Analysis of a digital technique for frequency transposition of speech DiGirolamo, Vincent

Analysis of a digital technique for frequency transposition of speech DiGirolamo, Vincent
Calhoun: The NPS Institutional Archive
Theses and Dissertations
Thesis Collection
1985-09
Analysis of a digital technique for frequency
transposition of speech
DiGirolamo, Vincent
http://hdl.handle.net/10945/21142
KKOX LIBRARY
NAVAL POSTGRADUATE S
MONTEREY, CALIFORNIA
DUui.u/
C
'3
NAVAL POSTGRADUATE SCHOOL
Monterey, California
THESIS
ANALYSIS OF A DIGITAL TECHNIQUE FOF
FREQUENCY TRANSPOSITION OF SPEECH
by
Vincent uibiro xarno
eotember 1985
Thesis Advisor
Paul H. Moose
Approved tor public release: distribution is uniimita-
.
T223455
SECURITY CLASSIFICATION OF THIS PAGE (*nan Data
Entared)
READ INSTRUCTIONS
BEFORE COMPLETING FORM
REPORT DOCUMENTATION PAGE
1.
4.
REPORT NUMBER
TITLE
2.
GOVT ACCESSION NO
(and Subtitle)
3.
RECIPIENT'S CAT ALOG NUMBER
5.
TYPE OF REPORT
7.
AUTHORfs,)
4
PERlOO COVEREC
Master '3 Thesis
September 1985
Analysis of a Digital Technique for
Frequency Transposition of Speech
6.
PERFORMING ORG. REPORT mjMBER
8.
CONTRACT OR
Ofl»N T NUMBER'!;
Vincent DiGirolamo
9.
PERFORMING ORGANIZATION NAME AND ADDRESS
10.
Naval Postgraduate School
Monterey. California 93943
11.
CONTROLLING OFFICE NAME AND ADDRESS
PROGRAM ELEMENT. PROJECT, TASK
AREA a WORK UNIT NUMBERS
12.
REPORT OATE
13.
NUMBER OF PAGES
15.
SECURITY CLASS,
SeDtemoer 1985
Naval Postgraduate School
Monterey, California 93943
99 oaqea
U.
MONITORING AGENCY NAME
a
ADDRESSf//
different from Controlling Office)
(of thle report)
Unclassif lee
15«.
DECLASSIFICATION DOWNGRADING
SCHEDULE
16.
DISTRIBUTION STATEMENT
(of this Report)
Approved for public release; distribution unlimited
17.
DISTRIBUTION STATEMENT
18.
SUPPLEMENTARY NOTES
19.
KEY WORDS
(ol the abstract
(Continue on reverse side
If
entered
In
Block
20, If different from Report)
necessary and Identity by block number)
Frequency Transposition, Speecn, Hearing Impaired, Pole-Shifti
ReflectJ
Predictive
Coding,
Linear
Content,
Frequency
Coefficient, LPC Parameters, Speech Processing
20
ABSTRACT
(Continue on reverse side
II
necessary and Identify by block number)
Frequency transposition is the process of raising or lowering
The
hearing
the frequency content (pitch; of an audio signal.
impaired community has the greatest interest in the applications
Though several analog and digital
of
frequency transposition.
frequency transposing hearing aid systems have been built
processing
tested,
this thesis investigates a possible digital
an
of
rr-domain.
the
in
shifting,
alternative.
Pole
DD
FORM
1
JAN
73
1473
EDITION OF
S
1
N 0102- LF-
NOV
65
IS
0U- 6601
OBSOLETE
SECURITY CLASSIFICATION OF 'HIS PAGE (When Data
Sntermj
SECURITY CLASSIFICATION OF THIS
20.
»»..ae (Whan
Data Ent«r«<0
(Continued)
autoregressive (all pole) model of speech was proven to be a
Since linear
viable theory for changing frequency content.
analyze and
predictive coding (LPC) techniques are used to code,
synthesize speech, with the resulting LPC coefficients related to
the coefficients of an equivalent autoregressive model,
a linear
relationship between LPC coefficients and frequency transposition
is explored.
This theoretical relationship is first established
using a pure sine wave and then is extended into processing
speech.
The resulting speech synthesis experiments failed
to
substantiate the conjectures of this thesis.
However,
future
research avenues are suggested that may lead toward a viable
approach to transpose speech.
-
•
•
." J
red
SECURITY CLASSIFICATION OF THIS P*GEf»b»n Dalm Enftmd)
Approved for public release; distribution is unlimited
Analysis of a Digital Technique
for Frequency Transposition of Speecn
by
Vincent DiGirolamo
Lieutenant, United States Navy
B.S., United States Naval Academy, 1S78
//
Submitted in partial fulfillment of tne
requirements for the degree of
MASTER OF SCIENCE IN ELECTRICAL ENGINEERING
from the
NAVAL POSTGRADUATE SCHOOL
September 1S85
ABSTRACT
raising
is the process of
transposition
Frequency
signal.
the frequency content (pitch) of an audio
lowering
or
hearing impaired community has the greatest interest in
The
Though several
the applications of frequency transposition.
analog and digital frequency transposing hearing aid systems
have
built and tested,
been
possible digital processing alternative.
z-domain,
the
speech
of
proven
was
frequency
content.
techniques
are.
with
an
viable
theory
used to code,
changing
coding
related
between
LPC
transposition is explored.
first
established
extended
into
using
(LPC)
coefficients
the
to
coefficients of an equivalent autoregressive model,
relationship
of
analyze and synthesize speech,
coefficients
LPC
model
for
linear predictive
Since
a
Pole shifting, in
autoregressive (all pole)
to be a
resulting
the
investigates
thesis
this
a
linear
frequency
and
This theoretical relationship is
a
processing
wave
pure sine
speech.
The
and
resulting
then
is
speech
synthesis experiments failed to substantiate the conjectures
of
this
suggested
However,
thesis.
that
transpose speech.
may
lead
future
toward
research
a
viable
avenues
are
approach
to
IABLE_OF_CONTENTS
I.
II.
III.
INTRODUCTION
A.
BACKGROUND
B.
FREQUENCY MODIFICATION
C.
A
9
10
NEW TECHNIQUE FOR FREQUENCY TRANSPOSITION -- 11
MODELING SPEECH PRODUCTION
13
A.
INTRODUCTION
13
B.
THE SPEECH PRODUCTION MODEL
14
C.
DIGITAL FILTER REPRESENTATION
17
LINEAR PREDICTION THEORY
19
A.
WHY LINEAR PREDICTION?
19
B.
LPC THEORY
20
C.
PARAMETER ESTIMATION
21
1.
Method of Least Squares
21
Autocorrelation Method
22
LINEAR PREDICTIVE CODING OF SPEECH
25
a.
IV.
9
A.
INTRODUCTION
25
B.
LPC ENCODING PARAMETERS
27
Unvoiced Decision Making
1.
Voiced
2.
Gain
27
3.
Pitch Period
29
4.
Reflection Coefficients
31
5.
Spectral Analysis
36
a.
/
Formant Frequencies
27
36
SPEECH SYNTHESIS
C.
V.
VI.
DIGITAL FREQUENCY TRANSPOSITION
39
A.
INTRODUCTION
39
B.
POLE SHIFTING IN THE Z-PLANE
39
C.
A
42
NEW PROPOSITION
1.
Statement of Theory
42
2.
Sine Wave Experiment
43
3.
Sine Wave Experimental Results
44
SPEECH PROCESSING EXPERIMENT
45
A.
INTRODUCTION
45
B.
VOICED/UNVOICED PHRASES
45
C.
DATA PROCESSING
46
1.
Speech Data
46
2.
Determining Reflection Coefficients
47
-
3.
4.
D.
a.
Trend Analysis
48
b.
Graphical Correlation
48
Spectral Analysis of Reflection
Coefficient Patterns
.
2.
50
Correlation Between Phrases With
Different Pitches
50
Voiced/Unvoiced Observations
50
CONCLUSIONS AND RECOMMENDATIONS
A.
49
Spectral Analysis for Frequency Content -- 49
SUMMARY OF EXPERIMENTAL RESULTS
1
VII.
38
CONCLUSIONS
51
51
1.
Complexity of Speech
51
2.
Physical/Mathematical Relationship
51
6
Periodic/Pseudo-Periodic Differences
3.
52
RECOMMENDATIONS
B.
APPENDIX A
-
53
REFLECTION COEFFICIENT DETERMINATION FOR
FREQUENCY VARIED SINEWAVE PROGRAM
A
55
APPENDIX B
-
REFLECTION COEFFICIENTS K1-K6 (NOISELESS)
APPENDIX C
-
REFLECTION COEFFICIENTS K7-K12 (NOISELESS)
APPENDIX D
-
REFLECTION COEFFICIENTS K1-K6 (S:N=10:l)
APPENDIX E
-
REFLECTION COEFFICIENTS K7-K12 (S:N=10:l)
APPENDIX
F
-
DATA ACQUISITION SYSTEM
SECTION
1
-
SYSTEM DESCRIPTION
61
SECTION 2
-
CIRCUIT DIAGRAMS
67
SECTION 3
-
SOFTWARE FLOW CHART
68
SECTION 4
-
ASSEMBLY LANGUAGE PROGRAM
69
SECTION 5
-
INTELHEX TO DECIMAL CONVERSION PROGRAM
74
--
57
-
58
59
--
60
61
APPENDIX G
-
GRAPHICAL REPRESENTATION "READY"
75
APPENDIX
H
-
GRAPHICAL REPRESENTATION "SO WHAT"
76
APPENDIX
I
-
GRAPHICAL REPRESENTATION "SNEEZE"
77
APPENDIX J
-
SPEECH RESOLUTION SUMMARY
78
APPENDIX
K
-
SPEECH REFLECTION COEFFICIENT PROGRAM
79
APPENDIX L
-
6 PLOTS OF
REFLECTION COEFFICIENT PATTERNS
FOR SPEECH WAVEFORMS
81
APPENDIX
M
-
TREND ANALYSIS RESULTS
87
APPENDIX
N
-
FFT PROGRAM FOR PATTERN RECOGNITION
S9
APPENDIX
0-6
PLOTS OF FFT RESULTS
91
LIST OF REFERENCES
97
BIBLIOGRAPHY
98
INITIAL DISTRIBUTION LIST
99
7
ACKNOWLEDGMENTS
work is the result of the efforts and
This
many
people.
attribute
I
the
support
accomplishment
of
this
of
research and every sacrifice that lead to it's completicn to
my
loving
over the past two and a half years,
encouragement
not have done it.
I
and
could
I
love you.
thanks and appreciation goes to Professor Moose
My
an interest in this topic and allowed me to pursue
took
That means a lot to me.
my way.
who
your support
Without
Diane.
wife,
acquisition,
and
its completion,
To
my
I
it
Also, to Professor Madan
worked with me on my many schemes
patiently
who
data
for
endured with me as my second reader until
am grateful.
friends,
Dennis
Poulos and
Alan
labored with me in my many academic trials,
I
Farmer,
cherish
who
many
fond memories and am not quick to forget, fair winds.
Finally,
Val
Johnson
transforming
the
your
in the most critical moments of this research,
and
Bob Scott provided me with the
and transfering the speech data necessary
speech processing portion of this
rescue
means
efforts,
I
research.
wouldn't be writing
Though words are not always adequate,
for
Without.
these
words.
your efforts will
always be appreciated, God speed, until we meet again.
8
of
INTRODUCTION
I.
A.
BACKGROUND
Adjusting
frequency
the
content
or pitch of a signal
is a topic researched within the audio field.
impaired
community
has
the
greatest
The
hearing
interest
the
in
applications
of
techniques.
This is due to their need for auditory speech-
frequency
modification
transposition
or
processing aids.
speech-processing
Auditory
groups:
those
speech
signal,
aids are divided
into
two
of
the
which involve nonradical processing
with
the
speech still intelligible
person with normal hearing,
to
and those which involve radical
re-coding of the speech signal [Ref
.
l:pp. 547-557]
An example of radical recoding involves such systems
cochlear
implants
processed
into
a
where
series
interprets as sound.
surgically
inserted
completely
different
hearing.
most
systems
normal
the
of
speech
vibrations
signal
that
the
as
is
brain
Individuals who have this type of aid
in
their
cochlear
language than
a
must
person
learn
with
a
normal
Examples of nonradical processing aids include the
widely
frequency
a
used
lowering
amplifier aids and
devices
or
the
frequency
less
familiar
transposition
Some aids may amplify
Moat hearing aids amplify sound.
or soften certain frequencies,
from the aid on one ear to the aid on the other
in either case,
purpose,
primary
is to amplify everything
however,
drive
aid
an
lowers
which
the
frequency
we
someday
interested in developing an algorithm that may
are
Their
ear.
In this thesis,
they are capable of sensing.
sound
while others transmit
content
and
preserves the intelligibility of the speech signal.
B.
FREQUENCY MODIFICATION
Pickett
methods
CRef
that
.
2:pp.
191-194]
have been used
for
categorizes
frequency
two
lowering:
1.
Frequency transposition, where a portion of the
signal is separated out and resynthesized in a lower
frequency band.
2.
Frequency division, where the frequency
signal is reduced by a fixed ratio.
All
of the methods involve
signal
distortion.
distortion tends to increase with greater frequency
Here
basic
we
are concerned primarily with the idea of
of
the
Signal
shifts.
moderate
frequency transposition, where the signal is shifted without
major distortions in the information content.
The earliest known suggestion of frequency lowering
by Perwitschky (1925).
was
The earliest transposing hearing aid
wa3 built and tested by Johansson (1955).
Since then, there
have
and
been
considering
several
other systems
built
the advances and trends of current
10
tested,
but
technology.
research
in the area of frequency transposition
of
speech
utilized
analog
has not been productive.
Frequency
techniques
transposition
such as frequency modulation (shifting an
band to a lower band);
of
a
systems have
frequency division
tape recorded signal);
playback
slow
(a
and digital techniques such as
sampling distortion (omitting segments of recorded
doppler (the delaying of the incoming
and
methods
these
speech),
Though
signal).
have been developed and extensively
approach
digital
the
upper
presented
here
tested,
produce,
may
all
together, different results.
confirms
Pickett
frequency
shifting
extensively
enough
CRef
1933
2:p.
.
obtaining
the
algorithms
have
not
usable
for
explored
been
recommendations
to make
The
possibilities
practice
for
research needs in this area include
new information on the potential for digital
general
cues can be sent in this
optimum parameters,
finding
way,
re-
finding
exploring the principles of transposition,
coding,
which
.
that
and examining what system can be
the
built
that meets our general and specific needs.
C.
A
NEW TECHNIQUE FOR FREQUENCY TRANSPOSITION
Recently,
Hall
CRef.
3:p.
56]
postulated
that
poie
shifting in the z-domain using an auto-regressi ve (all pole)
model
of
lowering.
for
frequency
He used linear predictive coding (LPC)
techniques
speech
may be a possible
11
option
to
the speech to determine if pole shifting was
process
viable
positive
were
results
experimental
His
option.
a
because he was able to create a change in pitch on the input
speech segment.
This
thesis
is an extension of
ventures
beyond
the
domain
frequency
research.
Hall's
directly with the linear predictive time domain
was
postulated
that a linear relationship
works
and
model,
It
model.
It
between
exists
frequency content and the reflection coefficients determined
using LPC
Once this theory has been postulated,
.
experiment
processing
speech
a
was undertaken to determine
if
the
introduced,
the
conjectures made were plausible.
In
report linear prediction
this
particular
algorithms
explained,
and
Identical
phrases
levels
Possible
segments
by
the
used
experimental
of speech,
same speaker,
patterns
to
is
process
research
spoken
was
at
are sampled
existing between
the
data
are
carried
out.
the
different
and
pitch
processed.
different
pitch
of speech and their linear predictive coefficients
are analyzed.
The
results of this research indicate that there is
linear
relationship
content
of speech and the LPC reflection coefficients,
recommendations
linear
that
exists
between
the
are made for continued analysis
no
frequency
and
concerning
predictive coding and the frequency transposition of
speech
12
MODELING_SPEECH_PRODUCTION
II.
A.
INTRODUCTION
order
In
synthesis,
it
elements
the
understand
combine
speech
produce
to
model
of
speech.
illustrated
below
as
NUSflL
and
the
basic
The
most,
the production of
model used to explain
human
reproduction
to consider some
useful
"is
that
elementary
is
to
speecn
Figure
1.
lRflLI
VQCflL TRflCT
Human Speech Production System [Ref
Figure
The
lungs
1
.
4: p.
42]
.
produce the air flow necessary to begin
the
tongue, mouth,
lips
generation of sound.
The vocal cords,
and nasal tract combine their different properties to
the
airflow
to produce the speech waveform we hear.
