How fast can modern speech app's convert speech to text?

Saturday, October 3, 2009

How fast can a modern speech recognition app'
translate speech into text with no mistakes?

In my case, 384 wpm, because that is the very
fastest that I can speak, and still have another
human understand what I am saying.

Neither my hardware nor my software limited me,
only the rate at which I talked limited me.

My computer is a run of the mill model that has
a 2.4 GHz Intel core 2 duo CPU. Ram is 4 GB.

OS is Vista Ultimate, with all the patches.

Speech recognition app' is Dragon NaturallySpeaking
version 10.0 - the full "medical" version, which
is far from being ideal to run this sort of
speed test.

Reports from speech experts claim that the
lower cost regular $900 version of Dragon
is speedier.

I dictated 24 repetitions of:

"now is the time for all good men
to come to the aid of their country"

....which took me 60 seconds to speak 384 words,
or 6.4 words per second.

By contrast, Fran Capo, the worlds fastest
female talker, can speak twice as fast as I can,
speaking more complex words.

She was timed at 603.32 words in 54.2 seconds,
which works out to be slightly over 667 wpm.



It did not seem to matter whether I gulped a
breath at the end of each repetition, or spoke
two repetitions and gulped a bigger breath.

I could not break the 384 wpm barrier either way.

No text mistakes, of course.

Just to be certain, I ran two more one-minute tests,
and again there were not any text errors in the
resulting text.

If I had grossly mispronounced any word, of course
there would be a text mistake, but I am very careful
to speak distinctly when I am running this test,
even though I am speaking as rapidly as I can.

My guess is that at least half the people who are
modestly proficient with Dragon would be able to
duplicate the results I achieved, regarding both
speed and accuracy of this test.

BTW, when I was dictating my 60 second speed test
with Dragon, no text appeared on the screen until
I stopped dictation at the end of the 60 seconds.

This is normal, with all modern speech apps that I
have used in this sort of test.



Next I will put WSR to the test. (Vista Speech)

Should be interesting to see how it fares,
compared to Dragon.

I anticipate it should perform about the same
as regards speed, I do not know about accuracy,
however. That remains to be seen.

Mark-

 

On Sat, 03 Oct 2009 13:39:35 -0700, Mark Conrad
<mconrad@earthlink.invalid> wrote in
<031020091339358920%mconrad@earthlink.invalid>:
<!--coloro:blue--><span style="color:blue <!--/coloro-->
>How fast can a modern speech recognition app'
>translate speech into text with no mistakes?
>
>In my case, 384 wpm, because that is the very
>fastest that I can speak, and still have another
>human understand what I am saying.
>
>Neither my hardware nor my software limited me,
>only the rate at which I talked limited me.
>
>My computer is a run of the mill model that has
>a 2.4 GHz Intel core 2 duo CPU. Ram is 4 GB.
>
>OS is Vista Ultimate, with all the patches.
>
>Speech recognition app' is Dragon NaturallySpeaking
>version 10.0 - the full "medical" version, which
>is far from being ideal to run this sort of
>speed test.
>
>Reports from speech experts claim that the
>lower cost regular $900 version of Dragon
>is speedier.<!--colorc--><!--/colorc-->

My guess™ is that Google Voice is considerably faster when
transcribing voice mail to text, and my experience is that it's also
considerably more accurate.

--
Best regards,
John <http: avasgroup.com>

 

In article <q0ffc5dko4n6bvdf56hmgpsn1qvngpkipl@4ax.com>, John Navas
<spamfilter1@navasgroup.com> wrote:
<!--coloro:blue--><span style="color:blue <!--/coloro--><!--coloro:green--><span style="color:green <!--/coloro-->
> >How fast can a modern speech recognition app'
> >translate speech into text with no mistakes?<!--colorc--><!--/colorc-->
>
> My guess™ is that Google Voice is considerably faster when
> transcribing voice mail to text, and my experience is that it's also
> considerably more accurate.<!--colorc--><!--/colorc-->

Interesting, I will look into it.

Only modern general-purpose speech recognition apps
I have tried so far are:

1) Dragon NaturallySpeaking 10.0 (4 versions)
2) WSR (Vista Speech)
3) MacSpeech Dictate 1.5.2 (2 versions)

Mark-

 

Mark Conrad <mconrad@earthlink.invalid> wrote in news:031020091339358920%
mconrad@earthlink.invalid:
<!--coloro:blue--><span style="color:blue <!--/coloro-->
>
> Saturday, October 3, 2009
>
> How fast can a modern speech recognition app'
> translate speech into text with no mistakes?
>
> In my case, 384 wpm, because that is the very
> fastest that I can speak, and still have another
> human understand what I am saying.<!--colorc--><!--/colorc-->

384 ? Yesterday, it was 600 wpm. I call you out on that and now yo upost
references and your tests ?

