COMMENT ⊗   VALID 00012 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Artificial Intelligence, Speech Recognition 
C00003 00003	Introduction
C00011 00004	The Analytic Engine
C00023 00005	The modern Digital Computer
C00030 00006	The Introduction of Software
C00036 00007	Artificial Intelligence
C00043 00008	Robots and Scene Analysis
C00054 00009	Language Translation
C00058 00010	Why Speech Recognition
C00061 00011	Speech Recognition or Speech Understanding
C00067 00012	Conclusions
C00068 ENDMK
C⊗;
Artificial Intelligence, Speech Recognition 
   and other Esoteric Uses of Computers

		Outline

1.  Introduction
2.  Early Beginnings of Computers
3.  Charles Babbage and the Difference Engine
4.  The Analytic Engine
5.  The Modern Digital Computer
6.  The Introduction of Software
7.  Artificial Intelligence
8.  Game Playing
9.  Robots and Scene Analysis
10. Language Translation
11. Why Speech Recognition
12. Speech Recognition vs Speech Understanding
13. Conclusions
Introduction

The modern  digital computer has  become so  much a part  of our scientific  and
industrial life that it seems very presumptious for me to begin my talk today by
first  describing  the  computer  and  stating  its  limitations.    I  do  this
deliberately  because I  want  to dispell  some  of the  mystery  that seems  to
surround the  computer, at least in the popular press.  Were I to launch at once
into a discussion of some of the more esoteric uses of computers, there would be
a real danger  that I would augment this air of  mystery rather than dispell it.
There is really nothing very hard to understand about computers.

Computers are indeed marvels of engineering perfection and their fantastic speed
and accuracy makes it  possible for us to get  them to do a prodigous  amount of
simple arithmetic  in a very short period  of time.  Unless we  take the time to
understand how  just  ordinary people  are  able to  use  computers to  do  very
complicated things,  we can easily be  mislead in to thinking  that the computer
must indeed possess some mystical powers.

Early beginnings

Computers  are, of course not  nearly as new  as most people think.   I could go
back say to  Pascal, who in  1642 and at  the age of  19 built the first  simple
digital calculating machine.  There  were others, Leibnitz for example who built
early mechanical calculating devices.   Unfortunately these early machines  were
never very dependable and they fell into disuse.   Some 200 years were to elapse
before  mechanical devices of  this general sort  were to be  reliable enough to
find general use.

These early machines  and their  modern descendents  were all of  the type  that
required  human intervention  at every  step.   It seems  reasonable to  inquire
whether  it might  not be possible  to do  a complete  calculation without human
intervention.  The first suggestion that such a machine could  be made came more
than  150 years ago  from an  English mathematician Charles  Babbage.   In fact,
Babbage understood rather  clearly all or  substantially all of  the fundamental
mathematical  principles  that  are embodied  in  the  modern digital  computer.
Babbage's contemporaries  marvelled at  him  although few  understood him.    He
failed to  realize his  dreams both  because of this  lack of  understanding and
because the inadequate state of general enginering knowledge at the time.

Charles Babbage

Charles Babbage was a most peculiar person and I could easily spend all the time
that is  at my disposal in talking about him.  Let me begin in 1812 when Babbage
was all of 20 years old.  The story goes that he was sitting in his rooms at the
Analytical Society, which he had  helped found, looking at a table of logarithms
which he knew to be full of mistakes, when the idea occurred to him of computing
all such  tabular functions by  machinery.  The  French government  had produced
several tables  by a new method in which a  very few mathematicians outlined the
general procedures that were to be used to make the desired calculations, then a
somewhat  larger  group  of  less   proficient  mathematicians  broke  down  the
operations  involved into  simple stages,  each one  of which involved  only the
operations of simple  addition or  simple substraction.   These tasks were  then
assigned to some 80 men who  only knew these two simple operations.  Babbage was
siezed with the idea that this final  labor could be done entirely by  machinery
which would be faster and certainly mush more reliable.

