When I
>> listened to the hts-engine synthesized alice samples, I could observe
>> where ever the phoneme /r/ appeared, synthesized speech was highly
>> Please suggest me how can to address this issue.
>> Thank you,
>> - Anil
- [hts-users:03108] Issue in hts synthesized alice samples, Anil John M
- [hts-users:03109] Re: Issue in hts synthesized alice samples, Xi Wang