Introduction To DSPXilinx

The DSP Primer 4
Introduction to DSP
Return Return
DSPprimer Home DSPprimer Notes
August 2005, University of Strathclyde, Scotland, UK For Academic Use Only

THIS SLIDE IS BLANK
Top
The DSP “Revolution” 4.1
• 1980s: Special-purpose DSP microprocessors introduced.
• 1990s: Processing power of DSP (micro-)processors has increased

by an order of magnitude, and price decreased.
• The (digital) reliability, repeatability, and programmability of DSP has

widely displaced analogue systems in both consumer and industrial
markets.
1000 DSP enabled 3G Mobile

(MIPS millions instructions/sec)
consumer Multimedia
Processing Power
applications Teleconferencing
100
Digital Video
Digital Mobile Comms
Multimedia
10
Computing
Biomedical DSP
Digital Telephony CD-Audio
1
1970 1980 1990 2000 time (years)
August 2005, For Academic Use Only, All Rights Reserved

Notes:
Digital Audio: Early 1980s DSP systems such as CD-audio did not require high level of processing power. Data
was mainly be read from the CD and output via a digital to analogue converter (DAC). More recently CD Audio
systems have been equipped with sound effects systems, recording capabilities and so on; however processing
requirements are still relatively low for such operations. The Modem: In the 1990s the computer “fax-modem”
became ubiquitous and from 1990 the speed of modems increased from a nominal 2400 bits/sec to a standard
of 57200 bits/sec by the end of the decade. This communication was over the same copper wires that have
existed in the ground for some 50 or more years; so why the order of magnitude increase in speed? The answer
is again, DSP, or more specifically adaptive DSP algorithms in the form of the LMS (least mean squares). Via
adaptive DSP enabled echo cancellation and data equalisation, bandlimited telephone channels were used with
various signalling methods to reach data speeds approaching the theoretical limit by essentially using DSP to
correct for any distortion that may have been present in the channel.
Digital Subscriber Loops: The last couple of years of the 1990s saw the introduction of DSL (digital subscriber
loops) which bring data rates of millions (M) bits/sec to the home over conventional telephone lines. In summary
DSL is using DSP techniques to allow the high frequency portions of a copper wire twisted pair (MHz) to be
utilised. Prior to the requirements for fast data communications, most telephone lines were bandlimited between
300-3400Hz - sufficient for voice, but rather limiting for data communications - this is the reason that “voiceband”
modems have a clearly calculable limit based on the available teleco bandwidth of a few kHz. The introduction
of DSL equipment to telephone exchanges will bring a new lease of life to the copper wire infrastructure.
The application of mobile multimedia will allow consumers to communicate via teleconferencing (audio and
video), and transmit documents (email/fax) from a small hand held communicator. Very high levels of
processing power will be required for the audio/video coding/compression algorithms and for the DSP
communications strategies. Third generation mobile communications (3G) will allow data rates of up to 2Mbits/
sec to be available from hand held wireless devices (laptop/mobile communicator). To achieve these high rates
a “chip” rate of some 5Mbits/sec to and from the basetation will be used. In order to process this speed of data
transfer, a modulation scheme called CDMA (code division multiple access) will be used. DSP strategies for
pulse shaping, channel equalisation, echo control, speech compression and so on, will require very powerful
DSP devices capable of 1000’s millions of instructions per second will be required. 2002 is the roll-out year....
Top
DSP in 2000’s - Towards Software Radio 4.2
• The ideal DSP enabled Software Radio receiver directly convert to/from
RF frequencies (typically GHz, (G = 109).
• Initial implementations will work with IF (mixed down from GHz to a few
MHz) frequencies.
f s > 2f RF
RF Antenna @ GHz
Output
Anti- ADC
Alias
Video,
Audio,
f s > 2f RF DSP & Data
Input
DAC
Video,
Linear RF Amplifier Audio,
& Data

Notes:
For the received signal, all downconversion/demodulation is done after the wideband ADC sampling (so called
zero-IF or homdyne receiver). As in current 2nd generation mobile, all baseband processing (echo cancelling,
speech coding, equalisation, despreading, channel decoding) is done in the DSP processor.
For current mobile 2nd generation technologies, a desirable software radio would be one that could operate in
the 800 to 900MHz spectrum and can adapt between AMPs, GSM, DAMPs, CT2 and so on by simply changed
the software on the DSP processor.
The next generation of UMTS, although being driven by cooperating worldwide standards is likely to be a major
beneficiary from the implementation of software radio strategies. Currently multiband radios are available
(particularly in the USA) that cover both GSM at 900MHz and DECT at 1300MHz, although this is essentially
accomplished by having separate hardware rather than as a programmable SR architecture.
For next generation CDMA with 5MHz bandwidth if the ADC used 4 times bandpass sampling, then more than
20Msamples per second are required, with an analogue front end capable of integrating at a frequency range
extending as high as 2GHz. At least 18 bits of resolution is likely to be required. Currently such devices are not
available. The DAC would require to produce more than 20Msamples/sec at up to 18 bits, and input to a linear
RF amplifier.
Some of the ideas for software radio include algorithms toolbox’s, where for example there is separate software
for each standard stored within a handset or mobile terminal. Alternatively there may be a toolbox of algorithms
(such as QPSK, equalisers etc) which can be called with appropriate parameters depending on the actual
standard (GSM, W-CDMA etc) being used.
There is also the concept of over-air download whereby a mobile terminal can download required software to
the generic hardware of the mobile terminal. This, however, is a concept that is still a long time away.
Top
Amplification and Conditioning 4.3
• The voltage from a signal sensor is very small in magnitude.

A microphone may produce voltages of the order of 10-6 volts.
Similarly for ECG sensors, vibration sensors etc.
• Prior to recording the signal or reproducing with an actuator an amplifier

should signal condition by linearly amplifying the signal by an
appropriate factor.
Vin Vout 1 Volt
10-3 Volts
Voltage
Voltage
60dB
time time
Amplifier x 1000
• The above amplifier adds 60dB of gain (20log101000 = 60)
V out
V out = 1000V in therefore ----------- = 1000
V in

Notes:
A system is said to be linear if the output can formed as the convolution of the input and system transfer function
x(t) Linear y(t)

System
time time
In general for a linear system y ( t ) = f ( x ( t ) ) , if,
y1 ( t ) = f [ x1 ( t ) ]
then by superposition: y 1 ( t ) + y 2 ( t ) = f [ x 1 ( t ) + x 2 ( t ) ]
y2 ( t ) = f [ x2 ( t ) ]
One of the simplest ways to test the linearity of a system is to input a pure tone sine wave, and if, at all
frequencies, the output is only a pure tone at the signal input frequency (i.e. no harmonics are produced), then
the system is truely linear.
Amplification is often presented as a logarithmic measure of the power amplification ratio ( P out ⁄ P in ) given the
large linear dymanic range. Recall that P ∝ V 2 , then:
A dB = 10log 10 ( P out ⁄ P in ) = 10log 10 ( V out ⁄ V in ) 2 = 20log 10 ( V out ⁄ V in ) decibels (dB)
Therefore if an amplification A = 1000 , then the power amplification is 60 dB. Similary an attenuation of a factor
of 1000, (or a gain of 0.001) corresponds to -60dB of gain, or 60dB of attenuation! The placement of the -ve
sign needs some care given the antonymity of the words gain and attenuation.
Top
Amplifier Distortion 4.4
• An amplifier which introduces unwanted artifacts, is said to be non-

linear and is, of course, very undesirable as it may mask signal
components of interest.
Vin Vout Distortion
-3
10 Volts
Voltage
Voltage
60dB
time time
Amplifier x 1000
• The above amplifier is non-linear and actually outputs the input signal
plus a 3rd order harmonic:
V out = ( 1000 x V in ) + 10 x ( V in ) 3
• Unlike noise it is essentially impossible to remove the effects of

distortion. Therefore we try to minimize the possibility of distortion by
using suitable components.

