Re: [Cmusphinx-devel] Adapt acoustic for multiple accents

[Nickolay V. S. <nsh...@ne...>]

В Вск, 10/04/2011 в 08:39 +0100, Glenn Pierce пишет:
> Hi 

Hello Glenn

> I have an application where I wish to detect a number of questions and
> answers that will contain
> sentences like
> "I want to *"
> "I like you"
> "What can you do"
> 
> These will be passed to a robot system to perform an action.
> These are the generic sentences. If someone says "I want to talk about
> animals" I plan to change the dictionary to be animal specific so I
> can get more accurate responses.
> 
> So far my accuracy hasn't been great on the generic terms. I have
> improved it a little by using a British English acoustic model I
> found.
> 
> I was hoping to improve it further using by adapting the acoustic
> model. However, am I right in think this would only improve accuracy
> for my accent? Could I improve accuracy for multiple accents for my
> small dictionaries ?

The proper way to approach this problem is to setup a test database.
Once you'll be able to say - here is my model, here are my files, here
are my scripts, I run them and get only 50% WER it will be possible to
advise you something specific. Until that it's sort of early to discuss
everything else.

For small vocabulary accuracy of recognition should be very good if
there is no bugs in the decoder.