1

INTRODUCTION

1.1SignalsandInformation
1.2SignalProcessingMethods
1.3ApplicationsofDigitalSignalProcessing
1.4SamplingandAnalog to DigitalConversion

ignal processing is concerned with the modelling, detection,

identificationandutilisationofpatternsandstructuresinasignal
process.Applicationsofsignalprocessingmethodsincludeaudiohi-
fi,digitalTVandradio,cellularmobilephones,voicerecognition,vision,
radar,sonar,geophysicalexploration,medicalelectronics,andingener al
anysystemthatisconcernedwiththecommunicationorprocessingof
information. Signal processing theory plays a central role in the
developmentofdigitaltelecommunicationandautomationsystems,andin
efficientandoptimaltransmission,receptionanddecodingofinformati on.
Statisticalsignalprocessingtheoryprovidesthefoundationsformodel ling
thedistributionofrandomsignalsandtheenvironmentsinwhichthesignal s
propagate. Statistical models are applied in signal processing, and in
decision-makingsystems,forextractinginformationfromasigna lthatmay
benoisy,distortedorincomplete.Thischapterbeginswithadefinition of
signals,andabriefintroductiontovarioussignalprocessingmethodologie s.
Weconsiderseveralkeyapplicationsofdigitalsignalprocessing inadaptive
noise reduction, channel equalisation, pattern classification/recognition,
audiosignalcoding,signaldetection,spatialprocessingfordirectiona l
receptionofsignals,Dolbynoisereductionandradar. The chapter conclude s
withanintroductiontosamplingandconversionofcontinuous-timesignals
to digital signals.
S

H HH H E EE E LL LL LL LL O OO O

2 Introduction

1.1SignalsandInformation

Asignalcanbedefinedasthevariationofaquantitybywhichinform ation
isconveyedregardingthestate,thecharacteristics,thecomposi tion,the
trajectory,thecourseofactionortheintentionofthesignalsource .Asignal
isameanstoconveyinformation. Theinformationconveyedin a signal may
beusedbyhumansormachinesforcommunication,forecasting,decision-
making,control,explorationetc.Figure1.1illustratesaninformations ource
followedbyasystemforsignallingtheinformation,acommunication
channelforpropagationofthesignalfromthetransmittertothere ceiver,
andasignalprocessingunitatthereceiverforextractionofthe information
fromthesignal.Ingeneral,thereisamappingoperationthatmaps the
information I(t)tothesignal x(t)thatcarriestheinformation,thismapping
function may be denoted as T[]and expressed as

)] ( [ ) ( t I T t x = (1.1)

For example, in human speech communication, the voi ce-generating

mechanismprovidesameansforthetalkertomapea chwordintoadistinct
acousticspeechsignalthatcanpropagatetotheli stener.Tocommunicatea
word w,thetalkergeneratesanacousticsignalrealisati onoftheword;this
acousticsignal x(t) maybecontaminatedbyambientnoiseand/ordistort ed
byacommunicationchannel,orimpairedbythespea kingabnormalitiesof
thetalker,andreceivedasthenoisyanddistorted signal y(t).Inadditionto
conveyingthespokenword,theacousticspeechsign alhasthecapacityto
conveyinformationonthespeakingcharacteristic, accentandtheemotional
stateofthetalker.Thelistenerextractsthesein formationbyprocessingthe
signaly(t).
Inthepastfewdecades,thetheoryandapplication sofdigitalsignal
processinghaveevolvedtoplayacentralroleint hedevelopmentofmodern
telecommunication and information technology system s.
Signalprocessingmethods are central to efficient communication, and to
thedevelopmentofintelligentman/machineinterfac esinsuchareasas
Information
source
Informationto
signalmapping
Signal
DigitalSignal
Processor
Channel
Noise
Noisy
signal
Signal&
Information

Figure1.1 Illustrationofacommunicationandsignalprocessingsystem.

SignalProcessingMethods 3

speechandvisualpatternrecognitionformultimedi asystems.Ingeneral,
digitalsignalprocessingisconcernedwithtwobro adareasofinformation
theory:

(a) efficientandreliablecoding,transmission,re ception,storageand

representation of signals in communication systems, and
(b) the extraction of information from noisy signal s for pattern
recognition, detection, forecasting, decision-makin g, signal
enhancement, control, automation etc.

Inthenextsectionweconsiderfourbroadapproach estosignalprocessing
problems.

1.2SignalProcessingMethods

Signalprocessingmethodshaveevolvedinalgorithm iccomplexityaiming
foroptimalutilisationoftheinformationinorder toachievethebest
performance.Ingeneralthecomputationalrequireme ntofsignalprocessing
methodsincreases,oftenexponentially,withtheal gorithmiccomplexity.
However,theimplementationcostofadvancedsignal processingmethods
hasbeenoffsetandmadeaffordablebytheconsiste nttrendinrecentyears
ofacontinuingincreaseintheperformance,couple dwithasimultaneous
decrease in the cost, of signal processing hardware .
Dependingonthemethodused,digitalsignalproces singalgorithmscan
becategorisedintooneoracombinationoffourbr oadcategories.Theseare
non parametricsignalprocessing,model-basedsignalpr ocessing,Bayesian
statisticalsignalprocessingandneuralnetworks. Thesemethodsarebriefly
described in the following.

1.2.1Non parametricSignalProcessing

Non parametricmethods,asthenameimplies,do not utiliseaparametric

modelofthesignal generation or a model of the st atistical distribution of the
signal.Thesignalisprocessedasawaveformora sequenceofdigits.
Non parametric methods are not specialised to any parti cular class of
signals,theyarebroadlyapplicablemethodsthatc anbeappliedtoany
signalregardlessofthecharacteristicsorthesou rceofthesignal.The
drawback of these methods is that they do not utili se the distinct
characteristics of the signal process that may lead to substantial
4 Introduction

improvementinperformance.Someexamplesofnon parametricmethods
includedigitalfilteringandtransform-basedsigna lprocessingmethodssuch
astheFourieranalysis/synthesisrelationsandthe discretecosinetransform.
Somenon parametricmethodsofpowerspectrumestimation,in terpolation
and signal restoration are described in Chapters 9, 10 and 11.

