Greetings to all !

I would like to start working on Greek language prompts and support.

I have been "lurking" into the forums for several days now, reading about localization etc (mainly the Hebrew and Italian threads)

I have taken the first step and translated the texts for the applet, welcome page and applet usage/upload instructions to Greek.
It's in 3 text files, so my first question is where i should upload/send them ?
Should i post the text here in the forum ?

Also, i am faced with the same problem as other languages had. No gramar etc.
Assuming i will able to get sufficient sound files, could someone help with that ?

Regarding the collection of samples a few more questions.

Is there any "golden ratio" for the number of men/women/ young/adult/senior voices required ?

Should the texts used to produce the prompts contain the most "common" words or this is not of an issue ?
People have suggested translating the existing applet prompts as a start, but these might not contain common,everyday words etc

Should people read/record the same text for better accurancy or it does notmake any difference?

Sorry (in advance) if these questions look very "101" but i do not have a background in speech processing/recognition, and from what i understood from the rather limited reading i did the past few weeks, the "devil is in the details"

Stelio

Hi Stelio,

>I would like to start working on Greek language prompts and support.

Great, thanks!

>I have taken the first step and translated the texts for the applet,

Please post your translations in the new Greek forum I just created.

>welcome page and applet usage/upload instructions to Greek.

For your web page translations, I've added privileges to your account so that you can edit the content directly on the new Greek section of the VoxForge web site.

See this FAQ entry for help with WebGUI (the CMS we currently use)

>Also, i am faced with the same problem as other languages had. No

>gramar etc.

>Assuming i will able to get sufficient sound files, could someone help

>with that ?

I am assuming that you mean pronunciation dictionary... when you get a few hours of speech, we can start looking into that.  To get a feel of what is required, you might want to try running through the VoxForge acoustic model creation tutorial (you need a bit of scripting experience to understand everything).

>Is there any "golden ratio" for the number of men/women/

>young/adult/senior voices required ?

In theory, you need to collect speech that best models your target application - so if you target application is going to be used by many seniors, then you need to collect more speech from seniors.  In reality, you can't really be too choosy about the kind of speech you might collect using the VoxForge app. 

One thing you do want to avoid is having more 2-3  hours of speech from a single person... it can cause over-training and reduce recognition accuracy.

>Should the texts used to produce the prompts contain the most

>"common" words or this is not of an issue ?

You want to try to get good monophone and triphone coverage.  If you know your domain ahead of time, then it is good to include words from it as well.

>People have suggested translating the existing applet prompts as a

>start, but these might not contain common,everyday words etc

I agree with you... I think it is best to use words and phrases that best reflect your target domain.  However, you still need good monophone and triphone coverage (not necessarily all of them, but the definitely most common ones...). 

You can spend way too much time on this... For English, I just used the sentences used to train the Festival text-to-speech engine - you might be able to do the same if you can get access to the training phrases for a Greek TTS system (if there are no Copyright restrictions...), or just use some public domain texts from Librivox or some similar repository.

>Should people read/record the same text for better accurancy or it

>does notmake any difference?

You need a balance... for English we have a pretty good coverage for English monophones, but less good coverage for triphones using only about 1200 prompt sentences.  But the acoustic model training process can estimate less common triphones, and reduce the overall size of the acoustic model, with little impact on recognition.  Start with 200-300 prompt lines, and as we get more speech, you can add more prompts...

>Sorry (in advance) if these questions look very "101" but i do not

>have a background in speech processing/recognition, and from what

>i understood from the rather limited reading i did the past few weeks,

>the "devil is in the details"

No worries, thanks for helping VoxForge!

Ken

Hi Stelio,

Though at this stage there is no need to yet for a word list with phonetic transcriptions, at some point there will be. Before you make your own, be sure to search on the web. It is very likely that someone has already made such a list in a university project for instance. Perhaps not published on the web, but you can see (better than we, because most of us do not speak the language) if anyone in Greece is working on speech recognition. If so, ask them if they are willing to release their wordlist under an open source license such as the GPL or the LGPL. That can save a lot of work.