13
shape
B.
THE SPEECH PRODUCTION MODEL
Evans
CRef
.
relates the
40-45]
4:pp.
This is standard
functions to mechanical models.
and
widely
a
accepted
several
approach
practice
production
speech
to
human
the lungs are
the
excitation
source for the vocal and nasal tract areas.
An
excitation
modeling.
He
states
that
source can either be modeled as a pulse train generator or a
random number generator when reproducing speech.
In the case of voiced sounds <ie.
consonants, vowels or
nasal sounds), the air released by the lungs is periodically
modulated by vibrations from the vocal cords,
the excitation model in this case is a
Thus
velum.
generator.
glottis,
In
the case of unvoiced sounds (ie.
and
pulse
sss,
sh,
fff) which require no vibrations to be produced, the modeled
excitation source is a random number generator.
Both
form
excitation sources produce a
that we recognize as speech.
That is,
the period of
the wave form varies with time depending on the sound
of voiced or vibrated sounds.
model
point
being
This phenomena is most obvious in the production
produced.
time
wave
quasi -per iodic
more
Figure 2,
of the human speech process,
clearly.
a
general discrete-
illustrates
Here we have represented
this
the
vocal
labeled
pitch
tract model as a time-varying digital filter.
Note
period.
that
the pulse train has an input
This
input
determines
14
when the pulses
will
be
z
-
O
I
u
Q
y-
cr
0C
UJ
I
u
^
n
j
^
a
e
>-«
;
z
-«
c
i
^
i
E
en
o
ta
a
UJ
u
-•
a
cr
UJ
Z
>
z
D
\
UJ
CD
cr
£
a
a
uJ
ci
z
H<
cr
C3
cr
2:
o
»a
cr
UJ
z
UJ
ID
Q
U
o
Ul
Discrete-time Model for Speech Production CRef. 4:p. 43]
Figure
15
2.
periodicity.
what
from the pulse generator and at
emitted
This is only necessary for voiced speech.
speech is a continuous stream
unvoiced
The
The flow
commonly referred to as white noise.
numbers
random
of
of
random numbers may produce a seemingly quasi-periodic sound,
however,
since they are usually of such short duration,
consider
the sound to be continuous and constant,
and
we
not
periodic.
waveform has a specific amount of
speech
Each
energy.
The energy contained within each utterance of a set duration
will be referred to as gain <G)
body
its
also
It
voiced or unvoiced decision is
the
or gain is assigned,
drives
the vocal tract model.
Kaiser
current
Bell
of
thinking
refocused
physics
in
made
In a phone
is
Laboratories,
he
the area of speech
behind
movement to
a
the
and
an
interview
with
mentioned
that
reproduction
attention on this portion of the
its
there
that
by
the scaled excitation function
energy
James
reproduction
aids
the intensity or inflection of the voice signal.
indicating
Once
quality.
or
This is what gives speech
.
different
more
clearly
physical
has
model
and
describe
the
contributors
of
speech.
This
energy
vocal tract model is driven by the excitation
function and controlled by time varying vocal
parameters.
These
tract
to
model
and
tract
vocal tract parameters adjust the vocal
yield the
desired
16
output
waveform.
By
replacing
varying
the
vocal tract model with an
equivalent
digital filter that models the vocal tract
response,
time-
model's
we are able to step right into the next phase
of
synthetic speech reproduction.
C.
DIGITAL FILTER REPRESENTATION
Although speech is modeled most efficiently by poles and
zeros,
may
it
regressive
large
by
auto-
an
(all pole) filter if the order of the filter
For example,
enough.
filter
be modeled accurately
also
accurately
will
Therefore,
most
model
transfer
the
auto-regressive
tenth order
a
audible
function (H(z))
filter in Figure 4. is shown as Eq
.
is
the
of
sounds.
digital
1-1.
G
H<z)
(2-1)
=
P
1
-
^
K=
a k z*
I
where p is the order of the filter, G i3 the gain, and
ak
is
the filter coefficient.
G and a k are the time-varying vocal tract parameters for
For a given segment of time (i.e.,
10 milli-
seconds) the vocal tract parameters are constant.
However,
this filter.
stringing
produce
a
these
segments together in rapid
succession
to
one second interval of speech, the parameters will
change 100 times.
This is why they are referred to as time
varying; they vary over a short period of time.
17
The
type
arbitrary.
counts.
and
For
It
they
is
the
filter
used
concept behind
in
the
the purposes of this research,
attributes
because
digital
of
Figure
diagram
themselves well
coding implementation.
18
to
is
that
the properties
of a time-varying lattice filter
lend
2
linear
are
best
predictive
LINEAR_PREDICTION_THEQRY
III.
A.
WHY LINEAR PREDICTION?
Although spectral analysis is
studying signals,
from
number
a
a
well-known technique for
its application to speech signals suffers
serious limitations
of
arising
from
the
nonstationary as well as the quasiperiodic properties of the
speech
than
wave.
By modeling the speech wave
spectrum,
its
avoid
we
problems
the
rather
itself,
inherent
in
frequency-domain methods.
instance,
For
traditional
analysis
Fourier
methods
require a relatively long speech segment to provide adequate
spectral resolution.
As a result,
rapidly changing speech
events cannot be accurately followed [Ref
predictive coding is applicable to
Linear
research
of
perception.
processing
problems
One
including
the
of
technique
276-294]
5: pp.
.
wide
a
range
production
speech
main objectives
speech
any
in
is the synthesis of speech
and
which
is
indistinguishable from normal human speech.
Atal
noted
that
information-carrying
altering
the
much
can
structure
properties of the
be
of
learned
speech
speech
stated that LPC techniques can serve as
a tool
He
.
also
for modifying
[Ref
are exactly the intentions of this thesis:
19
the
selectively
by
signal
the acoustic properties of the speech signal
These
about
.
5:p.276]
to modify
signal by investigating the properties
speech
the
the
of
information carrying structure.
of this chapter is a summary
remainder
The
The major portion of this
prediction
theory.
extracted
from
Makhoul's
prediction
CRef.
6:pp.
approach,
intuitive
tutorial
124-143],
review
linear
of
section
is
linear
on
and will be based on an
with emphasis on the clarity of
ideas
rather than mathematical rigor.
B.
LPC THEORY
In applying time series analysis, each continuous signal
is sampled to obtain a discrete-time signal s(nT>,
s<t)
also
known as a time series, where n is an integer variable and T
is
The sampling frequency is
then
Note that s(nT) will be represented as s n in
this
the sampling interval.
f s =l/T.
discussion
The
system
signal
s n is considered to be the output
some unknown input u n such that the
with
of
some
following
relation holds:
P
sn
where
a^,
=
bi,
hypothesized
sn
past
"
^
8-
akSn-k
and
+
G j£
the gain G are the parameters
system.
(3-1)
b].u n
of
This equation says that the 'output'
is a linear combination of past outputs and present
inputs.
That is,
the
the 3ignal s n is predictable
20
and
from
linear combinations of past outputs and inputs.
Hence
the
name linear prediction.
C.
PARAMETER ESTIMATION
the all-pole model,
In
given
as
a
we assume that the signal s n is
linear combination of its past values and
current input u n
some
:
P
an
which
yields
=
the
^
"
following
(3-2)
Gu n
aksn-k
frequency
domain
transfer
function
G
H(z)
(3-3)
=
P
1
Given
a
a^z-k
IZZ
particular signal s n
determine
the problem is to
,
coefficients
predictor
the
+
(a^)
and the gain
some
in
G
manner
1
.
Method_of _Least_Sguares
Here
unknown,
assume
we
that
the
input
un
which is the case of speech analysis.
is
Therefore,
the signal s n can at best be approximately predicted
linearly
weighted
approximation of
sn
summation
be s n
»
of
where
21
past
samples.
totally
from
Let
a
the
p
sn =
^
-
(3-4)
a^sn-k
Then the error between the actual value s n and the predicted
value s n is given by
P
en
quantity
The
an
=
"
3n
^
+
an
=
method of least squares the parameters
a
residual
is also known as the
en
(3-5)
*kSn-k
{aj<}
In
.
are obtained
result of the minimization of the expected value or
of
the error squared term,
each
Ep
the parameters.
of
£
=
Ep is the
prediction error, averaged over all
Ep
=
with
a
~
&
n,
mean
with respect to
minimum
mean
square
and is represented by
Un
+
£"
a^ s n -k i
of the signal
set of unknowns can be
a
as
P
r
definition
any
For
equations
£<e n 2 >
(e n 2 ),
the
sn
,
solved
a
set
for
of
the
predictor coefficients which minimize Ep.
There are two distinct methods for the estimation of
these parameters,
covariance
namely the autocorrelation method and the
method.
Makhoul
CRef.
method
is
6:pp.
Both
methods are clearly described
126-127].
the preferred method,
summarized here.
22
Since the
by
autocorrelation
only that method will
be
Autgcgrrelatign_Method
a.
Here
we assume that the error Ep is
over an infinite duration.
minimized
Since
+ 00
R(i)
autocorrelation
the
is
£_
=
(3-7)
s n s n+i
function
of
the
signal
3n
,
Equation 3-6 reduces to
Ep
where
is
R<0)
+
^
a|<
(3-8)
R<k)
is the total energy of the input signal and R(k)
R(0)
autocorrelation
the
Figure
=
matrix of the
input
signal
(see
3)
S0S1
SiS 2
S2S 3
Sp-lSp
Rfcl
Rl.2
«2i3
SiS2
S0S1
SlS2
Sp-2Sp-i
"li2
R0i
1
«1i2
Ro-2,p-i
S2S3
SiS2
S0S1
Sp-3Sp-2
"2i3
"1,2
fyil
Rd-3, p-2
RP-2.P-1
Rd-3»P-2
Rfci
Sp-iSp
Sp- 2 Sp-i
Sp- 3 Sp- 2
...
sesi
Rp-liP
Autocorrelation Matrix
Figure
It
is a symmetric toeplitz matrix
matrix is one in which
are equal).
interval,
3.
(a
toeplitz
all the elements along the
diagonal
Since the signal s n is known over only
one
popular
method to control the size
23
a
finite
of
the
toeplitz
matrix
function
wn
.
is to multiply the signal s n
by a
This yields a slightly different signal
window
s' n ,
which is zero outside the finite interval.
In any case,
means
for
solving
the autocorrelation matrix is the
several
of
the
coefficients needed to analyze and synthesize
following
chapter discusses,
speech.
in greater depth,
coefficients are and how they are obtained.
24
predictive
linear
The
what those
LINEAR_PREDICTION_OF_SPEECH
IV.
A.
INTRODUCTION
mentioned earlier,
As
there are several ingredients or
time-varying parameters that are needed to generate
using
When
predictive
linear
coding
speech.
techniques,
three
ingredients are essential: gain or energy, pitch period, and
the
reflection
filter
parameters
coefficients or
.
Figure
4 illustrates the fact that,
specified frame length,
depending
on
the
these ingredients must change every
On a frame-by-frame basis the incomming signal
10 to 20 ms.
is
envelope
spectral
the pitch period and
processed to obtain the gain,
the
reflection coefficients kl, k2,...,kN.
construct
voiced
period and the gain parameters are
pitch
The
excitation function for production of
an
unvoiced speech.
or
This
driving
or
is
spectral
envelope parameters determined from the
output
is
frame
one
of
synthetic
are produced
CRef.
7 pp
:
.
337
-
either
by
the
analysis.
speech,
stringing several frames of speech together,
to
excitation
input to a filter which is configured
function
The
used
and
by
audible sounds
345].
Analysis of the speech signal is done by calculating tne
LPC
model
parameters
for each 10
ms
time
frame.
chapter will discuss these essential parameters.
25
This
Sfcp
vo UJ
Q
4
t
Uj
KJ*
^
o3
<x.
«5
1 o
O
O
^*
5*
I ^
LPC Model of the Human Voice
Figure 4.
26
CRef.
7:p.
338]
B.
LPC ENCODING PARAMETERS
1
.
¥oiced_/_Unvoiced_Deci3ion_Making
Some
vocal cords,
sounds
require the vibrations induced
while others do not.
require an excitation from the vocal
that
lips.
Unvoiced sounds are generated by
in the case of 's'
order
or
'f.
the
Voiced sounds represent
those
as
by
a
cords
steady flow of air
decision must be made
A
or
to properly excite the digital filter to produce
in
the
desired sounds.
According
unvoiced
to
Atal
[Ref.
5:p.
the
280]
voiced/
decision is based on the ratio of the mean-squarea
value of the speech samples to the mean-squared value of the
prediction
error
This
samples.
ratio
considerably
is
voiced
smaller
for unvoiced speech sounds than for
sounds.
Typically, this ratio is a factor of 10.
Voiced Decision:
ECs n ]
Unvoiced Decision: E[s n 3
This
speech
10 ECe n ]
>
<
10 ECe n
]
decision will determine whether to excite
the
digital filter with an impulse function or white noise, each
having a particular gain or energy.
2-
5§iQ_Cgmputatign
In
prediction
explaining
we
the least squares method
assumed
that
27
the
input
was
of
iinear
unknown.
Equation 3-5
can be rewritten as
P
3n =
ak
sn-k
Gu n
=
®n
That is,
•
output will be different than s n
the
input
input
output
is
signal
is
For any other input
proportional to the error signal.
the
(4-1)
®n
that will result in the signal s n as
un
where
that
^
Equations 3-2 and 4-1 we see that the only
Comparing
signal
"
Therefore the energy of
.
input signal must be equal to the energy of the
signal s n
the
output,
.
Since the filter H(z) is fixed, it is clear from the
the total energy in the input signal
above
that
equal
the total energy in the error signal,
by Ep.
Again, Makhoul
CRef
.
6:p.
128]
Gu n
must
which is given
is the primary source
for this information and he provides additional mathematical
background in determining the resultant gain equation
P
G2
where
G^
=
is
Ep
=
R(0)
+
^
KM
(4-2)
ak R(k)
the total energy in the input
and
R(k)
is,
again, the autocorrelation matrix.
The
determines
input
classification of a sound as voiced or unvoiced
the input to the filter H(z).
However
Gu n is white noise or a series of impulses,
is calculated from the same equation.
28
if
the
the gain
3
Eitch_Period
-
period
The
excitation
[Ref
each
pulse is referred to as the pitch period.
Atal
since
it
that
describes
2793
determining pitch period.
here
elapses
between
5:p.
.
time
of
is
different
two
methods
His second method is
based
on
the
for
summarized
predictive
linear
representation of the speech wave.
In this method,
of
each
pitch period,
except for
a
sample at the beginning
every sample of the
voiced
waveform can be predicted from the past values
.
speech
The method
of determining pitch period is relatively simple.
Once
the
prediction error of the speech signal
is
determined through linear predictive processing, the largest.
or
peak
values
are
(Figure
noted,
excitation
determine
the
initiated
from the excitation source.
picking
times
that
5).
pulses
This
procedure was found to be effective in
pitch period as developed in Reference 7.
29
These
points
should
simple
be
peak-
determining
Pitch Period Estimation Using Peak Picking
Fiaure 5.
30
Ref l§ctign_Coef f icients
4-
Earlier
coefficients
the
was
mentioned
that
reflection
the
related
determined using LPC are directly
polynomial
section
it
coefficients of an all
show
will
the
relationship
pole
This
model.
between
to
them
and
illustrate how the reflection coefficients are determined.
that we are looking for an estimated
Recall
which
is
Eqns.
3-4
weighted
the
and
3-5).
sum of past
system
The autoregressi ve
output
outputs
(AR)
model
(see
in
Figure 6 illustrates this process.
Autoregressi ve Model
Figure 6.
The
Ep
.
goal of LPC is to adjust the a^'s to
Achieving
minimize
it involves solution of a linear system
31
of
equations,
lattice
Figure
structure
7)
and leads
Levinson's algorithm,
using
the
to
AR model we are most interested in
(see
The mathematical development for this may
.
found in Parker CRef
.
9:pp.
be
110-112]
Lattice Structure Analysis Model
Figure
Lattice
structuring
reflection coefficients,
K's
7.
requires the determination
hereafter referred to as
of an n-th order Lattice filter transfer
related
filter
to
The
K.
function
are
the polynomial coefficients of an nth order
transfer
function
through
following
the
of
AR
matrix
equation
<N-1)
<N)
<N + 1)
(N)l
K
32
-a
(4-3)
where
- T(N)
ryy
<N«-1)
Ryy
<N+1>
-
Ryy(O)
matrix
The
1
Z
-
K
ryy
T(N)
a^
-
(4-4)
<N)
.
the laat
ia
<N)
«k
ryy
column
autocorrelation matrix mentioned earlier.
to
the
Ryy
The notation has
been slightly altered from Parker's presentation
112]
of
CRef
9:p.