1.2.2Model-BasedSignalProcessing

Model-basedsignalprocessingmethodsutiliseapar ametricmodelofthe
signalgenerationprocess.Theparametricmodelnor mallydescribesthe
predictablestructuresandtheexpectedpatternsin thesignalprocess,and
canbeusedtoforecastthefuturevaluesofasign alfromitspasttrajectory.
Model-basedmethodsnormallyoutperformnon parametricmethods,since
theyutilisemoreinformationintheformofamode lofthesignalprocess.
However,theycanbesensitivetothedeviationsof a signal from the class of
signalscharacterisedbythemodel.Themostwidely usedparametricmodel
isthelinearpredictionmodel,describedinChapte r8.Linearprediction
modelshavefacilitatedthedevelopmentofadvanced signalprocessing
methodsforawiderangeofapplicationssuchaslo w bit ratespeechcoding
incellularmobiletelephony,digitalvideocoding, high resolutionspectral
analysis, radar signal processing and speech recogn ition.

1.2.3BayesianStatisticalSignalProcessing

Thefluctuationsofapurelyrandomsignal,orthe distributionofaclassof
randomsignalsinthesignalspace,cannotbemodel ledbyapredictive
equation,butcanbedescribedintermsofthestat isticalaveragevalues,and
modelledbyaprobabilitydistributionfunctionin amultidimensionalsignal
space.Forexample,asdescribedinChapter8,ali nearpredictionmodel
drivenbyarandomsignalcanmodeltheacousticre alisationofaspoken
word.However,therandominputsignalofthelinea rpredictionmodel,or
thevariationsinthecharacteristicsofdifferent acousticrealisationsofthe
same word across the speaking population, can only be described in
statisticaltermsandintermsofprobabilityfunct ions.Bayesianinference
theoryprovidesageneralisedframeworkfor statist ical processing of random
signals,andforformulatingandsolvingestimation anddecision-making
problems.Chapter4describestheBayesianinferenc emethodologyandthe
estimation of random processes observed in noise.
Applications of Digital Signal
Processing
5

1.2.4NeuralNetworks

Neuralnetworksarecombinationsofrelativelysimp lenon-linearadaptive
processing units, arranged to have a structural res emblance to the
transmissionandprocessingofsignalsinbiologica lneurons.Inaneural
networkseverallayersofparallelprocessingeleme ntsareinterconnected
withahierarchicallystructuredconnectionnetwork .Theconnectionweights
aretrainedtoperformasignalprocessingfunction suchaspredictionor
classification. Neural networks are particularly us eful in non-linear
partitioningofasignalspace,infeatureextracti onandpatternrecognition,
andindecision-makingsystems.Insomehybridpatt ernrecognitionsystems
neuralnetworksareusedtocomplementBayesianinf erencemethods.Since
themainobjectiveofthisbookistoprovideacoh erentpresentationofthe
theoryandapplicationsofstatisticalsignalproce ssing,neuralnetworksare
not discussed in this book.

1.3ApplicationsofDigitalSignalProcessing

Inrecentyears,thedevelopmentandcommercialava ilabilityofincreasingly
powerfulandaffordabledigitalcomputershasbeen accompaniedbythe
developmentofadvanceddigitalsignalprocessinga lgorithmsforawide
varietyofapplicationssuchasnoisereduction,te lecommunication,radar,
sonar,videoandaudiosignalprocessing,patternr ecognition, geophysics
explorations,dataforecasting,andtheprocessing oflargedatabasesforthe
identificationextractionandorganisationofunkno wnunderlyingstructures
and patterns. Figure 1.2 shows a broad categorisati on of some DSP
applications.Thissectionprovidesareviewofsev eralkeyapplicationsof
digital signal processing methods.

1.3.1AdaptiveNoiseCancellationandNoiseReducti on

Inspeechcommunicationfromanoisyacousticenvir onmentsuchasa
movingcarortrain,oroveranoisytelephonechan nel,thespeechsignalis
observedinanadditiverandomnoise.Insignalmea surementsystemsthe
information-bearing signal is often contaminated by noise from its
surrounding environment. The noisy observation y(m) can be modelled as

6 Introduction

y(m) = x(m) + n(m) (1.2)

where x(m) and n(m)arethesignalandthenoise,and misthediscrete-

timeindex.Insomesituations,forexamplewhenus ingamobiletelephone
inamovingcar,orwhenusingaradiocommunicatio ndeviceinanaircraft
cockpit,itmaybepossibletomeasureandestimate theinstantaneous
amplitudeoftheambientnoiseusingadirectional microphone.Thesignal
x(m) maythenberecoveredbysubtractionofanestimat eofthenoisefrom
the noisy signal.
Figure 1.3 shows a two-input adaptive noise cancell ation system for
enhancementofnoisyspeech.Inthissystemadirec tionalmicrophonetakes
DSP Applications
Information Transmission/Storage/Retrieval
Informationextraction
Signal Classification
Speech recognition, image
and character recognition,
signaldetection
Parameter Estimation
Spectral analysis,radar
andsonarsignal processing,
signal enhancement,
geophysics exploration
ChannelEqualisation
Source/ChannelCoding
Speech coding, image coding,
datacompression,communication
overnoisychannels
Signal and data
communicationon
adversechannels

Figure1.2 Aclassificationoftheapplicationsofdigitalsignalprocessing.

y(m) = x (m)+n(m)
n(m+)
x(m)
^
n(m)
^
z z
. . .
Noise Estimation Filter
Noisy signal
Noise
Noise estimate
Signal
Adaptation
algorithm
z
1
w
2
w
1
w
0 w
P-1
1 1

Figure1.3 Configurationofatwo-microphoneadaptivenoisecanceller.
ApplicationsofDigitalSignalProcessing 7

asinputthenoisysignal x(m) + n(m) ,andaseconddirectionalmicrophone,

positioned some distance away, measures the noise n(m+ ). The
attenuationfactor andthetimedelay providearatherover-simplified
modeloftheeffectsofpropagationofthenoiseto differentpositionsinthe
space where the microphones are placed. The noise f rom the second
microphoneisprocessedbyanadaptivedigitalfilt ertomakeitequaltothe
noisecontaminatingthespeechsignal,andthensub tractedfromthenoisy
signaltocanceloutthenoise.Theadaptivenoise cancellerismoreeffective
incancellingoutthelow-frequencypartofthenoi se,butgenerallysuffers
from the non-stationary character of the signals, a nd from the over-
simplified assumption that a linear filter canmode lthediffusionand
propagation of the noise sound in the space.
In many applications, for example at the receiver o f a
telecommunicationsystem,thereis noaccesstotheinstantaneousvalueof
thecontaminatingnoise,andonlythenoisysignal isavailable.Insuchcases
thenoisecannotbecancelledout,butitmaybere duced,inanaverage
sense,usingthestatisticsofthesignalandthen oiseprocess.Figure1.4
showsabankofWienerfiltersforreducingadditiv enoisewhenonlythe