Notes:
Unlike in the above example, it is of course very unlikely that you will ever actually known the true non-linear
equation of an amplifier. If you do, then you might be in a position to remove/address some of this non-linearity.
However even the simplest non-linearity can be very difficult to remove. As an example, consider a system that
adds a 2nd order harmonic according to the equation:
2
V out = V in + 0.001V in
Try to solve this equation for V in as a function of V out . ....not easy, and in fact there is not a unique solution.
(Consider even the a simple squared equation:
V a = V b2 , then V b = ± V a which is a non-unique solution.
Non-linearities can cause serious problems in DSP by losing or masking desirable signal components.
It is probably true to say that every amplifier is non-linear to some extent. However if the power levels of the
non-linear signal components are very low, then for pragmatic implementation purposes the amplifier can be
considered linear; (the definition of “very low” of course depends on your application). In the above example
where the non-linear second harmonic component has 1/5 of the voltage (1/25th of the power) of the
fundamental, this is very high and the amplifier is at best “very” non-linear. Systems may often be classed as
“weakly non-linear”, “moderately non-linear”, or “strong non-linear”.
See system c:\DSPedia\intro\non-linear_amplifer.svu for implementation of the above slide and

see c:\DSPedia\intro\non-linear_amplifer_speech.svu for a speech example.
Top
Signals and Noise 4.5
• Most acquired signals are corrupted by some level of noise which can
cause information to be lost.
• Signal processing techniques are often used in an attempt to remove

or attenuate noise.
• Most noise can be considered as additive (linear superposition) which

can be address by linear filtering techniques.
Voltage
time
Voltage
Signal
time
Voltage
Microphone
Signal + Noise
time
Noise

Notes:
One of the key strategies for DSP is too remove/attenuate noise from a signal of interest. There are some
situations where it is very easy to remove the noise, for example in cases where the the signal and the noise
are very disimilar with respect to some signal feature. If speech is corrupted by, say, a low rumbling single
frequency, then removal is straightforward. However in situations where the signal of interest is very similar to
the noise then things are less straightforward. One example where this type of problem occurs is the “cocktail
party” noise effect. If a speech signal is corrupted by other speech signals, then extracting the desired signal
can be very difficult. iIn general removing the noise from a signal requires that we know some fundamental
characteristics about the signal and noise, i.e. the frequency range, typical power levels etc.
We should try to be clear about the difference between noise and distortion. Noise is usually the name given to
an interfering signal, and which is usually in most cases additive noise. Using signal processing techniques we
can address the effects of this noise and perhaps try and attenuate the noise using linear filtering or other
techniques in order to improve the signal to noise ratio. Distortion on the other hand is caused by some non-
linear process occuring in the signal acquisition or processing chain, and usually there is nothing that can be
done to address distortion after it has occured.
Additive Noise: Consider an acoustic signal source s ( t ) received by a microphone which is corrupted by a
nearby acoustic noise source n ( t ) . The composite signal recorded at the receiver is denoted y ( t ) . The simplest
form of “additive noise” that could occur would be y ( t ) = s ( t ) + n ( t ) A slightly more complex version of additive
noise is y ( t ) = s ( t ) + An ( t ) where the noise is attenuated by a factor of A perhaps because of the acoustic
propogation path. Even more realistically the additive noise may take the form y ( t ) = s ( t ) + An ( t ) + Bn ( t – t 0 )
given that the noise may arrive at the receiver via a number of paths. And in its more general form, the received
signal will take the form:
∞
y ( t ) = s ( t ) + ∫ n ( t )h ( t – τ ) dt
0 where h ( t ) is the “impulse response” of the acoustic path from noise source to
receiver. Therefore although we may know exactly what the noise source is emitting, to “remove” this noise from
y ( t ) also requires that we have “information” about the acoustic transfer path.
Top
The Noise/Distortion Chain.... 4.6
• Consider the various levels of noise and distortion added in a digital

mobile communications link:
Environment noise Digital quant- Digital compression Modulation

(e.g. car engine) isation distortion distortion/artifacts Distortion
Micro- Analogue to DSP Speech Modulation/

phone Digital Converter Compression Transmission
Atmospheric
noise/ Multipath/
Doppler/ Other
users
Loud Digital to DSP Speech Reception/
Speaker Analogue Conv. Decompress Demodulation
• DSP must minimize the amount of noise/distortion input to the chain,

and where possible attenuate other sources.

Notes:
Environment Noise: Noise coming from vehicle engine, wind noise etc. We may be able to address the noise
at the microphone using DSP algorithms and techniques for linear or adaptive filtering. At the receiver, linear or
adaptive filtering or perhaps active noise control could be introduced to improve the signal to noise ratio.
Quantisation Distortion: Quantisation distortion or noise is introduced by the analogue to digital converter
(ADC). Ideally we use as many bits in the ADC converter as possible for quality, but as few bits as possible to
keep bandwidth requirements low. To attempt to improve quality for a given low number of bits we may use
psychoacoustic quantisation noise shaping, or dithering techniques.
Speech Coding/Compression Noise: To keep bandwidth requirements low we require to compress the
signal. By compressing the signal we aim to maintain the signal quality and intelligibility but will accept some
level of distortion or loss of fidelity.
Modulation: The modulation/demodulation process will introduce various levels of noise and distortion due to
the various stages of modulation and filtering that are required.
Atmospheric/Multipath Noise: When the electromagnetic mobile signal is transmitted general signal reflection
problems, interference from other users etc, will introduce a level of noise. Therefore it is desirable that the
digital coding scheme is as resistant to noise as possible.
Top
Signal to Noise Ratio 4.7
• Taking the logarithm of the linear signal power to noise power ratio
(SNR) and multiplying by 10 gives the measure of decibels or dBs.
( Signal Power ) P signal

SNR = 10 log ---------------------------------------- = 10 log ---------------
( Noise Power ) P noise
• Recalling that Power ∝ Volts 2 , then:
V 2 signal V signal
SNR = 10 log ------------------ = 20 log ---------------
V 2 noise V noise
SNR = 10dB, P signal = 10 × P noise and V signal = 10 × V noise

Very low quality telephone line
SNR = 60dB, P signal = 1000000 × P noise and V signal = 1000 × V noise
Audio cassette deck

More....

Notes:
Note there are many forms of “decibel” and although all have specific definition, it is often the case that the
meaning of dB is implied rather than explicity stated. For acoustic signals in particular there are a number of
definitions of dB A, dBm, dB SPL, or dB HL. One of the most common uses of dB is Sound Pressure Level (SPL)
is specified in decibels (dB) and is calculated as the logarithm of a ratio:
I
SPL = 10 log  -------- dB
I 
ref
where I is the sound intensity measured in Watts per square meter (W/m2) and Iref is the reference intensity of
10-12W/m2 which is (or perhaps was!) the approximate lower threshold of hearing for a tone at 1000Hz.
Alternatively (and more intuitively given the name sound “pressure” level) SPL can be expressed as a ratio of
a measured sound pressure relative to a reference pressure, P ref , of 2 × 10 – 5 N/m 2 = 20 µ Pa:
I  P2  P
SPL = 10 log  -------- = 10 log  ---------- = 20 log ---------- dB
I  P2  P ref
ref ref
Intensity is proportional to the squared pressure. i.e. I ∝ P 2 . A logarithmic measure is used for sound because
of the very large dynamic range of the human has a linear scale of more than 10 12 and because of the
logarithmic nature of hearing. Due to the nature of hearing, a 6dB increase in sound pressure level is not
necessarily perceived as twice as loud in fact far from it. For example the difference in intensity between 110dB
and 116dB seems much greater than the same difference between 40dB and 46dB. The threshold of pain is
about 120dB. (See entry for Sone.)
It is worth noting that standard atmospheric pressure is around 101300 N/m2 and the pressure exerted by a
very small insect’s legs is around 10 N/m2. Therefore the ear and other sound measuring devices are measuring
extremely small variations on pressure.
Top
Generic Analogue I/O Signal Processing 4.8
• In general an analogue signal processing system can be defined as a

system that senses a signal to produce an analogue voltage,
“process” this voltage, and reproduce the signal to its original
analogue form.
• A public address system is an example of an analogue signal

processing system:
Amplifier Amplifier
TONE ... Blue Suede
Shoes....
BASS
VOLUME ... Blue Suede
Shoes....
... Blue Suede

