[mary-users] HTK not patched with HTS
Marcela Charfuelan
Marcela.Charfuelan at dfki.de
Fri Jul 15 11:37:47 CEST 2011
Hello Florent,
What I did was select long files/sentences, and check with wavesurfer
for example, where a comma might be meaningful,
I did that for 20 files approx. that was enough. So, make sure that if
you put a comma on the transcription that corresponds to a more or less
short pause on the audio file.
Regards,
Marcela.
On 07/15/2011 10:13 AM, fxavier at ircam.fr wrote:
> Hi satish,
>
> Indeed, there are no "ssil" at all in the emmh.align file. The only
> punctuation in the text files are "." at the end of sentences.
>
> So if I add some punctuations, it's ok? What do you mean by sufficient
> "ssil"? How many "ssil" do I have to consider to be enough? There are 510
> sentences (so 510 text files), do I have to add punctuations in all the
> text files?
> Thank you very much for your answers,
>
>
> Florent
>
>
>
>> Hi Florant,
>>
>> No, it is not normal. Apparently, the labeling process was aborted at 15th
>> iteration while separating 'short silence' model from 'pause' model. EHMM
>> distinguishes two kinds of silence/pauses: 'ssil' (phrase boundaries or
>> word
>> level boundaries) and 'pau' (utterance beginning and ending). At 15th
>> iteration, EHMM is trying to initialize 'ssil' model from the middle state
>> of 'pause' model (which is trained from 1st iteration).
>>
>> The reason for the failure could be anyone of the following:
>> 1. 'pau' model does not exist at 15th iteration due to lack of "pau"
>> labels
>> in EHMM's transcription (i.e. $VoicebuildingDIR/ehmm/etc/ehmm.align).
>> 2. 'ssil' (short silence) labels were not identified in EHMM's
>> transcription
>> (i.e. $VoicebuildingDIR/ehmm/etc/ehmm.align).
>>
>> In most of cases, the reason would be the second one. so lets check EHMM's
>> phone level transcription "$VoicebuildingDIR/ehmm/etc/ehmm.align" for
>> 'pau'
>> and 'ssil'. How many of them present in that file?
>>
>> If you did not find sufficient 'ssil's in that file, I think the sentence
>> transcription files ($VoicebuildingDIR/text/*.txt) may not have sufficient
>> punctuations ( , ? ! . ) as Marcela mentioned earlier.
>>
>> Hope that helps,
>> Best regards,
>> Sathish
>>
>>
>> On Thu, Jul 14, 2011 at 10:01 PM,<fxavier at ircam.fr> wrote:
>>
>>> Hi Satish,
>>>
>>> I saw some wave files were empty, that's why the process stopped early.
>>> However, now removing those files and their transcription, I run the
>>> ehmmlabeler, and it took around 2 houres which is a good sign. But still
>>> the ehmm/lab dir and no .lab outputed.
>>> The last lines of the log.txt are:
>>>
>>>
>>> Den == 0 - SID: 1 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Den == 0 - SID: 1 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Den == 0 - SID: 2 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Den == 0 - SID: 2 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Den == 0 - SID: 3 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Den == 0 - SID: 3 wrd: 0 GauDen:0
>>> Forcing means to be 0.01 and variance to be 1
>>> Note: Floored: word:state:gau:dim 29:147:1:0
>>> FLOORED : -2.28828e-11 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:1
>>> FLOORED : 1.92343e-13 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:2
>>> FLOORED : 7.87252e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:3
>>> FLOORED : 1.00861e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:4
>>> FLOORED : -6.1927e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:5
>>> FLOORED : -1.59955e-13 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:6
>>> FLOORED : -6.30745e-15 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:7
>>> FLOORED : -3.67623e-13 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:8
>>> FLOORED : -5.42379e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:9
>>> FLOORED : -6.57738e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:10
>>> FLOORED : 4.11719e-13 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:11
>>> FLOORED : 5.87447e-14 TO 0.001
>>> Note: Floored: word:state:gau:dim 29:147:1:12
>>> FLOORED : 4.52353e-13 TO 0.001
>>> Average Likelihood: (15) -30.2782 0.00828603
>>> Accepted sentences are: 510 / 510
>>> GENERATING SHORT PAUSE: AT 15 iteration
>>> Number of pause states are: 5
>>> Pau begin state: 145 end state: 149
>>> Short silence state: 1
>>> Source state number from Pause Model: 147
>>> I am trying to middle state of pause model short silence model...