.
be consistent with the preceding chapters of
this
development
Equations
4-3
4-4 have been included
and
presentation to show how the polynomial coefficients
are
(a^'s)
coefficients
<K's).
there is an easier and more direct method
towards
related
However,
this
in
determining K's.
Working
tho
A
the
reflection
brief development is presented here.
in the Z-domain,
function of the AR model is
and
33
we know that the
transfer
where
A
is A<z)
(z)
in reverse order.
Combining and reforming in matrix form, yields
(4-7)
or more simply
(4-8)
and
fi^-i-ru-K^'W"^)
(4-9)
Writing Equation 3-5 in the Z-domain yields
N
N
E(z)
Combining
4-10
=
(4-10)
S(z)
A(z)
with 4-8 and 4-9 and returning to the
time
domain, yields the following error equations.
£
(*)*e
(k-o-K
e
~<N)
(N+l)
where e
the
(k)
(4-12)
(k)
is the forward difference error,
backwards difference error.
34
and e
Equations 4-11 and
(k)
is
4-12
correspond to the lattice implementation in Figure
have been used to determine the K's of
They
7.
12th order model in
a
the sine wave and speech experiments which follow.
The
assigning
order
of
For
N.
the filter
speech,
determined
simply
is
anywhere from
6th to
a
by
12th
a
model has been found to be sufficient.
order
The reflection coefficients are determined every 10 to
20
appear
milli-seconds and when lined up side by side
present
a
spectral envelope,
to
(Figure 8).
INPUT SPEEO*
r
—
200
a# r • j r t
«
FUNDAMENTAL
FREQUENCY (f.J
I
a
«r
an
bo
•
/
s
ond
I
our g
tt>*-
t
r
I
%
"ni&i
O
7SOO
CAIN
.»(
SPECTRAL ENVELOPE PARAMETERS
l»,
.
--
».r,l
I.O
ir
*<|»j0j|UY#wt
£&.
$U>*fti'&*•"
-jflw&L^-.
l*
V
ft*r «s»
&
-1.0
0.5
r
L;jj.Uf.'?i^..ijii(.
A, a;
fe^; V ii^<^ J^iitiail
-o.s
Display of Analysis/Synthesis Parameters
CRef.
10:p.
16].
Figure 8.
in any case,
Determining the reflection coefficients,
is
a
straight forward calculation which is
35
an
attractive
feature of LPC.
the pattern these K's may produce in
It is
our experiment that we will be most interested in.
Spectral_Analysis
5.
convenient way to portray the frequency content of
A
is through the determination of formant frequencies.
speech
frequencies
Formant
are
most
the
prominent
frequencies
present in a speech waveform.
Formant frequencies are not required to produce
synthesized
speech.
decision,
gain,
In
other
pitch
words,
period,
given
the
and
LPC
voiced
the
reflection
one has enough information to reconstruct the
coefficients,
speech wave form.
However, the determination of the formant
frequencies aids us in depicting a frequency transposition.
a
Fgrmant_Freguencies
.
The complex roots of the denominator
the complex formants (bandwidths and frequencies)
are
to approximate the speech signal.
denominator
the
calculations
waveform;
N
polynomial
is
polynomial are obtained
samples of
on
namely
(s n
)
=
(si
from
a
short segment of
,
S2
number of samples,
the
The coefficients,
,
used
aj<
of
time-domain
speech
the
...sn}, where N>>p.
and p is the
,
order
of
Here
the
polynomial CRef. ll:pp. 364-3663.
Under
samples,
entire
sn
,
speech
the
assumption
that
the
waveform
are samples of a random gaussian process, the
sample is broken up into an equal number
36
of
samples
Each
which
we will refer to as
segments,
(Figure
segment is processed using the Fast Fourier
9).
Transform
(FFT) and then low pass filtered if desired.
(Ph)
•
l*
»
TspeecH seo>
t
2"°si>e£CH set.
3**wecH
n*
1
set.
FFT
LPF
&)£Q\/&M.i CO NT fNT
tpeetrt se&.
Fe*
i£>r*S.
IS6MENT
Flow Chart for Obtaining the Spectral Content
of One Complete Utterance
Figure 9.
The output of each segment contains the spectral
content of that segment.
to
yield
a
Each segment is sequenced together
time-varying frequency content profile
of
the
entire utterence with each segment containing its particular
frequency
prevalent,
form
content.
or
peak,
formant frequencies are the
most:
frequencies found in the speech
wave
The
.
37
SPEECH SYNTHESIS
C.
speech
A
signal
is
synthesized
parameters determined with LPC analysis.
block diagram of
A
synthesizer was shown in Figure 4.
speech
a
same
the
using
by
control
The
parameters supplied to the synthesizer are the pitch period,
a
binary voiced or unvoiced parameter,
speech
samples
or gain,
the rms value of the
and the predictor
reflection
or
coefficients.
at
The
pulse generator produces a pulse of unit
the
beginning of each pitch period.
amplitude
white
The
noise
generator produces uncorrelated uniformly distributed random
samples with standard deviation equal to
1
sampling
at each
The selection between the pulse generator and the
instant.
white noise generator is made by the voiced-unvoiced switch.
The
synthesizer control parameters are reset to
values
at
the beginning of every pitch period
their
new
voiced
for
speech and once every 10 msec for unvoiced speech.
The
the
amplitude
of the excitation signal is adjusted
amplifier G.
speech
The linearly predicted value s n
un
form the n-th sample of the synthesized speech signal.
signal
is
finally
continuous speech wave
the
mathematical
parameters.
A
low-pass
filtered
Atal
{s n }.
development
CRef.
needed to
to
provide
5:p. 280]
further here.
38
to
The
the
provides
synthesize
mathematical discussion will not be
the
of
signal is combined with the excitation signal
by
these
pursued
V
.
DIGITAL_FREQUENCY_TRANSPOSITigN
INTRODUCTION
A.
object
The
research
this
of
was
to
determine
algorithm that will digitally transpose speech using
predictive coding.
In this chapter.
linear
Hall's research CRef.
will be briefly discussed and summarized.
33
A
new theory
then be postulated and a simple experiment using
will
an
pure
sine waves will be presented to test the credibility of
the
Keep in mind that the real test will be the actual
theory.
processing of speech, this section simply sets the scene for
further study.
B.
*
POLE SHIFTING IN THE Z-PLANE
Only the highlights and summary of Hall's thesis will be
presented
before
produce
here.
His goal was to change the pole
reconstruction
the
(of the sampled
speech
output voice with different pitch
frequencies
while
retaining a natural sound and
information
CRef.
3:p 47]
locations
signal)
and
the
to
format
same
.
The autoregressive vocal tract transfer function used in
his research is represented by Equation 5-1.
39
1
H(z>
(5-1)
=
-
1
where
-1
-2TT(BW)TS
cos(2flFTs )z
2e
-4"iT<BW)Ts
-2
e
z
center frequency of the formant,
F is the
and BW
is
The pole locations associated
the bandwidth of the formant.
with this transfer function are:
z
=
A e
-
Je
Converting Equation 5-1 into polar form produces Equation
5-2.
H(z)
(Eqn 5-2)
=
-1
A
z
several mathematical manipulations and
Through
A and 8,
for
2
2A cos 8z
-
1
-2
solving
the following relationships for F and BW
are
determined:
F
8
=
BW
=
/
2
TT
(5-3)
T
(5-4)
(-In A )/ 2TT T
-2iT(BW) T
where
A
=
Assuming
(the
and
e
that
a
8
general
2
iT
FT
linear relationship exists
original frequency) and F
several
=
'
(the
modified
expressions are stated to
40
between
F
frequency)
illustrate
the
underlying
modification to the pole locations.
Note
that
the following equations are all linear relationships.
F'
most
6
F
(5-5)
'
=
c*BW
(5-6)
8'
—
*e
(5-7)
ex
(5-8)
BW
The
=
important consideration for
producing
these
relationships is guaranteeing that no unstable poles will be
created by shifting them outside the unit circle.
of
the
specifics
on Hall's development see
For more
Reference
3,
pages 49 and 50.
Two
experiments are illustrated in Hall's thesis.
They
are:
1.
Pitch was reduced by a factor of
and the
.58
formant
frequencies reduced by
.88
for
voiced
speech
2.
The same modification
unvoiced speech.
was done for
segment
a
Hall concluded that upon completion of the process
of
most
listeners agreed that, although the input speech was female.
the
also
modified output speech sounded typically male.
noted
that
lacking in quality,
tapes
which
although the audio
output
it was intelligible
recorded
audio
that
available for subjective evaluation.
41
are
was
somewhat
was
CRef 3:p.
output
It
73]
no
.
The
longer
predictive coding is a means to an end for Hall.
Linear
processes
mentioned
the the variables
modifies
He
the
speech with
LPC
(F,BW,0,A>,
programs.
computer
and
This
conversion between an autoregressive vocal track model and
most easily by
(implemented
model
LPC
configuration)
is
through Equations
possible
filter
lattice
a
a
(4-3)
and
(4-4)
The mathematics are simple.
that
is
the
relationship
representations of speech,
between
C.
the AR model and the LPC
To calculate one,
is to calculate the other.
§tatement_gf _Theory
•
LPC techniques can serve as a
As mentioned earlier,
tool
modifying the acoustic properties of
for
This
signal.
thesis postulates that
a
spectral
envelope of the speech
wave
frequency content of that wave form.
exists
which determine
form,
modifying
the
reflection
and
the
If this relationship
and the linear relationship is determined,
selectively
speech
the
linear relationship
exists between the reflection coefficients,
the
model,
NEW PROPOSITION
A
1
different
two
the
are closely associated with one another.
in a sense,
here
What is most important
then
coefficients,
by
the
frequency content will also be modified.
Is
reflection
relationship
between
the
coefficients (K's) and frequency
content?
The
there
a
linear
42
first
step
case.
Since speech is often represented as
different frequencies,
many
analyze
out
in our proof is to analyze the most
to be negative,
combination of
a
the simplest case would be
fixed frequency sine wave.
a
simplified
to
If the results
turn
complex
case
then exploring the more
(speech) would probably be futile.
Sine_Wave_Experiment
2-
considered
frequency a pure sine
wave
continuous energy and amplitude
a
may
be
which
signal
generate an audible pitch when it is within the 200 Hz
will
to
given
any
At
15 kHz audible range.
wave
speech
normal
audible pitch range is somewhere
the
forms,
When dealing with
between
200 Hz and 5 kHz.
A
for
computer program was written in Fortran CApp.
use on the IBM 3033
and
rate
incremented
the
,
to produce a sine wave for further
The resultant sine wave could be sampled at
analysis.
desired
Aj
frequency
of
the
wave
any
could
be
satisfy the range requirements of 200 Hz
to
5 kHz.
Once
wave
sine
coefficients
were
holding
file
frequency
calculated for
and
the sampling rate was determined and the
a
set,
reflection
the
10ms time frame,
plotted to determine if
a
a
relationship exsists
frequency and the nth order K's.
43
stored in
between
determine
To
each
Equations
frequency.
4-11
and
4-12
were
noise on the outcome.
for
used.
affect
runs were also made to determine the
Additional
3•
12 reflection coefficients (K's)
of
The results were promising.
Sine_Wave_Exger imental_Resul ts
Appendixes
B
and C illustrate the apparent
linear
relationship that exsists between frequency and the LPC
K's in a noiseless environment.
order
illustrate
<S:N
=
same relationship in a
that
nth
Appendixes D and
E
environment
noise
10:1)
It
would
exist
between
Noise
on
Noise
addition
appear
that a linear
relationship
the different frequencies of
seems to affect
sine
a
the other hand changes that linear
does
wave.
relationship.
K7 through K12
much
more
than Kl through K6
Considering
K,
are
the mathematics involved in calculating
these observations are reasonable.
affected
most by small changes in
Since the later K's
the
input
signal,
addition of noise will affect them more drastically than the
earlier stages
.
Though these observations are promising, they are by
no means conclusive.
frequency
considered.
existed,
If no correlation between the K's
another
Nevertheless,
scheme would have
speech
to
be
is the more complicated
signal that we consider in the next two sections.
44
had
and
VI.
A.
SPEECH_PROCESSING_EXPERIMENT
INTRODUCTION
Now
the fundamentals of linear predictive
that
have been presented and a theory of frequency
proposed,
itself.
is
it
obtain
To
correlation
necessary
to work directly
the information we
are
between reflection coefficients
coding
transposition
speech
with
seeking,
and
the
frequency
content, speech samples must be demonstrated.
Documentation
used
concerning
the data
acquisition
in this research to obtain speech samples is
as Appendix F.
system
provided
This chapter discusses the data itself, and
the processing of it.
B.
VOICED/UNVOICED PHRASES
Three phrases were chosen for their voiced and
characteristics as described in Chapters 2 and
1)
2)
3)
Each
make
phrase
things
They are:
"READY"
"SO WHAT"
"SNEEZE"
was repeated at
simple,
4.
unvoicea
a
different pitch
ana
the musical scale was picked to
harmonize a change in pitch with some type of reference.
to
he-?
In
other words, "READY" was first spoken in the middle-C range.
45
until it was finally spoken in the
and then in the D range,
high-C range.
of
This procedure yielded eight different pitches for
each
One male speaker provided the
data
the three phrases.
for
constant
For
for
each pitch and their
graphical
a
Additionally the period
three phrases.
all
representation
individual
of
utterances.
selected
the
utterances, refer to Appendices G, H, and
remained
speech
I.
Each phrase was chosen for content and can be classified
as voiced,
strictly
a
unvoiced,
or a combination of both.
voiced word, whereas "SO WHAT" and "SNEEZE" are
combination of voiced and unvoiced segments.
T
sounds
"SNEEZE"
"READY" is
in "SO WHAT" will be our
will
be
The S,WH, and
unvoiced
combined example
the
a
example,
the
as
data
ana
is
*
analyzed
C.
DATA PROCESSING
section
This
discusses
the
techniques
utilized
to
analyze the data and the observations made.
1
•
Speech_Data
The raw speech data was edited and displayed using a
generic display program.
a
maximum
resolution
lower
range
of
of
256
The data is 8 bit information with
equally
spaced
values.
each utterance varied with the
frequencies
tended to have less gain or
46
pitch.
energy
The
The
and
therefore did not use all the 256 range values available.
summary of the ranges is provided in Appendix
The
periods
of each phrase
were
J.
different.
differences between the same utterance at different
varied
33 much as 20 msec.
periods are given in Table
PERIOD
sec
"XXX"
1.
NO. SEGMENTS
N
NO. DATA PTS./SEG
(10 msec SEG)
.30
30
100
"SO WHAT"
.40
40
100
"SNEEZE-
.38
38
100
utterances,
sampling
so
pitches
1.
"READY"
The
The
short summary of the average
A
TABLE
UTTERANCE
A
rate
was
10
kHz
for
ail
the number of data points in each
of
the
10
msec
for
each
segment is 100.
2
Determining_Ref lect ign_Cgef f lcients
•
Once
the
starting
point
is
determined
utterance, the reflection coefficients are calculated for 10
msec segments of speech CApp.
analyzed
K]
.
Successive segments are
to yield their respective reflection
47
coefficients
4-11
Equations
using
and 4-12,
as
were
sine
the
wave
calculations.
for
plotted
coefficients Kl through K6 were
Reflection
resultant
each of the 24 utterances and several of the
curves are included as Appendix L.
a
l£®Qd_Analy_sis
»
graphical trend analysis of the plotted
A
was undertaken to detect any obvious patterns.
summary
of
The details
However,
M.
those observations leads us to
a
conclusion
the
there were not any trends of any significance noted as
that
a
analysis is included as Appendix
that
of
data
function of pitch.
Gr^phical_Cgrrelatign
b*
One graph was held stationary as a reference and
the
others
obvious
ups.
there
that no correlation was
noted
any
elaborate
is nothing more
There
was
between
to
them.
though at times there were 2 or 3 points which matched
Even
up,
match
than
report
passed over it to see if
were
other 28,
the
seemed
or 38 points did not.
be no distinction between
to
portions
36,
of
conclusion
the speech wave.
that
the
48
and
This process leads
various speech
uncorrelated
voiced
segments
there
Also
are
unvoiced
to
the
highly
3
Spectral_Analysi3_gf _Ref lectign_Cgef f icient_Pat terna
•
was
It
temporal
seemed
noted during the trend
analysis
patterns presented by the reflection
periodic.
the
that
coefficients
At first it was believed that this could
possibly reflect the pseudo-periodic nature of speech or the
excitation source.
Spectral
analysis
was implemented using a
subroutine to compute the FFT of each pattern.
is
included
as
Appendix
N and
results are provided as Appendix
summary
In
relatively flat.
frequencies
4
•
several
The program
examples
the
of
0.
of the spectra turned
all
Fortran
out
be
to
This indicates that there are no prominent
within
the reflection coefficient sequences.