Shoes....
Controls
... Blue Suede
Shoes....
ANALOGUE SIGNAL PROCESSING

Notes:
Analogue systems have more than just the flexibility to amplify and filter. Before the availability of low cost, high
performance DSP processors, analogue computers were used for analysis of signals and systems. The basic
linear elements for analog computers were the summing amplifier, the integrator, and the differentiator. By the
judicious use of resistor and capacitor values, and the input of appropriate signals, analogue computers could
be used for solving differential equations, exponential and sine wave generation and the development of control
system transfer functions.
C
R
-
–1 t
V out = --------- ∫ V in dt
+
Vin
RC 0
Integrator
R
C
-
+
dV in
Vin V out = – R C ------------
dt
Differentiator
R1 Rf
V1
R2
V2 -
R3 +
Rf Rf Rf
V3
V out = ------- V 1 + ------- V 1 + ------- V 1
R1 R2 R3
Summer
Top
A Generic Input/Output DSP System 4.9
• A single input, single output DSP system has the following

components:
INPUT OUTPUT
TRANSDUCER ACTUATOR
fs fs
Anti- Recon-
Alias DSP structio
ADC Processor DAC
Filter n Filter
Signal Conditioning Signal Conditioning
Analogue Digital Analogue
Return More....

Notes:
Analogue components still play a very important part in DSP. In particular the input and output stages from and
to the real world are of course analogue. Where possible however the design requirements and specification of
the analogue components is being simplified in favour of more digital complexity, i.e. oversampling type
strategies. Many systems will now use much higher than necessary sampling rates in order that the “complexity”
of the analogue components can be reduced.
DSP systems can be classed as one of three general types:

(1) Real time Input/Output, e.g. a DSP enabled communications link;
(2) Real time Input only, e.g a speech recognition system;
(3) Real time Output only, e.g. a CD audio reproduction system.
The anti-alias filter is important in order to ensure that aliasing distortion is not introduced to the DSP system.
The reconstruction filter is important to ensure that reconstruction high frequency noise is not present in the
output signal.
In a DSP system the analogue voltage is converted from a real world signal, to a voltage, to binary numerical
values. DSP systems use binary numbers, or base 2 and usually with 2’s complement (to allow +ve and -ve
number representation.) Base 2 is used because it is easy to design electronic devices that only have two digits
corresponding to two voltage levels. Binary adders, multipliers, memory elements and so on are widely
available and together form the core element of every computer’s ability to perform high speed arithmetic. Every
DSP microprocessor/ASIC uses binary arithmetic. 2’s complement arithmetic used by most DSP processors
which allows a very convenient way of representing negative numbers, and imposes no overhead on arithmetic
operations. In two’s complement the most significant bit is given a negative weighting, e.g. for the 16 bit number.
1001 0000 0000 0001 2 = -2 15 + 2 12 + 2 1

= -32768 + 4096 + 1 = -28671
Top
Generic Analogue Communications 4.10
• For most baseband telecommunications a voltage signal is

transmitted over a cable.
Line interface Line interface
Channel
... Blue Suede
Shoes....
... Blue Suede

Shoes.... BASEBAND ANALOGUE COMMUNICATIONS
Scotland USA
• A simple example is a telephone. The acoustic signal is converted to

a voltage which is then directly transmitted over a twisted pair of wires
to be received at a remote location.
Notes:
In a telephone system the line interface is some form of driver/amplifier that is capable of converting the
analogue voltage of the speech into a signal with sufficient power to be transmitted from transmitter to receiver
(or transmitted to an intermediate point such as an exchange for subsequent switching).
Generic Analogue Radio Communication
For radio based analogue communications a modulator is required to convert from voltage to an
electromagnetic radio frequency (RF) signal.
An example of (one way) analogue communication is an FM radio station (modulation to around 100MHz) or
first generation mobile telephones which had 30kHz of bandwidth available for a speech channel.
Modulation Demodulation
Radio
Channel
... Blue Suede
Shoes....
... Blue Suede

Shoes.... RADIO ANALOGUE COMMUNICATIONS
Top
Digital Data Communications 4.11
• Modern communications systems require that digital information is

transmitted and received.
fspeech
USA
ADC
fchannel
Switch
DSP - coding Re-
DAC
& modulation const
Digital Line Interface
Data In RF Modulation
Channel
fspeech Scotland
fchannel
DAC
Switch
DSP - decoding Anti-
ADC
& demodulation Alias
Digital RF Demodulation
Data Out Line Interface

Notes:
Data communications requires that digital information (1’s and 0’s) is coded/modulated to produce an analogue
signal that can be transmitted over a cable. If the information is to be sent over radio, then a transmitter and
receiver is also required.
The digital coding/modulation may compress or coded the data in some way, then translate into QPSK or other
QAM format. A final analogue signal is then produced by the DAC which is then suitable for sending directly
over a cable, or it can be modulated by a carrier frequency (RF (radio frequency) or centre of the baseband).
The sampling rate f channel is at least twice the channel bandwidth to satisfy the Nyquist sampling criteria.
At the receive side, the analogue signal is demodulated from RF, or otherwise, and then converted to digital by
the ADC prior to the final stage of decoding/demodulation.
Note that with the generic input/output DSP system we have the signal “domain” sequence: analogue-digital-
analogue:
fs Digital fs
Analogue Anti- ADC DSP Re-

DAC Analogue
Alias const
whereas with the DSP communications system we have the signal “domain” sequence: digital-analogue-
digital:
fchannel fchannel
DSP Re- Anti- ADC DSP

ADC DAC DAC
const Alias
Digital Analogue Digital
Top
Analogue to Digital Converter (ADC) 4.12
• An ADC is a device that can convert a voltage to a binary number,

according to its specific input-output characteristic.
Binary Output
fs
127 01111111
96 01100000
8 bit 1
64 01000000 0
0
32 00100000 1
ADC
1
-2 -1 1 2 Voltage 1
-32 11001000 Voltage Input 0
Input 1
-64 11000000
Binary
-96 10100000 Output
-128 10000000
•The number of digital samples converted per second

Generic DSP is defined by the sampling rate of the converter, fs Hz.

Notes:
Viewing the straighline portion of the device e are tempted to refer to the characteristic as “linear”. However a
quick consideration clearly shows that the device is non-linear (recall the definition of a linear system from
before) as a result of the discrete (staircase) steps, and also that the device clips above and below the maximum
and minumum voltage swings. However if the step sizes are small and the number of steps large, then we are
tempted to call the device “piecewise linear over its normal operating range”.
Note that the ADC does not necessarily have a linear (straight line) characteristic. In telecoms for example a
defined standard nonlinear quantizer characteristic is often used (A-law and µ-law). Speech signals, for
example, have a very wide dynamic range: Harsh “oh” and “b” type sounds have a large amplitude, whereas
softer sounds such as “sh” have small amplitudes. If a uniform quantization scheme were used then although
the loud sounds would be represented adequately the quieter sounds may fall below the threshold of the LSB
and therefore be quantized to zero and the information lost. Therefore non-linear quantizers are used such that
the quantization level at low input levels is much smaller than for higher level signals. A-law quantizers are often
implemented by using a nonlinear circuit followed by a uniform quantizer. Two schemes are widely in use: the
A-law in Europe, and the µ -law in the USA and Japan. Similarly for the DAC can have a non-linear
characteristic..
Binary Output
Voltage Input
Top
Digital to Analogue Converter (DAC) 4.13
• A DAC is a device that can convert binary numbers to voltages,

according to its specific input-output characteristic.
Voltage Output fs
Binary
2 Input
1 8 bit
10000000
10100000
11000000
11001000
1 0
0
1 DAC
127
1 Voltage
32
64
96
1
-128
-96
-64
-32
Binary Output
00100000
01000000
01100000
01111111 0
Input
-1 1
-2
Generic DSP