.
.
.
y(0)
y(1)
y(2)
y(N-1)
Noisy signal
y(m)=x(m)+n(m)
x(0)
x(1)
x(2)
x(N-1)
^
^
^
^
I
n
v
e
r
s
e

D
i
s
c
r
e
t
e

F
o
u
r
i
e
r

T
r
a
n
s
f
o
r
m
.
.
.
Y(0)
Y(1)
Y(2)
Y(N-1)
D
i
s
c
r
e
t
e

F
o
u
r
i
e
r

T
r
a
n
s
f
o
r
m
X(0)
X(1)
X(2)
X(N-1)
^
^
^
^
WN -1
W
0
W
2
Signal and noise
powerspectra
Restored signal
Wienerfilter
estimator
W
1
.
.
.
.
.
.

Figure1.4 AfrequencydomainWienerfilterforreducingadditivenoise.
8 Introduction

noisysignalisavailable.Thefilterbankcoeffici entsattenuateeachnoisy
signalfrequencyininverseproportiontothesigna ltonoiseratioatthat
frequency.TheWienerfilterbankcoefficients,der ivedinChapter6,are
calculatedfromestimatesofthepowerspectraoft hesignalandthenoise
processes.

1.3.2BlindChannelEqualisation

Channelequalisationistherecoveryofasignaldi stortedintransmission
throughacommunicationchannelwithanon-flatmag nitudeoranon-linear
phaseresponse.Whenthechannelresponseisunknow ntheprocessof
signalrecoveryiscalledblindequalisation.Blind equalisationhasawide
range of applications, for example in digital telec ommunications for
removalofinter-symbolinterferenceduetonon-ide alchannelandmulti-
pathpropagation,inspeechrecognitionforremoval oftheeffectsofthe
microphonesandthecommunicationchannels,incorr ectionofdistorted
images,analysisofseismicdata,de-reverberation ofacousticgramophone
recordings etc.
Inpractice,blindequalisationisfeasibleonlyif someuseful statistics of
thechannelinputareavailable.Thesuccessofab lindequalisationmethod
dependsonhowmuchisknownaboutthecharacterist icsoftheinputsignal
andhowusefulthisknowledgecanbeinthechannel identificationand
equalisationprocess.Figure1.5illustratestheco nfigurationofadecision-
directedequaliser.Thisblindchannelequaliseris composedoftwodistinct
sections:anadaptiveequaliserthatremovesalarg epartofthechannel
distortion, followed by a non-linear decision devic e for an improved
estimateofthechannelinput.Theoutputofthede cisiondeviceisthefinal

Channel noise
n(m)
x(m)
Channel distortion
H(f)
f
y(m)
x(m)
^
Error signal
-
+
Adaptation
algorithm
+
f
Equaliser
Blind decision-directed equaliser
H
inv
(f)
Decision device
+

Figure1.5 Configurationofadecision-directedblindchannelequaliser.
ApplicationsofDigitalSignalProcessing 9

estimate of the channel input, and it is used as th e desired signal to direct the
equaliseradaptationprocess.Blindequalisationis coveredindetailin
Chapter 15.

1.3.3SignalClassificationandPatternRecognition

Signalclassificationisusedindetection,pattern recognitionanddecision-
makingsystems.Forexample,asimplebinary-state classifiercanactasthe
detectorofthepresence,ortheabsence,ofaknow nwaveforminnoise.In
signalclassification,theaimistodesignaminim um-errorsystemfor
labellinga signal with one of a number of likely classes of signal.
Todesignaclassifier;asetofmodelsaretrained fortheclassesof
signalsthatareofinterestintheapplication.Th esimplestformthatthe
models can assume is a bank, or code book, of wavef orms, each
representingtheprototypeforoneclassofsignals .Amorecompletemodel
foreachclassofsignalstakestheformofaproba bilitydistributionfunction.
Intheclassificationphase,asignalislabelledw iththenearestorthemost
likelyclass.Forexample,incommunicationofabi narybitstreamovera
band-passchannel,thebinaryphaseshiftkeying(B PSK)schemesignals
thebit1usingthewaveform A
c
sin
c
t andthebit0using A
c
sin
c
t .
Atthereceiver,thedecoderhasthetaskofclassi fyingandlabellingthe
receivednoisysignalasa1ora0.Figure1.6 illustratesacorrelation
receiverforaBPSKsignallingscheme.Thereceiver hastwocorrelators,
eachprogrammedwithoneofthetwo symbols represe nting the binary states

Receivednoisysymbol
Correlator for symbol "1"
Correlator forsymbol "0"
Corel(1)
Corel(0)
"
1
"

i
f
C
o
r
e
l
(
1
)

C
o
r
e
l
(
0
)
"
0
"

i
f
C
o
r
e
l
(
1
)
<
C
o
r
e
l
(
0
)
"1"
Decision
device

Figure1.6 Ablockdiagramillustrationoftheclassifierinabinaryphase-shiftkeying
demodulation.
10 Introduction

forthebit1andthebit0.Thedecodercorrel atestheunlabelledinput
signalwitheachofthetwocandidatesymbolsands electsthecandidatethat
has a higher correlation with the input.
Figure1.7illustratestheuseofaclassifierina limitedvocabulary,
isolated-wordspeechrecognitionsystem.Assumethe reare Vwordsinthe
vocabulary.Foreachwordamodelistrained,onma nydifferentexamples
ofthespokenword,tocapturetheaveragecharacte risticsandthestatistical
variationsoftheword.Theclassifierhasaccesst oabankof V+1models,
oneforeachwordinthevocabularyandanaddition almodelforthesilence
periods.Inthespeechrecognitionphase,thetask istodecodeandlabelan
M
ML
.
.
.
Speech
signal
Feature
sequence
Y
f
Y|
M
(Y|M
1
)
Wordmodel M
2
likelihood
of M
2
M
o
s
t
l
i
k
e
l
y
w
o
r
d
s
e
l
e
c
t
o
r

Feature
extractor
Wordmodel M
V
Wordmodel M
1
f
Y|
M
(Y|M
2
)
f
Y|
M
(Y|M
V
)
likelihood
of M
1
likelihood
of M
v
Silencemodel
M
sil
f
Y|
M
(Y|M
sil
)
likelihood
of M
sil

Figure1.7 Configurationofspeechrecognitionsystem,f(Y|M
i
)isthelikelihoodof
themodelM
i
givenanobservationsequenceY.

