Re: speech recognition

[http://dummy.us.eu.org/robert (robert)]

OK!

 > From: "Mike McQueen" <http://www.gte.net/~mike.mcqueen>
 > Date: Wed, 16 Jul 2003 10:47:27 -0700
 >
 > >  > > Without training, it's probably 80%.  But, with training, it's about
 > 95%.
 > >  > > (I used to regularly get 99% accuracy.)
 > >  >
 > >  > Impressive numbers.  Maybe with a 10 word vocabulary.  But nowhere near
 > that
 > >  > with a 5000 word vocabulary.
 > >
 > > Actually, the commands I used were from about 200 word vocab.  Going
 > > outside that vocab seriously degraded performance.
 > >
 > > But, speech recognition from Scansoft/Dragon is quite good these days.
 > > Probably Microsoft's sucks.
 > 
 > Yes MS's probably does suck.  But I'm not even necessarily referring to MS's
 > software.  This is just data I assimilated by being in that group and
 > reading/osmosizing industry articles and data.  Granted, the tools with
 > which I worked were designed for applications like call centers, where there
 > is no voice "training" (must accept general public voices), and there may be
 > background noise, sneezing, coughing, stuffy noses, funny accents, poor
 > English skills, etc.  In that environment, even the numbers 0-9 are barely
 > pushing 80% if i recall correctly.
 > 
 > Also, a 200 word vocab is a toy.  That's not real speech recognition.  Yes I
 > could probably find some use for it, but it's not what I'm talking about
 > when I say "speech recognition isn't *there* yet".  I'm talking about free
 > form dictation of a large vocabulary.  Basically, understanding spoken
 > English.