Spectral*_Analysis_f gr_Freguency_Cgntent
Spectral analysis to determine the frequency content
of
each utterance,
as described in Chapter 4,
have
would
been useful had a pattern or linear relationship shown up in
the observations mentioned.
Since
are no patterns or correlations
there
worth
mentioning, exploring the specific frequency content of each
utterance
would not benefit us.
between each frequency, or
the
f
difference
relative
is approximately 32 Hz.
range of the utterances was chosen to
The
with
A
The
musical scale from middle-C to high-C
difference).
Had
a
relationship
49
been
(a
coincide
256
discoverea.
Hz
as
then
proposed,
more in-depth spectral analysis
a
the
of
input speech would have been in order.
D.
SUMMARY OF EXPERIMENTAL RESULTS
1
C2££©i§ti2Q_l§tween_Phrases_With_Dif f erent_Pitches
•
The
did
between identical phrases
exist
Three
pitches.
the
of
5
relationships
spoken
four categories
Chapter
in
yielded more obvious results if
have
should
relationship postulated
linear
different
at
mentioned
above
yielded negative or uncorrelated results.
V2ic^d_/Unvgiced_0bservatigns
2.
Though
techniques
there
available
may be other
to analyze
or
this
sophisticated
more
data,
mentioned above were sufficient to show that
a
the
methods
voiced phrase
«
was
no
more
Since
uncorrelated
relationship
correlated than an unvoiced phase.
the
results
were consistently
negative
leads us to some conclusions about the
between
frequency
coefficients
50
content
and
or
actual
reflection
VII
A.
.
CONCLUSIONS_AND_RECOMMENDATIONS
CONCLUSIONS
new theory to transpose frequency was postulated
A
Initial results, using sine waves, seemed positive
tested.
lead to a further 3tudy using
and
and
experiment
preceding
showed
subsequent
and
waveforms.
speech
analysis
The
speech
of
no apparent correlation between pitch and reflection
coefficient values.
These results may be attributed to
the
following reasons.
1
Complexity_gf _Speech
.
The
excitation, and spectral content.
of gain,
particular
combination
speech wave form is a very complex
attribute
and
analyze
To pick out one
for
it
a
particular
phenomenon, such as frequency content, may be unrealistic.
Speech
combination
has
historically
been
However,
of sine waves.
point in terms of the physics involved
this
speech.
•
a
in
rethink
generating
This leads to our next conclusion.
EbY.§i£§i/y§thematical_Relatignship
The
case,
as
slow progress in the
of speech processing has caused engineers to
field
2
modeled
there
experimental
results indicate
that,
in
this
is no obvious relationship between the physics
51
(pitch) of speech and the LPC mathematical representation of
speech
(reflection coefficients)
observation
This
sense
makes
reflection
since
coefficient determination is based on probabilistic methods,
feedback,
error
output
original signal.
resultant
the
resembles
the
Once the error signal passes through
the
stage
lattice
each
of
random input samples,
and
longer
no
first stage of the lattice network, its characteristics have
Reflection coefficients
been altered as much as 10 percent.
therefore
are
calculations
a
based
tool
on
determining
for
past inputs,
and
predicted
not
error
physical
a
interpretation of the signals content.
Just as engineers are in error when they refer to the
that successive reflection coefficients present
pattern
spectral
its
reflection
envelope,
coefficients
as
not
do
directly reflect the frequency content of the signal.
3
E§Ei2^isZE§®yd°.zE®£i2dic_Dif f erences
•
Simulation
coefficients
reflection
signals
(speech)
(sine
calculating
wave,
slightly
the
wave)
experimental
work
differently
with
than
results
show
with
that
periodic
pseudo-periodic
signals
.
In
sine
and
the reflection coefficients
the samples of one frequency are
from the previous frequency's
calculated
reflection
samples.
coefficients also
52
changed
for
a
very
Therefore
change
very
slightly.
an
This observation may be useful in the design
musical synthesizer,
LPC
adjustment is processed in
than music.
vibrations
complex
waveform.
sense
produce
random
that
speech
the
the rate and randomness at which those
change
reflection
necessary to
are
However,
vibrations
speech behavior is more
is pseudo-periodic in the
It
and
controlled environment.
a
the other hand,
On
content
where frequency
of
frequencies
coefficients
seems
having any
from
prevent
to
kind
the
linear
of
relationship with frequency content.
It is therefore the conclusion of this research that
the
relationship
coefficients
reflection
modifying
between frequency content of
sufficiently
is
coefficients in
reflection
order
speech
complex
and
that
transpose
to
pitch will not be practical.
B.
RECOMMENDATIONS
The
have
conclusions
relationship
between
present
reflection coefficients.
research
this
was
based
concerning pole shifting.
are
recommended
if
stated that there is
frequency
linear
no
content
and
Recall that the motivation behind
on
Hall's
research
ERef
Therefore the following
further
or more
extensive
actions
study
is
desired
1.
Continue Hall's research using LPC as a tool for
speech analysis/synthesis, but focusing attention on
the shifting of poles and not on the adjustment of
reflection coefficients.
53
2.
Use a data acquisition system that yields 12 or
bit resolution of the speech samples.
3.
containing
speech
base
larger data
Build
a
utterences at different pitch levels and have the
speakers be both male and female.
4.
Have the ability to match articulation patterns and
synchronize points where speech utterences begin and
16
end.
5.
Synthesize the input and processed speech to
for intelligibility of the utterences.
6.
Use more
sophisticated
recognition
It
is believed that the preceding
followed,
will
techniques
for
check
pattern
recommendations,
if
help substantiate or refute Hall's research
as well as the findings of this research.
The need for
an
adequate technique for frequency transposition still exists.
54
APPENDIX_A
REFLECTI0N_C0EFFICIENT_DETERf1INATigN_F0P_A
FREQUENCY VARIED SINEWAVE PROGRAM
=
determines the reflection coefficients for
program
This
a
12th order lattice filter model of a variable frequency sine
wave.
a.
>
UJ
>
fc>
'
J
—
>
k
u.
k
k
>
UJ
k
U
k
•-
V
V
k
V
—
<
J
<
UJ
V
>
k
3
O
•
O
>
U.
>
k
*
Uj
Z
>
k
>
z
u.
111
o
—
•,">
>
u.
UJ
>
v
>
*
i
Z
a
o
—
-
v
u
i
LU
k
-J
u.
k
U
*
—
4
z
<
y-
i>
N
cr
*
C
ii
«
u
.
<
u.
X
UJ
Q
u
>
k
Z
«
u
k
V
<
k
—
>
cr
k-
UJ
k-
i
c
«
UJ
cr
u.
in
a
a
to-
<
>
^
u.
V
i
3
o
* V
a
UJ
aj
—
Z
>
z
<fi
Z
—
<
k
>
i
in
k-
—
in
V
k
-•
u.
o
>
—
k
UJ
a
k
•
k
—
•
k
X
k<
z
z
a
—
w
*U
UJ
cr
a
UJ
J
3
o
O
O
z
cc
in
3
4
k
4
k
a
h y
4
a
>
<k
or
cr
<k
>
IL
<>
— O
k
<
<
k
z
4»
4k
k
Si
k-
»
z
4>
^
a
4
>
Xli'<CCi k
— > J »-*%
k
«„
«T
* <
«f
*u:z<
4k
J UJ X < Uj — *
<— — UJ0DZZ4 k
z u a z :c<
O-t- - Z 24
<
•- u. «r
4
>
<>
<>
UJ~04
2
•
>
LJ4k
*
>
C
•
k
•—
t*4k
;
•- 4>
j
U.
u
.
•<
I(k
k
I
k
^
2
•
k
«• 4k
k
k
* 4>
k
k
C
»0
4>
IT.
U4k
k
>
00 4k
•- 4V
>
o
k
3 4k
k
»
O
O» ««
k
f\J
k
>
k
k
V
>
,
k
,
4>
k
k
i.
4>
1/>U"
•
k
4k
T
k
».•"
V
uj
4k
<r
Z
iuc»»<
^» uJ
a z I *—
<
.2?J( C f X
<
luCk
i• z
- a —41.4 • O •
O uj 3 Z*-u.ft"»< z — O
. z o 2
4 V— O O L 3 < c a 3 —
—
j:c
a: 4-.—
n/! — — C — J
. — o
o cr h. y > — 2
—
J Z UJ uj U C U 2 — Ml,"< 1QJ2Z3-.H • — • Z
>ZZ 2 _l C C ^ O kHt'NC
— 3u*-U.U2< 3 — > »oj — •
ICCJ- UJCC* VU4 2Cr- V
ifl
k
4f
<4>
k
i«U4 k
u")
k
-L
k
k
|
4k
^S
~"
T.
k
C4>
L
k cr 4k
> UJ 4k
4k
k
»
—
3C
4k
2
k
Z 4k
C4k
C4k
2 4k
<
E
4>
'dj
i
4
p
k
O
4
L_.
IT
-»
—
•
n c 2 » 4k 4
-c
t-ZlJIkJC'J** 3. dJ • o• w
U' o —
KJ<IC 3cr 3
l.
< k cr
k
M—C
v az < z
LJ • X o
<u. u.
—
—c C
k
•
>C J"
tn <
• 3 2 Q O — —
ZO; uj O
—
in *J
k
T7< U. » •• J m T sr — •_ —
2
—
co
2
— U O UJ 3 3 23 a. 3 3
;-ccrv-*'k'<u.u.i/ic'n -C v > < IT LT k- oc c u.
k
feMj]
1
I
II
1'
i
1
ij}
<
ii
*
k
I
1
1
1
k
ifl
ii
ii
k
U<k
b
*" 4
3
k
t* 4k
4k
•
L.'4k
i
2
<
k
"» 4k
ii
4k
(.1
•
ft
k
•
f
k
u.
k
Z
3
-C
2
e> — *
X 1
M
4k
cr 4k N- <•
4 k „; 4k "^ -T
< k *- 4k
n
,
k
J
—X
c—
k
k
i k
l
4k
>
k
.
t
(\i
;
|
II
b
l
•
II
i
k
J
k
<
k
w
4k
^
k
_l
<
2
4k
«
4
k
<
4k
4v
k
3
C
k
e~Z
3—
4k
cr cr
k
k
—
<
<k
k
<\
—
u
-
<v
4>
i
—
k
<
t
•
k
<
2
U4k x
C4k —
k
/
^
3
k a.'
_l
>- JJ
H 3
-2
> I—
J — — •J n •• 2
3 < x 23
O U — 'S \J
k
fc-
»
B
»
r
4k 4k
k
!•
UJ
*
J
*
•
k
4
k
4
UUUUUUUUUUUUJUUUVI iJU
w"-
1
55
1
^
<UOu
4k
4k
4k
it
4k
4k
4k
4k
4k
4k
4k
4k
4k
4>
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4k
4>
4k
4k
4k
4k
4k
4*
4k
4k
4k
4k
4k
4k
4>
T
4k
•
4k
4k
4>
— "1
•
-1 —
1
•
—
a»
*
1
>
mm
mmm*
O2
OS
—
J<
m>
k HI
WW
^C»
s——
k
k
k
>
—
»
*
k l/1< k
< k
k
k
—
1/5 1/1
oa
k
— < 4> 4.
o<<
»
k
_ _j*»«»
*
-u< k
o— —— 2
Z
O j*» —
•—
II
II
-1
4« mt
*
-1
1
<
<
4
4
Z
^
Qt;^ U
<
it
>.
II
l_l
ul ul
M
(,*!
O
««j
UJ-ik
-1
1
•
•
w
c—
O—
1
zz
4k
-
«»«—
<>
ri i_
lb LJ
m
-
mm
-i
i/".
-*
X
3
Ml
1
Z
!_l ii
4k
a
•*»
•
1
.,
—
i
|
•"•
t
1 "1
•
C — 4k
4k _m i~
_. **
•
X ——
s—
_ _
if)
"
* ¥
—
—
•-4k
<M4k
= w *<k
— » — X a 4k
«»«M
4k
C•
4k
4k
4k
4k OJ <r sC
«aa
4k
4k * *
4k a • •
Z x
WW
1
fM
4k
4k
4>
Z
•
n
4k
4k
J.
4k
4k
4k
4k
1
_ _
5.14k
ZN
—
a—
—
—
Z<wO
tOtT
CO —
4>
"—
u.
-»
—i
r
1
*
-::
— — —
cw-
—
3
an
k
i
II
.
o< k
c< k
2 4k
(I «—«•»
O
—
4k 4k
-z
—
Zw
—
UJ
—
——
-1
—
-V.1
—3
III
4k
4k
4k
4k
1
*
—
<k
4k
>— —
—
— «— 4k
^
<>
Ul
HI c
X
a
—
t
n
M
C
a
4k
4k
4k
»!« — «w H *
4k
X
a«k
Ul
4k
4k
• • • » c c
C O O C X 0.4k
CNJ
* * ^ *
p
a
a
ILXLIL — — 4k
— c zz
"^
C4k
> 4k
03
<>
ozz
(M
—
—
WW 4k C O C O O«w O 4k Ca
ww — o o oc mm —.,._,
J< k aiww mo
< k • * *
n4k
• UJ U
4k C O O C
• • • • t • U CI
s • • •
mm — "i r» • a 4k
X
t-z: 7 m c r^c: rjcocoN<i —
«_
Z MU. c
4k
a
• •
• c — 4k oJ pu
< Miri uu S
OJ
O) UJ
xi n
-* *^ — — 4k
-)—
n.
—
—
2
3U
3 3 4k <r
k
01
^ 1 z — mm — — — — z < ~i 1 z Z <k w— —"«- — •»<* Z —
z .1 — —
— H * •o • •— o • •
— o • » — — 4k Uj tr ua
U.'4k — <
k C
O
k ojzq k- — — — 1 — — — <r z z z 3 nzZ" H- — 43ZZ t— » 4k — — m. — ka — 4k i-n
mm mm Z Z4k — — — — — — 4k z;;a
>
_
z r z a w mm
1_
z
M dJ IU O il UJ 3 w — -I
tSZDO
ou ji o — Oujuj Q 3<k cr k 222 s.k oa»- z
a in ii u *: u. c ciLruru.CtfZC<i UJtCHS U W « X T X * X 4k U U. UJ
IT-
|
——
C1
y
-^
4
—
1
1
1
'.1 (.1
1
1
II
II
4k
1
^
»
II
II
ii
ii
II
ii
II
II
II
1
1
II
II
II
1
I
II
1,1
'
il
II
—
,11-
"1
ii;
,_l
i/1
k
o<
— JUlj
V
1.1
o
M
o
n
56
4k
4k
4k
4k
4k
4k
4k
4k
4k
«u
o
o
(j
urn
o
O0
If!
>c
4k
APPENDIX B
-
REFLECTION COEFFICIENTS K1-K6 (NOISELESS)
CO
w
o
o
I—*
w
T
21
00
K0-
S3Q1VA XN3DIJi303
57
AEEINDIX_C
-
REFLECTIQN^CQEFFICIENTS K7-K12 (NOISELESS
o
u
c
ro
samvA
roiNaioiiiaoo
co
58
APPENDIX D
-
REFLECTION COEFFICIENTS K1-K6 <S:N=10:l)
CD
CD
O
o
J
W
»'o
oo
ro-
SamVA 1N3I0IJJ3O3
59
APPENDIX E
-
REFLECTION COEFFICIENTS K7-K12 <S:N=10:12
o
o
tu
o
g . =, o - CM
uu: * ApC^ClaJ*
o
X o
2° o
o
o
o
<3
o
o
o
I
fa
fa
W
o
u
o
w
fa
w
21
fl'O
>Q
00
>0-
S3fnVA 1N3I0I3J30D
60
00-
21-
APPENDIX_F
-
DATA_ACQUISITION_SYSTEM
INTRODUCTION
A.
There
the
are a vast number of data acquisition systems
market
Though this is the
today.
case,
system
the
originally
planned for the acquisition of thi3 data,
down
no hope of timely
with
When
repair.
all
on
broke
possiole
alternatives had been explored, it was decided that the oniy
way
a
to accomplish this portion of the research was to build
system capable of obtaining speech data samples.
section
This
will discuss the
hardware,
system,
and
software utilities that were combined to produce the desired
data
In an effort to provide the novice,
samples.
with the information needed to retrace these
as the expert,
steps,
anything worth documenting,
bibliography
as wei!
is
Additionally,
is.
provided in the main Bibliography of
a
this
thesis
B.
EQUIPMENT REQUIREMENTS AND SETUP
Figure
utterances
recorder
shows
10
were
the
recorded
experiment.
on
a
and stored for later use.