ApplicationsofDigitalSignalProcessing 11

acousticspeechfeaturesequence,representinganu nlabelledspokenword,
asoneofthe Vlikelywordsorsilence.Foreachcandidatewordt he
classifiercalculatesaprobabilityscoreandselec tsthewordwiththehighest
score.

1.3.4LinearPredictionModellingofSpeech

Linearpredictivemodelsarewidelyusedinspeech processingapplications
such as lowbitrate speech coding in cellular tele phony, speech
enhancementandspeechrecognition.Speechisgener atedbyinhalingair
intothelungs,andthenexhalingitthroughthevi bratingglottiscordsand
thevocaltract.Therandom,noise-like,airflowf romthelungsisspectrally
shapedandamplifiedbythevibrations of the glott al cords and the resonance
ofthevocaltract.Theeffectofthevibrationsof theglottalcordsandthe
vocaltractistointroduceameasureofcorrelatio nandpredictabilityonthe
randomvariationsoftheairfromthelungs.Figure 1.8illustratesamodel
forspeechproduction.Thesourcemodelsthelunga ndemitsarandom
excitationsignalwhichisfiltered,firstbyapit chfiltermodeloftheglottal
cords and then by a model of the vocal tract.
Themainsourceofcorrelationinspeechisthevo caltractmodelledbya
linearpredictor.Alinearpredictorforecaststhe amplitudeofthesignalat
time m, x(m) , using a linear combination of P previous samples

x(m1),L, x(m P) [ ]as

=
=
P
k
k
k m x a m x
1
) ( ) ( (1.3)

where

x(m) is the prediction of the signal x(m) , and the vector

] , , [
1
T
P
a a K = a isthecoefficientsvectorofapredictoroforder P.The

Excitation
Speech
Random
source
Glottal (pitch)
model
P(z)
Vocal tract
model
H(z)
Pitch period

Figure1.8 Linearpredictivemodelofspeech.
12 Introduction

predictionerror e(m),i.e.thedifferencebetweentheactualsample x(m)
and its predicted value

x(m) , is defined as

e(m) = x(m) a
k
x(m k)
k=1
P

(1.4)

Thepredictionerror e(m)mayalsobeinterpretedastherandomexcitation
ortheso-calledinnovationcontentof x(m) .FromEquation(1.4)asignal
generated by a linear predictor can be synthesised as

x(m) = a
k
x(m k) + e(m)
k=1
P

(1.5)

Equation (1.5) describes a speech synthesis model i llustrated in Figure 1.9.

1.3.5DigitalCodingofAudioSignals

Indigitalaudio,thememoryrequiredtorecordas ignal,thebandwidth
requiredforsignaltransmissionandthesignalto quantisationnoiseratio
arealldirectlyproportionaltothenumberofbits persample.Theobjective
inthedesignofacoderistoachievehighfidelit ywithasfewbitsper
sampleaspossible,atanaffordableimplementation cost.Audiosignal
codingschemesutilisethestatisticalstructureso fthesignal,andamodelof
thesignalgeneration,togetherwithinformationon thepsychoacousticsand
themaskingeffectsofhearing.Ingeneral,therea retwomaincategoriesof
audiocoders:model-basedcoders,usedforlowbit ratespeechcodingin

z
1
z
1
z
1 . . .
u(m)
x(m-1) x(m-2) x(mP)
a
a
2
a
1
x(m)
G
e(m)
P
Figure1.9 Illustrationofasignalgeneratedbyanall-pole,linearprediction
model.

ApplicationsofDigitalSignalProcessing 13

applicationssuchascellulartelephony;andtransf orm-basedcodersusedin
highquality coding of speech and digital hi-fi aud io.
Figure1.10showsasimplifiedblockdiagramconfig urationofaspeech
codersynthesiserofthetypeusedindigitalcellu lartelephone.Thespeech
signalismodelledastheoutputofafilterexcite dbyarandomsignal.The
randomexcitationmodelstheairexhaledthroughth elung,andthefilter
modelsthevibrationsoftheglottalcordsandthe vocaltract.Atthe
transmitter,speechissegmentedintoblocksofabo ut30mslongduring
whichspeechparameterscanbeassumedtobestatio nary.Eachblockof
speechsamplesisanalysedtoextractandtransmit asetofexcitationand
filterparametersthatcanbeusedto synthesis the speech. At the receiver, the
model parameters and the excitation are used to rec onstruct the speech.
A transform-based coder is shown in Figure 1.11. T he aim of
transformationistoconvertthesignalintoaform whereitlendsitselftoa
moreconvenientandusefulinterpretationandmanip ulation.InFigure1.11
theinputsignalistransformedtothefrequencydo mainusingafilterbank,
oradiscreteFouriertransform,oradiscretecosi netransform.Threemain
advantages of coding a signal in the frequency doma in are:

(a)Thefrequencyspectrumofasignalhasarelati velywelldefined
structure, for example most of the signal power is usually
concentrated in the lower regions of the spectrum.
Synthesiser
coefficients
Excitation e(m)
Speech x(m)
Scalar
quantiser
Vector
quantiser
Model-based
speech analysis
(a) Source coder
(b) Source decoder
Pitch and vocal-tract
coefficients
Excitationaddress
Excitation
codebook
Pitchfilter
Vocal-tract filter
Reconstructed
speech
Pitchcoefficients
Vocal-tract coefficients
Excitation
address

Figure1.10 Blockdiagramconfigurationofamodel-basedspeechcoder.
14 Introduction

(b) Arelativelylowamplitudefrequencywouldbemaske dinthenear
vicinityofalargeamplitudefrequencyandcan the refore be coarsely
encoded without any audible degradation.
(c) The frequency samples are orthogonal and can be cod ed
independently with different precisions.