Notes:
Type of ADCs and DACs
ADCs (Quantizer):
Successive Approximation ADC - Uses a DAC and comparator to determine voltage
Dual Slope ADC - Uses a capacitor connected to a reference voltage, and time taken for capacitor to charge is
counted by a digital counter.
Flash ADC - Accurately trimmed ladder of resistors.
Sigma-Delta - Oversampling single bit converter
DACs:
Multiplying DACs, accurately trimmed resistor is used to generate the output voltage via summing amplifier.
Sigma-Delta DAC - Single bit oversampled data
In this course we will specifically study different type of ADC or DAC. We will use them as functional
components.
Because of the DSP processing within sigma delta devices (oversampling and noise shaping), later in the
course we will look at sigma delta in more detail.
Top
Signal Conditioning 4.14
• Note that prior to a signal being input to an ADC, an amplifier will be

required to ensure that the full voltage range of the ADC is used - this
is referred to as signal conditioning.
• For the above ADC with a maximum input and output of 2 volts we
would require that the input signal to the ADC has a similar range:
Volts Volts
0.001 2
0.0005 1
-0.0005 time Amplifier -1 time

-0.001 x 2000
-2
•Depending on the output actuator, an amplifier, or at

Generic DSP least a buffer amplifier will be required.

Notes:
An analog signal with a magnitude larger than the upper and lower bounds± V maxof an ADC chip, will be clipped.
Any voltage aboveV max will be clipped and the information lost. Clipping effects frequently occur in amplifiers
when the amplification of the input signal results in a value greater than the power rail voltages.
Vmax
ADC
time time
-Vmax
Digital Output
-Vmax Vmax Vin
The DSP designer must ensure that for the particular application, clipping does not in general occur.
Top
Sampling 4.15
• The speed at which an ADC generates binary numbers is called the

sampling rate or sampling frequency, f s .
• The time between samples is called the sampling period, t s :
1-
ts = ---
fs
• Sampling frequency is quoted in samples per second, or simply as

Hertz (Hz).
• The actual sampling rate will depend on parameters of the application.

This may vary from:
10’s of Hz for control systems,

100’s of Hz for biomedical,
1000’s of Hz for audio applications,
Generic DSP
1,000,000’s of Hz for digital radio front ends.

Notes:
Note that the ADC may or may not have a zero output voltage level. For example an ADC/quantizer could have
a mid-tread or a mid-rise characteristic. Consider a 3 bit converter for the mid-tread and the mid-riser:
Binary Binary 011

011
010
010
001
Mid-tread 001 Mid-rise
000
000
-4q -3q -2q -q 0 q 2q 3q 4q Voltage -4q -3q -2q -q 0 q 2q 3q 4q Voltage
111
111
110
110
101
101
100
100
Note that the mid-riser does not have a zero output level, whereas the mid-tread does have a zero output level
but there are more levels above the zero level than there are below, this being a feature of two’s complement
arithmetic.
For an ADC with a very low number of bits, the differences between the mid-tread and mid-rise quantizer may
be noticeable, particularly in terms of the perceived quantisation noise. For example if a very small sine wave
of amplitude q/10 was input to the above mid-rise quantiser, then the output would always be zero, 000.
However inputting the same waveform to the midtread would produce a square wave of levels 000 and 111 at
the same frequency as the sine wave. Hence the mid-tread registers something of the input signal, but the mid-
rise has not!
Top
Sampling an Analogue Signal 4.16
• After signal conditioning the ADC can produce binary number

equivalents of the input voltage.
• If the ADC has finite precision due to a limited no. of discrete levels
then there may be a “small” error associated with each sample.
Voltage fs Binary ts
Value
0.25 4
3
0.125 2
1
0 ADC 0
-1
-0.125 time -2 time
-3
-0.25 -4
• The quantisation step size is 0.0625 volts. If an 5 bit ADC is used, then
the max/min voltage input is approx 0.0625 x 16 = 1 volt.
Generic DSP

Notes:
For example purposes, we can assume our ADC or quantiser has 5 bits of resolution and maximum/minimum
voltage swing of +1 and -1 volts. The input/output characteristic is shown below:
01111 (+15)
Binary Output
0.0625 volts
Vmin = -1 volts
Vmax = 1 volts Voltage

Input
10000 (-16)
Top
Reproducing an Analogue Signal 4.17
• Using a DAC at an appropriate sampling rate, we can reproduce an

analogue signal:
Binary ts fs Voltage
Value
4 0.25
3
2 0.125
1
0 0
-1 DAC
-2 time -0.125 time
-3
-4 -0.25
• Note that the output is a little “steppy” caused by the zero order hold
(step reconstruction);
....this artifact can however be removed with a reconstruction filter.
Generic DSP

Notes:
For example purposes, we can assume our DAC or quantiser has 5 bits of resolution and maximum/minimum
voltage output swing of +1 and -1 volts. The input/output characteristic is shown below:
Vmax = 1 volts
0.0625 volts
Voltage Out
10000 (-16)
01111 (+15)
Binary Input
Vmin = -1 volts
Top
First Order Hold 4.18
• Alternatively a first order hold could be attempted in the DAC. Here

the voltage between two discrete samples is approximated by a
straight line.
Analogue voltage
ts ts
time time
Zero Order Hold First Order Hold
• A first order apparently produces a “more accurate” reproduction of

the analogue signal. However implementation of a circuit to perform
interpolation is not trivial and turns out not to be necessary.

Notes:
A zero order circuit is essentially a form of capacitive element whereby the input voltage is held (almost!)
constant for one sampling period. This is a straightforward and low cost circuit to build and implement.
A first order hold voltage reconstruction does in fact produce a signal that is closer to the original in a mean
squared error sense. However building an electronic circuit that will generate a linearly increasing voltage
between two arbitrary input voltages is not trivial.
Hence we favour zero order hold. We can further note that the “problem” of the zero order hold (i.e the
squareness of the output voltage) can be corrected for later using a reconstruction and a sinx/x compensating
filter.
Top
Binary Data Wordlengths 4.19
• Data wordlengths for DSP applications, typically:
Fixed Point Wordlengths: Dynamic Range

8
• 8 bits −128 to +127 20 log 2 ≈ 48 dB
16
• 16 bits −32768 to +32767 20 log 2 ≈ 96 dB
24
• 24 bits −8388608 to +8388607 20 log 2 ≈ 154 dB
Floating Point Wordlengths (for arithmetic only):
• 32 bits (−1038 to +1038)

(24 bit mantissa, 8 bit exponent)
• Note that data input from an ADC, or output to a DAC will always be
fixed point, although the internal DSP computation may be floating
point.

Notes:
Not all applications will use ADCs or DACs that are 8, 16, or 24 bits. Some applications such as digital
communications typically use between 10 and 12 bits linear quantising ADCs, and thereafter use DSP
processors that use 16 bit wordlengths.
ADC DSP DAC

Processor
10 bits 16 bits 10 bits
DSP processors used to process the data will use data wordlengths of 8 (more a microcontroller), 16, 24 or 32
bits. 32 bits is usually floating point, 24 bit mantissa, 8 bit exponent.
It is worth pointing out at this stage that the no. of bits resolution of an ADC or DAC can be increased by using
oversampling techniques. For example a 3 bit converter, could be “oversampled” to produce a sampled signal
that has resolution of 8 bits! Oversampling techniques will be discussed later in the course.
In general the arithmetic “modes” of DSP processors are straightforward unlike general purpose processors. In
a general purpose processor, arithmetic modes of 16 bit integer, 32 bit integer, 32 bit floating point, 64 bit
floating point etc are likely to be available. This is not the case for DSP processors. Although various arithmetic
modes could be programmed, in general the mode is fixed point, where the sign bit is the MSB (most significant
bit) and a binary point is placed after the MSB. Therefore for a 16 bit processor which has an integer numerical
range of – 32768 to 32767 , the numerical range with the binary point at position 15 is – 1 to 0.9999 , i.e.
1.000 0000 0000 0000 to 0.111 1111 1111 1111 .
Binary data formats are discussed later in the course.

