What is DSP?

Input Signal Analog-toDigital Conversion

Computer

Digital Speech Processing Lecture 2

Digital-toAnalog Conversion

Output Signal

Digital
Method to represent a quantity, a phenomenon or an event Why digital?

Signal

Review of DSP Fundamentals

1

What is a signal?
something (e.g., a sound, gesture, or object) that carries information a detectable physical quantity (e.g., a voltage, current, or magnetic field strength) by which messages or information can be transmitted

What are we interested in, particularly when the signal is speech?

Processing
What kind of processing do we need and encounter almost everyday? Special effects?
2

Common Forms of Computing

Text processing handling of text, tables, basic
arithmetic and logic operations (i.e., calculator functions)
Word processing Language processing Spreadsheet processing Presentation processing

Advantages of Digital Representations

Input Signal

A-to-D Converter

Signal Processor

D-to-A Converter

Output Signal

Signal Processing a more general form of

information processing, including handling of speech, audio, image, video, etc.

Filtering/spectral analysis Analysis, recognition, synthesis and coding of real world signals Detection and estimation of signals in the presence of noise or interference
3

Permanence and robustness of signal representations; zerodistortion reproduction may be achievable Advanced IC technology works well for digital systems Virtually infinite flexibility with digital systems Multi-functionality Multi-input/multi-output Indispensable in telecommunications which is virtually all digital at the present time
4

Digital Processing of Analog Signals

xc(t) A-to-D x[ n] Computer y[ n] D-to-A yc(t)

Discrete-Time Signals
A sequence of numbers Mathematical representation: x = {x[n ]}, < n < Sampled from an analog signal, xa (t ), at time t = nT , x[n ] = xa (nT ), < n < T is called the sampling period, and its reciprocal, FS = 1/ T , is called the sampling frequency FS = 8000 Hz T = 1/ 8000 = 125 sec FS = 10000 Hz T = 1/10000 = 100 sec FS = 16000 Hz T = 1/16000 = 62.5 sec FS = 20000 Hz T = 1/ 20000 = 50 sec

A-to-D conversion: bandwidth control, sampling and quantization Computational processing: implemented on computers or ASICs with finite-precision arithmetic basic numerical processing: add, subtract, multiply (scaling, amplification, attenuation), mute, algorithmic numerical processing: convolution or linear filtering, non-linear filtering (e.g., median filtering), difference equations, DFT, inverse filtering, MAX/MIN, D-to-A conversion: re-quantification* and filtering (or interpolation) for reconstruction
5

Speech Waveform Display

plot( );

Varying Sampling Rates

Fs=8000 Hz

Fs=6000 Hz

stem( ); Fs=10000 Hz

Varying Sampling Rates

Quantization
x[ n] can be quantized to one of a finite set of values which is then represented digitally in bits, hence a truly digital signal; the course material mostly deals with discrete-time signals (discrete-value only when noted).
out
7 6 2.4 1.8 1.2 0.6 0.3 0.9 1.5 2.1

Fs=8000 Hz

Quantization: Transforming a continuouslyvalued input into a representation that assumes one out of a finite set of values in The finite set of output values is indexed; e.g., the value 1.8 has an index of 6, or (110)2 in binary representation Storage or transmission uses binary representation; a quantization table is needed

Fs=6000 Hz

5 4 3

Fs=10000 Hz

2 1 0

A 3-bit uniform quantizer

Discrete Signals
6

Sinewave Spectrum
quantize

sample
y

Sampled Sinusoid 5sin(2nT)

6

Analog sinusoid, 5sin(2x)

-2

-4

-6

2 0 5 10 15 20 25 30 35 40 0

Discrete sinusoid round[5sin(2nT)]

n
6

-2

y
-4 4 -6 2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0

Quantized sinusoid round[5sin(2x)]