Thenumberofbitsassignedtoeachfrequencyofa signalisavariable
thatreflectsthecontributionofthatfrequencyto thereproductionofa
perceptuallyhighqualitysignal.Inanadaptiveco der,theallocationofbits
todifferentfrequencies is made to vary with the t ime variations of the power
spectrum of the signal.

1.3.6DetectionofSignalsinNoise

Inthedetectionofsignalsinnoise,theaimisto determine if the observation

consistsofnoisealone,orifitcontainsasignal .Thenoisyobservation
y(m) can be modelled as

y(m) = b(m)x(m) + n(m) (1.6)

wherex (m)isthesignaltobedetected, n(m)isthenoiseand b(m) isa

binary-valuedstateindicatorsequencesuchthat b(m) =1indicatesthe
presenceofthesignal x(m) and b(m) = 0 indicatesthatthesignalisabsent.
Ifthesignal x(m) hasaknownshape,thenacorrelatororamatched filter
.
.
.
x(0)
x(1)
x(2)
x(N-1)
.
.
.
X(0)
X(1)
X(2)
X(N-1)
.
.
.
.
.
.
X(0)
X(1)
X(2)
X(N-1)
Inputsignal BinarycodedsignalReconstructed
signal
x(0)
x(1)
x(2)
x(N-1)
^
^
^
^
^
^
^
^
n
0
bps
n
1
bps
n
2
bps
n
N-1
bps
T
r
a
n
s
f
o
r
m

T
E
n
c
o
d
e
r
D
e
c
o
d
e
r
.
.
.
I
n
v
e
r
s
e

T
r
a
n
s
f
o
r
m

T
-
1

Figure1.11 Illustrationofatransform-basedcoder.

ApplicationsofDigitalSignalProcessing 15

canbeusedtodetectthesignalasshowninFigure 1.12.Theimpulse
response h(m)ofthematchedfilterfordetectionofasignal x(m) isthe
time-reversed version of x(m) given by

1 0 ) 1 ( ) ( = N m m N x m h (1.7)

whereN is the length of x(m) . The output of the matched filter is given by

=
=
1
0
) ( ) ( ) (
N
m
m y k m h m z (1.8)

The matched filter output iscomparedwithathresh oldandabinary

decision is made as

=
otherwise 0
threshold ) ( if 1
) (

m z
m b (1.9)

where

b(m)isanestimateofthebinarystateindicatorsequ ence b(m),and

itmaybeerroneousinparticularifthesignalto noiseratio is low. Table1.1
listsfourpossibleoutcomesthattogether b(m)anditsestimate

b(m) can
assume.Thechoiceofthethresholdlevelaffectst hesensitivityofthe
Matchedfilter
h(m) =x (N 1m)
y(m)=x(m)+n(m)
z(m)
Threshold
comparator
b(m)
^
Figure1.12 Configurationofamatchedfilterfollowedbyathresholdcomparatorfor
detectionofsignalsinnoise.

b(m)
b(m) Detector decision
0 0 Signal absent Correct
0 1 Signal absent ( Missed)
1 0 Signal present ( False alarm )
1 1 Signal present Correct

Table1.1 Fourpossibleoutcomesinasignaldetectionproblem.

16 Introduction

detector.Thehigherthethreshold,thelesstheli kelihoodthatnoisewould
beclassifiedassignal,sothefalsealarmratefa lls,buttheprobabilityof
misclassification of signal as noise increases. The risk in choosing a
threshold value can be expressed as

( ) ) ( ) ( Threshold
Miss Alarm False
P P + = = R (1.10)

Thechoiceofthethresholdreflectsatrade-offbe tweenthemisclassification
rateP
Miss
() and the false alarm rate P
False Alarm
().

1.3.7DirectionalReceptionofWaves:Beam-forming

Beam-formingisthespatialprocessingofplanewav esreceivedbyanarray
ofsensorssuchthatthewavesincidentatapartic ularspatialangleare
passedthrough,whereasthosearrivingfromotherd irectionsareattenuated.
Beam-formingisusedinradarandsonarsignalproc essing(Figure1.13)to
steerthereceptionofsignalstowardsadesireddi rection,andinspeech
processing for reducing the effects of ambient nois e.
Toexplaintheprocessofbeam-formingconsidera uniformlineararray
ofsensorsasillustratedinFigure1.14.Theterm linear array impliesthat
thearrayofsensorsisspatiallyarrangedinastr aightlineandwithequal
spacing dbetweenthesensors.Considerasinusoidalfarfie ldplanewave
withafrequency F
0
propagatingtowardsthesensorsatanincidencean gle
ofasillustratedinFigure1.14.Thearrayofsensor ssamplestheincoming

Figure1.13 Sonar:detectionofobjectsusingtheintensityandtimedelayof
reflectedsoundwaves.
ApplicationsofDigitalSignalProcessing 17

waveasitpropagatesinspace.Thetimedelayfor thewavetotravela
distance of d between two adjacent sensors is given by

=
d sin
c
(1.11)

where cisthespeedofpropagationofthewaveinthemed ium.Thephase

difference corresponding to a delay of is given by

c
d
F
T

sin
2 2
0
0
= = (1.12)

whereT
0
istheperiodofthesinewave.Byinsertingappro priatecorrective

W
N1,P1
W
N1,1 W
N1,0
+

0
1
N-1
Array of sensors
I
n
c
i
d
e
n
t

p
l
a
n
e

w
a
v
e
Array offilters
Output
.
.
.
.
.
.
. ..
W
2,P1
W
2,1
W
2,0
+
.. .
z
1
W
1,P1
W
1,1 W
1,0
+
. ..
d

dsin

z
1
z
1
z
1
z
1
z
1

Figure1.14 Illustrationofabeam-former,fordirectionalreceptionofsignals.
18 Introduction

timedelaysinthepathofthesamplesateachsens or,andthenaveragingthe
outputsofthesensors,thesignalsarrivingfromt hedirection willbetime-
aligned and coherently combined, whereas those arri ving from other
directionswillsuffercancellationsandattenuatio ns.Figure1.14illustratesa
beam-formerasanarrayofdigitalfiltersarranged inspace.Thefilterarray
actsasatwodimensionalspacetimesignalprocess ingsystem.Thespace
filteringallowsthebeam-formertobesteeredtowa rdsadesireddirection,
forexampletowardsthedirectionalongwhichthei ncomingsignalhasthe
maximumintensity.Thephaseofeachfiltercontrol sthetimedelay,and can
beadjustedtocoherentlycombinethesignals.The magnitudefrequency
response of each filter can be used to remove the o utofband noise.

