Hi lua,

thanks for the translations!

I will incorporate them into the next release of the Speech Submission Java Applet. 

We also need Italian prompts.  Do you have access to any public domain or GPL Italian texts?

thanks,

Ken 

Hi lua,

I just found some Italian texts on Project Gutenberg.  Please let me know which text might be appropriate for segmentation into prompts.

thanks,

Ken 

There are also text here http://www.liberliber.it/biblioteca/index.htm

 Any suggestion on a method to choiche segmentation texts?

Hi lua,

It depends how much time you want to spend on this  :)

To get the best possible coverage of phones for you language, you should select prompts that cover all the phonemes for your language.  And select a variety of words that provide as good triphone coverage.

For now, just select a text (public domain or GPL) that you can break up into 10-15 word prompt lines.  We'll need at minimum 50 lines ... preferably 100.

Will you be translating the Intro Page to the Speech Submission applet also?

thanks,

Ken 

 

Yes we plan also to translate the submission page.

What do u think to maintain applet string translation with standard .po of gnu gettext so the string could be managed outside the code?

http://www.gnu.org/software/gettext/ 

http://www.gnu.org/software/gettext/manual/html_node/Java.html  

----------------------------------------------------------------------------------------- 

About submission: 

This is the italian IPA with some example http://en.wikipedia.org/wiki/International_Phonetic_Alphabet_for_Italian

 We have some voluteers in our laboratory that want to record some contribute.

Whe want to ask some things:

1- Is it better to record 50 lines from more people or 100 lines for few people?

2- Is it better to record the same lines for all the volunteers or change sentences for each one?

3- Is it possible to submit it to flac instead of raw wave? 

 Thanks,

Lua 

in the meantime we post the submit page translation:

<div class="content">
<a name="idTy6e1fNnnwAJmnHT6H4VdQ" id="idTy6e1fNnnwAJmnHT6H4VdQ"></a>




    <h2>Leggere le frasi ed inviare le registrazioni</h2>







    <table border="1" width="450"><tbody><tr><td><a href="/home/read2/java-security-warning">Informazioni sulla finestra a comparsa dell'avviso di sicurezza di Java</a></td></tr></tbody></table>
<h3>Prima di iniziare a registrare le frasi<br /></h3><p>Posizionare il microfono in modo che non registri il respiro.&nbsp; Durante la registrazione, bisogna tentare di ridurre qualsiasi rumore non riguardante il parlato (es. sbattere le labbra, prendere fiato,..) o rumori di sottofondo. &nbsp; Non sospirare finché non si è premuto il bottone Ferma.&nbsp; <br /></p>
<p>Il riquadro nero vicino alla parte basse del tuo schermo visualizzerà la forma d'onda dell'audio che si sta registrando.
 Si prega di effettuare <a href="/home/read2/microphone">una registrazione di prova</a> per assicurarsi che il volume del microfono non si troppo alto o troppo basso.</p><h3>Registrare le frasi mostrate.&nbsp;</h3><p>Per ogni frase su ogni linea, registrare il parlato come segue:</p><ol><li>Premere il bottone <b>Registra</b>,</li>
 <li>fare una pausa di mezzo secondo,,</li>
 <li><b>Leggere</b> la frase corrispondente,</li>
 <li>fare una pausa di mezzo secondo, e dopo<br /></li>
 <li>premere il pulsante <b>Ferma</b>.</li>
 </ol>Se si fa un errore, premere nuovamente Registra per registrare ancora la frase.<br />
<p>Una volta completate le registrazioni di tutte e 10 le frasi verrà attivato il bottone Carica.&nbsp;  Premere il bottone <b>Carica</b> per inviare l'intero contributo all'archivio VoxForge con un unico file zip.</p>
<p><b>Ripetere</b> la procedura (sono consigliati invii multipli!) &nbsp;<br /> </p>
<h3>Altre Opzioni per Inviare il Parlato</h3><a name="idBjVP0QQemKdgiEk6Vbcs7A" id="idBjVP0QQemKdgiEk6Vbcs7A"></a>
 
 
 
 
 
 
 
