[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[hts-users:04225] Re: Modulation Spectrum-based post-filter
Hi,
I'm the author of the paper of MS-based post-filter.
I agree with your revision of the code.
> To me, when comparing results of CDGV+MSPF vs. CDGV, it is hard to
> notice much difference, and the little difference that is there makes
> the MSPF results arguably even slightly worse.
One of the reasons I think is the kind of the spectral parameter.
My paper ( http://isw3.naist.jp/~shinnosuke-t/pdf/globalSIP2014.pdf )
uses mel-cepstral coefficients and confirmed the quality gain (Fig. 10).
In the other spectral parameters, the effect will be different as Global
Variance (GV) is so.
Let me know if you have any questions or problems.
Best,
- Shinnosuke Takamichi
On 2015/02/18 21:14, Merlijn Blaauw wrote:
Hi,
I think in the HTS 2.3beta demos' MSPF code, the following
# apply weight
$line .= "$VOPR -l " . ( $mspfFFTLen / 2 + 1 ) . " -s
$gendir/$base.$type.mspec_dim$d | ";
$line .= "$SOPR -m $mspfe{$type} | ";
$line .= "$VOPR -l " . ( $mspfFFTLen / 2 + 1 ) . " -a
$gendir/$base.$type.mspec_dim$d > $gendir/$base.p_$type.mspec_dim$d";
should be something more like
# apply weight
$line .= "$VOPR -s $gendir/$base.$type.mspec_dim$d | ";
$line .= "$SOPR -m $mspfe{$type} | ";
$line .= "$VOPR -a $gendir/$base.$type.mspec_dim$d >
$gendir/$base.p_$type.mspec_dim$d";
Afaik VOPR A +/- B where A and B are M-by-N matrices will add/subtract
first row of B from all rows of A if option -l N is used.
When the default -l 1 is used, VOPR will add/subtract element-wise (each
row of B from each row of A).
Without this change setting mspfe{'mgc'} = 0.0 gives me very obvious
artifacts (cyclic with MSPF shift period).
Has anyone else tried the MS-based post filter?
To me, when comparing results of CDGV+MSPF vs. CDGV, it is hard to
notice much difference, and the little difference that is there makes
the MSPF results arguably even slightly worse.
Thanks,
Merlijn
--
Shinnosuke Takamichi, Ph.D student,
Nara Institute of Science and Technology,
Email: shinnosuke-t@is.naist.jp