1.3.8DolbyNoiseReduction

Dolbynoisereductionsystemsworkbyboostingthe energy and the signal to

noiseratioofthehighfrequencyspectrumofaudio signals.Theenergyof
audio signals is mostly concentratedinthelowfre quencypartofthe
spectrum(below2kHz).Thehigherfrequenciesthat conveyqualityand
sensationhaverelativelylowenergy,andcanbede gradedevenbyalow
amountofnoise.Forexamplewhenasignalisrecor dedonamagnetictape,
thetapehissnoiseaffectsthequalityofthere cordedsignal.Onplayback,
thehigherfrequencypartofanaudiosignalrecord edonatapehavesmaller
signaltonoiseratiothanthelowfrequencyparts. Thereforenoiseathigh
frequenciesismoreaudibleandlessmaskedbythe signalenergy.Dolby
noisereductionsystemsbroadlyworkontheprincip leofemphasisingand
boostingthelowenergyofthehighfrequencysigna lcomponentspriorto
recordingthesignal.Whenasignalisrecordedit isprocessedandencoded
using a combination of a pre-emphasis filter and dy namic range
compression.Atplayback,thesignalisrecoveredu singadecoderbasedon
acombinationofade-emphasisfilterandadecompr essioncircuit.The
encoderanddecodermustbewellmatchedandcancel outeachotherin
order to avoid processing distortion.
Dolbyhasdevelopedanumberofnoisereductionsy stemsdesignated
DolbyA,DolbyBandDolbyC.Thesediffermainlyi nthenumberofbands
andthepre-emphasisstrategythatthattheyemploy .DolbyA,developedfor
professionaluse,dividesthesignalspectruminto fourfrequencybands:
band 1 is low-pass and covers 0 Hz to 80 Hz; band 2 is band-pass and covers
80Hzto3kHz;band3ishigh-passandcoversabov e3kHz;andband4is
alsohigh-passandcoversabove9kHz.Atthe encod er the gain of each band
ApplicationsofDigitalSignalProcessing 19

isadaptivelyadjustedtoboostlowenergysignalc omponents.DolbyA
providesamaximumgainof10to15dBineachband ifthesignallevel
falls45dBbelowthemaximumrecordinglevel.The DolbyBandDolbyC
systemsaredesigned for consumer audio systems, an d use two bands instead
ofthefourbandsusedinDolbyA.DolbyBprovides aboostofupto10dB
whenthesignallevelislow(lessthan45dBthan themaximumreference)
and Dolby C provides a boost of up to 20 dB as illu strated in Figure1.15.

1.3.9RadarSignalProcessing:DopplerFrequencySh ift

Figure1.16showsasimplediagramofaradarsyste mthatcanbeusedto
estimatetherangeandspeedofanobjectsuchasa movingcaroraflying
aeroplane.Aradarsystemconsistsofatransceiver (transmitter/receiver) that
generatesandtransmitssinusoidalpulsesatmicrow avefrequencies.The
signaltravelswiththespeedoflightandisrefle ctedbackfromanyobjectin
itspath.Theanalysisofthereceivedechoprovide ssuchinformationas
range, speed, and acceleration. The received signal has the form
0.1
1.0 10
-35
-45
-40
-30
-25
R
e
l
a
t
i
v
e
g
a
i
n
(
d
B
)
Frequency(kHz)
Figure1.15 Illustrationofthepre-emphasisresponseofDolby-C:upto20dB
boostisprovidedwhenthesignalfalls45dBbelowmaximumrecordinglevel.

20 Introduction

]} / ) ( 2 [ cos{ ) ( ) (
0
c t r t t A t x = (1.13)

whereA (t),thetime-varyingamplitudeof the reflected wave , depends on the

positionandthecharacteristicsofthetarget, r(t)isthetime-varyingdistance
oftheobjectfromtheradarand cisthevelocityoflight.Thetime-varying
distance of the object can be expanded in a Taylor series as

L & & & & & & + + + + =

3 2
0
! 3
1
! 2
1
) ( t r t r t r r t r (1.14)

where r
0
isthedistance, r& isthevelocity, r& & istheaccelerationetc.
Approximating r(t)withthefirsttwotermsoftheTaylorseriesexp ansion
we have
t r r t r & +
0
) ( (1.15)

Substituting Equation (1.15) in Equation (1.13) yie lds

] / 2 ) / 2 cos[( ) ( ) (
0 0 0 0
c r t c r t A t x = & (1.16)

Note that the frequency of reflected wave is shifte d by an amount

c r
d
/ 2
0
& = (1.17)

ThisshiftinfrequencyisknownastheDopplerfre quency.Iftheobjectis
movingtowardstheradarthenthedistance r(t)isdecreasingwithtime, r& is
negative,andanincreaseinthefrequencyisobser ved.Converselyifthe

r=
0
.5T
c
cos(
0
t)
Cos{
0
[t-2r(t)/c]}

Figure1.16 Illustrationofaradarsystem.

SamplingandAnalogtoDigitalConversion 21

objectismovingawayfromtheradarthenthedista ncer(t) is increasing, r& is

positive,andadecreaseinthefrequencyisobserv ed.Thusthefrequency
analysisofthereflectedsignalcanrevealinforma tiononthedirectionand
speed of the object. The distance r
0
is given by

c T r = 5 . 0
0
(1.18)

whereTistheround-triptimeforthesignaltohittheo bjectandarriveback
at the radar and c is the velocity of light.

1.4SamplingandAnalogtoDigitalConversion

Adigitalsignalisasequenceofrealvaluedorco mplexvaluednumbers,
representingthefluctuationsofaninformationbea ringquantitywithtime,
spaceorsomeothervariable.The basic elementarydiscrete-timesignalis
the unit-sample signal (m)defined as

(m) =
1 m = 0
0 m 0



(1.19)

wheremisthediscretetimeindex.Adigitalsignal x(m)canbeexpressedas
the sum of a number of amplitude-scaled and time-sh ifted unit samples as

x(m) = x(k)(m k)
k=

(1.20)

Figure1.17illustratesadiscrete-timesignal.Man yrandomprocesses,such
asspeech,music,radarandsonargeneratesignals thatarecontinuousin
Discretetime
m
Figure1.17 Adiscrete-timesignalanditsenvelopeofvariationwithtime.

