[hts-users:03742] Re: Problems in the HTS 2.3alpha implementation of the LSP postfilter

[Pieter Scholtz <pieterscholtz@xxxxxxxxx>]

Ok, I think I understand the process a bit better now.

What is the significance of the spectral energy estimation in the LSP postfilter implementation? Is it detailed somewhere?  My initial tests seem to indicate that the new postfilter is considerably more stable than an implementation I have used previously that was based on [1]. I assume the energy normalisation has something to do with this?

It is impressive how well the postfilter performs in terms of quality and stability in both MGC-LSP and LSP cases. However, I am wondering whether the impulse response length cannot be shortened in computationally constrained environments? Is there a compelling reason to not just use a value for -M that is closer to the frame period, e.g. 160? I will test it in any case.

Regards
Pieter

[1]	Zhen-Hua Ling, Yi-Jian Wu, Yu-Ping Wang, Long Qin, Ren-Hua Wang, “USTC System for Blizzard Challenge 2006 an Improved HMM-based Speech Synthesis Method,” Blizzard Challenge 2006 Workshop, 2006.


On 25 Apr 2013, at 9:26 PM, Pieter Scholtz <pieterscholtz@xxxxxxxxx> wrote:

> Whoops, sorry, problem is only that mgc2mgc is just *really* slow with -M 4095(!) due to $fl being set to 4096 for some reason. I see the default value in HTS is now 576, the same in hts_engine and 384 for HTS_EMBEDDED. Nevermind then :)
> 
> 
> On 25 Apr 2013, at 3:39 PM, Pieter Scholtz <pieterscholtz@xxxxxxxxx> wrote:
> 
>> Hi
>> 
>> I have found a few problems in the Perl implementation of postfiltering_lsp in the Training.pl script of the HTS 2.3alpha SLT demo:
>> http://hts.sp.nitech.ac.jp/archives/2.3alpha/HTS-demo_CMU-ARCTIC-SLT.tar.bz2
>> 
>> I noticed my machine spending an unusual amount of time on the SPTK-based synthesis in the $WGEN1 branch, and especially the mgc2mgc routine and found two instances of the following error in postfiltering_lsp:
>> 
>> -   $line .= "$MGC2MGC -m " . ( $ordr{'mgc'} - 1 ) . " -a $fw -c $gm -n -u -M " . ( $fl - 1 ) . " -A 0.0 -G 1.0 | ";
>> +   $line .= "$MGC2MGC -m " . ( $ordr{'mgc'} - 1 ) . " -a $fw -c $gm -n -u -M " . ( $ordr{'mgc'} - 1 ) . " -A $fw -C $gm | ";
>> 
>> I have attached a patch to fix these two instances. However, the patch does not resolve all the issues. I now get the following errors:
>> 
>> Synthesizing a speech waveform from arctic_a0005.mgc and arctic_a0005.lf0...[No. 1] is unstable frame
>> [No. 2] is unstable frame
>> [No. 3] is unstable frame
>> [No. 4] is unstable frame
>> ...
>> for all frames and then a stream of
>> x2x : warning: input data is over the range of type 'short'!
>> x2x : warning: input data is over the range of type 'short'!
>> x2x : warning: input data is over the range of type 'short'!
>> x2x : warning: input data is over the range of type 'short'!
>> 
>> The resultant synthesised waveforms are just noise (and I have applied this fix too: http://hts.sp.nitech.ac.jp/hts-users/spool/2012/msg00407.html). If the postfiltering routine is bypassed, the synthesised waveforms sound fine. The hts_engine-based synthesis in the $ENGIN branch, on the other hand, works perfectly. I also tested the voice manually with hts_engine using various postfiltering coefficients and everything sounds as expected, without any errors or warnings.
>> 
>> I replaced the packaged 48 kHz waveforms with the original 32 kHz versions from http://festvox.org/cmu_arctic/dbs_slt.html, with the appropriate settings and FRAMESHIFT=160. The vocoder configuration I am using is MGCORDER=24 GAMMA=2 FREQWARP=0.55 LNGAIN=1 PSTFILTER_LSP=0.7.
>> 
>> I am using SPTK 3.6, hts_engine 1.07 and HTS 2.3alpha compiled to 64-bit binaries on OS X 10.8.3.
>> 
>> I plan on testing with GAMMA=1 FREQWARP=0.0 soon and will report back.
>> 
>> Regards
>> Pieter
>> 
>> <Training.pl.diff>
>