[hts-users:00307] Re: HTS and STRAIGHT

["Heiga ZEN (Byung Ha CHUN)" <zen@xxxxxxxxxxxxxxxx>]

Hi,

lenemo@xxxxxxxxxxxx wrote:

>   ERROR [+2121]  HInit: Too Few Observation Sequences [0]

This error indicates that there is no training data for phoneme oui.

> I don't understand why this error is occuring. I've succesfully trained
> this data before without errors, but then without the aperiodicity ratio
> values. Do you think it can be something wrong with my aperiodicity
> values?

I don't think so.

> I've attached one file with my aperiodicity value and the cmp file.

I saw its inside and their values seemed to be good.
However, I have one doubt about its inside.

% HList -C hlist.conf -h ./ntnu_no_fondat1_JK_000124.cmp

----------------------- Source: ./ntnu_no_fondat1_JK_000124.cmp ------------------------
  Sample Bytes:  372      Sample Kind:   USER
  Num Comps:     93       Sample Period: 5000.0 us
  Num Samples:   192      File Format:   HTK
------------------------------------ Samples: 0->-1 ------------------------------------
0:      -9.006 -10.635 -13.587  -8.071  -8.866  -0.062  -1.137  -1.258  -0.615   0.108
        -0.081   5.035  11.562  -0.716   1.698   2.420   1.431   0.388   0.351   0.077
         0.182   0.210   0.169   0.040   0.119   0.251   0.080  -0.044   0.009   0.133
         0.094   0.062   0.180   0.105   0.084   0.105   0.090   0.070  -0.055  -0.029
         0.172   0.089  -0.054   0.045  -0.010  -0.016   0.029  -0.002  -0.025  -0.026
        -0.001  -0.075  -0.074  -0.074  -0.021  -0.022   0.013  -0.090   0.113   0.075
         0.003   0.022  -0.031  -0.010   0.048  -0.671  -0.280   0.098   0.051   0.191
         0.320  -0.116   0.226   0.121   0.216  -0.030   0.306   0.385   0.424   0.079
         0.024  -0.019   0.048  -0.257   0.003  -0.014  -0.129  -0.010   0.279   0.116
      -10000000000.000-10000000000.000-10000000000.000

% dmp +f ./ntnu_no_fondat1_JK_000124.ap | head

0       -9.00644
1       -10.6348
2       -13.5866
3       -8.07139
4       -8.86584
5       -9.10827
6       -9.25451
7       -9.06382
8       -9.04467
9       -7.9085

The above information indicates that the aperiodicities are appended at the head of each feature vector.
However,

/home/studenter/lenemo/HTS_Kristin/HTS-demo-STRAIGHT/HTS-demo/proto/s5_m1_c75_a15_p3.prt
. . .
 States   :   2  3  4  5  6 (width)
 Mixes  s1:   1  1  1  1  1 ( 75  )
 ( 75  )
 Num Using:   0  0  0  0  0
 Mixes  s2:   1  1  1  1  1 ( 15  )
 ( 15  )
 Num Using:   0  0  0  0  0
 Mixes  s3:   2  2  2  2  2 (  1  )
 ( MSD  <=  1 )
 Num Using:   0  0  0  0  0
 Mixes  s4:   2  2  2  2  2 (  1  )
 ( MSD  <=  1 )
 Num Using:   0  0  0  0  0
 Mixes  s5:   2  2  2  2  2 (  1  )
 ( MSD  <=  1 )
 Num Using:   0  0  0  0  0
 Parm Kind:  USER

this information indicates that your model proto-type indicates that you want to insert aperiodicities between spectrum 
and F0.
So there is an inconsistency!

You should fix your model topology (proto-type) or training data.
I think this inconsistency did not cause your current problem, however, without fixing this problem, it doesn't work.

In my opinion your label file 
(/home/studenter/lenemo/HTS_Kristin/HTS-demo-STRAIGHT/HTS-demo/FonDat1_JK/labels/mono_JK.mlf) does not contain phone 
boundaries for phone oui.
Please check it.
Running HInit with deeper trace option will help you to solve this problem.

Best regards,

Heiga Zen (Byung Ha Chun)

--
------------------------------------------------
 Heiga ZEN     (in Japanese pronunciation)
 Byung Ha CHUN (in Korean pronunciation)

 Department of Computer Science and Engineering
 Nagoya Institute of Technology
 Gokiso-cho, Showa-ku, Nagoya 466-8555 Japan

 http://kt-lab.ics.nitech.ac.jp/~zen
------------------------------------------------