<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body bgcolor="#FFFFFF" text="#000000">
Dear Anderson, <br>
<br>
The hts_engine used during training, is the original provided by
HTS, so it does not use the mixed excitation that we have
implemented, therefore it sound quite different, I have left that
step on the training procedure just to verify that it is completed
successfully.<br>
<br>
In Mary we have a ported version of the HTS engine to java, which
uses our own version of mixed excitation and of course MARY context
features. On the configuration file of the voice you can modify some
options to tune a bit your voice, so it does not sound so metallic
or rough;<br>
# Parameter beta for postfiltering
<br>
voice.cmu-slt-hsmm.beta = 0.1
(the same effect described in the
original hts_engine)<br>
<br>
you can also tune a bit the number of iterations for GV:<br>
# Information about Global Mean and Variance PDFs
<br>
voice.cmu-slt-hsmm.useGV = true
<br>
voice.cmu-slt-hsmm.maxMgcGvIter = 200
<br>
voice.cmu-slt-hsmm.maxLf0GvIter = 200<br>
<br>
Regarding the other problem, could you provide more information,
which version of MARY are you using, in which OS, etc.<br>
<br>
Regards, <br>
Marcela.<br>
<br>
<br>
On 04/21/2012 06:19 PM, Anderson de Oliveira Monte wrote:
<blockquote
cite="mid:CADyZqaV-DdsziYot_7PFY3BeD5j_tnLvznRi_VQuVUhu4G7LRA@mail.gmail.com"
type="cite">Hello Marcela,
<div><br>
</div>
<div>Thanks for the settings. Now I have another problem to solve.</div>
<div><br>
</div>
<div>The voice that MARY synthesizes sounds rougher than the one
synthesized by the hts-engine during the HTS script training.</div>
<div><br>
</div>
<div>In addition to that, an exception occurrs when I try to save
the audio of this voice in .wav and .aiff files, it's only
possible to save it in au. files.</div>
<div><br>
</div>
<div>In the attachments you will find:</div>
<div><br>
</div>
<div>- Examples of the voice synthesized by hts-engine
(gen_voz0001.wav) and by MARY TTS (MARY_voz0001.au);</div>
<div>- The dialog box of the exception;</div>
<div>- A text file containing the error message of the exception.</div>
<div><br>
</div>
<div>Is there any configuration setting that I'm still missing?</div>
<div><br>
</div>
<div>Regards,</div>
<div><br>
</div>
<div>Anderson<br>
<br>
<div class="gmail_quote">2012/4/16 Marcela Charfuelan <span
dir="ltr"><<a moz-do-not-send="true"
href="mailto:Marcela.Charfuelan@dfki.de" target="_blank">Marcela.Charfuelan@dfki.de</a>></span><br>
<blockquote class="gmail_quote" style="margin:0 0 0
.8ex;border-left:1px #ccc solid;padding-left:1ex">
<div bgcolor="#ffffff" text="#000000"> Dear Anderson de
Oliveira Monte,<br>
<br>
The default settings are for 16Kz,<br>
HMMVoiceConfigure.fftLen 512<br>
HMMVoiceConfigure.frameLen 400 (0.025 sec)<br>
HMMVoiceConfigure.frameShift 80 (0.005 sec)<br>
HMMVoiceConfigure.freqWarp 0.42<br>
<br>
so if you want to train with 22050, please take into
account the following settings:<br>
<br>
HMMVoiceConfigure.fftLen 512<br>
HMMVoiceConfigure.frameLen ~ 550 (0.025*22050)<br>
HMMVoiceConfigure.frameShift ~ 110 (0.005*22050)<br>
HMMVoiceConfigure.freqWarp 0.45<br>
<br>
Frequency warping factor:<br>
8000 FREQWARP=0.31<br>
10000 FREQWARP=0.35<br>
12000 FREQWARP=0.37<br>
16000 FREQWARP=0.42<br>
22050 FREQWARP=0.45<br>
32000 FREQWARP=0.45<br>
44100 FREQWARP=0.53<br>
48000 FREQWARP=0.55<br>
<br>
Regards,<br>
Marcela.