Babbage proposed to use a  simple method known as the method  of differences and
so  he called  his  proposed machine  a  Difference Engine.   Many  of  you are
doubtless familiiar with this method.

Babbage  constructed  a  small  working  model  of  his  Difference  Engine  and
demonstrated it  in 1822 and  then proposed to build  a much larger  machine the
would  work to twenty decimal places and  sixth-order differences, with which he
said, useful tables could be  constructed.  Babbage was first supported  in this
endeavor  by  the government  and  by  the Royal  Society  but  Babbage had  the
propensity of letting his ideas run away with him.  For example, had he  settled
for a  machine working to  ten decimal positions  and 2 or  at most 3  orders of
differences it  would have revolutionized the contemporary  table making, but no
he must  try for 20  places and  6th order  differences.   The first  difference
engine used  extensively for  real work did  not appear until  1913 and  it only
worked to  15 decimal places and used first order differences.  Babbage was that
far ahead of his time.
The Analytic Engine

In 1833, while construction  of the Difference Engine was suppended  for a year,
Babbage  conceived the idea of  his Analytic Engine, which  he realized would be
far more versital than  his Difference Engine. It would  be able to perform  any
calculation whatsoever. It was to be the first digital computer as we understand
the  term today. Of course, it was not  to use transistors or even vacuum tubes,
such things were  still 100 years away.   It was  to be a completely  mechanical
device, made up of cogwheels  and levers, but it was to embody substantially all
of the basic ideas that make the  modern digital computer such a useful  device.
In terms of the earlier analogy to the French Government's method the Engine was
to  take over  not  only the  task of  the  80 computers  but also  that  of the
intermediate group of mathematicians who controlled the distribution of tasks to
the actual workers and who compiled the final results.

Babage was to spend the rest of his  life in an attempt to build such a machine.
The art  of working metal to close tolerances had not reached the state in which
the cogwheels and levers could be produced in the quantity and to the tolarances
required, so Babbage spent years  in inproving lathes and gear cutting tools. He
employed as his shop foreman  the man who later  as Sir Joseph Whitworth  became
famous as one  of the best precision engineers  in the world and  who introduced
the first series of standard machine screw threads.

While Babbage  wrote more that 80 books and papers,  his writings were very hard
to understand.  and he was so  concerned with the Analytic Engine that he  quite
failed to document  it adequately.  Were  it not for the writings  of an English
woman  the Countess  of Lovelace we  would know but  little of  his basic ideas.
Lady Lovelace can also perhaps  be called the first programmer as she  wrote and
published a  a complete program  for computing the  Bernoulli numbers by  a very
sophisiticated method.

Babbages machine was to contain four distinct parts, the first, which he  called
the mill, was to be a unit capable of doing the normal operations of arithmetic.
Modern  computers  also  contain such  a  unit.   In  the  smaller  machines the
arithmetic organ can only  add and subtract while in  the larger machines it  is
capable of doing multiplication and division.

The second part was to be a store, that  is a part of the machine used to retain
numbers  that are needed during the process  of computation.  Babbage planned th
store 1000 numbers each of 50 decimal digits.  John Von  Neumann, who is usually
credited with  much of the  credit for the  modern digital computer,  was firmly
convinced that  this was  all the  memory that  would ever  be needed.    Modern
digital computers now routinely store many more numbers.

The numbers stored in the machine were to be of  two types, the first type being
the data  that was to be  manipulated by the machine, and  the second type being
actually a  list  of the  operations  to be  performed.  We  call this  list  of
operations  the  program  and the  elements  of  which  it  is made  are  called
instructions.  by these terms.  But more on this later.

We call this part of  the machine the memory but I think that  Babbage's term is
much to  be prefered as it does not  contain any anthropomorphic implications as
does the term memory.  One can with equal justification call a piece of paper on
which one has written a note to one's self a memory.