22 Introduction

timeandcontinuousinamplitude.Continuoussignal saretermedanalog
becausetheirfluctuationswithtimeareanalogous tothevariationsofthe
signalsource.Fordigitalprocessing,analogsigna lsaresampled,andeach
sampleisconvertedintoan n-bitdigit.Thedigitisationprocessshouldbe
performedsuchthattheoriginalsignalcanbereco veredfromitsdigital
versionwithnolossofinformation,andwithashi ghafidelityasisrequired
inanapplication.Figure1.18illustratesablock diagramconfigurationofa
digitalsignalprocessorwithananaloginput.The low-passfilterremoves
outofbandsignalfrequenciesaboveapre-selected range.Thesample
andhold (S/H) unit periodically samples the signal to convert the
continuous-time signal into a discrete-time signal.
The analogtodigital converter (ADC) maps each co ntinuous
amplitudesampleintoan n-bitdigit.Afterprocessing,thedigitaloutputof
theprocessorcanbeconvertedback into an analog signal using a digitalto
analog converter (DAC) and a low-pass filter as ill ustrated in Figure 1.18.

1.4.1Time-DomainSamplingandReconstructionofAn alog
Signals

Theconversionofananalogsignaltoasequenceof n-bitdigitsconsistsof
twobasicstepsofsamplingandquantisation.Thes amplingprocess,when
performedwithsufficientlyhighspeed,cancapture thefastestfluctuations
ofthesignal,andcan be a loss-less operation in that the analog signal can be
recoveredthroughinterpolationofthesampledsequ enceasdescribedin
Chapter10.Thequantisationofeachsampleintoan n-bitdigit,involves
some irrevocable error and possible loss of informa tion. However, in
practice the quantisation error can be made negligi ble by using an
appropriatelyhighnumberofbitsasinadigitala udiohi-fi.Asampled
signalcanbemodelledastheproductofacontinuo us-timesignal x(t)anda
periodic impulse train p(t)as

Analog input
y(t)
LPF &
S/H
ADC
DAC
LPF
y(m) x(m) x(t)
Digital signal
processor
x
a
(m) y
a
(m)

Figure1.18 Configurationofadigitalsignalprocessingsystem.
Sampling and AnalogtoDigital
Conversion
23

=
=
=
m
s
mT t t x
t p t x t x
) ( ) (
) ( ) ( ) (
sampled

(1.21)

where T
s
isthesamplingintervalandthesamplingfunction p(t)isdefined
as

p(t) = (t mT
s
)
m=

(1.22)

Thespectrum P( f )ofthesamplingfunction p(t)isalsoaperiodicimpulse

train given by

=
=
k
s
kF f f P ) ( ) ( (1.23)

whereF
s
=1/T
s
isthesamplingfrequency.Sincemultiplicationof twotime-
domainsignalsisequivalenttotheconvolutionof theirfrequencyspectra
we have

=
= = =
k
s
kF f f P f X t p t x FT f X ) ( ) ( * ) ( )] ( ). ( [ ) (
sampled
(1.24)

wheretheoperator FT[.] denotestheFouriertransform.InEquation(1.24)

theconvolutionofasignalspectrum X( f )witheachimpulse ) (
s
kF f ,
shifts X( f ) andcentresiton kF
s
. Hence,asexpressedinEquation(1.24),
thesamplingofasignalx (t)resultsinaperiodicrepetitionofitsspectrum
X( f ) centred on frequencies K , 2 , , 0
s s
F F . When the sampling
frequencyishigherthan twice the maximum frequenc y content of the signal,
thentherepetitionsofthesignalspectraaresepa ratedasshowninFigure
1.19.Inthiscase,theanalogsignalcanberecove redbypassingthesampled
signalthroughananaloglow-passfilterwithacut -offfrequencyof F
s
.Ifthe
samplingfrequencyislessthan2 F
s
,thentheadjacentrepetitionsofthe
spectrum overlap and the original spectrum cannot b e recovered. The
distortion,duetoaninsufficientlyhighsampling rate,isirrevocableandis
24 Introduction

knownas aliasing.Thisobservationisthebasisofthe Nyquistsampling

theoremwhichstates: aband-limitedcontinuous-timesignal,withahighe st
frequencycontent(bandwidth)of BHz,canberecoveredfromitssamples
provided that the sampling speed F
s
>2B samples per second.
Inpracticesamplingisachievedusinganelectroni cswitchthatallowsa
capacitortochargeupordowntothelevelofthe inputvoltageonceevery
T
s
secondsasillustratedinFigure1.20.Thesample- and-holdsignalcanbe
modelledastheoutputofafilterwitharectangul arimpulseresponse,and
Time domain
Frequency domain
Impulse-train-sampling
function
Sample-and-holdfunction
x(t)
t
X(f)
f
f
F
s
F
s
=1/Ts
0
Ts
x
p
(t)
Xp( f )
sh(t)
SH( f )
X
sh
(t) |X( f )|
f
f
f
t
t
S/H-sampled signal
Impulse-train-sampled
signal
B B
2B
...
...
0
*
= =
Ts 0
*

=
=
0
t
t
F
s
/2
...
. ..
.. .
. .. . ..
... ...
F
s
/2
F
s
F
s

=
=
k
s
kF f f P ) ( ) (
0
F
s
/2 F
s
/2

Figure1.19 Sample-and-Holdsignalmodelledasimpulse-trainsamplingfollowed
byconvolutionwitharectangularpulse.

Sampling and AnalogtoDigital

Conversion
25

with the impulsetrainsampled signal as the input as illustrated in

Figure1.19.

