[hts-users:02045] Re: Problem Feature Extraction

[Geoffrey Wilfart <geoffrey.wilfart@xxxxxxxxx>]

Hi,

I'm sorry, the coefficients do not seem right to me.

You might try to write down the complete command you're using, from the original wav file, until the extraction of those MGC. 

I guess you're using dmp to print out those coefficients, right ?

Regards,

Geoffrey

2009/6/24 wangxiang <wangxiang68@xxxxxxxxx>

My mgc extraction program extracts the mgc feature:
Is the mgc coefficients normal?  They are extracted as the same configuration.

$mgcep -g 3 -m 24 -a 0.42 -e 1 -f 1e-500  -l 512 > wav0001.mgc

wav0001.mgc:

-30.7302
44.503 
129.839
203.888
399.677
431.963
392.866
541.827
745.529
877.982
980.848
1183.57
1334.44
1439.76
1471.56

wav0002.mgc:

1.00958e+08 
1.48241e+08 
1.25264e+08 
1.39607e+08 
1.2099e+08  
-6.45573e+06
-4.2926e+07 
-1.65103e+08
-3.30876e+08
-3.67122e+08
-4.51146e+08
-4.90574e+08
-3.66214e+08
-2.42457e+08
-9.05429e+07
1.85796e+08 
4.25594e+08 
5.50507e+08 
7.34185e+08 
7.64299e+08 
2.08303e+06 
6.06196e+06 
1.58646e+07 
4.56373e+07 
4.88302e+07 
6.96161e+07 

2009/6/23 Geoffrey Wilfart <geoffrey.wilfart@xxxxxxxxx>

I don't think there's any restriction in the range of -e option, as long as it's positive.

If you're value of -e is too small, it's possible that the matrix determinant in theq function remains under the lower limit set by the -f option (default is 0.000001). 

You might therefore increase the value of your -e option, or decrease the one of the -f option.

For instance, you may try with -e 1.0, to convince yourself the option works, and then decrease the value as long as you don't get an error.

Hope that helps,

Geoffrey

2009/6/23 wangxiang <wangxiang68@xxxxxxxxx>

Thanks, but revise -e seems not to work :)
Can you tell me what is the Numerical range of -e?

2009/6/23 Keiichi Tokuda <tokuda@xxxxxxxxxxxx>

This can happen when all samples in an input frame have the
value of zero, and can be avoided by setting a small number to
-e option in the command line of mgcep.

Regards,
Keiichi

From: wangxiang <wangxiang68@xxxxxxxxx>
Subject: [hts-users:02038] Re: Problem Feature Extraction
Date: Tue, 23 Jun 2009 01:20:15 +0800
Message-ID: <64a692590906221020k6af36059sf646cdbb9aec8643@xxxxxxxxxxxxxx>

> I read the code of SPTK and find when the theq() condition is not true, the
> mgcep exits.
>
> I just wonder why some wave can't apply the theq condition?
>
> When I listen to these waves, they seem to be normal ! Can anyone give me
> some paper about this theory?
>
> 2009/6/22 wangxiang <wangxiang68@xxxxxxxxx>
>
> > Then it stops when error, and the units is not complete.
> > So can you give me some suggestion to solve it ?
> > Thanks!
> >
> > 2009/6/22 Geoffrey Wilfart <geoffrey.wilfart@xxxxxxxxx>
> >
> > Hi,
> >> It will be harmful, since the feature extraction stops when the error
> >> raises. This means that you have incomplete features (and later cmp) files.
> >> You might then get not enough or no data for some units.
> >> Regards,
> >> Geoffrey
> >>
> >> 2009/6/22 wangxiang <wangxiang68@xxxxxxxxx>
> >>
> >>   MGCORDER=12 MGCLSP=1 GAMMA=3 FREQWARP=0.42 LNGAIN=1
> >>> >
> >>> While extracting using mgc-lsp
> >>>
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0289.raw
> >>> > mgcep : Error in theq() at 0th iteration!
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0290.raw
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0291.raw
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0292.raw
> >>> > mgcep : Error in theq() at 0th iteration!
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0293.raw
> >>> > Extracting MGC-LSP coefficients from raw/cmu_us_arctic_slt_a0294.raw
> >>> >
> >>> > Does anybody tell me what happened? Is there any harmful for following
> >>> > training?
> >>>
> >>> It means that determinant of Toeplitz+Hankel matrix, which is a
> >>> coefficient matrix of a linear system in MGC analysis, is too small (<
> >>> 1.0e-6).
> >>> You can modify this threshold via -f option as a smaller threshold
> >>>
> >>> Does anyone know the reason for this warning  and  Is it harmful to the
> >>> model training?
> >>>
> >>> Thanks!
> >>>
> >>> --
> >>> Xiang Wang
> >>>
> >>>
> >>
> >
> >
> > --
> > Xiang Wang
> > Institue Of Acoustics, CAS
> >
> >
>
>
> --
> Xiang Wang
> Institue Of Acoustics, CAS

--

Xiang Wang
Institue Of Acoustics, CAS

--

Xiang Wang
Institue Of Acoustics, CAS