Another option would be to crawl an online dictionary that has phonetic transcriptions available under an open source license. Of course you could also ask the organisation that has created the dictionary in that case.

Robin

Hi Robin,

"Though at this stage there is no need to yet for a word list with phonetic transcriptions"

I disagree with your opinion.  For example, if you have a small dictionary (let's say with at least 10 words), you can import the dictionary into Simon, and then start training.  Simon can compile the speech model (the internal compilation process is very similar to the VoxForge tutorial).  And then you can try whether Simon would be able to recognize your voice.

The proof is that I was able to import a small sample Hebrew dictionary, and to compile a speech model.  The same should be possible with Greek, of course.

Robin, why don't you try to import the Voxforge Dutch dictionary into simon? It contains over one million dutch words (very big dictionary). I think that it is in Sphinx format (if I remember correctly). So it shouldn't be a problem to train a few Dutch words with simon.

Hi SteliosK,

You should focus on getting a Greek pronunciation dictionary.  If you don't find one on the Web (try Google "Greek pronunciation dictionary", and similar expressions), then you can try the following:
 
1.  Create a word list of 10-100 words, and store it under the name "greek-words-source.txt".  It is better if you choose long words, not short words for your sample Greek pronunciation dictionary.

2.  On Ubuntu, type the following command into the terminal: ~$ espeak -f greek-words-source.txt -v el -q -x --phonout="greek-pronunciations-target.txt"
This will create a text file with Greek pronunciations. 

3.  Maybe it is possible to combine the source text documents  (greek-words-source.txt) with the target text document (greek-pronunciations-target.txt) using the command: ~$ paste greek-words-source.txt greek-pronunciations-target.txt > words-pronunciations-combined.txt

 (Unfortunately, this didn't work out for me for the German language because eSpeak didn't transcribe all words of my word list. I don't know the reason for this issue.)

4.  Transform the text file (words-pronunciations-combined.txt) into the PLS format.  I can help you with that. Use an XSLT stylesheet to transform the eSpeak phonemes into IPA (use the IPA symbols as described in the Wikipedia).  I am using saxonb-xslt, it works very well (no problem with crap symbols, there won't be a problem with the Greek special characters).  I can give you the necessary advice so that you can get quick results without having to learn XSLT/XPath.  This is a very good approach for the creation of your own pronunciation dictionary.

5.  Import your sample Greek IPA/PLS dictionary into Simon, and train 10 words with Simon.  Then try whether recognition is possible. There might be this specific issue.

I can help you with the first steps of the creation of a Greek pronunciation dictionary (I am familiar with ancient Greek, so I don't have a problem with the Greek alphabet). You need a Greek pronunciation dictionary if you want to get quick results.

Regards, Ralf

Ralf Thanks !

That's really very helpfull info, that would have taken days if not weeks to find myself.

I'll do steps 1 to 3 and post them here for further processing,fixes etc.

Hi SteliosK,

Yes, please post when you are making progress with the Greek pronunciation dictionary.

You don't have to reinvent the wheel. You just have to get familiar with the workflow.

I have made the experience that eSpeak produces sufficient results for the German language. It would be interesting to know whether this also applies for the Greek language.

By the way, if someone is interested in building a pronunciation dictionary for ancient Greek (spoken with German accent), I would be interested to help.

Regards, Ralf

Haven't done much, as i got swamped with daily work.

I am planning to have a few hours this week to deal with this matter (and also fix the Greek transcription texts for the applet)