The third and very essential  part of Babbage's machine was to be  a device that
could  compare  two  numbers  and  that  could  alter  the  source of  the  next
instruction depending upon  which of these two  numbers was the  larger.  It  is
this part  of the machine  that sets the computer  apart from other  devices and
makes it  possible for it to do so many useful things.  It is simply the ability
to compare two numbers with each other and to note which is the larger.

It is here that the modern computer jargon gets in the way of understanding.  We
refer to this as a decision unit and talk about the computer deciding what to do
next.   One can  with  equal justification  say that  the room  thermostat is  a
decision device  that decides to turn on the heat  when the temperature falls as
to say that  the computer decides  anything.  The  only difference between  this
computer  function and  the thermostat  is  one of  degree.   The  input to  the
thermostat  is fixed and is the temperature of  the room while the input to this
portion of the computer can be any two numbers given to  the machine or computed
by it. At  different times during the computation these  numbers can be measures
of quite different things.  Calling such a simple act the power of judgement  is
to cloak the operation with an anthropomorphic term which  all to often leads to
confusion and cause one to ascribe human judgements to an inantimate machine.

The  fourth an  final  part  of Babbage's  machine  was  to be  an  input-output
mechanism which was  to allow the operator to feed numbers and instructions into
the machine and  to extract the results  of the calculations  from it.   Babbage
proposed to use the punched card as developed by Jacquard for this purpose.  And
remember,  this was in 1833!  Babbage also  proposed to use the punched card for
the storage of intermediate results  when ever the accumulation of  such results
exceeded the  capacity of his store.   He planned to have the  machine punch out
these results for the operator to file  externally and to sound a bell when  one
of these externally  stored numbers was needed  and to display a  calling number
which  would enable the  operator to extract  the desired card  from an external
file and feed it into  the input mechanism of the  machine.  Of course we  today
use  magnetic tapes  and magnetic  disks  for this  purpose and  the machine  is
capable of spinning through the tape or searching the disk for the desired item.
Strange enough many installations still depend on a human  operator to install a
desired tape or disk when  the magnitude of the required data store exceeds some
limit.

Babbage expected his  machine to be mainly used  for numerical calculations, but
his vision extended well beyond this.  For example he considered the problem  of
programming his  machine to  play chess and  concluded that  this would be  well
within its  capabilities.  Alphabetical information also did  not seem to him to
pose any special  problems, one simply  assigned numbers to  the letters of  the
alphabet  and then  proceded  to process  these  numbers.   As  an aside  modern
computers still work in exactly this way.  Of course one usually does not try to
add and subtract such numbers but they may be moved  about and compared, perhaps
to match a name with a previously stored record used to accumulate data relating
to that particular person etc.

Well that is all there is to the modern digital computer, An arithemetic unit, a
store  which  we call  a  memory,  a  number  comparer, which  we  dress  up  in
anthropomorphic  terms as a decision  unit, and some input-output  devices.  The
power of this device depends upon the ingenuity that people have brought to bear
to the  problem of  making this  essentially simple  device do  very complicated
things.
The modern Digital Computer

Now we skip 100 years while technology is catching up with Babbage's demands.

Two fairly independent developments led to the modern computer. One of these was
the very  important advances in electronics  which made it possible  to do by by
electrical means things that  either could formerly  only be done by  mechanical
devices  or could  not  be done  at  all.   The second  was  the development  of
punched-card  accounting machines which lead to  a proliferation of input-output
devices.  The technology  was waiting when the  impetus of the second  world war
forced the development of faster computational methods.

During the war several different groups of people rose to the challenge and made
tremendous progress in  the final realization  of Babbage's old  ideas.  It  is,
however, not  clear as  to how  much this  work was  sparked by  a knowledge  of
Babbage's work and how much of it was a rediscovery.