-2

-4

-6

x
y

10

15

20

25

30

35

40

-2

quantize

-4

sample
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

SNR is a function of B, the number of bits in the quantizer

11 12

-6

Issues with Discrete Signals

what sampling rate is appropriate
amplitude

The Sampling Theorem

Sampled 1000 Hz and 7000 Hz Cosine Waves; Fs = 8000 Hz 1 0.5 0 -0.5 -1 0 0.2 0.4 0.6 time in ms 0.8 1 1.2

6.4 kHz (telephone bandwidth), 8 kHz (extended telephone BW), 10 kHz (extended bandwidth), 16 kHz (hi-fi speech)

how many quantization levels are necessary at each bit rate (bits/sample)
16, 12, 8, => ultimately determines the S/N ratio of the speech speech coding is concerned with answering this question in an optimal manner
13

A bandlimited signal can be reconstructed exactly from samples taken with sampling frequency
1 = Fs 2fmax T or 2 = s 2max T
14

Demo Examples
1. 2. 3. 5 kHz analog bandwidth sampled at 10, 5, 2.5, 1.25 kHz (notice the aliasing that arises when the sampling rate is below 10 kHz) quantization to various levels 12,9,4,2, and 1 bit quantization (notice the distortion introduced when the number of bits is too low) music quantization 16 bit audio quantized to various levels:

Discrete-Time (DT) Signals are Sequences

Maple Rag: 16 bits, 12 bits,

10 bits, 8 bits,

6 bits, 4 bits;

15

x[n] denotes the sequence value at time n Sources of sequences: Sampling a continuous-time signal x[n] = xc(nT) = xc(t)|t=nT Mathematical formulas generative system e.g., x[n] = 0.3 x[n-1] -1; x[0] = 40
16

Noise: 10 12 bits

Impulse Representation of Sequences

A sequence, a function

Some Useful Sequences

unit sample [n] =
1, n = 0 0, n 0

x[ n ] =

k =

x[k ] [n k ]
a1 [ n 1]

Value of the function at k

real exponential

x[n] = n

a3 [ n + 3]

a2 [ n 2]

a7 [ n 7]

x[n ] = a3 [n + 3] + a1 [n 1] + a2 [n 2] + a7 [n 7]
17

1, n 0 unit step u[n ] = 0, n < 0

sine wave x[n] = A cos( 0 n + )

18

Variants on Discrete-Time Step Function

u[n] u[n-n0]

Complex Signal
x[n ] = (0.65 + 0.5 j )n u[n ]

u[n0-n]

n n signal flips around 0

19

20

Complex Signal
x [n ] = ( + j )n u[n ] = (re j )n u[n ] r = 2 + 2

Complex DT Sinusoid x [n ] = A e j n
Frequency is in radians (per sample), or just radians
r

= tan1( / )
x [n ] = r e
n n j n

once sampled, x[n] is a sequence that relates to time only through the sampling period T

u [n ]

Important property: periodic in with period 2: j ( 0 + 2 r )n j 0n

r is a dying exponential e j n is a linear phase term

21

Ae

= Ae

Only unique frequencies are 0 to 2 (or to +) Same applies to real sinusoids

22

Sampled Speech Waveform

xa(t)
MATLAB: plot

Signal Processing
Transform digital signal into more desirable form
x[n] y[n]=T[x[n]] x[n] y[n]

MATLAB: stem

x[n] xa(nT),x(n)

single inputsingle output

single inputmultiple output, e.g., filter bank analysis, sinusoidal sum analysis, etc.

T=0.125 msec, fS=8 kHz

Trap #1: loss of sampling time index 23 24

LTI Discrete-Time Systems

x [ n]

LTI Discrete-Time Systems

Example: Is system y [n ] = x[n ] + 2 x[n + 1] + 3 linear?
x1[n ] y1[n ] = x1[n ] + 2 x1[n + 1] + 3 x2 [n ] y 2 [n ] = x2 [n ] + 2 x2 [n + 1] + 3 x1[n ] + x2 [n ] y 3 [n ] = x1[n ] + x2 [n ] + 2 x1[n + 1] + 2x2 [n + 1] + 3 y1[n ] + y 2 [n ] Not a linear system!
Is system y [n] = x[n ] + 2 x[n + 1] + 3 time/shift invariant?

