[hts-users:02551] Re: Utterance size problem

[ali azimizadeh <ali.azimizadeh@xxxxxxxxx>]

the duration of file is 1m and 8s but it's just an except. i wanted to say the maximum size, typically my utterances are very shorter than this. I agree with you but there is a question. in Persian typically sentences have 20-30 word. these sentences are selected from festival chunking (by punctuation) output. if the utterances are splitted, they wouldn't be the same as real. and also semantically, they are incomplete. do you think this case doesn't decrease the sythesis quality? 
because we destroy some possible utterances.

Regards,
-Ali

On Fri, Jul 9, 2010 at 3:19 PM, Simon King <Simon.King@xxxxxxxx> wrote:

On 9 Jul 2010, at 12:16, ali azimizadeh wrote:

> my biggest wave file is about 2.5MB.

What's the duration of that file - it must be pretty long. We tend to find that very long files don't always work very well, presumably because EM training (especially in a flat-start situation) does not converge to a good alignment between the initial (poorly trained) models and the observations. In that case, the simple solution is to manually split long files into several shorter ones.

Simon


--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.