Two specific  groups made so much progress  that they deserve especific mention.
One of these groups was at Harvard University where Professor H.H.Aikin with the
help of IBM designed and built the  Mark 1 computer. When completed in 1944 this
machine did  not have all of capabilities envisioned by Babbage in that its only
choice was to continue  or to stop and it  could not modify its own  program  in
the manner described by Lady Lovelace.  Couriously enough Aikin followed Babbage
in visuallized  that the  chief function  of  his machine  would be  to  compute
tables. Many volumes  of such tables were produced  and printed.  Most  of these
tables were  never used because as soon  as computers became generally available
it was found to be much simplier to recompute any desired functional  value than
it was to  store the necessary tables and  spend the time to look  up the needed
value.

The  second  major group  of  people  was  at  the Moore  School  of  Electrical
Enginering  at   the  University  of  Pennsylvania   where  Dr  J.W.Mauchly  and
Dr.J.P.Eckert designed and built what  was perhaps one of the largest  machines,
that is largest physically, that has ever  been built.  This development is also
noteworthy  in that  it engaged  the interest  of John  Von Neumann, who  was an
already well known  physist and mathematician.   Von Neumann and his  associates
Arthur Burks  and Herman Goldstine were  the first modern  workers to understand
the digital computer in the way that Babbage and Lady Lovelace understood it and
of course they had  the advantage of 100 years of  technological developments to
make  it possible  to realize these  ideas.   Von Neumann's reports  were widely
circulated and led  to a proliferation  of computer projects,  at first in  many
Universities.  It was not long  until commercial firms became interested and the
race was on.

Now I am going to skip at once  to the modern computer of the 1970's which  uses
the  most  advanced  technology.   I  will  not  attempt  to  explain  the  many
interesting developments that make it possible to compress most of the functions
of the computer onto a very few chips of silicon, little chips of single crystal
material each approximately  one quarter of an  inch square, nor will  I be very
specific  as the the precise  capabilities of specific computers.   This is a an
entirely different subject.  But I  do want to give you a feel for  the kinds of
things that are being done and can be done in exploiting the capabilities of the
present day  computer  and  I want  to  introduce  you to  some  of  the  modern
vocabulary of computers.

There is, of course, the work-a-day functioning of  the computer in handling our
accounts at  the bank, in computing our pay checks, that  is for those of us who
are fortunate enough to get pay checks, and in doing all those things  that make
the computer industry one of the fastest growing segments of our economy.  But I
will  have to  skip over  all of this  if I  am to  say anything  about what was
advertized as being the subject of my talk.

I have  taken pains to  debunk the  computer and to  show that  it is simply  an
inanimate collection of  materials that can only do what we tell  it to do.  How
then can we  talk about having  it do such  complicated things as  understanding
speech?  Well it isn't easy.  It all depends on  the sharing of the task by many
workers.
The Introduction of Software

The early  workers,  and in  this we  must  include Babbage  and Lady  Lovelace,
realized that the problem of programming a computer was indeed a horrendous task
and it would be foolish to  require every programmer to start from scratch  this
every time he wanted  to get the computer to do the simplest  task.  Many things
are  done over and over again.  It  seemed reasonable to collect the programs to
do these  simple things into  a library and  to make  this library available  to
every one.   It had  become the fashion to  call the physical  components of the
computer the computer hardware and so it seemed quite natural to some people  to
refer to this as computer software.

Computer software,  as it is  used today, is  of two basically  different types.
The first type is  what is usually known as the operating system.  The operating
system is simply a program, and  usually a very large program, which  interfaces
the  computer hardware  with  all users  and  allows the  users  to express  his
requirements  in a much more  concise manner than  he would be able  to do if he
were dealing with the bare machine.  In most  modern computer installations, the
operating system is  so intimately tied to the computer  hardware that it cannot
be dispensed with and users  think of the computer  and its operating system  as
simply the computer.  The operating system attends  to task of providing service
to many  users, either by providing queques into which  the users may feed their
programs or by so called time sharing in which the computer can switch  from one
job to another so fast as to give each user the illusion that it working for him
alone.