Top
Sampling - How Fast? 4.20
• To intuitively derive the sampling theorem, consider first a pure sine

wave of frequency 100Hz:
period = 1/100
Amplitude
Voltage
0.01 0.02 0.03 time

Phase
• In order to ensure that we retain all of the information in the signal

what sampling rate should be used? (no quantisation)
Sampler
Voltage In Voltage Out
Generic DSP

Notes:
Note that the sampling being performed in this slide does not quantise the signal. We are assuming the
existence of a sampler only at this stage.
In sampling this signal, we are clearly trying to ensure that we retain the information regarding the signal’s
amplitude, frequency and phase, the three components which characterise the signal.
As an alternative to representing a sine wave as an equation, e.g. “Sine wave amplitude spectra
Amplitude
y ( t ) = 10 sin ( 2π2300t + θ ) for a 2300 Hz sine wave of amplitude 10
volts and phase θ we can use a simple frequency domain representation
of a (sine wave amplitude) spectrum. (Note that this simple spectrum
does not represent phase information.) 2300 freq/Hz
If a signal is composed of a “few sine waves” then the spectrum may be “A few sine waves””
Amplitude
represented as illustrated, and if the signal is somewhat more aperiodic
and we can only identify the “average” or typical frequency content, then
we may use a continuous type spectrum
1000 2000 3000 freq/Hz
In effect the sine wave is the simplest type of time varying signal that
exists, and ANY real signal can be produced from a sum of suitable sine “Continuous Spectrum””
Amplitude
and cosine waves. Later in the course when we discuss Fourier analysis
it will be shown that any periodic signal can be decomposed into a sum
of sine waves of suitable frequencies, amplitudes and phases. These
frequency components actually form a mathematical basis for describing 1000 2000 3000 freq/Hz
the periodic signal. Hence in attempting to derive a sampling theorem it
is entirely appropriate that we start by finding a suitable rate for a fundamental single sinusoidal wave, and
extend the theorem from this point.
Top
Sampling - Too Fast? 4.21
• Sampling at fs = 800Hz, i.e. 8 samples per period:
Voltage
0.01 0.02 0.03 time
Appears to be a “reasonable” sampling rate.
• Sampling at fs = 3000Hz, i.e. 30 samples per period:

Voltage
0.01 0.02 0.03 time
Perhaps higher than necessary sampling rate?

Notes:
Sampling - Too Slow? 4.22
• Sampling at fs = 100Hz, i.e. 1 sample per period:

Voltage
0.01 0.02 0.03 time
Signal interpreted as DC!
• Sampling at fs = 80Hz, i.e. 1 sample every 0.8 of a period:

Voltage
0.01 0.02 0.03 time
Most of the signal features are “missed”.

Top
“Suitable” Sampling Rate 4.23
• From inspection of the above 100Hz digital waveforms at the four

different sample rates:
• fs = 800Hz seems a reasonable sampling rate;
• fs = 3000Hz is perhaps higher than necessary;
• fs = 100Hz is too low and fails to correctly sample the

waveform, and loses the signal parameter information;
• fs = 80Hz is too low and fails completely
• From mathematical theory the minimum sampling rate to retain all

information is:
greater than 2 x fmax
where fmax is the maximum frequency component of a baseband,

bandlimited signal.

Notes:
If we sample at a rate that is much higher than the signal frequency/bandwidth, such as sampling a 100Hz sine
wave at 1600Hz, then the signal that can be viewed by simply joining the samples looks clearly like a sine wave.
Amplitude
1
T s = ------------
1600
Ts
N = 15
0 1 2 ... sample,k
1-
-------- th sec
100
However when we sample the 100Hz sine wave at just above Nyquist, i.e. f s = 204Hz then this is high enough
to mathematically retain all information about ALL sine wave components that may be present from 0 to 100Hz
(actually 102 = f s ⁄ 2 ). But observing the join the samples view, it does not necessary look correct...... but it is.
Amplitude
N = 15
0 1 2 ... sample,k
1-
-------- sec
204
1-
-------- th sec
100
Top
Signal Frequency Range Terminology 4.24
• Nyquist frequency/rate: The Nyquist frequency, f n is identified as

twice the maximum frequency component present in a signal.
• Baseband: The lowest signal frequency present is around 0 Hz:
Magnitude
⇒ f n = 2f b
f b = bandwidth
0 fb frequency/Hz
• Bandlimited: For all frequencies in the signal f l < f < f h :

Magnitude
f l = lowest freq
⇒ f n = 2f h
f h = highest freq
fb = fh – fl 0
fl fh frequency/Hz

Notes:
More precisely the Nyquist sampling theorem says that if a signal x ( t ) was sampled at greater than twice the
maximum frequency component present, then the signal can be exactly recovered from the sample values
using the interpolation rule (or convolving a (non-causal) sinc function with the sampled signal):
∞
sin ( πt ⁄ t s )
x(t) = ∑ x ( kt s )h ( t – kt s ) h ( t ) = ---------------------------
πt ⁄ t s
k = –∞
voltage
time
IIn practice true sinc interpolation is not possible to do (due to temporal contrants), hence as we will see later
we usually perform zero order hold followed by low pass filtering to recover the original signal.
Top
Nyquist Sampling Rate 4.25
• If a baseband, bandlimited signal is composed of “sine waves” up to

a frequency f b Hz, then
Nyquist frequency, f n = 2f b
Magnitude
fb frequency/Hz
• In we require to sample this signal and retain all information, then the
sampling rate, f s must be chosen as:
f s > f n i.e. f s > 2f b
• This frequency is often refered to as the Nyquist sampling rate,

(distinct from the Nyquist frequency!).

Notes:
If a signal bandwidth is from 1,000,000 Hz to 1,010,000 Hz: f bw
Magnitude
we might be tempted to suggest that the sampling frequency
(according to the Nyquist criteria) is
f s = 2f bw = 2, 020, 000 Hz .
1000000 1010000 freq/Hz
However if we realise that we could “demodulate” this signal
Magnitude
f bw
to baseband and then sample, it would seem that we can find
some mathematical justification that the direct sampling rate “Demodulated Signal”
be only f s = 2f bw = 20, 000 Hz . Because the signal is
bandlimited, if we sample at 20000 kHz then the signal from 1 10, 000 freq/Hz
- 1.01 MHz will alias down to baseband and all signal
information retained.
Therefore we are effectively downsampling the signal by a factor of 50 and then sampling. Hence it is
appropriate to state that the sampling frequency required to sample any signal can be f s = 2f bw rather than
that predicted by the Nyquist criteria. The above signal could be sampled at 20000Hz (Note in this example for
simplicity we have carefully chosen the maximum frequency in the bandlimited signal and bandwidth to be
integer multiples of each other, in practice when this is not true the required sampling rate will be greater than
2f bw in cases where demodulation to baseband is NOT performed, i.e. the demodulation is implicitly done by
aliasing.)
This general process is called undersampling and is an area of considerable “DSP” activity at present with
respect to “software radio” for communication systems where the electromagnetic signal of interest (e.g. a
mobile phone in the high 100’s MHz’s) is demodulated to an intermediate frequency of around 1 MHz and then
sampled at a frequency rate of the order of 2f bw rather than 2f max . It is important to point out however that the
sampler required to accomplish undersampling, requires a bandwidth capability equivalent to f max given that
the obtained sample is essentially an integration over 1 ⁄ f max .
Top
Aliasing 4.26
• When a (baseband) signal is sampled at a frequency below the