[n]

LTI System

y[n] h[n]

Linearity (superposition):
{ax1[n] + bx2 [n]} = a{x1[n]} + b {x2 [n ]}

Time-Invariance (shift-invariance): x1[n] = x[n nd ] y1[n] = y[n nd ] LTI implies discrete convolution:
y[n] = x[k ]h[n k ] = x[n] h[n] = h[n] x[n ]
k =
25

y [n ] = x[n ] + 2x [n + 1] + 3 y [n n0 ] = x [n n0 ] + 2 x [n n0 + 1] + 3 System is time invariant!

Is system y [n ] = x [n ] + 2 x [n + 1] + 3 causal?

y [n ] depends on x[n + 1], a sample in the future System is not causal!

26

Convolution Example
1 x[n] = 0 0n3 otherwise 1 h[n ] = 0 0n3 otherwise
x[n],h[n]

h[n]

x[n]

01234 5 n=0

01234 5 n=1

n=2

What is y [n ] for this system?

Solution : y [n ] = x [n ] * h[n ] =
m =

01234 5

h[m] x[n m]
0n3 4n6 n 0, n 7
01234 567
n y[n]

01234 5 n=3

01234 5 n=4

01234 5 n=5

n 1 1 = (n + 1) m =0 3 = 1 1 = (7 n ) m = n 3 0

01234 5 n=6

01234 5 n=7

01234 5 n=8

27

01234 56

01234 567

01234 5678

28

Convolution Example
The impulse response of an LTI system is of the form: h[n ] = a n u[n ] x [ n ] = b n u [n ] | a |< 1 | b |< 1, b a and the input to the system is of the form: Determine the output of the system using the formula for discrete convolution.
SOLUTION: y [n ] =
m =

Convolution Example

a
n m =0

u[ m ] b n m u[ n m ]
m

= bn

b m u [n ] = b n

m =0

(a / b )

u [n ]

1 (a / b )n +1 b n +1 a n +1 = bn = u [n ] 1 (a / b ) b a

29

30

Convolution Example
Consider a digital system with input x [n ] = 1 for n = 0,1,2,3 and 0 everywhere else, and with impulse response h[n ] = a u[n ], | a |< 1. Determine the
n

Linear Time-Invariant Systems

easiest to understand easiest to manipulate powerful processing capabilities characterized completely by their response to unit sample, h(n), via convolution relationship

response y [n ] of this linear system.

SOLUTION: We recognize that x [n ] can be written as the difference between two step functions, i.e., x [n ] = u[n ] u[n 4]. Hence we can solve for y [n ] as the difference between the output of the linear system with a step input and the output of the linear system with a delayed step input. Thus we solve for the response to a unit step as: a n a 1 u [n m ] = u[n ] 1 1 a y [n ] = y1[n ] y1[n 4] y 1[ n ] =
m =

y [n ] = x[n ] h[n ] =

k =

x[k ] h[n k ] = h[k ] x[n k ]

k =

y [n ] = h[n ] x[n ], where denotes discrete convolution

x[n]

h[n]

y[n]=x[n]

* h[n]

u[m] a

n m

basis for linear filtering

31

used as models for speech production (source convolved with system)

32

Signal Processing Operations

x1[n] x1[n]+x2[n] x2[n] x[n] x[n]