The  second type of software is actually a  library of programs that one is free
to use or not  use as fits ones needs.  Some of these programs are  really quite
basic, for  example one type  is called an editor.  When one stores  data into a
computer one must specify where in the computer memory the data is to be stored.
The modern  editor program  allows the user  to type  into the computer  without
worrying about  such a mundane matter.  Then there  are programs that are called
compilers that, in  effect do a  translation job from  a user oriented  language
language into  the machine language, that  is into the basic  sequence of simple
operations  of adding and subtracting numbers  and of making comparisons between
these numbers. These programs are  so basic that again the user comes  to regard
them as being in effect a part of the machine.

In connection with  the general subject of software, you  may be interested in a
brief mention of  the so called  Bundling controversy.   In the  early days  the
software package  became so  useful and  indeed so  necessary that the  computer
manufacturers  adopted the practice of  including the software  package with the
machine.  As  the cost of producing  this software package  began to grow,  some
users thought that they  could produce their own software at  less cost than the
amount  they were forced to pay the  manufacturers. Indeed there came into being
separate companies who would undertake the task of writing speciallized software
packages for users. Finally a real controversy developed, the matter was refered
to the courts and the  manufactures were forced to  Unbundle, that is to  charge
separately for the software from the hardware.
Artificial Intelligence

First as to what  we mean by Artificial Intelligence.   This is a term  that was
coined by  John McCarthy in the early days of  such developments to include work
with computers where we are attempting to get the computer to do  certain things
that people can do, and playing chess  is an example, which when people do these
things we assume and rightly or wrongly, that they are using their intelligence.
It  was  never intended  to  suggest  that  the  computer  would  be  exhibiting
intelligence what ever that term means but only that the task would be performed
without human intervention other than that required to write the program in  the
first place.  One  way of  looking  at the  matter is  that  one exercises  ones
intelligence ahead  of time in the process of writing  the instructions.  But do
not jump to the conclusion that one must anticipate the exact situation that the
computer program will  face, in terms of  the chess problem that  one will store
answers  to  every board  situation that  will ever  be  encountered in  a game.
Rather the instructions are  generallized and expressed in such a  way that they
can  still be  applied to  situations that  have  never been  seen before.   The
instructions must be  very mush  more detailed than  those one would  give to  a
person who  is just learning the  game,since the computer is  much more childish
than say a 6 year old but they must be of the same general nature.

A great deal of work was done in the early days in trying to get the computer to
play games, not just to follow the rules of the game but to play  well enough to
challenge a human opponent.  While this proved to be a very interesting thing to
do,  the purpose was never to  achieve any specific goal  as far as game playing
skill per se,  but games seemed  to offer a  simplified subset of problems  from
those  encountered in  real life  situations. It  was thought  that it  would be
easier to deal with such a  subset at the start.   It was, of course easier  but
the magnitude of  this task was  but poorly understood.   Many wild  claims were
made  as to the  results that were  soon to be  achieved. I  myself decided that
chess would be a bit difficult and so I chose to work on checkers, But even here
I was wildly over optimistic.

I was at  the University of Illinois at  the time.  This was in  1947 before the
day of  the commercially available computer.  We had decided to build a computer
at Illinois and  we had talked  the trustees out  of quite a  sum of money  with
promises that we could build  a really good machine with this amount. We started
to design it and soon realized that we  would need much more money than we  had.
So what to do.  A few of us got togather and decided that the thing to do was to
build a small prototype right away and to get it to do something exciting by way
of demonstrating our competance.  If  I remember corectly this was in the  early
fall of  1947.  It happened  that the World's  champion checker match was  to be
held in a neighboring town of Kankakee and it seemed like a good idea to write a
program to  cause  our  proposed machine  to  play  checkers and  challenge  the
champion to a  game and hopefully beat him.  So I,  thinking that this should be
an easy thing  to do I accepted  the job.   Obviously our computer was  not even
built in  time and now some 28 years later my  checker program is still not able
to beat the world's  champion altho it once  played him to  a draw in one  match
when he was not playing too carefully.