Nyquist rate, then we “lose” the signal frequency information and
aliasing is said to have occured.
• Aliasing can be illustrated by sampling a sine wave at below the

Nyquist rate and then “reconstructing”. We note that it appears as a
sine wave of a lower frequency
(aliasing - cf. impersonating).
• Consider again sampling the 100Hz sine wave at 80Hz:

Voltage
0.01 0.02 0.03 time
Reconstructed signal has a freq. of fs - fsignal = 20Hz

Notes:
Clearly if a signal has frequency components greater than f s ⁄ 2 then aliasing will occur. Aliasing will manifest
as distortion of a signal; for example if a 6000 Hz tone is input to a DSP system (without anti-alias and
reconstruction filters) and sampled at 10000Hz, the sampled signal is interpreted as a 4000Hz tone. Clearly this
is non-linear behavior! (One of the simplest ways of testing the linearity of any system at a particular frequency
is to input a pure tone. If the output is not a pure tone (it may contain harmonics) then the system is NOT linear.)
Consider the following figure where we sample a 9000Hz tone at f s = 10000Hz:
9000Hz sine wave
Amplitude
0 time/s
0.0001 0.005 0.001

Aliased 1000Hz sine wave 1 - 1 -
-------------- s ----------- s
10000 9000
1
------------ sec
1000
Clearly this is above f s ⁄ 2 = 5000Hz and the 9000Hz will alias. The diagram illustrates that when we
reconstruct this signal we get a 1000Hz sine wave. Note that the phase has also shifted, and is now 180o out
of phase, compared to the phase of the input at 9000Hz.
From a knowledge of f s and the input frequencies it is straightforward to establish the frequency of the aliased
components for any input components above f s ⁄ 2 . This is presented in the next two slides.
Top
Aliasing Examples 4.27
• Consider the output from the following three systems:

f s = 1000Hz
100Hz + 100Hz +
250Hz + “Reconst” 250Hz +
400Hz w(t) Sampler wo (t) 400Hz
f s = 1000Hz
900Hz + 100Hz +
750Hz + “Reconst” 250Hz +
600Hz x(t) Sampler xo (t) 400Hz
f s = 1000Hz
1100Hz + 100Hz +
Voltage In Voltage Out 250Hz +
1250Hz + “Reconst”
1400Hz y(t) yo (t) 400Hz
Sampler

Notes:
If you are viewing the output signal only, after sampling then a suitable reconstruction you cannot be sure what
the input signal given that aliasing may have occured.
Clearly the above system sampling at 1000Hz would be best limited to only have input signals with frequencies
below 500Hz.
Of course sometimes aliasing can be exploited in situations where perhaps we are aiming to demodulate a
signal. As we will discuss in the Software Radio section of notes, direct digital downconversion could be used
to “demodulate” (or alias down) from between 60 and 70kHz to 10kHz by taking only 20000 samples per second
(although the front end of the ADC must still be capable on integrating over a time interval that is commensurate
with the signal bandwidth).
10kHz
Amplitude
fs/2 fs
10 20 30 40 50 60 70 freq/Hz
If this signal is appropriately bandlimited, the output signal will alias to the same “shape” at baseband
frequencies:
Amplitude
10kHz
fs/2 fs
10 20 30 40 50 60 70 freq/Hz
Top
Aliased Spectra 4.28
Sampling frequency, fs = 1000Hz
Magnitude
Signal spectrum
w o ( t ) = 100 sin 2π100t + 50 sin 2π250t + 25 sin 2π400t
NO ALIASING
fs = 1000 freq/Hz
Aliased spectrum
Magnitude
Signal spectrum
x o ( t ) = – 100 sin 2π900t – 50 sin 2π750t + – 25 sin 2π600t
ALIASING
fs = 1000 freq/Hz
Aliased spectrum
Magnitude
Signal spectrum
y o ( t ) = 100 sin 2π1100t + 50 sin 2π1250t + 25 sin 2π1400t
ALIASING
fs = 1000 freq/Hz

Notes:
If we sample a sum of a few sine waves signal,
Amplitude
w ( t ) at 1000 Hz we adhere to the Nyquist
criteria. We can represent the signal to be
sampled as a simple (sine wave amplitude)
spectrum: 1000 freq/Hz
w ( t ) = 100 sin 2π100t + 50 sin 2π250t + 25 sin 2π400t
Of course if we had sampled the the signal
Amplitude
Aliased spectrum
x ( t ) (i.e. sine waves at 900Hz, 750Hz and Signal spectrum
600Hz at 1000 Hz), then because Nyquist
criteria is not preserved, all of these sine
waves will alias respectively to 100, 250 and 1000 freq/Hz
400 Hz components. x ( t ) = 100 sin 2π900t + 50 sin 2π750t + 25 sin 2π600t
Similarly if we sampled the signal y ( t ) (i.e. Aliased spectrum
Amplitude
sine waves at 1100Hz, 1250Hz and 1400Hz)
Signal spectrum
at 1000 Hz, then because Nyquist criteria is
not preserved, all of these sine waves will
alias respectively to 100, 250 and 400 Hz 1000 freq/Hz
components. y ( t ) = 100 sin 2π1100t + 50 sin 2π1250t + 25 sin 2π1400t
We can then of course note the “pattern” of spectrum aliasing as shown below:
Amplitude
Baseband spectrum
fs ⁄ 2 fs freq/Hz
Top
Anti-Alias Filter 4.29
• Prior to the analogue to digital converter (ADC) all frequencies above

fs/2 must be blocked or they will be interpreted as lower frequencies,
i.e aliasing.
Analogue fs
Analogue Anti-Alias Filter
Processor
input voltage
To DSP
−
+
etc.... ADC
Magnitude
Magnitude
Attenuation
fs/2 frequency freq fs/2 frequency

fs/2
• The anti-alias filter is analogue (ideally a brick wall filter), cutting off
just before fs/2 Hz.
Generic DSP

Notes:
Clearly brick wall filters cannot be implemented, and therefore the analog filter designer must compromise to
produce a filter “enough” roll-off, attenuation etc. The filter designed should also ideally have linear phase.
Therefore the work of the anti-alias designed is less than trivial!
Roll off rate in dB/decade or dB/octave
Attenuation (dB)
0 -3dB cut off
Stopband attenuation
-40
fc frequency
3dB cut off frequency
V out V out
0dB corresponds to an attenunation of 1, i.e., if 20 log ------------ = 0 , then ------------ = 1 and therefore V out = V in .
V in V in
V out V out
- = – 40 , then ------------ = 0.01 and therefore V out = 0.01V in .
-40dB corresponds to 20 log -----------
V in V in
Top
Reconstruction Filter 4.30
• The analogue reconstruction filter at the output of a DAC removes the

baseband image high frequencies present in the signal (in the form of the
steps between the discrete levels).
y(t) Voltage
time time
Analogue
Reconstruction
Filter Reconstruction filter
“Steppy/staircase” removes the high
output ananlogue − frequency signal
+
voltage from DAC Attenuation etc....
components. .
Magnitude
Magnitude
fs/2 freq
fs/2 freq fs/2 freq
Generic DSP

Notes:
Note that the (time) impulse response (discussed later) of the perfect brick wall filter is in fact a sinc function
Note that the (time) impulse response (discussed later) of the Roll off rate = ∞
perfect brick wall filter is in fact a sinc function:
Attenuation
0dB
sin ( πt ⁄ t s )
---------------------------
πt ⁄ t s
which starts at t = – ∞ and ends at t = ∞ . Hence the existence of fc

frequency
the sinc interpolation process.
-2ts -ts 0 ts 2ts time

Top
Zero Order Hold (ZOH) “Filter” 4.31
• Note that the operation of zero order hold can be interpreted as a

simple “reconstructing” frequency filtering operation:
Binary ts f Voltage
Value s
4 0.02
3
2 0.01
1
0 0
-1 ZOH
-2 time -0.01 time
-3
-4 -0.02
H ZOH ( f ) 1
Frequency shaping characteristic of ZOH circuit
Attenuation
2/π
fs /2 fs 2fs 3fs 4fs 5fs freq
• The step reconstruction therefore causes a “droop” near fs/2.

