Speech Recognition - Is 90% accuracy achievable?

[Guest]

Hi

I'm running XP Pro with Office Basic 2003 (Word,Excel,Outlook) and I have
been trying to get Speech Recognition to work. I have 2 questions :-

1) I can get SR to function ok in Outlook, both in Dictation and Command
mode. However, in Word or Excel I can only get Command mode to work; I can
activate the menus and options with reasonable accuracy, but when I switch to
Dictation mode(by any of the available methods) the thing continues to behave
as if it is in Command mode. Any ideas why?

2) I have done the initial speech training exercise, plus one "additional"
one (Aesop's fables). I find this gives me about 30% accuracy when running
in Dictation mode(in Outlook). I'm happy to do more training to improve this
if I can believe that the effort will be worthwhile. Is 90% accuracy (or
more!!) realistically achievable, or is the technlogy not really up to it?
How much training would this take? I'd be interested in people's experiences.
(BTW - I have taken on-board all the advice in the helps on being in a
quiet room, microphone positioning, using normal speech etc etc so you can
take that as read)


Many Thanks

Rob

 

[Bob I]

90% was achievable 10 years ago, so more "training" should get the
system a lot better at recognizing your particular speech patterns.

 

[Guest]

Thanks Bob

So it's not just a toy then!!.......I'll get some more training in

Cheers
Rob

 

[_RR]

90% was achievable 10 years ago, so more "training" should get the
system a lot better at recognizing your particular speech patterns.

robw wrote:

(to the OP

Some systems can get 90% with no training. Of course that's a pure SR
app that requires a preset BNF vocabulary, and it can't be expected to
understand syntax unless you program it.

Scansoft has been The Borg of speech recognition (assimiating every SR
company it finds), so their site may be a good resource.

 

[Don Taylor]

I'm running XP Pro with Office Basic 2003 (Word,Excel,Outlook) and I have
been trying to get Speech Recognition to work. I have 2 questions :-

1) I can get SR to function ok in Outlook, both in Dictation and Command
mode. However, in Word or Excel I can only get Command mode to work; I can
activate the menus and options with reasonable accuracy, but when I switch to
Dictation mode(by any of the available methods) the thing continues to behave
as if it is in Command mode. Any ideas why?

2) I have done the initial speech training exercise, plus one "additional"
one (Aesop's fables). I find this gives me about 30% accuracy when running
in Dictation mode(in Outlook). I'm happy to do more training to improve this
if I can believe that the effort will be worthwhile. Is 90% accuracy (or
more!!) realistically achievable, or is the technlogy not really up to it?
How much training would this take? I'd be interested in people's experiences.
(BTW - I have taken on-board all the advice in the helps on being in a
quiet room, microphone positioning, using normal speech etc etc so you can
take that as read)

I don't have experience with the version included with Office.
But I have put in a few hundred hours working with various
Dragon versions, and some of what I've seen might help you.

Microphone quality and sound card quality make a great difference.
Some of the noise cancelling microphones or the higher quality
microphones can be FAR better than the inexpensive toy microphones.

Microphone position appears to be one of the single biggest factors.
Tiny changes in position can make very large differences in the
recognition quality.

Even seemingly small backgrount noises that come and go can make
big differences. I had excellent accuracy and then when nothing
seemed like it had changed the accuracy dropped and Dragon seemed
to loose its mind. Finally tracked that down to a little mini-
fridge sitting over the wall behind me. As it turned off and on
Dragon would work and not work. And this was with what was
supposed to be a good noise cancelling mike, apparently it was
picking up the hum of the fridge and not cancelling that.

Dragon has a quick "audio check" and running that every time you
start up a session can tell you whether you have got the mike
in the wrong position or something else is wrong.

Developing a consistent clear pronunciation style and never
changing that can make a substantial difference. If you get
tired and cranky because Dragon isn't recognizing everything
correctly then the recognition accuracy really goes down hill.

The Scansoft web site does not seem to be helpful at all to me.
But there are other more helpful resources if you scan the web.

 

[_RR]

Microphone position appears to be one of the single biggest factors.
Tiny changes in position can make very large differences in the
recognition quality.

Even seemingly small backgrount noises that come and go can make
big differences. I had excellent accuracy and then when nothing
seemed like it had changed the accuracy dropped and Dragon seemed
to loose its mind. Finally tracked that down to a little mini-
fridge sitting over the wall behind me.

You're right about mics and noise immunity, Don. What type of mic
were you using when you had the fridge problem?

I haven't had much luck with Andrea or other mics that are played up
as 'speech recognition' mics. The newer Plantronics DSP-series mics
have worked well for me, though.

OP: There are a couple Plantronics DSP's at the top of this page:
http://tinyurl.com/4ldcc

I can't vouch for the other mics shown.

 

[Don Taylor]

You're right about mics and noise immunity, Don. What type of mic
were you using when you had the fridge problem?

At that moment it was a Radio Shack "Hands Free Headset" 33-3012.
It had been highly recommended for its noise cancelling abilities.
But it is SO flexible, and the mike boom is a little long, that I
find it infuriatingly difficult to get it in exactly the same
position every time. Maybe I was just sitting in a position that
didn't cancel the sound coming over/around/through the wall from
that fridge when that happened. Who knows, and you can't ever get
the real answer or solution to what caused a problem, just like Windows.

I peeked over the dividers, went around and unplugged the fridge,
as an experiment, and Dragon regained its senses. Now I just have
to find a solution that won't get me yelled at.

I haven't had much luck with Andrea or other mics that are played up
as 'speech recognition' mics. The newer Plantronics DSP-series mics
have worked well for me, though.

The Andrea ANC models, 600, 700 and 750 have been recommended.
I had pretty good luck with the headset provided in a really old
version of IBM ViaVoice, no brand on it. But some of the low-end
cheap Emkay that have been included with Dragon have been pretty poor.

OP: There are a couple Plantronics DSP's at the top of this page:
http://tinyurl.com/4ldcc

I can't vouch for the other mics shown.

Based on my experience, and that of others, speech recognition is
more a process of training you how to use it than it is a process
of training it how to use you.

There are some relatively short web pages out there which claim to
give you the instructions to get 95% or better recognition after a
couple of weeks of work.

 

[Guest]

Rob,

Did you ever get or figure out an answer to your question number 1? I'm
having the same problem with dictation in Word. Would love to be able to use
it there if I could get it to work. I can only get SR to work right in
powerpoint, although I have speech checked in the other programs.

Thanks!

Val