I would like to tell you about my checker program but if I am to be  able to say
anything at all about other AI endeavours I  will have to pass this by.  You may
be  interested to know that there is still  a great deal of effort being devoted
to writing chess playing programs and there is an computer chess tournament held
each  year in  connection  with a  national  computer meeting  at which  various
programming groups  compete.    The  best programs  are  hardly  better  than  a
reasonably good amateur.
Robots and Scene Analysis

Another form of A.I.  research that still bulks  large in the overall effort has
to do with the control of robots.  A long range goal might be an automobile that
would steer itself down the  road, Or we might  be interested in a device  which
would do the  entire process of assembling  some sort of device.   Almost all of
such  operations envolve the use of visual  inputs to control the operations and
it is just  here that we  encounter the greatest  difficulties. We start with  a
television camera  to gather  the necessary input.   We can  use two of  them to
simulate our two eyes.   Of course we must have  some sort of mechanical  device
controlled by electric  motors to do the final  manipulations.  By far  the most
important  part is  a  computer program  that must  prescribe  the step  by step
process that has to be  followed.  We must  start with a television picture  and
extract some  meaningful representation  of objects  in three  dimentional space
that  can be stored  within the  computer as lists  of numbers.   These lists of
numbers must  be manipulated  by the  computer,  always in  terms of  performing
arithmetric  operations on them  and of  comparing one  number with  another, to
eventually cause  the computer  to dispatch  signals,  again in  the form  ao  a
sequence of numbers, signals to control some output device that translates these
numbers into action.

This turns  out to be a  task largely of data  reduction. The television cameras
merely report the relative light intensity at every point in an array of perhaps
one quarter  of a  million elements as  a function of  time, with  a measurement
being  made perhaps 30 times a second.   This is a nearly unmanageable amount of
data and we  have the task of  writing a list  of operations based on  comparing
specific  numbers with each  other that  will compressing this  into a  very few
numbers that can then be used to control some output device. If the object being
viewed is a person that  has unexpectedly appeared on the road ahead  we are not
interested  in  the  color and  relative  abundance  of the  hair  on  his head.
Never-the-less this information and more  is still presented to the computer  by
the camera.   In a sense, we ask the  computer to do a much harder  job than our
central  nervous system is asked to do when we  do the same task because some of
the data processing  is done  within our eyes.   Motion in  the visual field  is
detected  quite independent of  the details  of the  image, particularly  in the
peripherial region and reflexes cause the eye  to be moved to bring theimage  of
the area in question into  the fovia.  The more we learn about  how to program a
computer to  do even very simple  tasks the more we marvel  at the complexity of
human body.

Were one person to  undertake this task of programming  the computer to do  even
the simplest task it would be his entire life's work.  Fortunately, the task can
be  shared, and  it in  fact has  already been  shared by  many people  who have
contributed to the library of programs that are available for  general use.  One
person concerns himself  with adapting an Editor program  which someone else has
written so that it is  easier to write tha basic  commands that are going to  be
used  to direct  the  television  camera, that  is  to  make adjustment  of  the
direction,  tilt,  and focus,  Another  person concerns  himself  with  the data
reduction problem as it relates to the camera's output, and so it goes.

We are finding that one must supply  much more data to the computer than was  at
first thought necessary.  The human being  when faced with a typical task brings
to  bear many ideas and facts  about situations in general  that may not seem to
have any direct bearing on the  problem at hand but that he does in  fact use. I
can  quote one example  that was  revealed in  some of the  early work  on scene
analysis.  The  problem is how  do we judge  the distance to  objects seen in  a
photoghaph or perhaps on television where binocular vision is of no help.