Selected
4-channel
,
8-track
The analog to
<A/D) circuit was built and driving software written.
circuit
was
interfaced with the
61
Zenith-100
speech
rape
digxta.
This
microcomDuter
the Prolog 7804-Z80A Processor
through
Counter/Timer
Card
and the 8255 Parallel Peripheral Interface (PPI) microchip.
2.100 Co«mr\T
MICROPHONE
ItUi'n-iUf
^2^
5033
Data Acquisition 3-Dimensional Flow Chart
Figure 10.
Once
it
the data was captured in the Prolog's 32K
was uploaded to the Zenith-100,
stored in Intel-Hex data files.
from
the
Zenith
formatted
buffer,
via ZMDS software,
and
The files were transferred
disk,
via
Osborne
an
microcomputer, and placed on Kaypro formatted disks.
A
into
Kaypro 10 microcomputer converted the hexadecimal data
decimal data using Microsoft Basic (MBASIC)
software.
Edited versions of these files were then transferred to
IBM-3033 main frame computer for data processing.
62
the
C.
ANALOG TO DIGITAL CIRCUIT
chip that provides the analog to digital conversion
The
is
the AD-570.
rates
It provides 8-bit information at
to 33K samples/second.
up
sampling
rate
was
our
For
set at 10K since the
sampling
the
purposes,
majority
the
of
frequency content is below 5 kHz.
The
circuit
diagram
CApp.
illustrates
F.2]
tne
interfacing between the 8255 PPI chip and the Host computer.
coordinates all of the necessary
8255
The
handshaking
in
driving the AD-570 chip.
was necessary to amplify the signal prior to entering
It
the
AD-570,
available.
256
amplitudes
It was also necessary to provide an
adjustable
to
obtain
full use
of
the
DC-offset to assure a unipolar input (i.e.
the middle value
had to be adjusted to be level 128 instead of level 0)
Also, the signal was filtered prior to data acquisition,
through
frequency
However,
the
use of a Butterworth filter
cutoff
during
This helps smooth
of 5 kHz.
the
designed
processing
necessary to filter it again.
of
the data
with
the
it
a
data.
may
be
These additional circuits are
also provided as Appendix F.2.
D.
MICROCOMPUTER INTERFACE
The
flow chart,
provided as Appendix F.3,
the Z-80 assembly language program.
illustrates
Appendix F.4,
that was
needed to drive the A/D circuit and collect the speech data.
63
The program,
step by
was also useful in testing,
A2D.ASM,
step, the proper operation of the circuit.
The Z-80A micro-processor is at the heart of the
software designed to drive it is assembled
and
the
the
Macro Assembler <M80) and linked to the Prolog
Link software <L80)
using
system
using
station
For more information on
.
these
procedures refer to the Bibliography.
1
Sampl ing_Rate
.
sampling
The
rate
step
the microprocessor goes
that
specific amount of time.
as a T state.
arbitrary.
is
It
a
In assembly language programming
function of the software.
each
not
is
through
takes
a
We will refer to a measure of time
Each T-state equals the inverse of the
rate interfaced with the Z-80 chip.
clock
Since we are using a
4
MHz clock, one T-state equals 250 nano-seconds
Every
including
the
command
line
command
'No
in
assembly
the
Operation'
several T-states to accomplish its task.
in
the
interval
or
NOP,
program,
requires
We are interested
of time it takes from one sample
to
the
next, and then we modify the software accordingly.
This
program has a delay loop in it (labeled DELAY)
to slow down the data acquisition to 10K samples/second.
it did not have the delay loop in it,
at 23K samples/second.
If
it would easily sample
Since each utterance was limited to
64
leas than one second,
10K samples is workable and does
not
present prohibitive record lengths.
DATA FILE SETUP AND MANIPULATION
E.
Once
32K
the data is collected and stored in
buffer,
it
is
uploaded onto
a
Zenith
Prolog's
the
100
formatted
floppy disk and stored in an appropriately titled HEX
A
sample
of
a
typical segment of
data
is
file.
provided
as
Figure 11.
:1C5AE0G0 £071 £0717171797776797271717? ££52 IS
:ie5A7 ee0e5££e6£6£eS5535e7I7F?2£07C7r7I7I93
:1053000e£e£l£0£0£l£0£177£e77£27r7I7C7I7lA2
:10£310ee7I7C7E7E7I7S£e£0£07IE07?3'J7I7r7igj
:ie5£2Z2e7I7E7E7C777S777Fe3Se7IS0e78IcA£77A
:ie5E2000Se£££652?t7C7E7i:7E7B7g7I£0£2£l£e6A
:10534000SC£lc2£2£07I7I7C7C7I7C7L7I7I53£3cE
Hexadecimal Data File Segment
Figure 11
The file is in Intel-Hex format.
each
line.
full
of
following '10' tells us that the line
The
data.
The next four digits indicate
location in the buffer.
location represents
Following
a
The colon starts
a
nine
is
memory
the
Every two bits following the memory
byte of information.
double 0,
there are 16 records of data, and
then a checksum byte at the very end.
first
off
For our purposes the
digits and the last two digits are of
no
use.
An Osborne microcomputer was used to transfer data
from
The Intel-Hex file is already in ASCII format.
the
Zenith 100 formatted floppy disk to
65
a
Kaypro
formatted
floppy disk.
Since the data is needed in integer form to do
the necessary processing,
in
a program was written
Microsoft Basic Language (MBASIC),
CApp.
F.5],
to convert the
data
files from hexadecimal into the equivalent integer values.
Finally,
the
data is ready for processing.
Since the
software was already written on the IBM-3033 to process
display
the data,
and
it was sent there via a 1200 baud modem,
and processed
66
APPENDIX F.2
CIRCUIT DIAGRAMS FOR THE DATA ACQUSITION SYSTEM
Figure 12.
Pin Out of the 8255 Programmable Peripheral Interface
.u**.wF
.
—w/-^WA
U>J>.
i«*
jmi
I
Figure 14.
Figure 13.
Adjustable Gam and
DC-Offset Circuit
Low Pass, 2 Pole
But terwor th Filter
67
APPENDIX F.3
(sTARrj
ser up
eo>MTe\
X.KUTI
An^e
ffzss p*i
SfT
PlM I*
t& cyit
eaaaaa
p"»^^ a '^ CT
.
i>«lMT
ftfiJAb-^
PolljwO-
B«u_
We
7r«e
tit*
LCr»FF
«SS
HO
P*(WT
L06Q D *TA
IMT7J
JifWR
Peiw-r
SQFTWARE_FLOW_CHART_FOR_THE_ASSEMBLY_LANGUAGE
DATA ACQUISITION PROGRAM
68
APPENDIX_F\4
This
program
M
*
ft
*
»
*
*
»
«
ft
*
ft
*
ft
»
»
*
*
*
O
(so
•
O .J to
2: o «
WK2.
1.
O.M
V1MW
im
O Ei 3
<
-
>-*
.z
t*.
k««C
xnw
fr* e-» </j
•—
ft
fc-«
6->
*
ft
ft
4
ft
ft
^
•
z:
Wm£3 oW
QUrj
£-•
^. ce *;
•«:-' a=
OO
(x4
«
o
-a-
a.
e-i
w
Z«
ft
•*•
1-4=
•—<
#
*
pi
CO •<
ft
fc->
.<;
^
ps
_;
«*; £-i
aw3
«
pM
e->
.3.
1
ft
O C\JOOO0UOOOlH<*:>—
•
ft
X
Ot"
^
o
ft
fe-"
ft
*
.fr
ft
ft
ft
4
*
Q
HM
^
»
E-«
ft
ft
ft
Z
«*;
3
PL,
to
£»,
ft
35
e-«
»
»_i «*;
to
ft
ft
6-
O
34
+ t +
P=
eUK53I2XW«
(
f*.o-iCu(x.cx,ai<sjcjGiQ<3UD
»
Cija-.P-iCucuCiiC53Ci3oca<SicaZft
CCPQ
ft
ft
OIO
O
ft
OC
ft
PU
ft
*
ft
ft
-».
GU
t-«
>•
ft
•»
ft
ft
ft
ft
*
4
»
*.>.....-..
«*:
z:
mu
>-
1
o
«*:*:
ee
O
WW^OwSOlOKO
i*kmi/)k^:i->2>-'<«:
«»:«*:
ee
fc-'t-
z*r<
3
coco
uj
z
1—•
mm
o
-i
p-nrtr
fc-
>~
Uj t->
»-•>-•«*:•--•—
1
•«:
mm
'
t
.-.-...-.-.-•-.-.-.-.-•-
ft
ft
ft
ZZ
ft
*
ft
ft
ft
ft
ft
Oft
6->
W
<WU«(xt-'W
ft
cu
pj
h-»
l-«
4
"O
Z W*(« (^N O ^ NH u)
1QZ0O3
-•"-•
fl,
#l/)i-«CO^O>—tZi—
ft
fe-»ftft
ft
ft
—J >
ft
—
;zz*
hkkkmqwko
puOOOp^o^Jt-^asP^M^iiJ*
£-<6-tt-'Uk*tl-«l-M</5
i
-»
»-Ji-J
>—
••
*
1^4
1*1 ft
h-j
*
^j
j^j
^q^;!-^.*:^:^^*:^-.^;^^!^.
#
ft
*
35#
»
••
«:
ft
ft
-1
ft
«-•
>-
Oft
•
ft
*:
1*4
k-»
^
V3
S<3.
*
#
a.d.a.o-ia.i.uoa4u»-
•»
ft
ft
ft
333D3Z)3DDDDa*4
CTO'O'O'Or-O'Cr'O'O'iD'O'O'O-ft
^x|HMUJMt«(>4|Mt«Mu|JUWwCi)4
ft
£D
W
6->Oi_
to
1
>-.
D«ft
COCLiO
•-!*-•£:
c-0
6-»
'
u.
>3ift*-<"<
##6-ik>Mfc-»Wt^»*^
sooxi<ia =
CGftfto-)E-'i=:-r
U --«:0
-•—•6-
ft
tri
8->
p->
Wft
ft «
MftftCJt-)
ft
#t-»(-4ft
Wft
4
*z:E-.«ft
««;>-.t-'
ft.
4
cji—<>—••—11—i«t.Cv- r;u)^>-*:iJ-)6-'ft
•»1/}«*;;_U*
ft
•-**
1
,
ft
Pft
O
«
r-4C\JrO
ft
mm
ft
z4
o * H^
ft
.*
"•*•
<
L
000*
i-iCEO
*
2: i_
-
,
ft
mm P5
MM
•
to
K," w# Hit-r<tr"-'-;t- *E-'HU6-'
< H
U-"f
h» u*r'f":z<r"H#
Z ''Ji* -'ui
Oft MMMMf,
I
UUM
;'
"')
z^ffii
ic»
3v »!,# «iiU,ft,C>-"-'/)l-i->t:-i
#
ft
(
ft
-
<
a.
ft
'
E-.
10
•
oczd
ft
—
,-m
^:
O 6* —
z o<*o*«
*2cu
to
o
—
>.
*•
^fji—ii^i
hh O
Z >- mm N w ^
«c
.£.
4
*
4
cd'
.i;
UK
t-.
<to 4
* n n
.•.....,. ^ 4
-> —.
a«*2:»i-l>-a
r
2
wwwfr* Z o
_l
_
W >— pe ;=
Z3'
(E J
«^U'-w«<
»e-'Uj:3r-:*i
ft
CeH-
•
c_>
.
st
-*
ft
«
#
Z '«J»Ja:»u»(>< *
ft^J0-.«»;WOOOOO«COt-«tJ-ti<Jft
IEAu<
O (wt-ez
kmuXift-eajt-e-t-"
3iH«CE«>-*ttZZ<DSZO'-i#
1—
ftMKOjcua<(XiaL.z:c_3cjcnc_>i-J^;tvi->
»
ft
.»
-*
£3 s;
.< pa mm
4
£-•
1
e-
t-iO
ft
ft^:^:e-«
t->4
-ct-iOk-icx^;^^
ft
O
"—
t-4
ft
—
3
us^cc
V«
i~w
5C
c^-oeP<t<f-o
—
?-c-4
(EWfiK
u« — <x
c-
1—4t-
i-«
1/)
km k»
1
a,
|Mi.UOOC»
WHHv)i.wu
csi*
mwo< uui«
CO
•-•
c^-
.-1
6-»-si
c«4
O
ee
mm
Z (4 m ^ D l/J 1« ft
h4ffia:«p-iM[fi,i;ft,e'<0(«ft
&m *
k-joqoo o o o mm
*
<_>
eu
ft
ki-j
»
ft
ft
ft
•-•
.9.
e->
m-« ix; <_>
ft
2:
u->
a-
ft
*
fc->*
#
».
»
o
««
Km
ps
pM
ft
ft
ft
ZT
WW Wwo w « u Z -« ww= O
MinunoK • * •i.utu
ft
ft
£-•
e-
OS
c-i
ft
«<;
Pm
s»
x:
»
»
*
*
W
ft
ft
pa
tO
<
«.
Z3
pu
•->
l*t
*
(->
Q_ vi
ft
ft
CJJ.-J
-•-!
vi
VI
ft
ft
C-4
»
»
uao
«
fn
»
ZZW
ZStO&uOift
o
W
O
CJH-icjWft«b-««
o
o «
ZHW 33* » OW
w •mt*<t-u4i*Ci.'* «
t-4
ft
P5
M
t««
2
~
2:6fe-Z
coo
WJ
ft
ft
o
ft
O^WftO
ps*i-i
«
t~i
ft
»
*
*
=>
ft
ft
ft
=3
«»*
W
e-»
*
*
W
ft
ft
ft
ft
ft
•*
data.
ft
ft
ft
*
speech
to acquire
ft
ft
OOM
ft
designed
is
If
ftO
ft
ASSEMBLY_LANGUAGE_PRQGRAM
»
VJ
x
ft
-
1
*«
ft^-«
w w^
mumouutuuttu
C5ftft
3
*<_r
u
.^ .- «ft
p=
ft
.-......-
69
T-
......
^hOJcO
-c
—
t
M
-*:
—
•
z.
>—
>•
O
52
*—
CC
K>-
6-t
«*:
l-M
z
z
»—
Mm
ImMQ
m4
Z Ph
MM W
MM
D3«W
m£ OO MM <H</]
O
Ud
MM
CO
M
-C 3 < 3
O- Cm O
CM
Eh Z
PS CC
CO
PC PM O
O
O MM Eh
O CO
9 Z Mm
— O
P5
mm MM >~
Ma
?~ >* O
MM
PC PC
C- Mm
O
O USii <i£ O
W PC
21 5Z O
>H MM
l-M
E->
MM
l-M
3
O
»"3
MM
o
9
to
CO
MM
2:
t_>
Z
PC
PQ mm
«<
9O
mm CO
E->
9
Cm
6-»
O
<-h
<c CO
iC
:=>
«*:
CO fcH cc
CO PC MM
=»
en mm
cc mm 2; Cm
Mm C_3
Will
35
£h mm 33 UJ —
35 £h •AJ *;
-S Eh
CJU
<c
6Eh
1—•
O
O
i_
MM"
O
J
O
e-« 5=- fc—> cc
>i CO •— CO «OJ
MM Eh
6-t £- Cm f* aj
K *«
MM MM
i_ i_
21
ac mm
Mm pc 6- 33
(x^
ju
mmMM
£-. >«
MM"
MM
C3 MutncrtD CC
«:
CO
«i
hi
CO
pc pc mm cr
MM
CM -! CO 1—
£-1
21
>pc «*; «*;
uM
a)
lT) CM MM trin
U) PC t*Z -C
CJ3
i:
;'j
•— t— CU
MM .»
aj
6-i t-> mm
MU 35
<-4 5Z PC PC 1—
&-• 6>m
>-i t— «c -a;
crj
CC t-> 6->
mm
cu Cm CO CO Cm -«a •— Cm
-O
Cm _l
MM
PC
!3>
CM
O
2:
MM
35
CC
2Z
MM
£h
CO
>CO
Mm
Mm
OO
OO
O
Cm
l-M *
frn
**.
O
Eh
»
-
Z
—
£-•
PC
Pm
33
CO
t-i
•
Ma
D3
Em
fc-« fc-«
>->
MM
MM
MM
Z
z
•-H &-•
l-H
CO
MM PC
Z*
Pm co O O
OO
Mm
><
CO PC
=> MM
Q>
MM
<
2
WOm(C<
mW
CO
w « 3 < CO
Z —
f
oz
3WW3ZW003
*2
C
1
3
O
2
ww
•
<
O
1
Z
I
_
O
•
1
c^-
£-•
MM
>-
PC
<—>
MM
MM Cm
52
l-M
5Z
MM
p— 33 zn
l-M
b-i
£h
O
~z
1—
CC
MM mm i-J
2; rD CC
H-* CO •<
O
£h
30 e-« c->
C-) PC -c 33
MM
Pm
CO CC 2: CC
MM MM 21 ^P
3C 35 <~J 33
CO
Eh
HM
Mm
Urn ^: i—i
O
O
6-c
PC
O
^
O
e-
z
O
MM
"-« e-»
>-> CO
CO
mm MM E" e-i
&<J mm *C Cm
•
C<1
•*
'MM"
ZU
cu .-m
"3 1—
1
t->
-1.