<div>
<div><br>
<br>
On 04/13/2012 05:52 AM, Anderson de Oliveira Monte
wrote: </div>
</div>
<blockquote type="cite">
<div>
<div>
<div>Hello,</div>
<div><br>
</div>
<div>I was able to build and install in MARY TTS a
Brazilian Portuguese HMM Voice using wav files
that were downsampled to 16000Hz.</div>
<div><br>
</div>
<div>Now I'm trying to build another voice with the
same wav files in their original sampling rate
(22050Hz). I've made two attempts, each one with
the following sets of external programs:</div>
<div><br>
</div>
<div><b><u>RECOMMENDED SET (by MARY's HMM-based
voice building tutorial):</u></b></div>
<div><b>HTS-2.1</b></div>
<div><b>HTK-3.4</b></div>
<div><b>HDecode-3.4</b></div>
<div><b>SPTK-3.2 </b></div>
<div><b>hts_engine_API-1.01 </b></div>
<div><b>========================================================================</b></div>
<div><b><u>ALTERNATIVE SET:</u></b></div>
<div><b>HTS-2.1.1</b></div>
<div><b>HTK-3.4.1</b></div>
<div><b>HDecode-3.4.1</b></div>
<div> <b>SPTK-3.3 </b></div>
<div><b>hts_engine_API-1.03</b></div>
<div><br>
</div>
<div>With the RECOMMENDED set, I was able to build
and install the voice in MARY without erros.
However, it's impossible to understand what it
says (I believe it has something to do with the
alignment step, although no error was reported
during the whole voice building process). With the
ALTERNATIVE set, I was able to build a nice voice,
better than the one in 16000Hz (I checked it at
the directories under
"hts/gen/qst001/ver1/1mix/"), but errors ocurred
when the training process tried to synthesize
waveforms using the hts-engine (not only in
version 1.03, but also in versions 1.01 and 1.06).
In addition to that, I tried to install and use
this voice in MARY TTS, but errors also occurred
when the MARY SERVER tried to load it.</div>
<div><br>
</div>
<div>I would like to know if anybody has already
built a HMM voice in 22050Hz in MARY. If yes, how
did you solve such issues and what versions of the
mentioned programs did you use? Is there any
version of MARY TTS with improvements or bugfixes
to solve these issues?</div>
<div><br>
</div>
<div>All the best,</div>
<div><br>
</div>
<div>Anderson</div>
</div>
</div>
<pre><fieldset></fieldset>
_______________________________________________
Mary-users mailing list
<a moz-do-not-send="true" href="mailto:Mary-users@dfki.de" target="_blank">Mary-users@dfki.de</a>
<a moz-do-not-send="true" href="http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users" target="_blank">http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users</a>
</pre>
</blockquote>
<br>
<br>
<pre cols="72">--
_______________________________________________________________
Marcela Charfuelan, Researcher, DFKI GmbH
Projektbuero Berlin, Alt-Moabit 91c, D-10559 Berlin, Germany
Phone: <a moz-do-not-send="true" href="tel:%2B49%20%280%2930%2023895-1821" value="+4930238951821" target="_blank">+49 (0)30 23895-1821</a>
URL : <a moz-do-not-send="true" href="http://www.dfki.de/%7Echarfuel/" target="_blank">http://www.dfki.de/~charfuel/</a>
_______________________________________________________________
Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
Firmensitz: Trippstadter Strasse 122, D-67663 Kaiserslautern
Geschaeftsfuehrung:
Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster (Vorsitzender)
Dr. Walter Olthoff
Vorsitzender des Aufsichtsrats:
Prof. Dr. h.c. Hans A. Aukes
Amtsgericht Kaiserslautern, HRB 2313
_______________________________________________________________
</pre>
</div>
</blockquote>
</div>
<br>
</div>
</blockquote>
<br>
<br>
<pre class="moz-signature" cols="72">--
_______________________________________________________________
Marcela Charfuelan, Researcher, DFKI GmbH
Projektbuero Berlin, Alt-Moabit 91c, D-10559 Berlin, Germany
Phone: +49 (0)30 23895-1821
URL : <a class="moz-txt-link-freetext" href="http://www.dfki.de/~charfuel/">http://www.dfki.de/~charfuel/</a>
_______________________________________________________________
Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
Firmensitz: Trippstadter Strasse 122, D-67663 Kaiserslautern
Geschaeftsfuehrung:
Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster (Vorsitzender)
Dr. Walter Olthoff
Vorsitzender des Aufsichtsrats:
Prof. Dr. h.c. Hans A. Aukes
Amtsgericht Kaiserslautern, HRB 2313
_______________________________________________________________
</pre>
</body>
</html>