[hts-users:02013] Re: Problem of 1-order data stream in training and generation

[Keiichiro Oura <uratec@xxxxxxxxxxxxxxx>]

Hi,

It seems that this problem is not 1-order data stream problem.
Window.pl cannot calcurate correctly the first and final delta in principle.
You should remove the beginning and end 3 frames of the value in *.cmp.

Regards,
Keiichiro Oura

vancy wrote:
> 
> Hi all,
> 
>  
> 
> I have encountered a strange problem while using hts-2.1. The problem 
> is: when I set a separated data stream with order=1, windim=3 (window 
> file like the standard demo script) in training data, I get a wrong 
> generate parameter value at the beginning and end 3 or more  frames. The 
> other frames seem correct. For example
> 
> The training data is like:
> 
>  -13.0472       -10.5036       -10.6337       -10.5734       
> -10.7509       -10.6383 ……-10.5473       -10.5321       -10.6249       
> -10.8142       -10.4404       -10.6346
> 
> The generate data is:
> 
> -1.4591       -5.6805       -8.5919       -11.9597       -13.4174       
> -18.5253………-12.8879       -14.6535       -14.5206       -9.3618       
> -6.4378       -1.1382
> 
> I use window.pl in demo script to add window. The delta and delta-delta 
> computation for the beginning 3 frames is like
> 
> -13.047  -5.252  15.591
> 
> -10.504   1.207  -2.674  
> 
> -10.634  -0.035   0.190 
> 
> ...
> 
> I cannot fix the problem. Whether the algorithm or the code or the data 
> are not appropriate. Has anyone tried to train and generate 1-order data 
> stream ever before?
> 
>  
> 
> Thanks very much!
> 
> 
> -- 
> Best regards!
> 
> Xi Wang (汪曦)
> 
> Beijing University of Posts and Telecommunications
> Pattern Recognition & Intelligent Systems Lab