Do you rememebr.......
<!--coloro:blue--><span style="color:blue <!--/coloro-->
> And the joke is???
>
> I can dictate 600 words a minute with DNS. What can you do?
>
> Mark-<!--colorc--><!--/colorc-->

Did you forget saying that ?

 

Mark Conrad wrote:<!--coloro:blue--><span style="color:blue <!--/coloro-->
> Saturday, October 3, 2009
>
> How fast can a modern speech recognition app'
> translate speech into text with no mistakes?
>
> In my case, 384 wpm, because that is the very
> fastest that I can speak, and still have another
> human understand what I am saying.
>
> Neither my hardware nor my software limited me,
> only the rate at which I talked limited me.
>
> My computer is a run of the mill model that has
> a 2.4 GHz Intel core 2 duo CPU. Ram is 4 GB.
>
> OS is Vista Ultimate, with all the patches.
>
> Speech recognition app' is Dragon NaturallySpeaking
> version 10.0 - the full "medical" version, which
> is far from being ideal to run this sort of
> speed test.
>
> Reports from speech experts claim that the
> lower cost regular $900 version of Dragon
> is speedier.
>
> I dictated 24 repetitions of:
>
> "now is the time for all good men
> to come to the aid of their country"
>
> ...which took me 60 seconds to speak 384 words,
> or 6.4 words per second.
> <!--colorc--><!--/colorc-->
That's quite a test. I might mention that I can type approximately 5,000
characters per minute--IF they are all the letter n. Try it it something
and little more daunting. One suggestion--try something that contains
many references to "Jesus" and "cheeses".
Allen

 

In article <2badnVC757sWB1XXnZ2dnUVZ_uidnZ2d@giganews.com>, Allen
<allent@austin.rr.com> wrote:
<!--coloro:blue--><span style="color:blue <!--/coloro-->
> That's quite a test. I might mention that I can type approximately
> 5,000 characters per minute--IF they are all the letter n.<!--colorc--><!--/colorc-->

Yep, that is why I mentioned in my post that if I screwed up
and mispronounced even one word in those 384 words,
that I would not have got the 100% raw accuracy which I
actually achieved, in 3 seperate consecutive runs of the test.

I also posted enough details of the test for _you_ to duplicate
the same results I achieved, only requirement being that you
are fairly proficient with Dragon.

BTW, I could have used ordinary one syllable prose instead of
repeating the sentence, but that would be much harder for
any doubting Thomas to duplicate, while still maintaining
a 384 wpm pace.


<!--coloro:blue--><span style="color:blue <!--/coloro-->
> One suggestion--try something that contains
> many references to "Jesus" and "cheeses".<!--colorc--><!--/colorc-->

Yeah, homonyms and near homonyms, the curse of modern
speech apps, they would have sunk my demo.

There actually was a "near-homonym" in my demo.

"all good men" can be confused with "Augmentin",
the brand name for the antibiotic potassium salt
potassium clavulanate combined with amoxicillin.


The above limitations of even the best speech apps is why
I got such a big kick out of the asinine statement by "andy t"
regarding the scientist who used that speech app to decipher
the 603wpm of the fastest female talker.

That scientist is limited by the same technology as the
rest of us, even if he spends $100,000 on a specialized
speech app program.

Present technology uses HMM (Hidden Markov Model)
which can not handle homonyms very well.

Maybe 30 years from now speech apps might be able
to use monster AI programs to handle homonyms,
but don't hold your breath waiting.


My sluggish computer, and your probably faster
computer, can easily handle 480wpm.

Don't believe it?

Practice speaking two reps of the sentence:

"now is the time for all good
men to come to the aid of their cat"

....in four seconds.

It can be done with a concerted amount of practice,
running the two sentences together without a breath.

2 x 16 x 15 = 480wpm


It is a short hop from there to get a really fast
modern PC to handle the 603wpm of Fran Capo,
the worlds fastest female talker.

Mark-

 

"Mark Conrad" <mconrad@earthlink.invalid> wrote in message
news:041020090818446432%mconrad@earthlink.invalid...<!--coloro:blue--><span style="color:blue <!--/coloro-->
> In article <2badnVC757sWB1XXnZ2dnUVZ_uidnZ2d@giganews.com>, Allen
> <allent@austin.rr.com> wrote:
><!--coloro:green--><span style="color:green <!--/coloro-->
>> That's quite a test. I might mention that I can type approximately
>> 5,000 characters per minute--IF they are all the letter n.<!--colorc--><!--/colorc-->
>
> Yep, that is why I mentioned in my post that if I screwed up
> and mispronounced even one word in those 384 words,
> that I would not have got the 100% raw accuracy which I
> actually achieved, in 3 seperate consecutive runs of the test.<!--colorc--><!--/colorc-->

That's nothing. I can balance a pint of beer on the end of my cock while
reciting Shakespeare's MacBeth.

GtG