Followed your instructions and here is what i got so far

αποτέλεσμα,     ,apot'elesm,a
δημιουργηθεί,     D,im;uRjiT'i
παράδοξο,     pa*'aDoks,o
οικογενειακές,     ,ikoj,en^ac'es
μεταβάλλονται,     m,etav'alod,e
γονιμότητας,     Q,onim'otit,as
αύξηση,     'afksis,i
συμβιώσεων,     simv;'ose,on
διαζυγίων,     D;,azij'i;on
νοικοκυριών,     n,ikoc,i*i;'on
γεννήσεων,     jen'ise,on
παραδοσιακές,     p,a*aD,os;ac'es
παραμένουν,     p,a*am'enun
μεγάλο,     meQ'alo
ιδιαίτερο,     iD;'ete*,o
πρόβλημα,     p@-*'ovlim,a
βιομηχανίας,     v;,omixan'i;as
παραμένει,     p,a*am'eni
κατακερματισμένη,     k,atac,eRmatism'eni
αποτελούμενη,     ,apotel'umen,i
επιχειρήσεις,     ,epiCi*'isis
συντριπτική,     s,id@-*iptic'i
πλειοψηφία,     pl^,opsif'i;a
απασχολούν,     ,apasxol'un
σημαντική,     s,imadic'i
παρατήρηση,     p,a*at'i*is,i
σύγχρονη,     s'iNx@-*on,i
συγγενών,     s,iNgen'on
μειώνεται,      m;'onet,e
περισσότερο,     p,e*iss'ote*,o
ακολουθούνται,     ,akoluT'ude
παλαιότερα,     p,ale'ote*,a
περισσότερα,     p,e*iss'ote*,a
σημερινός,     s,ime*in'os
εορτασμός,     ,eoRtasm'os
υπαινίσσεται,     ,ipen'isset,e
μητρότητα,     mit@-*'otit,a
ενδιαφέρουσα,     ,enD;af'e*us,a
παραγωγική,     p,a*aQ,ojic'i
διαζύγιο,     D;az'ijo
τέσσερα,     t'esse*,a
πληρώνονται,     pli*'onod,e
υποτεθεί,     ,ipoteT'i
πολιτική,     p,olitic'i
καταφέρνει,     k,ataf'eRni
δημοσιονομικά,     D,imos;,onomik'a
οργανώσει,     ,oRQan'osi
δημόσια,     Dim'os;a
διοίκηση,     D;'icis,i
προχωρήσει,     p@-*,oxo*'isi
μεταρρυθμίσεις,     m,etaRiTm'isis
ταχύτερες,     taC'ite*,es
μεγεθύνει,     m,ejeT'ini
σημαντικά,     s,imadik'a
παραγωγική.     p,a*aQ,ojic'i

from a first look they look OK. The only thing i noticed is that the greek r (Ρ) is translated to an * , not sure if that's what it should be.

Now i need your help with the XSLT as from a first look, it has a rather steep lurning curve to master the tools.

Stelio

Hello SteliosK!

Sorry for not replying to your post earlier. I didn't saw your answer.

The phonemes that you created are looking OK. Probably, the * is a symbol for the r (there are several kinds of r).

By the way, I have created a Greek pronunciation dictionary with 500.000 Greek words. I used this XSLT style-sheet to transform the Greek eSpeak phonemes into IPA phonemes with the following command:

$ saxonb-xslt -ext:on -o:greek-dictionary-with-ipa-phonemes.xml -xsl:espeak2perfectipa-greek.xsl -s:greek-dictionary-with-espeak-phonemes.xml

Thanks to XSLT you are able to do exact transformations (it would be possible to develop an XSLT style-sheet that transforms the dictionary into SPHINX format). You can unpack Ralf's Modern Greek dictionary (GPLv3). Please take a look at the phonemes that are included in this PLS dictionary, e.g.: ŋ, ð, ɣ, θ

The stress information (eSpeak provides stress information for primary (') and secondary stress(,)) is being omitted - we don't need stress information for our pronunciation dictionary. We just need the correct phonemes.

I recommend that you do the following:

1. Install simon. The installation shouldn't be too complicated.

2. Install HTK.

3. Import Ralf's Modern Greek dictionary into simon. Then you can try to start training a few Greek words. Take a look into the simon handbook to learn about the details of this user friendly software.

4. Then you try to compile a Greek speech model with simon. If you are successful, you can use simon for the recognition of a few Greek words. To get an impression how simon recognizes German words, watch my video Dictating more than 70 German words under Ubuntu. Your goal should be to produce a similar result dictating Greek words.

Of course, there is a problem with the Greek phonemes: not all of them are imported correctly by simon. We will have to think about how we can fix that. So there are known errors, but the main concept should work.

Regards,

Ralf