How to build little vocabulary voice command acoustic model / language model to work in VK?

[Rio]

Hello,

I'd like to use Mandarin Chinese model on VocalKit. But those files
are simply too big (33MB for am, 212MB for lm!!). Knowing that, I
think I need to strip down the original one, or have to build my own.

Since my goal is to build a voice command app, I am wondering if
anyone has any experience with VK before? The needed commands will be
less than 10, all short words. Only the words above certain
confidential level would be recognized and those below the level are
simply ignored.

So, how can I do it? Would it be ok to build a AM/LM simply based on
these commands with different men/women's voice? How large it should
be?

thanks and best regards,

Rio

[Sanil Pillai]

You can follow the instructions here to build your own acoustic model -> http://cmusphinx.sourceforge.net/wiki/languagemodelhowto

--
You received this message because you are subscribed to the Google Groups "VocalKit" group.
To post to this group, send email to voca...@googlegroups.com.
To unsubscribe from this group, send email to vocalkit+u...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/vocalkit?hl=en.

[Brian King]

I need to update the README with that Language Model Link! That is
the best guide for getting started with a new Language Model.

You should be able to reduce the size of your Language Model
significantly with a new Language Model.

You may be able to find some other Acoustic Models on voxforge or on
the CMU Site that are smaller too. One thing you may want to verify
is that you're using 8k models and not 16k models. I personally would
not build my your own acoustic model just to save space - it's a good
chunk of work.

[Yehlin Lee]

Thank you Brian,

I am also wondering if it worth the effort to build my own light weight AM in performance wise.

I've checked a commercial SDK called vocalia (http://www.creaceed.com/vocalia/ ) , which has acceptable accuracy and speed even on my old iphone 3G. On the other hand, another project called Allears using pocketsphinx ( http://itunes.apple.com/au/app/allears/id361365954?mt=8 ) is almost unusable.

Did anyone has any experience to use a smaller, non generic, but voice command only model with VK or pocketsphix?

cheers,

Rio

[Samuel Kitono]

Actually if you want to only build command recognition it is recommended that you use grammar option. Much more accurate and faster in my opinion. I have tested this function on vocalkit and it works.

Sent from my iPhone

[Yehlin Lee]

Thanks Samuel,

But my focus language would be Mandarin Chinese. The Mandarin models in public domain for general purpose are simply too large that I have no choice but to build a smaller one, if the expected performance is acceptable.

Rio

[Samuel Kitono]

Well then I thin you can reduce the size of your language model to DMP format. usually a 212 MB can be reduced to around 50MB by converting it to DMP.