Pushing MacSpeech to its limits, frustrating
Mark Conrad
Just for fun, I wanted to push the low-cost version
of MacSpeech to its limits, to see what it is
capable of doing.
Below my signature line is the resulting (corrected) text
from my verbal dictation, using a standard low-priced
version of MacSpeech Dictate 1.5.5 ($200)
Used standard mic and USB adapter that came
with MacSpeech.
MacSpeech was pulled out of its box, installed in a
two-year old MacBook Pro with 4 GB of ram, running
Snow Leopard.
MacSpeech app was given minimum training,
about 20 minutes, to "learn my voice".
Next I imported 21 specialized medical words into
MacSpeech, which I knew that the $200 version
did not contain, words such as "annulopapillary".
Then I imported a few simple text macros.
For example, when I spoke the term "WL Gore",
MacSpeech would print:
"(W. L. Gore & Associates, Flagstaff, AZ)"
...without the quotes, of course.
Doctors worldwide know of outfits like W. L. Gore
who sell medical hardware, so they keep simple macros
in their speech apps to save time dictating addresses.
All the above chores took a few minutes, so now I
have a raging bull of a speech app that can handle
medical dictation, just like the MacSpeech $600 medical
version, right?
Well, not really. I tailored the $200 app to handle
just ONE lonely medical article, which I was getting
ready to dictate into MacSpeech for the very first time.
If I had wanted to get the $200 version to handle any
medical article that I threw at it, I would have had to
spend a year to customize the low-cost version.
Not practical.
Mark-
Okay, below is the _corrected_ text that resulted
from my first-time dictation of the complex article.
Of the approx' 600 words, there were a total of 7 text
mistakes, varying from simple mistakes like "there"
instead of the correct "their" ...
... to more serious mistakes like "valve you are"
instead of the correct word "valvular".
That is a raw accuracy of 98.8%, about normal for
complex medical dictation, using modern speech apps.
... corrected dictation results immediately below:
--
New Surgical Procedure for Ischemic/Functional Mitral
Regurgitation: Mitral Complex Remodeling
Hirokuni Arai, MD, PhD*, Fusahiko Itoh, MD, Takeshi
Someya, MD, Keiji Oi, MD, PhD, Kiyoshi Tamura, MD,
PhD, Hiroyuki Tanaka, MD, PhD
Department of Cardiothoracic Surgery, Tokyo Medical and
Dental University Graduate School of Medicine, Tokyo,
Japan
* Address correspondence to Dr Arai, Department of
Cardiothoracic Surgery, Tokyo Medical and Dental
University Graduate School of Medicine, 1-5-45 Yushima,
Bunkyo-ku, Tokyo, 113-8519, Japan Email:
hiro...@tmd.ac.jp
On-pump beating heart mitral complex remodeling was
performed without aortic clamping. The mitral valve was
exposed through a left atriotomy posterior to the
interatrial groove. Interrupted 2-0 braided horizontal
mattress sutures without pledgets were placed around
the annulus to optimize exposure of the subvalvular
apparatus. Secondary chords to the anterior leaflet
from both papillary muscles were carefully separated
from primary chords with a nerve hook and were divided.
Two pairs of 5-0 and 4-0 Gore-Tex sutures (W. L. Gore &
Associates, Flagstaff, AZ) were each placed to both
fibrous portions of the anterior and posterior
papillary muscle tips, buttressed with pledgets of
autologous pericardium. Two pairs of the free arms of
the 5-0 Gore-Tex sutures were twice passed through the
free edge of the middle portion of the anterior leaflet
about 5 mm from the margin, from ventricular to atrial
side. Suture length was adjusted to be the same length
as the corresponding marginal chords, and the sutures
were tied.
Each pair of the free arms of the 4-0 Gore-Tex sutures
was passed through the posterior annulus at sites
around the border of the lateral and middle portions
and middle and medial portions of the annulus,
respectively (annulopapillary suture), and was also
passed through corresponding sites in the annuloplasty
ring (Carpentier-Edwards Physio; Edwards Lifesciences,
Irvine, CA). The 26-mm semi-rigid annuloplasty ring was
then seated. The annulopapillary sutures were pulled to
retract the papillary muscle tips closer to the
annulus, to the point at which leaflet coaptation
occurred in the plane of the mitral annulus during
systole, to visually confirm no residual MR. Suture
lengths were determined, and the sutures were tied.
To avoid air embolism, a vent cannula with a
pressure-monitoring catheter (TOYOBO Co Ltd, Osaka,
Japan) was inserted into the left ventricular apex and
was connected to the suction circuit equipped with a
small reservoir chamber (Senko Medical Instrument Mfg
Ltd, Saitama, Japan). During the final adjustment of
the annulopapillary suture length, this chamber was
filled with blood, and the height of the fluid level of
this chamber was adjusted to load the left ventricle.