Notes:
The frequency response of the ZOH circuit can be calculated by simply noting that the impulse response is:
h ZOH ( t )
1
Impulse response of ZOH pulse
time
ts
and finding the frequency response (as shown in diagram in above slide) via a Fourier transform;
ts jπft s – j πft s sin πft s – j πft

1 – j 2πft s e –e – j πft s
H ZOH ( f ) = ∫ 0
1.e – j 2πft dt = --------------- [ 1 – e
j2πft s
] = ------------------------------------ e
2j ( πft s )
= ------------------ e
πft s
s
Hence the ideal reconstruction filter should compensate for the “sinx/x” “droop” at frequency f s ⁄ 2 by a factor of:
2 ⁄ π = 0.637 ≡ 20 log 0.637 = – 3.92 dB .
Therefore the magnitude response for “perfect” reconstruction (a H

ZOH ( f )
linear phase response is also ideally required) is a filter that 1 Frequency response of
compensates at f s ⁄ 2 with a gain of 1/0.637 = 1.569; above this ideal reconstruction filter.
value, all frequency components should be attenuated (in a brick
wall fashion!). fs ⁄ 2 frequency
In practice of course such an analogue filter will not be produced, and as far as possible a close to ideal brick
wall filter will be used. In order to compenstate for the droop, we can actually introduce a digital filter before the
DAC to amplify the signal with the above inverse droop characteristic! And finally in the modern DSP world,
oversampling will be used to reduce the analogue complexity (see later....)
Top
Anti-alias and Reconstruction 4.32
• Anti-alias and reconstructions filters are analogue, i.e. made from

resistors, capacitors, amplifiers, even inductors.
• Ideally they are both very sharp cut off filters at frequency fs/2. In
practice the roll off will be between 6dB/octave (a simple resistor and
capacitor) to 96dB/octave (a 16th order filter).
G(f)
Magnitude
Filter gain roll-off
Pass
band
Cut-off frequency freq
• Steeper roll-off is more expensive, but clearly for many applications,

good analogue filters are essential.
• In a DSP system the accurately trimmed analogue filters could

actually be more costly than the other DSP components: i.e. DSP
processor, ADC, DAC, memory etc.

Notes:
One of the key changes in signal processing over the last few years has been the introduction of oversampling
(and sigma delta) DSP systems. The key strategy behind oversampling is to reduce the cost of the analog
components at the expense of more DSP processing and higher sampling rates. With the reduction in cost of
reliable easy to fabricate DSP systems, more modern DSP implementations use oversampling ADCs and
DACs. Oversampling is discussed later in the multirate section of the course.
Top
Perfect Nyquist Sampling 4.33
• The Nyquist sampling theorem states that a (baseband) signal should

sampled at greater than twice the maximum frequency component
present in the signal:
f s > 2 × f max
• The sampled signal can then be perfectly reconstructed to the original

analogue signal with no added noise or distortion.
s(t) fs v(n) ts
0.25 4
Voltage
Binary value
0.125 2
Volts 1
0 to Real 0
-1
-0.125 time, t Number -2 sample, n
-3
-0.25 -4
v ( n ) = s ( nt s ), for n = 0, 1, 2, …
Notes:
In the above example the “converter” produces a real number value, i.e an input voltage of 1.10233 is output
as the real number 1.10233. Therefore the input/output characteristic of this system is:
Real Number Output

16
1 Voltage
Input
Clearly such real number “converters” do not exist.

Top
ADC Sampling “Error” 4.34
• Perfect signal reconstruction assumes that sampled data values are

exact (i.e. infinite precision real numbers). In practice they are not, as
an ADC will have a number of discrete levels.
• The ADC samples at the Nyquist rate, and the sampled data value is
the closest (discrete) ADC level to the actual value:
s(t) ^v(n) ts
0.25
fs 4
Binary value
3
0.125 2
Voltage
1
0 ADC 0
time
-1 sample, n
-0.125 -2
-3
-0.25 -4
v̂ ( n ) = Quantise { s ( nt s ) }, for n = 0, 1, 2, …
• Hence every sample has a “small” quantisation error.

Notes:
For example purposes, we can assume our ADC or quantiser has 5 bits of resolution and maximum/minimum
voltage swing of +1 and -1 volts. The input/output characteristic is shown below:
01111 (+15)
Binary Output
0.125 volts
Vmin = 1 volts
Vmax = 1 volts Voltage

Input
10000 (-16)
In the above slide figure, the second sample, the true sample value is 1.589998..., however our ADC quantises
to a value of 2.
Top
Quantisation Error 4.35
• If the smallest step size of a linear ADC is q volts, then the error of
any one sample is at worst q/2 volts.
01111 (+15)
Binary Output
q volts
-Vmax Vmax Voltage

Input
10000 (-16)

Notes:
The quantisation error is straightforward to calculate from:
V max
q = --------------
2N – 1
where N is the number of bits in the converter.
The dynamic range of an N bit converter is often quoted in dBs:
Dynamic Range = 20log 10 2 N = 20Nlog 10 2 = 6.02N
Therefore an 8 bit converter has a range of
Binary 10000000 to 01111111, or in decimal -128 to 127 has a dynamic range of approximately 48 dB.
Top
Quantisation Error in a Signal 4.36
• The ADC output, v̂ ( n ) , can therefore be modelled as the perfectly

sampled signal, v ( n ) , plus a quantisation error signal denoted, e ( k ) :
v(n) ^
v(n)
2
q
1
0
time
-1
-2
e(n)
q/2
-q/2 time
v̂ ( n ) = v ( n ) + e ( n )

Notes:
The actual ADC can be represented by a sampler and a quantiser:
Sampler Quantiser Binary

Voltage In Voltage out Samples out
s(t) v(n) v̂ ( n )
Analogue to Digital Converter
The quantisation error of each sample is in the range ± q ⁄ 2 and we can model the quantiser as a linear additive
noise source.
e(n)
Sampler Binary
Voltage In Voltage out Samples out
s(t) v(n) v̂ ( n )
Analogue to Digital Converter
We can therefore use this as a simple linear equation model of our quantiser where v̂ ( n ) = v ( n ) + e ( n ) and
e ( n ) is a white noise source uncorrelated with the input signal v ( n ) . Note that for the actual quantiser it is not
possible to write a simple set of mathematical equations to describe/define its input/output. (Later in the course
(Dithering) we will see that the “uncorrelated white quantisation noise” assumption is not actually
generally true.....but thats for later and is a reasonable assumption for now.)
Top
Quantisation Error PDF 4.37
• Assuming the ADC rounds to the nearest digital level, the maximum
error of any one sample is q/2 volts.
• If we assume that the probability of the error being at a particular value

between +q/2 and -q/2 is equally likely then the probability density
function (pdf) for the error is flat.
p(e)
1/q
-q/2 q/2 e
(In practice the quantisation error will be somewhat correlated with the input signal,
particularly for low level periodic signals. This will be discussed and addressed later in the
dithering section).

Notes:
By taking a uniformly distributed random variable, we can produce a histogram of the output values to confirm
the probability density function (pdf).
For example, if a signal source outputs a random variable that is uniformly distributed over the range -1 to 1 and
we take outputs and plot a histogram for ranges of 0.2, we may note that after observing, say 10000 samples
the histogram of output values is:
Frequency of occurence of x(.)

1000
500
-1.0 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1.0 x(.)
The histogram is straightforward to interpret, e.g. in the last 10,000 samples, 983 have been in the range 0 to
0.2. In the limit if the signal is truely generated by a uniformly distributed random noise source, then as we
reduce the interval width to 0, and increase the no. of samples to an infinite (!) number, the pdf is as shown in
the slide.
Use SystemView to calculate the PDF of a Gaussian noise generator.