    <table width="100%" border="0" cellpadding="0" cellspacing="0"><tr><td class="content">
<a href="/home/other">
    <img src="/uploads/R6/5B/R65BBJz2Liq0ff6uI7fyJQ/audacity-small.jpg" alt="/uploads/R6/5B/R65BBJz2Liq0ff6uI7fyJQ/audacity-small.jpg" align="left" border="0" />
</a>             
                 
 
 
 
 
 
    <p><a href="/home/other">utilizzando l'editor audio Audacity</a>; <br /></p>
    <p />
 
 
 
 
     
 
 
     
 
 
 
 
    </td></tr></table>
 
 
 
 
 
 
 
 
 <br /><a name="id8ZvwRjh7Ri8QH1LfkFJ-JA" id="id8ZvwRjh7Ri8QH1LfkFJ-JA"></a>
 
 
 
 
 
 
 
    <table width="100%" border="0" cellpadding="0" cellspacing="0"><tr><td class="content">
<a href="/home/other">
    <img src="/uploads/cs/_u/cs_uH3kgGCCuGHh8wa7YHg/phone-small.jpg" alt="/uploads/cs/_u/cs_uH3kgGCCuGHh8wa7YHg/phone-small.jpg" align="left" border="0" />
</a>             
                 
 
 
 
 
 
    <p><a href="/home/other">usando il telefono</a>. <br /></p>
    <p />
 
 
 
 
     
 
 
     
 
 
 
 
    </td></tr></table>
 
 
 
 
 
 
 
 
&nbsp;<table style="height: 24px" border="1" width="450"><tbody><tr><td><p><a href="/home/read2/java">Guida alla risoluzione dei problemi per Java e per l'applet per l'invio del parlato.</a>.&nbsp;&nbsp;</p></td></tr></tbody></table><p>&nbsp;</p>



   
   
   
   







</div>
 

Hi Lua, 

>Yes we plan also to translate the submission page.

Thanks! 

>What do u think to maintain applet string translation with standard .po of

>gnu gettext so the string could be managed outside the code?

Excellent idea - I did not know much about Java when I first created the Speech Submission app, and just incorporated the translations into the code.   I've never heard of gnu gettext - thanks for letting me know.  I don't know if I will be able to add it to the next release of the Speech Submission app, likely the release following the next one.

>1- Is it better to record 50 lines from more people or 100 lines for few people?

The best acoustic model is one where you have many recordings (to get phone samples in many different contexts), from many different people (to get samples from people with different vocal tract characteristics and dialects), using as many different texts as possible (for triphone coverage), using many different microphones, in a wide variety of rooms and environments.  However, this is very costly.  To save money, you need to focus on how you want to use your acoustic model - i.e. what is your target application.  From this, you can then focus on collecting only the audio you need that will support your application.

Therefore it really depends what your target application will be.  If you are planning to have a very limited grammar used by many people, then getting many people to record the words in your grammar is the way to go (you will also need additional words to cover all the phonemes in your language, but the focus would be on your grammar words). 

If your target application is to be used by a few people with a larger grammar, then having more prompts recorded by fewer people is the way to go.

Note: I am assuming that you are planning to create acoustic models for grammar based speech recognition, which can be created with around 100 hours of speech (the Sphinx-group of recognizers use 140 hours of speech).  Dictation applications require an order of magnitude greater of speech (1000+ hours), and the creation of a statistical language model (created from a large text corpus with a million+ words).

>2- Is it better to record the same lines for all the volunteers or change sentences for each one?

For good triphone coverage, I would recommend that your volunteers read a different subset of sentences from a large selection of sentences.  This will allow better recognition of words that were not in your training set. 

But again, it depends on your target application, if your app only has a small grammar, then the same lines (from your grammar + additional words to make sure you are using all the phonemes in your language) from all your volunteers would be the way to go.

>3- Is it possible to submit it to flac instead of raw wave?

The SpeechSubmission app was derived from MoodlesSpeex - it used the Speex codec (lossy codec) to record the audio.  I changed it to use WAV, because I wanted uncompressed audio.  The longer-term plan is for the app to use FLAC (lossless compressed codec).

Ken 

Hi Lua,

thanks,

Ken 

So it will not be better to randomize submitted prompt from a larger text file senteces base for the java applet?