Well, obviously,  we know  something about  the general  characteristics of  the
objects  in the  scene.  To  make this  definite, suppose  the scene is  a group
photoghaph of people.  We  know that people are roughly  of the same size or  at
least that we can  judge their size from their age and their  age is revealed by
the relative  size of their heads to their bodies and by facial characteristics,
by the kind of clothes they wear etc.  Knowing the approximate relative sizes of
the people we can also use their apparent size in the photograph to also help us
in deciding who is on the front row.  We note further that some objects obstruct
the view of others so the the obstructed objects are not seen in their entirety.
But how  do we know that there  is anything hidden?  This  we judge by knowing a
great deal about the objects, in this  case about people.  We know, for  example
that heads,  not attached to  bodies are highly unlikely  to be in  the picture.
Oh,  of course this  is a possibility  but then the expressions  on the detached
head and indeed on the other faces would not be the bland  looks that people are
apt to assume  when they are posing for  a photograph.  And so it  goes.  We are
usually quite unaware of the clues that we  use but we put them all togather  to
form a hypothesis as  to the three dimentional characteristics of  the scene and
this is what we think we see.

The early workers tried  to make allowances for all  of these factors, and more,
but still their computer programs were woefully deficient.  In fact there is one
factor that we all know about but a factor that was so obvious that it was never
mentioned.   This is now called the Support Hypothesis. We know that we are in a
gravatational field and that people have to stand on  something and that without
some evidence to the contrary this something is apt to be a roughly level floor.
So the computer  has to be  told about  floors and gravitational  fields if  the
numbers which  it  computes for  the relative  positions  of the  people in  the
phoroghaph are to be very accurate.

Before I leave this  subject, perhaps I should give you an idea as to where this
work stands.   Tremendous  strides have  been made  in  understanding the  basic
problems,  demonstrations have  been  made of  computers  doing simple  assembly
tasks,  such for example, as assembling a fairly  simple water pump out of parts
deposited at random on  the table by a mechanical  arm which is supplied  with a
few simple hand tools but this takes a million dollar computer so it is hardly a
commersially viable idea.  There are some companies that are getting  interested
in such ideas and it may be sooner than we think before the costs will come down
to where things like this will be done.

Language Translation

One of the  early efforts along A.I.   lines was the work on  Automatic language
translation.    It  seemed to  some  of  the early  workers  that  it should  be
realitively  easy to  store  within  the  computer  a  very  large  two-language
vocabulary togather with  the information as to the syntax  of the two languages
and  to use this as a  tool for language translation.  In  fact, a great deal of
effort went into  this work and  some very ingenious  devices were developed  to
store  the required  vocabulary, which  turned out  to be  much larger  than was
expected.  One such complete system was,  in fact constructed, and I believe  is
still in operation.   Unfortunately, the quality of the  translations that could
be achieved was so very poor that the entire project must be rated a failure, at
least by any objective standards, and as a result it was given up, at  least for
the time.

Why  did it  fail?   It turns  out  that a  great deal  more is  required for  a
successful translation  to be made than knowledge of the two languages involved.
A good human translator brings to bear on the problem  a very complete knowledge
of the historical  and cultural backgrownd of the  two different language groups
involved.  A speaker or writer, in using language, depends upon the fact that he
is communicating  ideas to someone  having a common  heritage and he  conveys by
implication  much more  total information  than that  contained in  the abstract
meaning of  the words  that he  uses.   A translator  most  extract this  deeper
meaning from what is said, then he  must express this deaper meaning in words of
the  target language,  not by giving  all of  the background  knowledge that the
original words  evoked  in  the reader  but  now  depending upon  the  differing
cultural backgrown  of the new  reader he must, with  a minimum of  words try to
evoke the same general intelectual response.