Top
Quantisation Error Power (Statistical) 4.38
• Consider the noise power or variance (in a 1 ohm resistor for a 1Hz
sampling rate) of the quantisation error signal from:
∞ q⁄2 q⁄2 2
1 q
n adc = ∫– ∞ e 2 p ( e ) de = ∫– q ⁄ 2 e 2 p ( e ) de = ------ e 3
3q –q ⁄ 2
= ------
12
• The quantisation error, E ( f ) , will extend over the frequency range 0 to

fs /2. , i.e. the full baseband. If V ( f ) is the signal of interest that is being
quantised, then:
V(f)
Power Spectral
Density (W/Hz)
Signal
Low level signals E(f)
of interest
may be masked by Quantisation Noise
the quantisation (theoretical only)
noise. q 2 ⁄ 12
----------------
fs ⁄ 2
ˆ (f) = V(f) + E(f)
V fs/2 frequency (Hz)
Return

Notes:
The terms “quantisation power” and “quantisation error” are often used interchangeably within DSP. Strictly
speaking the quantisation error is the error between one quantised sample, and its true value. Extending this
we have the quantisation error signal which is a time versus voltage signal of the error samples - this signal
does, of course, not actually exist and is only one that we use for analysis purposes. Collectively we then often
refer to the quantisation error signal as the “quantisation noise”, or “quantisation noise signal”. Arguably the term
noise is usually restricted to a signal “noise”, which we might expect to be able to remove by filtering.
Quantisation “noise” is more correctly termed quantisation “distortion” - once a signal has been quantised there
is NO inverse process.
Quantisation Noise Power (Time Average - compare to above Statistical Average)
The noise power can also be calculated from a time average rather than a statistical average. Taking N samples
(where N is large) of the quantisation error, the average power is simply:
N–1
Signal Energy 1 q2
Noise Power = ------------------------------------- = ----
Time N ∑ 2
e ( k ) ≈ ------
12
for large N
n=0
e(n) q/2
N-1 sample
-q/2
Average
e2(n) 2
q /4 q2/12
N-1 sample
Top
Quantisation Spectra 4.39
• After quantisation the signal spectrum will be:
ˆ (f) = V(f) + E(f)

V [ recall v̂ ( n ) = v ( n ) + e ( n ) ]
• The quantisation noise will therefore obscure/mask frequency

components of interest in the true signal spectrum:
ˆ (f)
V
V(f)
Density (W/Hz)
Power Spectral
Quantised signal spectrum
Signal Power
E(f)
q 2 ⁄ 12
----------------
fs ⁄ 2
Vˆ ( f ) = V ( f ) + E ( f )
fs/2 fs/2
V ( f ) and E ( f ) freq/Hz freq/Hz
• Compare to previous spectrum on Slide 4.38 .

Notes:
For an N-bit signal, there are 2 N levels from the maximum to the minimum value of the quantiser:
Binary Output
2N – 1 – 1 2
Quantization q = ------
N
-
step size 2
-1 1
Analog Input
–2 N – 1
Therefore the mean square value of the quantisation noise power can be calculated as:
 2 ⁄ 2 N ) 2
(--------------------- 4
Q N = 10 log = 10 log 2 – 2 N + 10 log ------ ≈ – 6.02N – 4.77 dB
 12  12
Another useful measurement is the signal to quantisation noise ratio (SQNR). For the above ADC with voltage
input levels between -1 and +1 volts, if the input signal is the maximum possible, i.e. a sine wave of amplitude
1 volt, then the average input signal power is: Signal Power = E [ sin2 2πft ] = 1 ⁄ 2 . Therefore the maximum
SQNR is:
Signal Power 0.5 3

SQNR = 10 log ----------------------------------- = 10 log ------------------------------------ = 10 log 2 – 2 N + 10 log --- = 6.02N + 1.76 dB
Noise Power ( ( 2 ⁄ 2 N ) 2 ⁄ 12 ) 2
For a perfect 16 bit ADC the maximum SQNR can be calcuated to be 98.08 dB.
Top
Timing Jitter Error I 4.40
• Note that when a signal is sampled there may be some “jitter” on the
sampling clock which will cause additional sample error.
Voltage
No Sampling Clock Jitter
time
Voltage
With Sampling Clock Jitter
time
• With jitter each sampling instant may be slightly offset, and therefore
the sample value obtained and sent to the DSP will be in error.

Notes:
Top
Conclusions 4.41
• This session has introduced them fundamental concepts of DSP:
• Sampling Theorem;
• Quantisation error and noise model;
• Aliasing and Reconstruction;
• The Generic DSP System

Notes:

Introduction To DSPXilinx

Uploaded by

Copyright:

Available Formats

Introduction To DSPXilinx

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To DSPXilinx

Uploaded by

Copyright:

Available Formats

The DSP Primer 4

DSPprimer Home DSPprimer Notes

August 2005, University of Strathclyde, Scotland, UK For Academic Use Only

• 1980s: Special-purpose DSP microprocessors introduced.

• 1990s: Processing power of DSP (micro-)processors has increased

• The (digital) reliability, repeatability, and programmability of DSP has

1000 DSP enabled 3G Mobile

August 2005, For Academic Use Only, All Rights Reserved

August 2005, For Academic Use Only, All Rights Reserved

• The voltage from a signal sensor is very small in magnitude.

• Prior to recording the signal or reproducing with an actuator an amplifier

• The above amplifier adds 60dB of gain (20log101000 = 60)

August 2005, For Academic Use Only, All Rights Reserved

x(t) Linear y(t)

In general for a linear system y ( t ) = f ( x ( t ) ) , if,

A dB = 10log 10 ( P out ⁄ P in ) = 10log 10 ( V out ⁄ V in ) 2 = 20log 10 ( V out ⁄ V in ) decibels (dB)

• An amplifier which introduces unwanted artifacts, is said to be non-

• Unlike noise it is essentially impossible to remove the effects of

August 2005, For Academic Use Only, All Rights Reserved

V a = V b2 , then V b = ± V a which is a non-unique solution.

See system c:\DSPedia\intro\non-linear_amplifer.svu for implementation of the above slide and

• Signal processing techniques are often used in an attempt to remove

• Most noise can be considered as additive (linear superposition) which

August 2005, For Academic Use Only, All Rights Reserved

• Consider the various levels of noise and distortion added in a digital

Environment noise Digital quant- Digital compression Modulation

Micro- Analogue to DSP Speech Modulation/

• DSP must minimize the amount of noise/distortion input to the chain,

August 2005, For Academic Use Only, All Rights Reserved

( Signal Power ) P signal

• Recalling that Power ∝ Volts 2 , then:

SNR = 10dB, P signal = 10 × P noise and V signal = 10 × V noise

SNR = 60dB, P signal = 1000000 × P noise and V signal = 1000 × V noise

Audio cassette deck

August 2005, For Academic Use Only, All Rights Reserved

• In general an analogue signal processing system can be defined as a

• A public address system is an example of an analogue signal

... Blue Suede

August 2005, For Academic Use Only, All Rights Reserved

• A single input, single output DSP system has the following

Signal Conditioning Signal Conditioning

Analogue Digital Analogue

August 2005, For Academic Use Only, All Rights Reserved

DSP systems can be classed as one of three general types:

1001 0000 0000 0001 2 = -2 15 + 2 12 + 2 1

• For most baseband telecommunications a voltage signal is

Line interface Line interface

... Blue Suede

• A simple example is a telephone. The acoustic signal is converted to

Generic Analogue Radio Communication

... Blue Suede

• Modern communications systems require that digital information is

August 2005, For Academic Use Only, All Rights Reserved

Analogue Anti- ADC DSP Re-

DSP Re- Anti- ADC DSP

• An ADC is a device that can convert a voltage to a binary number,

•The number of digital samples converted per second

August 2005, For Academic Use Only, All Rights Reserved

• A DAC is a device that can convert binary numbers to voltages,