Voice recognition software

Eric S. Johansson esj at harvee.org
Sun Jul 3 00:30:44 UTC 2005


Dougie wrote:
> After a little searching, I came up with somethings that may point you 
> in a good direction:

these pointers are dead ends.  ViaVoice for Linux has been withdrawn 
from the market.  Even at its best, it was still a seriously flawed 
product.  I wasted probably 50 hours trying to make it work and failed. 
  My experience was not uncommon and from watching the mailing list 
traffic, I believe most people attempting to use it failed.

Sphinx 3 is a good product if you're looking for a small vocabulary, 
grammar oriented dictation product.  It's not what you would use to 
write e-mail.

There are serious issues with training and customizations for individual 
users but that's the nature of the research project

Another research project was Sphinx 4.  It is a very good implementation 
of a speech recognition engine.  Again, it is really aimed at small 
vocabulary grammar oriented dictation products.

I know the developers of both of these products.  They understand where 
we need to be and they admit they aren't there yet.  For example, when 
you get to 20,000 words in Sphinx 4, recognition rate takes four times 
real-time.  In other words, if it takes you 10 seconds to say something, 
it will take 40 seconds for the recognition results to show on your screen.

It gets worse as the dictionary size increases.  A good continuous 
speech recognition system typically has a dictionary on the order of 
80,000 to 120,000 words.

good speech recognition is hard.  The number of developers with the 
knowledge to do good speech recognition is dropping because most of them 
are leaving the field because of economic reasons.

It's not a good time to be handicapped.

--- eric





More information about the ubuntu-users mailing list