1—4 6->
>—
CC CO
Cm &m
«=c
^
MM MM
I-J MM
«* Cm
•«*;
C_> C_3 --5
Pm
Z
Mm Mm
pM MM «<
Em
33 £h
O
ZW^<
-a
« 1—1
-
PM 35
CC
i-l
mM
•-a
-
—
tM
Cm
CO
O
O Eh
35 i_ O Z
•
i—
l-H
t-i
o»
O
<SJ
•>
PC
52 Cm Cm
MM"
|_4
—
t-i
w
MM
M-) t->
Mm
35 21 MM
* m 1—
C7j
CC
Li) c<
Cm Cm
(S
C_)
•<
Z
W
O
O
•-5
i-J i-M
^-1
l-M
l-M .-J
M-J
HM
Q<pp<<<a,M
Cm <
CJ
MUMi-IUU
t
Z
MM Z
< Z
1—
tiiA^XXluu)
mi
c-
35
•< Cm CC Cm <: «;
rj i-J C3 CJ
O
*
*
*
*
*
«
*
•»
*
«•
*
»
*
*
*
*
PC
a:
1
-a-a-
*
*
MM
PC
Cm mm PC
CO
>CO
UJ
<M Cm
00 mm
35
i-l
O
«*;
-x;
pc
z<
u
z
MM
3=
C_>
Cm
z •
O z H-Z
mm
>•
_*
^e>
G MM
£V-
z:
O.
Eh
-1
l-M l-M
«*;
•«:
Eh
-< Pm 21
CC
MM
Pm
Cm
•< E-« iO
C3
PC UJ
MM
MM CM
CO Eh =» UJ,
Mi
N
Z
I—
6- *:
O
•-M
61
IT) C«J c_>
-«:
Z
O
O
• MM
MM
35
*; mm 6r->
»-M
O
C3UM.
MM O
MM
Cm
S»-
s»
z ca- Z
•*;
O
i-M
t-»
PC
ID MM Z3
6m
MM
ID
;\i
mm
cu z; pc
P2.
33
«c:
E-> E-<
e-i
Eh 6- MmCJ mm «r; km,
Cm
Cm
£-• 53 Cm
<M
>~3
33
CJ
< ^3
Z
rU
Cm CO MM
CO;
ft
»
ft
»
ft
ft
*
»
ft
ft
ft
*
ft
*
#
*
*
«c
•
MM'^
MM PC
Cm
MM
21 Cm
<SJ
•Cm
Sj
O
£m
MM
CO
_*
O
»-M
«c —
O-
<»; Eh -C
• MM ~
«—» C/J *-m
OC MJ CC.
CJ PC
mm 31
CM C_) CmCm -Cm-
u
M
CJ
««:'—««:-—
Pm
—3
MM
l-M
••
••
MM
Mm
«*:
ft
ft
*
*
ft
»
ft
*
ft
ft
ft
^> Mm
Cm 33 MM"
en
1
MM)
O
1
-M
•»
Eh
MM
z
• •a
70
1
ft
ft
ft
••
£-•
9
ft
<
C_)
MM
Eh
ft
ft
mm
•—
ft
Si-
*
*
*
*
*
#
*
>~3
*
.
iCi Er+
j»
ft
ft
O
2Z
t-
ft
*
Eco
—
Mm
PC
O
33 O
ft
ft
PC
MM.
2:
ft
ft
ft
ft
1
*
O
*
1
#
*
M-)
ft
*
*
*
9
O
ft
*
*
Mm
MM
Pm
21
«*
co
E-i
*
*
*
-»
ft
ft
*
ft
*
*
a* Vi ft
* v^ -»
* 2 »
&-•
z
ft
MM
35
-
•—1
CO
-<
li-
ft
->
z
MM
t-M
*
M-)
ft
ft
ft
ft
-li-
*
*
»
•H PC
-=
ft
ft
ft
ft
ft
ft
ft
a-
5
MM
21
•»
O
PC
52 33
tr-i
MM
"**
>—i
>— >-<
c—
PC MM GO MM MM Mm MM
mC
P>—
mm
Mm
Cm UC
b- Pm
*
*
*
«
*
4
*
•»
*
*
*
*
*
*
»
a#
*
*
*
*
*
«
#
*
*
O
O
O
mm
Eh
<
Cm
..
6-
9 Cm 33
O
MM
C_3
CO
IN
MM CO
cc tsi
3CJ
mm
CG
Z
O
cMm
mM
PQ
«C
CO
t->
(-4
«*. iv<
mm
pq
MM
O
MM
«c
>~
•-3
a")
X
CO
EOh
CC MM
MM)
o
o
Pm
OS
o
fcM
«*;
t-«
z
+
MM)
cc
CC
A
—
1
C^-
1
i-J
CO
NX
mm MM MC
Pm
ZH«
Z
Mm
o
Mm
Em
O
>« Z o
CM O
CJ MM
•^r
Em
mm
«C
pM
CC
mm
CO CC
l-M
l-M
1—4
c>- (E
o
"s.
3
O
Em
«c
MM mM
fc-
-e >—•
CJ
MM a. mM 5:
—i
cc
cj
>* t- CJ
CO (-»
t— «*. =l <*;
MM
•«: CC PL. (-4
Zo
E-i
«*:
Ui
CM >—
O
O
—
«*;
<-A
^-^
^—»
CJ
MM
CC
PM
MM tM
Mm
»-
6-«
<£.
MM
CJ
«;
6-«
O
3»
«<
6-
O
CC
6-
t—
z
Em
CO
MM
E-
CM
e-»
<
Mm
CO MM
PM
mM
Zz
OO
O
<
<
*
o
3
ZZ
z
O
O
zZ
CO MM
MM C=
»-•
fc-
O
z
»-—-«
^
-<:
Z
21
O
CJ
^
CO MM
Mm .<
M-4 MM
PC
t- PM
co pc
MM <c e-*
«c £-" &64
2:
c-« CU —I >— «c
—
6-«
t2:
>- «<:
c— mm z:
-a; «; Cm
CO
Mi MM
e- 2:
MM
KM
MM mJ
0=
MM
e- -.i M-t t—
WOffi
t—
s»
CO c-> CJ
Mi
MM
MM CC
Cm
CC MC ^H feH C3
e- CC
r-t Pm •* -d
Mm cm MM Cm
ZD M-t
— fc-> -» PM Mm MM i: i_
CC Mu NH
CD
CM CC j« Pm
MM t_J •-J >-)
a
Z
^
1
3
O
^
«OO
m
O
Z
z
f
z
O
O
1
p
II
OK
Z
t=>
Z
l-M
cc
MM CO
^-1 MM
E-eui
CO
MM
t— 32
M-
U)
a;
CM
CU
Mm
Z
O
O
mM
*;
z
O
•—
CC
CJ
^:
1—
MM
CO CJ CC eCJ
fr6- in mm
CC
PC MM
MM -t; Ui fM
CJ
M»
MM MM UJ
mJ CB
CJ
CJ MM
3
f
O
*M Mm Mm
1
6r-
cyj
«C
c/j
«*;
e-
t—
t-<
PC Cm
<*;
—
&-• MM
CO OT
«-c
z
O
e
z
—
•
O
CU
X
—
1
CJCJ
MM
CC -^
• MM
«*: MM
Z
Z
CC
^»
iH
<-* •*;
Mm Emm «c;
PM CC
• Oj
liJ
•
PP PC
«;
-CJ
CS3
OZ
X
MC
CJ Cm
M "«;nHHHHMu^MiU(BMOMMunn
mm
Z
OZ
^
<^C4 M.O
^
^M
MM
mM MM
-«:
O
71
MKJ
j*
u
—
«*:
fvl
1
E- i/a
cc CO
Lf)
z
z
i/a
^
|
^ Z
1
z
1
O
z
—
z
1
«x
M-»
<c
Cm CM CO Cm
CJ
rj
etM MM MM CO MM
CO U i_
CJ
M
l-M
«r
c-"
<C
<yi
MM CJ
•"•
w H
__,
CO CC
1— \n «c
»-!-.
i_ co
MM
EO
CO CC CJ j*
W
Mm 3:
Z
CJ
^J
1—
1
CO tM -C
t- MM
Pm CO
Mi
-j-i
=>
«<
z
1
Em
O
zJ
fcM
'
ft
~-
Z
«CJ CC - — &-•
Cm
• Mm *• OS MM Cm
-T
•< -^ MM Cm
CJ
1
3
O
-M
MM
OS
t_i
MM
CC Mi CO
&M
• * •— CJ
MM CC
-
Z3
M
U Z
i—«
c<
MM
>-M
6m_)mJ
EmmJmM
QJC<<O.M.QJJ<;<a.M
MOMQUUnMUMuuun
l
CO
MM
Cm
Em
CO
<«:
CO r-
z
^-J
1
£m
MM CC MM
Mm
CO CJ
MM
»mi Em
CJ Cm
CC MM CC
•Cm
mM
mM
mM Em
'S)
MM
c—
CO
•<
Cm mM
«: mm
CJ E-"
mm Mm
• » •—»
Cvl mM CC
CM CM
(-M
«r-
Z
O
—O
E- a
CO MM
•<
«c
CJ
^ tM
CC
PM MJ CJ
CJ 21
«*;
CO fan CJ
UJ
<n •-M
CO ^z MM
>—
<-C CO Mm
MM
MM
i_ MM cu
—
-*:
l-l
Cm C
MM
Cm
MM —
Pm
CC <C C
s—
MM
CJ •< zz
CO Em Em
=» l-u MM
MM O
cc: :n M« CJ
DC MM CJ
Mm
£-> CM CC
£S fcM
CM.
mC
mC
CJ
-a
£m an >* Mm
a: >-"
MM
CC •< 6-"
CC MM t/1
c^tM e-<
»— MC
fc-« MM
-<
<c tM CU
CJ «>;
M- M-l
6-> MM
CJ
E- Ui CJ
MM MM
«c CC
MM
CC MM CO
HM
«; MM
MM
Ui
M»
MM
.
•<
O^O
t- j;
5
>- &M CO
O Cm
z
O
*L
1—
MM
«c -i
MM
P=
E-
z
£m
CJ
>*
•<—• (M
-^-^
O
Z^
1*
PC
Pm
O
O
O ^M
O
CO CO
^»
Mm
2:
MM
MM CC
MM
CC MM
MM 1—
CO
MM
2: 2«C MM
PC 6-"
Z >—
O £m
Pm
Z
ZZcCZmMZZmMPmCCmM-<mmCMmmOmMCCCmCm
mM
o
Mm
-a:
—
33
mM
Mm
?»
<<
CC mm
05 MM
MM PM
E-> e- CO (X CC
»-i «c MM Mm
21
•—
CU J_;
-2: fcM
CJ
MM
CO
cu cc
PM C_3 co CO
CM
>— V*4
Mm MU Mm
cc :C mm cc
ie- t-> CJ MM
l-M
PC MM
l*i
CO CO
MM «« 1—
6-> Eh MM rD >5
CO
CC PM CJ MM
e-> t*l MM 6Mm Cm Cm
2. i_
MM PM MM r+
mq Mm
mm CC CC cn
Mm MM
2:
CJ MM (—1 Mm MM
CJ
:j
1*1 i-P mm
CJ M- »—
0=1
Pm-^
»'— X
Qz
_
CM
a>
c-
i-m
Pm CC Cm
-UJ OJ
z
if)
>-•
1-4
M«
OO-<
^m- Ca3
m
cn
«-t
pM
CJ CO
e->
3=
^
»n
CJ mM
E-
o
01
*
«
ft
*
ft
ft
ft
ft
ft
*
o
C3
r^j
•—•
CSi
c- CO Cu
u)
mm E-" CJ
--C uC «£
Mm
MM =• mm
J
—
•^
i—
CJ
U
c_>
S
tn
•a-
»—
-a-
*
•<
o
t-i a-
M-t
*
>«
ft
CJ> a-
MM
ft
ft
•a-
O
Z
^
—
:
&-•
—
«»;
<
•
1Z
O-:^
Cw
MH
j3
-m"
P4 ce
mu :=>
t-
t-H iz;
-*: t-i
E-<
<" eMM
a:
C_>
>—
-c
i-h
Ph
Ph CJ 0,
p*
-a.
'
-«
«*:
si-
-a-
*
ft
*
t-
I-)
5.
^
-U
C3
O
^.
-a-
-a-
k«<
•a-
«%
MM
b- z 5:
\
Oo
i—
•>
*
1—
a-
ft
CJ
3£
Mi
Z
CJ
MM
a•a-
•a-
<*:
•a-
•a-
*
*
t-
e->
•a-
PQ
:o
CJ
6H
>—
3«
iZ sz
•^
«;
^_
o
>—
t-M
E-i
CJ
o
c
<t: fc-
O CJ
6-<
U-)
PC
HD
Eh
-m
(X
Z
MM
«
*
•a-
Or.
u)
CO
t— CJ «*;
cc 1— E->
Cm P Cj
£-•
)-M
aa-
*
*
z
z
•r-t
»
o
o
«
en
•>
o
ft
*
*
«
CJ
»«
O
ft
ft
ft
«*. ft
*
a-
a-
i_
ft
*
•a-
a-
*
*
CJ
•a-
MM
•a-
O
«
ft
ft
a-
CJ
-»
-a-
CJ
CJ
MM
*
*
*
a-
a-
(— t-»
«*:
MM
*
*
*
*
*
*
jt
O
ft
•a-
•a-
*
ft
ft
ft
ft
ft
ft
ft
ft
ft
ft
ft
ft
M.
a-
ft
•a-
ft
ft
ft
ft
a-
••
^i
«c
z
«
z3
m
o
CJ
>—i
i. CJ
r2 cm
MM ««;
CO
ft
-a-
a-
t-i
o
*
#
#
*
•a-
a-
WD2
s
O
i_
a-
•a-
a-
•a-
£h
Cm P- mM
«< •-3 «-3 CC
CO
CO
MM
•«;
•»•
•— i_
««: <t
-a-^.
*
(->
cj
»—
«
»
*
*
•-J
•a-
*
Cj
e-«
»—
^-t
a-
fc->
c_>
a-
3
O
«
M-.
W
o
«c
«
-a-
CJ
CJ PS
Oj f-l CU
tt >A
EH
Z
••CO
CJ
-
mx
f—
— <c ^o CMZ MM
j»
««;
»
*
ft
*-^*
r-i
CJ u^
(— cu
MM
2:
*
•«•
•H-
CJ
o
CJ o
-a•X"
*
*
#
*
i_
o
A
a-
i_
<c "X.
CJ PS
aa-
S-- a-
i—
1
->
t-i
6-"
i-l
—
*
*—
ft
M
•a-
•a-
ft
(-4
MM
CM
ft
*
m
o
*
c_i
•a-
&H
CO
MM
»
Em
1
•
<
ft
-^-
—« CJ
emm en «
-X CJ o
^~- uj
*
»
ft
a-
—
a-
a-
a•»
*
#
*
*
a#
*
«
*
*
a*
*
*
•a-
o
»
ft
MM CJ
CJ t-<
-M
MM u-i oj
SI CJ CO CJ »-h J* CJ CC
>*-*
»
*
*
#
*
CO
«*;
»
•»
*
*
*
*
*
#
»
*
#
*
•a-
*
*
*
ft
*
*
*
>CJ
•<
MM
C-4
U= >H
«-c
"
Mm Eh
21
«*: -i; CJ Mm
t-» .*
MM MM
«*:
MM ^_
CJ t-Q co :*s fc-M
mm cc C*
t->
cj <iJ
;—• «< e- c*. CJ. MM
•—1 mm 1—
^J •—
hd
Mm a. cj cj eg e->
SH
*
*
*
*
#
CJ i.