The left ventricular systolic pressure was monitored to
keep it slightly lower than the systemic perfusion
pressure to avoid ejection through the aortic valve.
This new technique has been performed on 3 patients
with ischemic/functional MR. The patients were aged 61,
64, and 69 years; their ejection fractions were 0.34,
0.25, 0.32; their left ventricular diastolic diameters
were 62, 74, and 79 mm; and tenting heights were 11,
12, and 14 mm, respectively. Preoperatively, all
patients showed severe MR; perioperative
transesophageal echocardiography showed disappearance
of MR. Mitral valvular function has remained stable
during a mean short-term follow-up of 6 months (range,
1 to 12 months), with no or trivial MR noted.
Mark Conrad
Michelle Steiner <mich...@michelle.org> wrote:
> > Below my signature line is the resulting (corrected) text
> > from my verbal dictation, using a standard low-priced
> > version of MacSpeech Dictate 1.5.5 ($200)
>
> Why are you using an outdated version?
Because it was handy - there are still a lot of versions 1.5.5
sitting on the shelves of stores, waiting to be sold.
Version 1.5.6 is a minor upgrade, fixes some crash issues
that some uses had. (not me, however)
The _big_ breakthrough version was older version 1.5.0
which allowed users to add their own words to the default
vocabulary that ships with MacSpeech.
Those user-added words still do not work quite right,
but MacSpeech will fix them eventually.
The word "TOYOBO" for example, I was forced to
create a macro in order to get that word recognized.
(...as in: "To avoid air embolism, a vent cannula with a
pressure-monitoring catheter (TOYOBO Co Ltd, Osaka,
Japan) was inserted into the left ventricular apex..." )
BTW, sad situation in comp.sys.mac.apps last week,
a doctor was pleading fo advice about how to keep from
being drowned by paperwork.
You know me, brevity is my forte, in 25 words or less,
I advised the doctor to forget about MacSpeech, and use
Dragon NaturallySpeaking medical at $1,600.
,,,and I am accused of being a shill for MacSpeech.
Go figure.
Mark-
Mark Conrad
<mco...@earthlink.invalid> wrote:
> Of the approx' 600 words, there were a total of 7 text
> mistakes, varying from simple mistakes like "there"
> instead of the correct "their" ...
>
> ... to more serious mistakes like "valve you are"
> instead of the correct word "valvular".
>
> That is a raw accuracy of 98.8%, about normal for
> complex medical dictation, using modern speech apps.
As a follow-up to my posts, I think it is important to
mention that the MacSpeech app is capable of learning
from its mistakes.
That means a user does not have to settle for 98.8%
accuracy, 100% accuracy is often obtainable, no text
correction needed.
Correction mistakes by following the user-manual method
never worked too well for me, the same text mistakes
would occur over and over again, despite repeated
correction efforts.
So I devised my own unorthodox correction methods.
The following is how I achieved 100% accurate text
without needing to do any text correction whatever:
Again I installed MacSpeech from scratch, just as it came
from the box. I did the minimum 15 minutes of training.
Then I added 4 very small files from my previous "trained"
MacSpeech:
1) User Profile (30 MB)
2) Special Medical Words, 21 of them (16 KB)
3) Six of my macro commands (37 KB)
4) Sample text file (3 KB)
The sample text file was dragged into the
"Tools -> Vocabulary Training" window.
Now for the 2nd attempt at dictating that complex
medical article, so we can see if MacSpeech is actually
learning from my offbeat correction efforts.
I spent 8 minutes speaking in that 600 word article,
at a leisurely rate of 75 wpm.
100% accurate, not one comma, period, hyphen,
semicolon, colon, paragraph, or word out of place,
no text correction needed.
Now get this, just to be ornery I had my TV blaring away
loudly in the background, which had no effect whatever
on the accuracy.
No special microphone, just the one that came in the box
with MacSpeech.
Just a matter of configuring the mic properly.
Now I challenge others out there to duplicate what I did
so far as getting 100% accuracy in challenging conditions.
If an 80 old senile duffer like me can do it, I see no reason
at all why others out there can not do the same.
Let's fulfill Steve Jobs desire to get speech recognition
going on the Mac, which _he_ tried to get going in 1999.
<http://www.voicerecognition.com/news/05_10_99.html>
That effort fell through, Steve said because of
"unforeseen circumstances"; he never elaborated as to why.
No reason for OS X to be without modern speech recognition,
Windows has had it for over a decade, they even have it
included in their OS, for several years now.
Mark-