HI Ralf,

>The German speech submission application now works fine under all systems.

thanks!

Ken 

Hi Ralf,

One other thing ... are the prompts I used OK?  Would there be another set that would be better (assuming the characters are fixed)?

thanks,

Ken 

 

Hi Ken,

OK, so you knew already about the problem with the special characters of the German language.

In my opinion, all of my prompts should be OK. So if you want, you can implement all of my prompts (de1, de2, de3, ..., de100).  At the moment, I am preparing to submit more prompts.  It should be possible to build a not too bad first statistical model (language model or acoustic model - I don't know the difference) of the German language, at least I hope so.

I try to submit normal sentences of the German language.  Most of those sentences should be of a medium level - not too easy and not too complicated. That means I'm trying to cover a lot of situations, and a lot of words.  And those words should have a distribution that is typical for the German language.  To achieve this goal, it is necessary to submit much more prompts than I already have submitted.  I will continue the work.  And I hope that other speakers will follow.  This was a lot of work dictating them with Dragon NaturallySpeaking, and editing them.

There shouldn't be major mistakes in my prompts.  It would be good if other persons would use my prompts.  They don't have to create their own prompts.  I have done the first steps.  So this should be a good basis.

My prompts should build a whole unit.  So why not integrate all of them into the VoxForge speech submission application?

Greetings, Ralf

Hi Ralf,

thanks for the feedback. 

>language model or acoustic model - I don't know the difference

From the VoxForge Tutorial:

All Speech Recognition Engines ("SRE"s) are made up of the following components:

• Language Model or Grammar - Language Models contain a very large list of words and their probability of occurrence in a given sequence.  They are used in dictation applications.  Grammars are a much smaller file containing sets of predefined combinations of words.  Grammars are used in IVR or desktop Command and Control applications.   Each word in a Language Model or Grammar has an associated list of phonemes (which correspond to the distinct sounds that make up a word).
• Acoustic Model - Contains a statistical representation of the distinct sounds that make up each word in the Language Model or Grammar.  Each distinct sound corresponds to a phoneme.
  
• Decoder - Software program (like Sphink, Julius, HTK's HVite) that takes the sounds spoken by a user and searches the Acoustical Model for the equivalent sounds.  When a match is made, the Decoder determines the phoneme corresponding to the sound.  It keeps track of the matching phonemes until it reaches a pause in the users speech.  It then searches the Language Model or Grammar file for the equivalent series of phonemes.  If a match is made it returns the text of the corresponding word or phrase to the calling program. 

>So why not integrate all of them into the VoxForge speech submission application?

Unfortunately, we are getting to the point where I need to create separate builds of the SpeechSubmission app for each language, otherwise the size of the downloadable application will get to big.  I will add this an RFE in Trac.

Ken 

Hello Ken,

A few weeks ago, Timo had generated the first edition of the German pronunciation lexicon.  This was a very important step, we all know that.  How is it possible to use this pronunciation lexicon to create a first edition of the German acoustic model?  Is there any one who can do this job?

Creating separate builds of the speech submission application for each language is probably a lot of work.  It could be a workaround to eliminate those sentences which contain special characters of the German language.

Greetings, Ralf

Hi Ralf,

>How is it possible to use this pronunciation lexicon to create a first edition of the German acoustic model?

The VoxForge Tutorial shows how to do it for English.  You should be able to create a workable triphone acoustic model by doing step 1-9, using German prompts and pronunciation dictionary.

To be able to complete Step 10 and create tied-state acoustic models you need a German tree.hed script.  For more information on how to create a tree.hed file for a new language, see the following links:

• Step 10 - Making Tied-State Triphones (from the Tutorial, where you create your Acoustic Model manually)
• nsh's overview of how to create clustered triphone "questions" for Sphinx and HTK for new languages
• Ticket #153 - htk error on step 10, and a related thread in the forums
  
• my post on this Thread on "Error when compiling model" where I discuss Creating clustered triphone "questions"

>Is there any one who can do this job?

Unfortunately, I can't do this right now.  My current focus is segmenting all the LibriVox audiobook submissions -  some date back to June of last year  :(  , and squeezing in another release of the speech submission app (for Italian and Russian).  So it will be a while before I can look at this.

>  It could be a workaround to eliminate those sentences which contain

>special characters of the German language.

Thanks for the suggestion (I like easy workarounds ...) but there must be an easy way to address this in Java - some unicode settings that I have missed ... 

Ken 

Well, I could build a model this weekend, until that you probably need to install and try pocketsphinx either on windows or, better on Linux. About language model, filtering is a trivial step already done by language modelling toolkits, we'll return to this later when we'll have acoustic model but you only need to use one of them.

 

Hm, the only little problem exists, where can I download the audio? It's rather annoying to check every archive in Listen page.

Ok, I created a model from a third of audio data, you can download it here:

  http://www.mediafire.com/?2bmbsmmzrm5

it decodes numbers quite well