CC
e-< cj
co cj
>-" MM CC
CJ E- Ph
*:
PM MM
3J M-i MH
»-4
•a-
ft
ft
ft
*
»
«
*
«•
*
^C
z
»—
o
<
Z
*
#
#
#
*
*
*
«
#
*
*
*
»
#
*
#
*
#
*
*
*
ft
ft
ft
ft
ft
ft
*
aaa-
*
a*
aa->
*
a-
•a-
•a-
a
*
•a*
a
l-J P-H
«»; -M
•a-
>—
CJ cc
l-J fc-
«;
«
p— CJ
a-a-
*
••
e-
z
4fr
*
••
*
fc-t
•a-
MM
»—i
MM
«
*L
«
ft
ft
ft
ft
ft
a-
ft
•a-
ft
ft
ft
ft
•a-
aa-
a*
a-
*
•
<
"P-I
72
MM
MM kJ
MM
J
MM
t-M
- CC
MM MM CC CJ MM
*o
cc
JO
pev
-« o
z
Pm
•
cc
••O
«*; "
Z
P-.
fc-"
- t-3
•
CC
cc CJ MM
•»-H
MM
•
Eh »
P=i
C3
Pm cj t_5
»
21 MM
CJ CC^
CJ tT"i •
>*
«»
•
CJ^
»
-
•-H
«
tM
M
MM
»— CC CO
ZD CH fH
fcH
(—1
CJ
PM CC 6-«
t—
MM CC E-<
S»
-• CJ
CD CJ
• CJ CJ
MM
CC ?-• 'JJ
^^ MM
MM
*c MM
MM 6-H CC
fcn
mM
CD
ZL «c p4 «*:
E««;
MM 2^
MM
(.M MM
CJ
CC t«J
MM CJ
t«i
tH
CO JJ i*£.
MW
o
z
Z
O
^
Eh CO
-»; co
-i MM
MH CC
51 Pm
CC *
O
P-l fcH
Eh CO
MM
z
oo
CH
«
fc->
•
O
•
CJ
1MM
(
O £— CC MM
CJ Mm
M S O MM >—
W Oh
O
CC -c
MM
z •—-^ CJ
CO z
o (XOH
—I
Pm Em cj CO mm
CO z
CJ CD MM hJ
Pm
^4 -< MM
MM
M- CJ O Pm MM
CO Eh c- 2
mJ
CO MM
-. Z MM CJ -<
tH
•—
Cvj
E->
l-l
|i-H
fc-l
Pm
I
2: 7i
•< «c -<
kC CO Eh
«<
MM CJ
CC X,
a- <c CO
o
O
CJ
il wJ
•—•
£h
CD
t-M
«s; c-3
O MM MM
O CC
CC MM Pm
CJ CU mJ
?-<
<m
Z
^
-tO*
>»4
MM il <^". MM
PM CJ CC <c: cc CO en CC MM UD CO
cc
MM Mm
P> tH
•
Eh
CC
CC
CD
(tm CO CJ
MM Eh Dm fC
\
X
s
X
%
s
«
CJ
3
O
o —
O
•
MM
CJ
••
MM
<<
MM
f-Q
(-1
O
o
Co
« W
PMMMPrt
OP
mm cu pq
4kk
MM
Pm
•
PS
t-M
**t
••
-J PS
-
o
PS
-
o
*
M/V
Pm
v
*
•
MM-
-Mm
+1
"
PS
»-M
««: h-i
u
3
PS
O
-
t-M
1—4
6-t
•
*<U• W
t>
MM
in i. «»-•
II)mOH —
•*»
s
ou
*
z as
«
•-«
03
Wutt
c~
Mm
SB •— mm
HSf^iJS
>t»*M uq
•
t-4
•
CJ
2.
g-t
D3 MM
MM
6-i 35
•—•<*:
Mm i— £-i
•—J
i-M
O
p*
>->
»—
•
=
•
-3
i-M 6-> CJ>
-a;
ZO
o
•—
as
£3
O
P—•
O
fc-<
t-« c_> «*;
»
e- to
Z
W
«
i
t-«
-a: t-a «*:
e-« -*: Cm
<Z
cm <x o
•<
wz
•—
MM
z
«*;
;*
x; -»
Mm t—
i
a_ «« t—
x; &-. _m mm
t
w
o
•-»
<OK^
« o e^«
X
*
\
\
X
x
2:
O v.
o
>*
-
—
o ous
—
— CO
l—t
l—t
<_a
c_j
OS
CJ
•
>->
—
o
t—
»
MM
v
t-J
MM
»
Mm
P-.
>.
EH
t-J
»
—
o
«
z
_
o
O o
o
t^
t—
o 2
«q
EH
E-«
«*:
Mm 3s
21 EH
—
EH
O
p=: e1*1
Mm to
MM E-i
3t
3
O
>-
Pm
PM
U
O
O
CJ
£-•
•
M
to
C_)
«
B-«
P4
k«<
t—i
—
•
o
x
X
X
3z
MM
Pm
z
•-M
»-M
•t
\
t-M
Mm
Pm
•
as
CJ
i—
E3
•-3
1—1
ts*
as
»-M
»
PS
o
MM
M4
MM
z
e—
pq pa PQ pq Pm
CM MM Cm CQ 1-4
fc-«
X
X
Pm OU
Cm
(-4
.-J
-
C\J
GU
i-M
»
PS
o o*
l^> X
c-"
uu
6-
9
O
-_
•
fJ
PS
O
O
PQ pq po aa
(MpMP
PS
o
>-
to
O
J^
Ph
to
=>
»-^
—
o
••
»—
1-4
^-1
Z.
E->
<
«<;
P-,
»-4
M-t
i
Ph i—
"
» as
1
o•
If)
v
—
fc-"
Z Ui
O
t/j
MM
OS
Pm
mM
•
»
•< Mm
ZD
PS
z
o
Mm f
Z
Mo
o o
MM
PS O
«-<
t-M
z
Om
Mm
E-"
MM
PS
Z
o
o
o
CJ)
E->
xs
C_)
as
\
PS
CO
as
O
pm py
Cm
P4 P4
t-)
(-4
to
M4
O
e-t
~o
.- ft
MM
z;
US
Mm
i/i
-a:
CO
CO
MM
as
PM
o
<
CS
&. Cm
e- ti
I-M
«<;
•*T
IJJ
PS
Pm
»
6to
6as
<;
£CJ
«.
O
—
z
c
ECO
UJ xs
-M c_>
<c
as t—
CJ CO
Pm
MU
MM
o
O
«c
Cm
to
•
Q-.
o
(r-t
>
.-
•
WW CJ
t—
o
—
o
o
MM OS
MM Cm
i-»-l
S
•-4 £-•
i-M
=3
UM
M/V,
.
X
•
Mm
a-t
to
X
z
o
i-J-
M-i
•
Mm
—5
m
as
t_)
tU PM
CM
t-M
Mm
%
Pm
C4
o
o
to
CO
c_>
z:
z
MM
MM
PS
f«c
c_>
•
&-•
z
o
—1
(_>
M
»
l-u
•
r—
Mm as
PS C-J
^
O
9
Pm
><
PS
•
E-«
-
mm
r)
as
C_J
CJJ
•
<JU
o
IH
Z
—
U)
cv
to
to
MM
Pas
•-•
C_>
33
z
o
<*JV
\
•
t«H Mm
-
o
*—
l-M
•
*
Mm
z
o
^
•
*
P3 PS
e-
CJ»
fc->
-
MQ ^-.
as
£- as
to MM
Ma
»—
£-•
•
U) in
LO CJ
Mm
O
?-
o
l-M
••
6-i
t-M
•-»
Cm
O
Pm
CJ
• - to
'
o
Pm
i-4
•
*
•
to
PQ PQ PM
i— CQ MM
a^
-—
ys
c*>
!-—•
Cm
CO
tv
22
CO
Cm
>
C^
CJ
CD
o
o
»-M
«c
CO CO
MM I-M
Mm
M4
i.
C5
—i
-vr>-
•-M cx;
iZi
HM
•
Pm Pm
X
»—
3
O
MM
t-M
Mm
t-M
PM
Z
X
o•
o
X
• -*A-
»-M
PS
•
PS
«
%
i-J
E-«
Z
O
o
Pm
o
z
O
o
o
o
—
PS
MM
Mm
MM
ro
PQ
t—i
»
PQ
MM
33
Mm
PS
CJ
z
«c
M
O
M-t
Oz
as
Pm
—•
X
«<
to as
I— E3 >-• -•
S3
E-< >- «« «:
V
o•
utl
5-t
X
•
=D
»
PS tT
O
i-J
Pm
k4
>
PS
t—
r/5
to I— >h as
H- 1—4 uh
>->
o
6->
t-4
«c
ro PS
MM ZD
to
•
EH to •as
CJ
«C MM
PM as =3 •
Ph a
PS
mO >- PS
PM MM
a.
Ph 6-i
<*:
MM Mm
zz
s»
«c CJ «s: «c
to
<"">
t—
|jM
»
O—
•
1-3
ps
%
e-i
>-
6-< &-•
H/V
«c uu
to t—
»
pM
e-<
C/J
1^1
M/V
-Mm
CJ
C5
O
O
X
\
Z
H- X
-JE-iK
CVJ 6-»
EH
«C
i-J
1
MM
V
o c2:
%
to
—
as*
-
X
u»
-Mm
»
o• as
o
MM
KM
•
ps
•
•— MM MM
t-Q t-Q'•
i/v
3
.- um .-
o
73
1*4
Mm
CJ)
CO •-
;
APPENDIX.F^^-INTELHEX.Tg^DECIMAL.DATA.CONVERSION.PROGRAM
This
program
designed to read a data file that
is
is
in
Intelhex format and convert it to an integer file.
£
4
£1?
PRINT "INPUT FILE "'.INPUT I *
PRINT "OUTPUT FILE ": INPUT FO*
OPEN "0",£, FU*
OPEN "I", 1, Fit
INPUT #1, IN*
IF MID*< IN*, £, 1 ="0" THfcN CLOSE:UGrO 140
FOR 1=10 TO 40 STEP £
HX*=MID* IN*, i, 2)
T-
30
40
60
70
60
90 V/-=V(4L <"&H"+HX*)
95 IF 1=40 THtN PRINT ffu',USlNU"««#"iV% ELSE PRlNl *£, USING "###" ;V%
33 IF 1=40 THEN PRINT USING "*#*" ;V% ELSE PRINT USING "«tttt";V%;
100 NEXT I
1£S PRINT
130 GOfU 40
140 PRINT "DGNL" + CHK*(/)
)
(
74
APPENDIX
01UE
o-sr
O'QBt
G
-
Q-SCI
READY-E
O-M
0-9
O'O
30OLIN3bW
This is an example of the sampled utterence 'Ready'
75
APPENDIX H
-
SO WHAT-HIGH C
trtuz
This is an example of the sampiea utterence 'So What'
76
APPENDIX
roa
I
-
SNEEZE-F
O'SCI
This is an example of the utterence 'Sneeze
77
APPENDIX_J
This
table
-
lists
SPEECH_RESOLUTION_SUMMARY
actual
the
ranges
used
utterence out of a possible 256 levels (from
UTTERENCE
READY
RANGE
60
52
-
10
HIGH-C
35
10
25
10
-
MIDDLE-C
60
-
10
15
10
-
D
E
F
G
A
B
SO WHAT
to 255).
SCALE REFERENCE
MIDDLE-C
D
E
F
G
A
B
MIDDLE-C
D
E
F
G
A
B
HIGH-C
78
-
-
8 -
HIGH-C
SNEEZE
by
65
48
30
45
45
52
30
45
-
-
-
-
220
230
255
255
255
220
250
255
175
220
225
210
255
202
255
255
180
255
255
255
210
220
230
225
each
APPENDIX_K
SPEECH_REFLECTION_COEFFICIENT_PROGRAri
-
program
This
coefficients
to
4>
4k
for
12th order
a
>
Z
n
z
a
to
—
.
<>
c
x
*»
uj
k.
a
u
u
U.
UJ
4k
«0
-J
<>
II.
»
*
4>
Uj
Z
a
UJ
UJ
»
*>
K
<
<>
X
-J
<
4k
uj
<*
4>
<>>
"
X
**
>-
tt
it
4»
4k
it
it
»'
**
it
<»
O
IL'
Z
4>
to
<k
"
4k
4k
4k
**>
»
X
>t-
4>
3
«
X
u
4k
4k
Z
UJ
it
2:
4k
i>
tt
V.
uj
I-
C
it
*-
u.
4k
<
c
*'
4k
yl
X
r
*"
4>
Q.
k»
4k
4k
Z
4k
uj
4k
"*"
<»
1/1
•
c
c
Z
u
-
uj
k-
u.
u.
S
a
U
^
c
U
k-
4k
4k
4k
4k
**
4k
4k
4k
4k
4k
4k
4k
4k
4k
«-.
•
o
O—
C—
•
«O
OC
I
v
mi
•
«
N
^.
*
u.<
—
cu
1
u«
k- .
•
n
•
w
z
.
¥
«
UJ
o
z •
CO
u •
o
> c
k-
•
r »
«-*
•
•
—
")
M
U-
>uj
UC
2
Z3
•
•
— UJ3
3— #CJ
—
«Hw
n.
uj v-
Ov *
iij
a z c u «»
o — o: u
a < .u.^1
i
• .s
uj
T—«a
—
^
<\J
—*
*-
•
—-
L"
Q
Z
<
z
a
4k
<
»
n
—
4k
4k
z
•
j
as
•
•—>o:
O •
o — «-»
w"i
UJ
_
•
X
«
w.
UJ<>i/)y
U"!
4k
!i
4k
<>
Hi
a
a
<k
Z
UJ
C
xip
* *
¥
•
*»ir
O
—
.w
W
4k
w
—
o
i
4
O
~
<v
4k
uj
u
O
Z
—
Jk
•
—
C
•f>K
4k
4k
4k
<k
•
«a
—»
4k
*
Z
'
wo
U
i-
->
*:
4>
•"•
4»
of
4k
3
O
model
filter
lattice
u
u
UJ
4'
reflection
the
4k
4k
4k
<*
<>
<>
!>
<*
determine
to
It yields a new set of K's every 10 ms.
speech.
to
to
<>
designed
is
•••*
u,x
w
o
•
* < <
Ul
*
.^
-•
<? -
a.
-»
c
.o
—
< <
»UJ
i,*j!:
"I
w-ccxiLZZCJjo;
fcOUJuJk-XXZO
—z
II
II
_
•
»x
•*
vfl
<r
«^-<i::j<ci:
•
n
•
m:\i«— -ujuj!M — *>0
_c
*
•
— e-sa
r».
O — •O a •
c — Z'j*-— czc
o:
.
O xlU
II
.
II
—
<
<>!>—
•
<:
-\j-.-«—.—.-.—
^Tnnni
«-»m«i
* x
j
h. i
u
--<J
- U
;ji1l
XI
•
-«.<»- #
uj
il
——
u
-
-
— ^J
m
t
-
'-ii;:-«<T
-^jjjjjj-.
— ZZ
OJZU. -i
Z m
,
'I
•
II
,
il
>l
il
t|
It
u
o «ccn--ocjiiw if»"i*-e
c
J-uajJJjjj
-ak-rM .k>k.,fl<f «k>j t .».0i111-i1k>K«
<-*-ijjjjjj^: « z
-zz
—
z
-z
z m
uj- zo <<<<<< 3iuoaQ«-ao a a x«-a< o-oo- m:i*mcca«
J
u
o
||
«j
n
tk
H ->
_i
^^^'2^-^uucci<uu::L:kv^ii^.<
ii
JC
4k
4k
o
79
oo
a
o
c
3
O
•
(\J
M
^
"^
„.
f\J
*
Z
><
O
o
c
»
•
CM
•
•
m
oj
•M
rr
n
*
*
ar*
*
2 y — x
JJ'i JuJ
~~\
rvj
x
">
JJ J
o
n
-
-
*
*
< ? X
JJ
in
JiJ JO.
jnaa:aa:3Q:aQ:aiaz
:dz:z3zdz3zdzc
+Z_l_l_J_J_!_IJJ_J_J_l_l_Jt-C
jj ijj jjjjjjj:
•
-:<<<<<<<<<<<<< oZ
u u u u uuuuui u
i_
iw
'w>
rj u_
o
80
APPENDIX_L
-
REFLECTION_COEFFICIENT_PATTERNS_FOR
SPEECH WAVE FORMS
:::::•
'
'
:
•
r
Wo
ill:
;
!
•
I
.
CM
n
'
'.*,*.'•'
in
::•:•:
:
i
:
J
o
o
X
u
2
!
_e\»
c
i
t
i
•
i
:
:
:
i
j
-
i
c
i
i
t
i
i
[
)
:
:
i
1
1
:
i
o
!
{
c
J
c
i
i
I
'
2*1
6*0
Figure L.l.
90
CO
00
aaailNOVW 'AA20D
1
!
CO-
9'0-
80-
21-
N0IU31J3H
Sneeze-E Pattern of Reflection Coefficient Kl
81
T
80
00
aaruiNovw \ji3oo
CO
CO-
80-
1
8*0-
I
O
8*1-
Mono3ii3a
Figure L.2. Sneeze-E Pattern of Reflection Coefficient
K2
82
^i
i
;
1
o
CJ
\
p|fj
111?
