Speech to Text Support in OLPC
satya.komaragiri at gmail.com
Sat Sep 13 12:45:02 EDT 2008
2008/9/13 Guy Sheffer <guysoft at gmail.com>:
> Speech recognition is a pretty hard thing to put in an OLPC, it demands
> a lot of computation power. and you need to train a language model per
> language, and you need basically a lot of different people reading text.
> I tired getting one working in Hebrew using sphinx
> ( http://en.wikipedia.org/wiki/CMU_Sphinx ), however it seems to be hard
> because you need to read a lot of text in to the model, and even then
> its hard to make it actually working.
> If you still want to try getting this to work i suggest you read more on
> sphinx 3 (the numbers are not versions, so make sure its 3). That's the
> only open source system I know of that might get it work.
> There are still theoretical limitations, like the more words you put it,
> the higher the margin of error, and so on.
I realize sir that we need a very light weight model for XO. And
hence, I was contemplating using Julius instead of sphinx ( as I
mentioned in my first mail to the list)
Julius is written in C. It might help solving some of the problems we
face with Sphinx.
As far as the language models are concerned, Julius is currently
developed well for Japanese and their homepage claims that VoxForge
is working on developing it for other languages.
Repeating for easy reference:
http://julius.sourceforge.jp/en_index.php (View the About Models section)
> I hope to see that too! It would be a great thing not only for the
> OLPC!, but to everyone!
I'll try my best :)
> Guy Sheffer
> Devel mailing list
> Devel at lists.laptop.org
More information about the Devel