>>> I assumed short pause to be of 3 states rather it is 5
>>> Aborting...
>>> Sequential Processing Flag: 1
>>> Feature Path: /home/florent/PierreVoice/ehmm/feat
>>> Feature exten: ft
>>> NO of words: do not match: TW: 0 35
>>>
>>>
>>>
>>> Is it normal?
>>>
>>> Thanks,
>>>
>>>
>>> Florent
>>>
>>>
>>>
>>>
>>>
>>>
>>>> Hi Florent,
>>>>
>>>> Have you verified the last few lines of log.txt? Could you provide
>>> them?
>>>> That may help us to understand what could be the problem.
>>>>
>>>> Best regards,
>>>> Sathish
>>>>
>>>>
>>>> On Wed, Jul 13, 2011 at 6:46 PM,<fxavier at ircam.fr> wrote:
>>>>
>>>>> Hello Marcela,
>>>>>
>>>>> Thanks for your answer. Commentes are below:
>>>>>
>>>>>
>>>>>>> HTK HHEd exists
>>>>>>> HTK version: 3.4.1
>>>>>>> HTK is installed but it seems not patched with HTS, because HHEd
>>> does
>>>>>>> not
>>>>>>> have command CM.
>>>>>>> HTK and HDecode have to be dowloaded, patched with HTS and
>>> compiled
>>>>>>> again
>>>>>>> HTK patched with HTS is not installed
>>>>>>>
>>>>>>>
>>>>>>> However HTS is properly patched with HTK (moved the patch to htk
>>>>>>> directory, then ran:
>>>>>>>
>>>>>>> patch -p1 -d .< HTS-2.1.1_for_HTK-3.4.1.patch
>>>>>>> ./configure
>>>>>>> make
>>>>>>> make install
>>>>>> I think this is what you should do: patch HTK with the HTS code.
>>>>>
>>>>>
>>>>> Yes this is what I did of course, many times. HTS is properly patched
>>>>> with
>>>>> HTK for sure, however the check_install_external_programs -check
>>> still
>>>>> says it is not the case...
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>>> This could be the reason why when running the newvoiceimport
>>> tools,
>>>>>>> ehmmlabeler, the process only takes about 15 mn, everything seems
>>> to
>>>>> be
>>>>>>> ok
>>>>>>> according to the log.txt (features files are created), except the
>>>>> .lab
>>>>>>> that dont exist. Indeed, /lab directory is completely empty...
>>>>>> These are two different programs, ehmm does not have much to do
>>> with
>>>>>> HTK-HTS
>>>>>> we use ehmm for phonetic segmentation using force alignment.
>>>>>> I have experienced some problems with ehmm when the text files
>>>>>> (transcriptions) do not have any punctuation,
>>>>>> if that is the case maybe you can add some commas, or ?, !
>>>>>
>>>>> My text files already have ponctuations. Each sentence ends with a
>>>>> point.
>>>>> ehhm is properly installed:
>>>>>
>>>>> festvox ehmm exists
>>>>> festvox ehmm: /home/florent/MARYTTS/lib/external/bin/ehmm
>>>>> ehmm exist
>>>>> ok
>>>>>
>>>>> and I removed the /bin from the ehmmdir. Any ideas?
>>>>>
>>>>> Thanks in advance,
>>>>>
>>>>>
>>>>>
>>>>> Florent
>>>>>
>>>>> _______________________________________________
>>>>> Mary-users mailing list
>>>>> Mary-users at dfki.de
>>>>> http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users
>>>>>
>>>> _______________________________________________
>>>> Mary-users mailing list
>>>> Mary-users at dfki.de
>>>> http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users
>>>>
>>> _______________________________________________
>>> Mary-users mailing list
>>> Mary-users at dfki.de
>>> http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users
>>>
>> _______________________________________________
>> Mary-users mailing list
>> Mary-users at dfki.de
>> http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users
>>
> _______________________________________________
> Mary-users mailing list
> Mary-users at dfki.de
> http://www.dfki.de/mailman/cgi-bin/listinfo/mary-users
>
--
_______________________________________________________________
Marcela Charfuelan, Researcher, DFKI GmbH
Projektbuero Berlin, Alt-Moabit 91c, D-10559 Berlin, Germany
Phone: +49 (0)30 23895-1821
URL : http://www.dfki.de/~charfuel/
_______________________________________________________________
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
_______________________________________________________________
More information about the Mary-users
mailing list