3°
m
1
—
CO
o O
cv —
£
*
X
u
|
-j
© —
]
;-kW
_
m
;
;
fl?
|
1
i
21
80
Figure L.3.
1
i
i
90GO00
aaruiNovw \i43oo noiio31J3H
90
CO
6"0-
21-
Sneeze-E Pattern of Reflection Coefficient K3
83
Q
Z"*
..
.
N
W
O
u
CO
o O
N —
u
.w
1
21
80
1
90
GO
aaaxiMovw
Figure L.4.
1
00
\moo
CO-
80-
fl'O-
2*1-
NoiiDTiiaa
Sneeze-E Pattern of Reflection Coefficient K4
84
Q
3°
N
CM
—
X
w
j
. cv*
-o
1
21
80
Figure L.5.
1
1
'
1
90GO00
3aaiINDVW -JJ303 N0LL33133H
00
CO
«0-
2T-
Sneeze-E Pattern of Reflection Coefficient K5
85
N
cv»
W
3
u
^V^l
_w
2*T
80
9'0
GO
GO-
0*0
90-
80-
2'T-
aanilNOVW \*i303 N0I1331i3H
Figure L.6.
Sneeze-E Pattern of Reflection Coefficient K6
86
APPENDIX_M
following
The
lists
-
TREND_ANALYSIS_RESULTS
are
the
observations
made
on
the
reflection coefficient curves for each utterence.
"READY"
All
pitches have relatively flat
curves.
The
magnitudes vary slightly between +.3 and +1.0.
The higher
the pitch, the more defined the troughs are.
Kl
-
These
curves all had the unique feature of sloping
upward.
They generally ranged from -.4 to +.9.
No
other
correlation was noted.
K2
K3
-
-
A
negative sloping tendency characterised this
of
set:
curves
K4 - Each of these curves had a plateau.
did not fit in with this set at all.
Ready
however
3,
K5 - These curves seemed to stay within a similar range,
to -.7. Also several prominent peaks were uncorreiated
- No
correlations were
K6
drastically different.
noted,
however
Ready-B
.3
was
"SNEEZE"
Kl
-
Relatively flat curves.
K2
-
Highly uncorreiated curves.
Ranges from .3 to 1.0.
- Also highly uncorreiated curves,
K3
than K2.
K4
-
with
a
however,
- There seems to be a peak,
then a declining
Again MC and D don't
most of these curves.
observation and are generally flat.
-
fiat
the rest seem correiai.ee
MC and D are similarly fiat,
valley to an elevated flat plateau.
X5
K6
more
trend
in
fit
this
There are several peaks, then relatively fiat curves.
37
"SO WHAT"
Kl
-
Similarly flat patterns.
K2
-
Highly uncorrelated with no recognizable patterns.
There is
except A
prominent valley in all of the
observations
K3
-
K4
-
Highly uncorrelated with no recognizable patterns.
K5
-
Highly uncorrelated with no recognizable patterns.
K6
-
Highly uncorrelated with no recognizable patterns.
a
38
APPENDIX_N
program
This
frequencies
FAST_FOURIER_TRANSFORM_PROGRAM
-
determines
existing
there
if
within
are
coefficient
reflection
the
discrete
any
patterns.
<>
ir
ir
ir
O
UJ
U
«.'(•
Z
d.'
a
i>
*>
<
<
»-
ir
Q
O
i>
U
Uj
O
I
<>
""<
ta
O
I
o
<>
U.
3
yi
ai
»-
CL
uj
rt
o
o
%>
o
#
>-
ir
•
z
uj
—
u
o
;>~j
<>,
O
O
o
U.
It
it.
UJ
>
O
•
<V
o
o
UJ
<>
M
o
UJ
N
uj
uj
z
y)
1
H
o
O
o
o
o
o
o
<r
•>
*
"">
-»
*-
•
-•
If!
k
a
e
o
<r
-)
1
ir
c*
if!
o
o
»
•
ao
a
o
o
«»
or
q:
ir
>>
<
o
o
or
-*-»
- c
o
O
o
<>
u;
*
1
—
in
'.1
•
c
cm
•
X<CO« -» «
-
-.'«"-•*-»*~ — * ft <
<f
u
UJ
<
o
if
a u.
<i-
—
o
o
<
<
ir
—
—
k-
0.
u
o
U.
u.
uj
uj
o
octuj
2
o
i>
o
<r
O
—
c
UJ
K
<r
z
—
<
J
r
<r
uj
o
O
o
O
O
o
J
ir
<
•1
«-i
U.
u,
a.
o
<m
—
q
u
uj
a
O
uj
(U>-JZ
uj
uj
(/»
3l <»
I
••
UJO
•„-
• "J
CD-*
•
«">
*
—
-1
—
* x
*—
—
o—
-*
~z
-»
->
z
ir
•*
o—»o*»
ifl
-.
Z
-}
a"*
•:>
—
—
—
——
-~
<>
^-^~• — ^
^.**n
o^-**c — w^
«-*
iff
<\
w* <>
i/»
"">
'
_O
a
o—
—
-1l^
— C3
D^A»
-*
-
O t"
—.—
-»+«-»
>
—•<
+
3*
—-»*
C¥
Z <VO
*
^
—
•
iu -* a ;> • •
x — ••on—
•
3
O ^—
jm
a— 00 —
t>orz*-ci"(HHuN
<> — oo<!iD— ^»<—
<Hl|j"j2tt"«J«w
o
•- LU uJ *• z o »o
z«^^n?i
o j u
o<a:ujtro<-»irto»3»
«» z o jc a inoo — ir—
'^1
^*»—
— —
«~
— *c— co
^
3
«/>
—-—1
ci"^"^
0010rw—
— ccc
^oo-fi^w
—
<
..it
—
— u.ij<0— o»»o
» •* + *
-*cvio
occ
»co» — — Z O zoor^i>~o
OOvUjO*VujOZZO
h h h
o
o — a a .j— o 2 o o o —
•-•*)
ox z
uj
— io-«::' —
oasir in •- "3:^0::::a o dc
---ti.izti— -"i^i:
<nj <u. m> jji •ootmo-3
— * — co— m — n».o— •• — c»»i
x
<^c n
i*-- -<:::
iO"ni.-M+Z3*-->-01 «i Z3(> tjj-z2
o-rrz- — ^ a— — z —
25<vk< — rT <w*»z
z
O ^ «-U^33 0ZUJ3DDU-"i.i303ii3D0-iuuJ3iidJiiJ33u-J
<>or^<u.iriLr. n— a2i'iif)CttC3UQoZi,*iri'^L.cca"»u.fl)ua-.a<i
S'
1
^
s,
ir
*j
.
.
.
.
- .1
r.
<.-
-•
v-
sj
2
Uj
m
<
2
* z uO
^ z uj —
Q
i,
U. U,
3- u.
o
<
o
o
o
j
•
z
rr
33
v
t-
o
<
<>
<>
J
<v
<
Z
</>
UJ
u.'
z
<
a:
k
a
ir
ifi
l/"w
—— x
OZuOOOUCDOOw—
— h ZZ<ri/1«
O Q
r
*
->
—
CVJ-*
oJ
if
<>
—
->•
-
o
oojio<zua:zo3»
*<p»-UJX<U.— <>0~
u
<v
1
—
*
»
o
zai
if!
• -*
—
<>
«*
*
->
—
Cc
w •
<>
O
O
ft
O
on—
z
<>
<>
<>
<>
Z
O
c
z
if
ir
»-
z
<
i>
ir
fc-
D
—-•
<>
<>
ft
• CM
<,
<
ir
O
O
O
<>
o
it
o
o
-*
ir
ir
ir
•
O
O
O
o
•
<>
<>
U.
11
11
11
11
<•
11
1
1
11
11
11
,
u
L
1
t
\
O
U
o
O
o
if
i,
o
aj 2.
11
1
ir
ir
11
«>
O
O
ir
o
UUWUUUUUUUUUUUUUU
IT
89
00
fM
-
—
—
4ft
4ft
4ft
4ft
4*
4ft
4ft
4ft
4ft
4ft
4>
4ft
4ft
4ft
4ft
4ft
4ft
4ft
ft
<*
4ft
1
4ft
—
M4ft
t
^
*»
w—
* —4ft
\ft
• -4ft
CO
4ft
4»
4ft
»
—
Z
^J
4ft
u.
»ft
4/
•JJ4ft
4>
£
4ft
• •
<ft
1
—
Z
——
LUWI
—
w«~
u, <:
X
1
1
1
—
z
*»
1
—
* *
—
L
———
|
4ft
Z *
1
z z
—
— —
W—
mm
4ft
—
—
ii^
<•
1
—i
1
—
3
4ft
X
V — —
—3
•
4*
•
ZZ
4ft
o
1
4ft
4k
</>
tl
II
It
1
•
4»
II
II
U.
4ft
II
<>
II
4ft
—
nonuoo
1
1
1
»
II
joi<:u^
ft- <ft
-V
• CO
4ft
mu.
t
»4ft
—
•
•
X
4ft
nr\j
r*.
4ft
4ft
—
•
X X
4ft
•
c
•
*J
•
T
* *
cr,
•
4ft
r*
4ft
-o
• •
4ft
4ft
4ft
~;3>
4ft
-u.
•
4ft
m *>
•
4ft
— O—
4ft
>?
4ft
U»LU4ft
ft- mm 4ft
—
rs»
•
•
N
33U——
•
3—
33 2ft- «"
t
II
n
-«x <
"III
Z 2 COO
22<»33GOOft-Z
— —4ft
4ft
<ft
4ft
4ft
4ft
4ft
4ft
OO
o
3
u>
^<?
u
•
nx
4ft
O3
3 3<ft
f>
•
•
—X
• <\
n •
(bo
——
•
— CD
»4ft
——
—^
——
•
m
»4ft
Za
7 7
O iijuJ33o
3iii.ouu<> *i<'Ouiuu
II
•
4ft
4ft
4ft
t
•
•
<ft
4ft
Z2
mm — r — mm
O-E—
• • — vj
•U U)
o ru * < s Z MO 33
o
dJ UJ
3
Id mm
— 5 0—3U1 O — — ~ 3
— — — — 2 < — 1")ZZ
— 3» •——o
3—
z in + z z*- — £ * Z7»- a
— — Z Z4>
zz
—
i
X *
•
f\j
mo
mm
* X
4ft
•
4>
— -^
on
——
•
mx
—
<ft
——
4b
•
•—4ft
•
•
4ft
«ft
<ft
^
1
*
fNJ
m aj
— u.
*>4ft
—3
* —
— —
•
/•>
-<0
4ft
4>
4ft
-> "1
in
t
4ft
4ft
——
IT—
—X
t
• u.
4ft
4ft
4ft
UjN
•
n •
—x
-•4ft
——
4ft
%r
1
X
• <e
<ft
OX
mm <»
4>
»
-> mm
4ft
4ft
4ft
—I
•
—
-co
4ft
m~mm
")
•
WPJ
m
4ft
-1
~
4ft
4ft
L
"1 u.
•
4ft
4ft
-
•
4ft
4ft
«»
o
— 00
4ft
4ft
—
mm
30
330
4ft
4ft
u
u,u)uj
•
——
-
i*>
90
APINQIX_0
-
This
an example of the output from the FFT
is
determine
FREQyENCY_CgNTENI_0F_KlN2.
program
to
if there are any discrete frequencies present
in
the reflection coefficient patterns.
B
n
a
o
trooi
o*a»
O'QQC
o*aoz
O'QOt
Domrusww
Fiaure 0.1.
Reflection Coefficient K6 for
Utterence 'Ready-MC
91
ro
-8
a
\
'd
o
(
i
a
\
s
1
o
3
)
\
o
3
{
j
O
S
\
c
y
o
A
>
o
^
)
o
(
//
^
o
a
•
crcD«
o*at»
o*aoc
O'txz
o-ao(
3QOL[N3bW
Figure 0.2.
Reflection Coefficient K3 for
Utterence 'Ready-MC.
92
0*0
T-8
3
gt
sS
2
t—
(TOM
o-ao*
crocs
o'ooe
o-ooi
3Cni[N3**J
Figure 0.3.
Reflection Coefficient K4 for
Utterence 'Sneeze-MC.
93
o-»
-8
f
^
g -
"2
3
o
"3
'R
4o*ao»
0*00l»
o*ok
o'ace
0*001
XOUNSbW
Figure 0.4.
Reflection Coefficient Kl for
Utterence 'Sneeze-MC
94
O'O
8
~e u>
"2^
y
O
--s
0*00*
croc*
Figure 0.5.
o'oce
O'QQt
o-ooi
Reflection Coefficient K6 for
Utterence 'So What-MC.
95
*r»
I
-e
^
s
o
2
I
-a
traor
0'Q0»
Figure 0.6
O'QOC
O'QQC
O'OOI
Reflection Coefficient K4 for
Utterence 'So What-MC
96
0*0'
*»
2
LIST OF REFERENCES
"Seporof the Panel on Communicative Disorders National
Advisory Neurological
arc
Communica
Councii,"3|rr]3orv_^::]_:':r_'
Disorders and Stroke
Hearina Imoairea, IEEE Press, 1330.
1
Pickett.
"Frequency Lowering for Hearii
J.M.
Sensory_A ics_f or_the_Hear i ng_Imo^ i^"
IEEE
•
1
980
.
Hall, Cgm2yter_Mgdeling_of — Voice_Sicna
d2Q2y5L52i:?_Eitch_and_Formant_Fre'2\:err .5S,
Naval Postgraduate School, Monterey, CA
L97
G.T.
,
;
,
Evans, SDeech_Svnthesis_-_Mgdels_ar]
I9S3 Student Pacers.
.B.
EEE
.\
and 5
rtanauer
linear oreaiction oi
bv
- a 1
.
'ntnesis
Soeech Analysis,
:
r-
L
mear
Soeech Analysis,
I=.nE
j
r.
.
aknou
j.H.
-'
B
? _
.
S
.
.
9
_
.
Lth,
o
'•
,
Atai
.
•
'-
5
-
.
_
"
:
L
-
968
.
"
•
thesis,
j
r
i
•
:
-
.
>lytecn.
?tec_
_
sta]
.
rati
.
.
5
.
;
.
r
-
-
"
__
'
.
.
tne
?T,
Press.
5eDtember
.
j
.
.
LiY?_'Y!2i2®_
_
Ph E
,
es
oeec
^u t gm a t i c_Si
,
3
-
pr
r"
.i'
et
l_
"
!
z
.
.
BIBLIOGRAPHY
Flanagan, J. L. ana Rabmer, L. R., Soeech_Syntnesis,
Dowden Hutchingson & Ross, Inc., 1973.
,
Link_80_0peratgrs_Guide
,
Digital Research, Monterey, CA
,
i960.
^3Q£2_d§§§!B2i§£_Q2§£§22£§_!£yi!2®
Monterey, CA
i960.
•
Researcn
Digital
,
Prglog__7804_ZS0A__Processor - Counter /Timer _Card__tJser^s_Manua
Prolog Corporation. Monterey CA
1981.
_
,
Raoiner, L. R. ana Shafer R. w., Digital_Processing_pf
Speech Sionals, Prentice-Hai., 1978.
,
Shafer, R. W.
Press, 1979.
anc
~ o e ec n
Markei
Ana
3 Z§try'2_280 _CPU _Central_Prgcess ir.2_^2 12 _?r 251!
_;_
Ziloa Inc.. Santa Clara, CA
Zaks,
R.,
.
feoruarv 1981.
Proorainmirn the Z80,
Svoex inc.,
IS
=
'
2_=
'
-iL
2
L -
I
INITIAL DISTRIBUTION LIST
No.
1.
Library, Code 0142
Naval Postgraduate School
Monterey, California 93943-5100
2.
Department Chairman, Code 62
Department of Electrical and
Computer Engineering
Naval Postgraduate School
Monterey, California 93943
3.
Professor P. Moose, Code 62Me
Naval Postgraduate School
Monterey, California 93943
4.
Professor B. Madan
Department of Computer Science
and Engineering
I.I.T., New Delhi - 110016
INDIA
5.
Vinny DiGirolamo,
Fire Road Drive
3ay Shore. New York
Lt.
LfSN
3
6.
H706
Defense Technical Information Center
Cameron Station
Alexandria, Virginia 223>jh-6j.45
Cooies
1
qq-VM^
DiGirolamo
Analysis of a digital technique for frequency transportation
of speech.
/
2H>173
Thesis
P5 7 39
c.l
D 0? ro3 37TO
Analysis of a digital technique for frequency transportation
of speech.
-
?
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement