Microsoft Office Word Forum - WordBanter - View Single Post

Hi Mark:

On 29/3/06 10:12 PM, in article
, "Mark Stapel"
wrote:

Natural-language recognition is an extremely difficult problem. I have been
very impressed with the progress that has currently been made. But in all the
years that speech recognition has been out, I have yet find anyone on a
regular basis using any form of speech recognition (with the possible
exception of voice activated calling).

You have now. The majority of Court transcripts in Australia for major
cases are handled with voice-recognition. The court reporter listens to the
case on headphones and repeats what was said into their headset. They use
Dragon Naturally Speaking to type up the transcript, in close-to-real-time.

The technical challenges of speech recognition are intense. If you get a
group of 10 English speaking people, tell the first person ³Joe likes golf²
and by the time it gets relayed to the 10th person, it probably ends up as
³Tarzan and Jane². If people have such problems understanding one another, I
can only imagine the problems that a computer has.

That example does not prove your case :-) You are quite correct: it's an
intense technical challenge, but the Chinese Whispers experiment is about
memory. The challenge in speech-to-text is that humans are appalling
speakers :-)

You use Speech Recognition on the BBC News and chances are you'll yet 99.9
per cent accuracy. Try it on an Australian commercial newscaster and you
won't do quite so well :-)

I would like to hear more about ³Dragon Naturally Speaking now produces
real-time transcripts for many large court cases, running on a laptop!² In
regarding accuracy, problems, how much human intervention, etc. with this
program.

Careful: You'll get a sales call :-) One of the main providers of the
service is a mate of mine, and he's worse than a TV Evangelist :-)

This is "particular speaker" transcription. Each court reporter spends a
long time training the software. They have been selected for their ability
to speak clearly, and trained to dictate. Their accuracy is in the order of
one character error per legal page. But they do it all day for a living!

Even more problematic then the technical issues, is the life style issues
with speech recognition.

Yup! :-)

All of this makes this idea of auto suggestion look simple when comparing to
speech language recognition. All the pieces of the puzzle exist; just no one
has put it all together.

You don't send that many TXT messages, right?? :-) If you did, you would be
now be reading to kill auto suggestion :-) Seriously: handwriting
recognition has a big future, as does auto suggestion. However, those of us
who have learned to type find that typing is faster than either speaking,
handwriting, or autosuggestion.

I type about 30-35 wpm and I can imagine with a form of
auto suggestion my ³typing speed² could approach about 50 to 60 wpm.

Yeah: I'm not quite at the 60 mark myself, I don't think :-) However, I
challenge you to be quicker at typing AND correcting AND formatting with
autosuggestion. That's my point :-)

Last point, many people today are making their discision whether or not they
are buying a piece of software based upon the return on their investment. If
it is not significant, they will be just as happy using their old piece of
software. My hope is that this idea and others may be something that will
cause people to upgrade.

Put your money into "collaboration" and "decision support". Most of the
information we want has already been typed: it's sharing it around that's
today's problem :-)

Cheers

--

Please reply to the newsgroup to maintain the thread. Please do not email
me unless I ask you to.

John McGhie
Microsoft MVP, Word and Word for Macintosh. Consultant Technical Writer
Sydney, Australia +61 (0) 4 1209 1410