[hts-users:00081] Re: f0 extraction using pda on raw sound files

["Heiga Zen (Byung-Ha Chun)" <zen@xxxxxxxxxxxxxxxx>]

Hi Anders,

Anders Lundgren wrote:

> For instance, what low/high freq boundaries should be set 
> when extracting a male voice? 

Such kind of parameters depend on data.

> Also, are there any other parameters (ie 
> voiced/voiceless treshold) that could affect HTS performance?

They would affect the performance.
I always check the quality of analysis/synthesis (mel-cepstral vocoder) 
speech of original data to determine f0 extraction parameters.

> I created som f0 files using this utility and packed them to a binary 
> float little endian, but I receive the "ViterbiAlign: No path found in 
> 8'th segment" when training reaches "sil" (silence). I have successfully 
> trained using the exact same data, but with f0 contours taken from the 
> KTH "Snack" f0 extraction tool. The problem then is that many segments 
> in sentence-final position becomes partially unvoiced though there is no 
> evidence for this in the training data.

Could you count the number of voiced/unvoiced frames assigned to segment 
"sil" in whole training data?

Best regards,

Heiga Zen (Byung-Ha Chun)

--
 ------------------------------------------------
  Heiga Zen     (in Japanese pronunciation)
  Byung-Ha Chun (in Korean pronunciation)

  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
 ------------------------------------------------