Hi!
I made some experiments using the mlsacheck command and with the values 2 and 0 for the options -c and -r, respectively, the clipping problem is solved. I made the tests with the STC synthesis because, in this case, parameter generation and synthesis stages are clearly differentiated in the Training script. However, I would like to stabilize the parameters in the hts-engine synthesis too, so I guess I must modify the hts-engine code.
On the other hand, I also tried the solution proposed by Marc and it works for me too. Besides, with USEGV=0 the clipping problem is solved too but, as might be expected, the speech quality is higher when USEGV=1.
Regarding to the IMPLEN variable, I'm wondering the same as Marc.
Thank you very much for helping me!
Regards,
Carmen
Hi all!I've just found a solution to the case:- USEGV=0 → I get a segmentation fault in the stage “Start synthesizing waveforms (1mix)”In HTS2.3alpha, IMPLEN default value is 576.In HTS2.2, IMPLEN=4096 (for a 48kHz corpus I suppose).Using 2048 or 4096 for IMPLEN (the only ones I've tried so far) solved the problem on my system (48kHz, gamma=0, 50th order, STRAIGHT voc).My questions are thus:– Where does the value 576 comes form?– What is the rule to set the impulse length value?I checked all over the place, couldn't find an answer to these ones...Regards,Marc
On Tue, May 20, 2014 at 2:17 PM, Keiichiro Oura <uratec@xxxxxxxxxxxxxxx> wrote:
Hi,
Could you try to use mlsacheck command (in CVS repository of SPTK) for
generated parameters ?
This command is mel-cepstrum stabilizer.
Regards,
Keiichiro Oura
2014-05-20 20:15 GMT+09:00 Carmen Magariños Iglesias <cmagui@xxxxxxxxxxxx>:
> Hi,
>
> I forgot to mentioned it but I had already made a test reducing the volume
> of the data and it didn't solve the problem. However, someone suggested me
> to make a test using a different analysis. I was using mel-cepstral analysis
> when I got clipped audio files. Then, I made a test using LPC analysis and
> the audio files were not clipped anymore.
>
> Regards,
>
> Carmen
>
>
>
>
> 2014-05-20 2:39 GMT+02:00 Keiichiro Oura <uratec@xxxxxxxxxxxxxxx>:
>
>> Hi,
>>
>> Could you decrease the volume of training data?
>>
>> Regards,
>> Keiichiro Oura
>>
>>
>>
>>
>> 2014-05-19 21:27 GMT+09:00 Carmen Magariños Iglesias
>> <cmagui@xxxxxxxxxxxx>:
>> > Hi mailing list,
>> >
>> > I'm using HTS-2.3 alpha to synthesize speech in Galician language, but
>> > when
>> > I synthesize female voices, the audio files that I get are clipped. I
>> > tried
>> > to reduce the GVWEIGHT parameter but nothing changed in the audio files
>> > (they are still clipped). Then, I made other tests:
>> >
>> > - USEGV=0 → I get a segmentation fault in the stage “Start synthesizing
>> > waveforms (1mix)”
>> >
>> > - USEGV=1 NOSILGV=0 CDGV=1 → I don't get any error but the audio files
>> > are
>> > still clipped.
>> >
>> > - USEGV=1 NOSILGV=1 CDGV=0 → I get an error in the stage “Start
>> > converting
>> > mmfs to the HTS voice format” because the file “mgc.inf” doesn't exist
>> > in gv
>> > directory.
>> >
>> > Finally, I tried to run the HTS-demo_CMU-ARCTIC-SLT_2.3_alpha with
>> > USEGV=0
>> > and I get also a segmentation fault in the same stage.
>> >
>> > Any idea of where can be the problem?
>> >
>> > Regards,
>> >
>> > Carmen
>>
>