1.4.2Quantisation

For digital signal processing, continuous-amplitude samples from the

sample-and-holdarequantisedandmappedinto n-bitbinarydigits.For
quantisationto nbits,theamplituderangeofthesignalisdivided into2
n

discretelevels,andeachsampleisquantisedtoth enearestquantisation
level,andthenmappedtothebinarycodeassigned tothatlevel.Figure1.21
illustratesthequantisationofasignalinto4 dis crete levels. Quantisation is a
many-to-onemapping,inthatallthevaluesthatfa llwithinthe continuum of
aquantisationbandaremappedtothecentreofthe band.Themapping
betweenananalogsample x
a
(m)anditsquantisedvalue x (m) canbe
expressed as

[ ] ) ( ) ( m x Q m x
a
= (1.25)

whereQ[] is the quantising function.

Theperformanceofaquantiserismeasuredbysign altoquantisation
noise ratio SQNR per bit. The quantisation noise is defined as

) ( ) ( ) ( m x m x m e
a
= (1.26)

C
R
2
R
1
x(t)
x(mT
s
)
T
s

Figure1.20 Asimplifiedsample-and-holdcircuitdiagram.

26 Introduction

Nowconsideran n-bitquantiserwithanamplituderangeof Vvolts.The

quantisationstepsizeis =2V/2
n
.Assumingthatthequantisationnoiseisa
zero-meanuniformprocesswithanamplituderangeo f/2wecanexpress
the noise power as
[ ] ( )
3
2
12
) ( ) (
1
) ( ) ( ) ( ) (
2 2 2
2 /
2 /
2
2 /
2 /
2 2
n

E
V
m de m e

m de m e m e f m e

= =
= =

E
(1.27)

where f
E
e(m)
( )
=1/ is the uniform probability density function of the
noise. Using Equation (1.27) he signaltoquantisat ion noise ratio is given
by
n
P
V
V
P
m
m x
n SQNR
n
n
e
6 77 . 4
2 log 10 log 10 3 log 10
3 / 2
log 10
) (
) (
log 10 ) (
2
10
Signal
2
10 10
2 2
Signal
10
2
2
10
] [
] [
+ =
+
|
|

\
|
=
|
|

\
|
=
|
|

\
|
=

E
E
(1.28)

where P
signal
isthemeansignalpower,and istheratioindecibelsofthe
peaksignalpower V
2
tothemeansignalpower P
signal
.Therefore,from
Equation(1.28)everyadditionalbitinananalogt odigitalconverterresults
in 6 dB improvement in signaltoquantisation noise ratio.
Sampling and AnalogtoDigital
Conversion
27

Bibliography

ALEXANDER S.T. (1986) Adaptive Signal Processing Theory and

Applications. Springer-Verlag, New York.
DAVENPORTW.B.andR OOTW.L.(1958)AnIntroductiontotheTheoryof
Random Signals and Noise. McGraw-Hill, New York.
EPHRAIMY.(1992)StatisticalModelBasedSpeechEnhanceme ntSystems.
Proc. IEEE, 80,10, pp. 15261555.
GAUSSK.G.(1963)TheoryofMotionofHeavenlyBodies.D over,New
York.
GALLAGERR.G.(1968)InformationTheoryandReliableCommun ication.
Wiley, New York.
HAYKINS.(1991)AdaptiveFilterTheory.Prentice-Hall,E nglewoodCliffs,
NJ.
HAYKINS.(1985)ArraySignal Processing. Prentice-Hall, Englewood Cliffs,
NJ.
KAILATH T. (1980) Linear Systems. Prentice Hall, Englewood Cliffs, NJ.
KALMAN R.E.(1960)ANewApproachtoLinearFilteringand Prediction
Problems.Trans.oftheASME,SeriesD,Journalof BasicEngineering,
82, pp. 3545.
KAYS.M.(1993)FundamentalsofStatisticalSignalPro cessing,Estimation
00
01
10
11
+
x(mT)
0

2
+2
2V
Continuousamplitude samples
Discreteamplitude samples
+V
V
Figure1.21 Offset-binaryscalarquantisation

28 Introduction

Theory. Prentice-Hall, Englewood Cliffs, NJ.
LIM J.S. (1983) Speech Enhancement. Prentice Hall, Eng lewood Cliffs, NJ.
LUCKY R.W., S ALZ J. and W ELDON E.J. (1968) Principles of Data
Communications. McGraw-Hill, New York.
KUNG S.Y. (1993) Digital Neural Networks. Prentice-Hall , Englewood
Cliffs, NJ.
MARPLES.L.(1987)DigitalSpectralAnalysiswithApplica tions.Prentice-
Hall, Englewood Cliffs, NJ.
OPPENHEIM A.V. and S CHAFER R.W. (1989) Discrete-Time Signal
Processing. Prentice-Hall, Englewood Cliffs, NJ.
PROAKISJ.G.,R ADERC.M.,L INGF.andN IKIASC.L.(1992)Advanced
Signal Processing. Macmillan, New York.
RABINERL.R.andG OLDB.(1975)TheoryandApplicationsofDigital
Processing. Prentice-Hall, Englewood Cliffs, NJ.
RABINERL.R.andS CHAFERR.W.(1978)DigitalProcessingofSpeech
Signals. Prentice-Hall, Englewood Cliffs, NJ.
SCHARFL.L.(1991)Statistical Signal Processing: Detecti on, Estimation, and
Time Series Analysis. Addison Wesley, Reading, MA.
THERRIENC.W.(1992)DiscreteRandomSignalsandStatistica lSignal
Processing. Prentice-Hall, Englewood Cliffs, NJ.
VAN-TREESH.L.(1971) Detection, Estimation and Modulation Th eory. Parts
I, II and III. Wiley New York.
SHANNONC.E.(1948)AMathematicalTheoryofCommunication .Bell
Systems Tech. J., 27, pp. 379423, 623656.
WILSKYA.S.(1979)DigitalSignalProcessing,Controland Estimation
Theory: Points of Tangency, Areas of Intersection a nd Parallel
Directions. MIT Press, Cambridge, MA.
WIDROWB.(1975)AdaptiveNoiseCancelling:Principlesan dApplications.
Proc. IEEE, 63, pp. 1692-1716.
WIENERN.(1948)Extrapolation,InterpolationandSmoothi ngofStationary
Time Series. MIT Press, Cambridge, MA.
WIENER N. (1949) Cybernetics. MIT Press, Cambridge, MA.
ZADEHL.A.andD ESOERC.A.(1963)LinearSystemTheory:TheState-
Space Approach. McGraw-Hill, NewYork.