Portuguese

Nested
Material about Speaker Independent Acoustic Model
User: eduruas
Date: 11/17/2009 12:38 pm
Views: 13652
Rating: 15

Hi!


I'm member of a college group that performs researches about Speech Recognition, from a brazilian university. We got a database with phonetic models ready to be processed, and we have a good knowledge of the HTK toolkit. But we're having difficulties with the process of speech recognition with speaker independent acoustic model.

So I want to receive some indications of good material about this theme, because we haven't sucess in our attempt. We believe that a good reference material can help us in our research, and facilitate the understanding of the database that we have available.

Thanks.


Eduardo Ruas

Re: Material about Speaker Independent Acoustic Model
User: kmaclean
Date: 11/17/2009 2:04 pm
Views: 159
Rating: 15

Hi Eduardo,

>So I want to receive some indications of good material about this

>theme,

I am not sure what you are looking for...

  • are you just looking for general papers on how to create speaker independent acoustic models?  Google might help... or
  • how to create speaker independent acoustic models with HTK? VoxForge AM create tutorial might help... or
  • are looking for help on a specific problem - if so, please clearly describe the problem you are having.

Ken

Re: Material about Speaker Independent Acoustic Model
User: eduruas
Date: 11/18/2009 6:53 am
Views: 177
Rating: 13

Hi Ken!

Our research so far has been based on the VoxForge AM Create Tutorial, however this tutorial is to create a database. I already have a database of Monophones, but for some reason I do not know, I can not use it.

Google did not respond very well to our need, we could not find material about the use of an existing database, we found only methods to create a database. So I wanted some tips on a specific method of how to use a database already exists.

In short, we need indications of books and papers with use methods of Speaker Independent AM, if possible applied directly to HTK. Thus perhaps we could understand the theory and algorithm, in order to be able to use the available data, without having to create another database. If you have information on how to use trained data to create a dictionary, they are also welcome!

Thanks for the attention.


Eduardo

Re: Material about Speaker Independent Acoustic Model
User: kmaclean
Date: 11/18/2009 6:24 pm
Views: 557
Rating: 15

>Our research so far has been based on the VoxForge AM Create

>Tutorial, however this tutorial is to create a database.

I am not sure what you mean here... the VoxForge acoustic model creation tutorial is a tutorial to create _acoustic models_ (which are not usually referred to as a 'database') in HTK format , so they can be used with the Julius speech recognition engine.

I think you are looking for the VoxForge tutorial on how to _adapt_ an acoustic model

> already have a database of Monophones, but for some reason I do not

>know, I can not use it.

The acoustic model creation process can be tricky because a problem in an earlier step only show up as a problem in a later step, making it very difficult to figure out the cause... sometimes, it is easier to restart the whole process.

One thing to note is that the triphone creation steps in the VoxForge tutorial are for English only (see my post in the thread).

>books and papers with use methods of Speaker Independent AM, if

>possible applied directly to HTK.

Look up the the HTK book, Googling "HTK acoustic models" brings up many references, including this paper: On Developing Acoustic Models Using HTK.

However, don't think trying to understand the theory at this point will help - it is very involved mathematics/statistics... best to get comfortable with the HTK tools first.

Ken

Re: Material about Speaker Independent Acoustic Model
User: bharathi.ravishanker
Date: 6/28/2010 6:13 pm
Views: 217
Rating: 16

Hi Ken

 

The link that you have specified for creating speaker independent acoustic model

http://www.voxforge.org/home/dev/acousticmodels/linux/create/htkjulius/tutorial/download

says that the method is used for creating speaker dependent acoustic model in the tutorial first page.  Also scripts are available to create the hmm models which also specify at the end of the tutorial that they are suitable for only the specific speaker.

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

"Note: the hmmdefs file shown here was trained with my voice.  It will not work with your voice.  Your file will have a the same structure as this one, but the statistics will be different."

 

Thanks

Re: Material about Speaker Independent Acoustic Model
User: kmaclean
Date: 6/29/2010 1:03 pm
Views: 158
Rating: 15

>Also scripts are available to create the hmm models which also specify at

>the end of the tutorial that they are suitable for only the specific speaker.

The process described in the VoxForge Tutorials are for creating a speaker dependent acoustic model.  If you want to create a speaker independent acoustic model, you need to train an acoustic model with *lots* of speech from many *different* speakers.

Re: Material about Speaker Independent Acoustic Model
User: bharathi.ravishanker
Date: 6/29/2010 1:40 pm
Views: 141
Rating: 14

Thanks. So the difference is only with different speakers and the procedure is same!

Re: Material about Speaker Independent Acoustic Model
User: kmaclean
Date: 6/29/2010 1:54 pm
Views: 169
Rating: 16

>So the difference is only with different speakers and the procedure is same!

That is the approach I have been using on the VoxForge site.

I am sure there are subtle tweaks that can be done to improve on performance.  You might want to look at Keither Vertanen's HTK Wall Street Journal Training Recipe for another approach.

Re: Material about Speaker Independent Acoustic Model
User: bharathi.ravishanker
Date: 6/29/2010 7:23 pm
Views: 200
Rating: 15

Thank you. The method specified for adapting the speaker independent acoustic model for our voice can also be made speaker independent?The acoustic model prepared b VoxForge is huge and I would want to use it. I need my speech recognition system to recognize proper nouns(those not available in the dictionary too) which should also be present as words in the acoustic model right? So I need to create a few audio files and add them in the acoustic model available here, again this needs to be speaker independent. Any way out? Thanks

Re: Material about Speaker Independent Acoustic Model
User: kmaclean
Date: 6/30/2010 11:11 am
Views: 155
Rating: 15

>So I need to create a few audio files and add them in the acoustic model

>available here, again this needs to be speaker independent.

Try it out - if it does not work well, add more audio to the acoustic model from different speakers... repeat

PreviousNext