Hi all,

I would like to help in the spanish part of this interesting project. I can donate my voice, but I haven't found the spanish phrases to say... where are they????

I have seen that the pages aren't traslated to spanish yet. If you want I can help too.

Another question, spanish spoken is diferent in many countries. For example, mexicans will pronounce some words diferent than spanish from spain. An example from http://www.speech.cs.cmu.edu/sphinx/models/hub4spanish_itesm/proyectos/h4.dict, this is a mexican project. Then the word ACCIDENTADOS is traslated to the phonemes:

ACCIDENTADOS     A K S I D E N T A D O S

for me (a person from spain) will be A C C I D E... or even A K C I D E... In some closed ASR solutions that I have seen they are considered to diferent languages.

I think that this will complex the process of geting recordings for each "version" of the spanish. But may be rasonable to submit each voice recording specificating the country (may be in a future cound be enough data to make diferent versions of the spanish language).

Last question. How much voices do you need in order to generate one acoustic model for any language? The ideal should be 140 hours?

Thanks in advance

> I would like to help in the spanish part of this interesting project.

You are welcome!

>I can donate my voice, but I haven't found the spanish phrases to say... where are they????

http://www.voxforge.org/es/read

You might probably need to setup java in browser. Actually you can read any text you like.

> In some closed ASR solutions that I have seen they are considered to diferent languages.

It's reasonable

> But may be rasonable to submit each voice recording specificating the country (may be in a future cound be enough data to make diferent versions of the spanish language).

You can already do it by pointing pronunciation dialect.

> Last question. How much voices do you need in order to generate one acoustic model for any language? The ideal should be 140 hours?

The ideal is 2000 hours

Yesterday I entered in the read page, but only apeared grey box... I don't now why, but today it works fine.

2000 hours!!! they are a lot!!!

Another generical question, does anyone know where can I read anything about the process of generating one acoustic model for any voice recognition program (in order to understand the whole process).

Thank in advance

Well, you can read this howto

http://www.voxforge.org/home/dev/acousticmodels/linux/create/htkjulius/how-to

or htkbook or sphinx training tutorial, but I must warn you it's not a trivial process.

Hi nsr,

I have read some documents and now I know, more or less, how it works. The only thing that I don't already understand is the creation of the acoustic model.

Let's say that I would like to build a recognizer for a small word dictionary (as explanied in the page http://www.speech.cs.cmu.edu/sphinxman/scriptman1.html in the section "When you have a very small closed vocabulary (50-60 words").

Here they say that is possible to generate an acoustic model based in words, then the grammar and the language model will be very simple...

Let's say that I would like to configure sphinx2 (or any other one) in order to recognize 10 words saied by me. Then I record the 10 words in 10 wav files (or even record 20 or 30 wav files reapeating 2 or 3 times each word). What else should I do in order to generate the HMM information file.

Thanks in advace.

Hi ubanov,

NSH's reply to this got delete (and I thought I replied too, but not sure what happened to my post...) he said:

To train the model you need to record audio, create a dictionary and transcription and start training by running the script. Follow an4 tutorial to become more familar with the process.

http://www.speech.cs.cmu.edu/sphinx/tutorial.html

You should use this tutorial to create your the HMMs for your acoustic model - because it includes scripts to automate most of the process.  My post here: Creating Sphinx Acoustic Models might help you too...

Ken

Hi ubanov,

>> Last question. How much voices do you need in order to generate one acoustic model for any language? The ideal should be 140 hours?

>The ideal is 2000 hours

I think the 2000 hours of speech nsh is referring to is for dictation applications. 

The Sphinx group of recognizers use an acoustic model trained with 140 hours of speech (using 1996 and 1997 hub4 training data), but these are used for command and control type applications.

Ken

Hi ubanov,

>I have seen that the pages aren't traslated to spanish yet. If you want I can help too.

Thanks!  We need all the help we can get...

I've set up your account so that you can edit the Spanish pages.  See this FAQ entry: Editing Content with WebGUI  for more info.

If you have any questions, please don't hesitate to ask...

Ken

I have traslated all the articles that i have found. I can´t traslate the shorcuts entries.  The FAQ says:

"Shortcuts" are pointers to other objects that are displayed on the Page Layout they are located in. Usually this points to another article or forum from the English side of the VoxForge website.  You cannot change the object that the Shortcut points to, but you can "overwrite" some of its properties.  Using this feature you can overwrite the title (i.e. translate it to another language) of the object being pointed to.

Does it men that they can not be translated?

Hi Ubanov,

Thanks for the translations!

>Does it men that they can not be translated?

Yes and no...  :)

A Shortcut is just a lazy way for me to point to English content until we get a translation. 

Therefore, the object that the shortcut is "pointing to" cannot be modified (well, actually it can be modified... if you had the correct permissions, but the object being pointed to is actually located on the English side of the website...).  But I can just make a copy of the original content, and place it on the Spanish portion of the VoxForge website, and give your account the permissions you need to make whatever edits you want....

Which pages would you like to translate?

Ken