Workers in this  field are now concentrating  their attention on the  problem of
language understanding  and some real progress is being made  in this field.  In
this work, the computer is being used as a tool and a very valuable tool indeed,
but  most  of  the  progress  is  intelectual in  terms  of  a  deeping  of  our
understanding of the way people use language to convey ideas.
Why Speech Recognition

The final subject that I  want to talk about is Speech Recognition,  not that we
are  very far alone in  this endeavour but  because it promises to  be the major
factor to bring the use  of computers to the  common man, not as something  that
sends him bills and makes mistakes but as  a device that he can use to do things
directly  for him personally  like acting as  a private secretary  in typing his
letters, maintaining his personal record of his bank account  balance and of his
budget, reminding him  of his appointments and of  his mother-in-laws birth day,
making ot  his income  tax return, and  even ordering  groceries and paying  his
bills.   What prevents just ordinary people from  making direct use of computers
has been firstly  the cost  of using  them which is  coming down  rapidly and  a
certain mumbo jumbo that has grown up  around them and the need to abreviate the
language  used  for  programming  them  to  conform  with  the requirement  that
everything had to  be typed in.   With voice input  there will be the  necessary
pressure to get  rid of the mumbo gumbo and design  computer systems that can be
instructed in ordinary English and there will then be nothing in the way of just
ordinary people using the computer without any outside help.
Speech Recognition or Speech Understanding

During th 1950's there was an early attempt at getting the computer to recognize
speech.   Just as in the case of language  translation this effort was premature
and premature in this  case for several  reasons.  One of  these was simply  the
matter of computer speed.  A speech  recognition system that can  not operate in
real  time, that is  keep up  with the normal  rate of speech  production of the
usual speaker  is of  little use.  Or,  of course,  one might  tolerate a  brief
cogitation period, people frequently seem to require this but if the recognition
time is  say 1000  times  real time  well  that is  a  different story.    Early
computers, in spite of their speed were simply not  fast enough.  There was also
a  question of mathematical  techniques.  Computers  have increased  in speed by
several orders  of magnitude  since that  time, aand  some  very important  fast
methods have been  developed in doing certain types  of mathematical operations.
For those of you with a mathematical background I might mention the Fast Fourier
Transform.

These factors  were  and are  important but  just  as in  the  case of  language
translation  the trouble was  that the  work concentrated on  speech recognition
when the problem was rather that of speech understanding.  It turns out that  we
ourselves are hard put to recognitize a speech utterance  when there is not some
understandable content  to the speech.  The telephone  company was well aware of
this fact.    It has  long  been customary  to  test  the quality  of  telephone
equipment by making  listening tests using nonsense syllables.   The scores that
even trained  operators can achieve for the very best of telephone equipment has
always been surprisingly low.

The current work has therefore been directed toward  speech understanding rather
than  toward speech recognition.   The present  trend is  to use what  one group
calls the Hyphothesis and Test Paradigm.  This simply means that no very serious
effort is made to completely understand and  specific component of an utterance,
Rather  a  hypothesis is  formed,  based in  part  on what  the  subject  of the
discourse is all about and in part on the neighboring portions of the utterance.
This  hopythesis defines  a  limited  number  of phonetic  elements  that  might
reasonably be  expected to occupy the place in the  acoustic stream.  Having, in
effect, produced a list of this sort then an attempt is made to match the actual
utterance against these in  turn and pick out the one  that most nearly matches.
Sometimes, or  perhaps I should say usually, several different phonetic elements
match about equally  well so a  final decision is  postponed and the process  is
repeated on  an ajoining element, but  now with a some what  set of requirements
since something is now known about its neighbor.

Speech understanding has a long way to  go before it will be a work-a-day  tool.
Never-the-less real  progress is  being made  and it  is only  a matter  of time
before  it will be  a reality.   Some 4 years  ago the ARPA  organization of the
Department of Defense started a 5 year program with the objective of realizing a
workable system capable of understanding a vocabulary of 1000 words as spoken by
a limited number  of speakers in  talking about  a restricted subject.   It  now
seems that this objective will be reached.
di
Conclusions

I have ranged far and wide in this talk, I  fear, without a central enough theme
to permit any very valid summary.  My purpose has been to give you a feeling for
the problems of using computers  for unconventional tasks without adding to  the
all too prevelant sense of mystery that seems to surround the compute.  Perhaps,
by your questions you will reveal how far from the mark I have come.