Vista/Office 2007

Brian Ellis · 4 December 2006, 10:48

That "cool demo" is a cheat! DNS does not put in punctuation until you tell it to!

Jammrock · 4 December 2006, 09:49

I haven't used voice rec software in ages. I'm surprised that some of them have ~99% accuracy. That's pretty impressive. I may have to give DNS a spin one of these days.

404 - File or directory not found.

http://www.nuance.com/talk/

cool demo.

Brian Ellis · 4 December 2006, 09:17

Originally posted by gt40 View Post

Maybe because Microsoft can't get it right

http://www.liveleak.com/view?i=c463074f30

That does not surprise me! I've been using speech recognition for ~15 years. Most of the time, I have been using IBM products, starting with VoiceType 1 (on 21 diskettes for the English version and 20 for the French version). I then went on to SimplySpeaking Gold. Unfortunately, IBM never made the transition to XP, so I transferred to Dragon NaturallySpeaking, typically with ~99% correct recognition (on a good day). I gave the Vista VR a good try, with full training, and I don't think I ever got better results than ~75%, i.e., one word out of 4 is transcribed incorrectly (same machine as used for DNS). It is crap.

Of course, IBM and Dragon started developing their VR engines 20 years ago, while MS started only 3 or 4 years ago, so they have a lot of catching up to do. In fact, I was getting much better results with VoiceType 1-3 (1990-1995) than with Vista VR, although there was a major difference, in that VT had to be dictated with discrete words (a very slight pause between each word) whereas all the others allow fluent speech, generally a phrase at a time.

Think about sentences like, "Please write to Mr Wright right away, to tell him two o'clock is too late!". DNS and SSG inevitably get this kind of thing correct, despite the homophones. Vista VR doesn't.

What is worrying is that people will try it, decide it is useless and condemn VR as a gimmick that will never make it big time, while the good systems, like DNS, will automatically be condemned along with it.

OK, I'll be frank. If I start a 3000 word article, the accuracy may be great to start with. By the end, it will be relatively poor (~90-95%) because my voice gets tired and I tend to start slurring my words. Similarly, if I have a bad night or a cold, the accuracy is not as good as when I'm fighting fit.