Equivalent LTI Systems

x[n] h1[n] h2[n] y[n] x[n] x[n] h2[n] h1[n] y[n] h1[n] h2[n] y[n]

x[n]

h1[n]+h2[n]

y[n]

x[n]

x[n-1]

x[n]

h1[n]*h2[n]

y[n]

h1[n]*h2[n]= h2[n]*h1[n]
D is a delay of 1-sample Can replace D with delay element z 1
33

h1[n]+h2[n]= h2[n]+h1[n]
34

More Complex Filter Interconnections

Network View of Filtering (FIR Filter)

h2[n] h1[n] h3[n] h4[n]

x[n]

y[n]

y [n ] = x [n ] * hc [n ] hc [n ] = h1[n ] * (h2 [n ] + h3 [n ]) + h4 [n ]
35

D (Delay Element) z 1 y [n ] = b0 x [n ] + b1x [n 1] + ... + bM 1x [n M + 1] + bM x [n M ]

36

Network View of Filtering (IIR Filter)

x[n]

y[n]

z-Transform Representations

y [n ] = a1y [n 1] + b0 x[n ] + b1x [n 1]

37 38

Transform Representations
z-transform:
x[n ] X (z) X (z) = x[ n ] =
n =

Examples of Convergence Regions

1 0.9

x[n]z
1
C

infinite power series in z , with x[n] as coefficients of n term in z

n

1. x[n ] = [n n0 ] -- delayed impulse X (z ) = z n0 converges for | z |> 0, n0 > 0; | z |< , n0 < 0; z, n0 = 0 2. x[n ] = u[n ] u[n N ] -- box pulse

0.8 0.7 0.6 a m p litu d e

0.5 0.4 0.3 0.2 0.1

4 5 sample number

direct evaluation using residue theorem partial fraction expansion of X(z) long division power series expansion

1 0.9

0.8 0.7

2 j

X (z)z n 1dz

1 z N -- converges for 0 <| z |< 1 z 1 n =0 all finite length sequences converge in the region 0 <| z |< X(z) = (1)z n =
N-1

0.6 amplitude

0.5

0.4

0.3

0.2 0.1

4 5 sample number

X(z) converges (is finite) only for certain values of z:

3. x[n ] = a nu[n ] (a < 1)

1

n =

| x[n] | | z

|<

- sufficient condition for convergence

1 --converges for | a |<| z | 1 az 1 n =0 all infinite duration sequences which are non-zero for n 0 X (z ) = a n z n = converge in a region | z |> R1

0.9 0.8 0.7

0.6 amplitude

0.5 0.4

0.3

0.2 0.1

10

15

20

25 30 sample number

35

40

45

50

region of convergence: R1 < |z| < R2

39 40

Examples of Convergence Regions

0 -0.1 -0.2 -0.3 -0.4 amplitude -0.5 -0.6 -0.7

Example
If x [n ] has z-transform X ( z ) with ROC of ri <| z |< ro , find the z -transform, Y ( z ), and the region of convergence for the sequence y [n ] = a n x [n ] in terms of X ( z )
Solution: X (z) = Y (z) =
n =

4. x[n ] = b nu[ n 1]
1 n n

-0.8

-0.9 -1 -50

-45

-40

-35

-30

-25 -20 sample number

-15

-10

-5

1 X ( z ) = b z = --converges for | z |<| b | 1 bz 1 n = all infinite duration sequences which are non-zero for n < 0 converge in a region | z |< R2 5. x[n ] non-zero for < n < can be viewed as a combination of 3 and 4,giving a convergence region of the form R1 <| z |< R2 sub-sequence for n 0 => | z |> R1 sub-sequence for n < 0 => | z |< R2 total sequence => R1 <| z |< R2
41

x [ n ]z

R2 R1

n =

y [ n ]z

n = n

x [ n ]z n

n =

x[n](z / a)

= X (z / a)
42

ROC: | a | ri <| z |<| a | ro

z-Transform Property
The sequence x [n ] has z-transform X ( z ). Show that the sequence nx[n ] has z-transform z dX ( z ) . dz

Inverse z-Transform
x[ n ] = 1 2 j
C

X ( z )z

n 1

dz

Solution: X (z) =
n =

where C is a closed contour that encircles the origin

n

x [ n ]z

of the z-plane and lies inside the region of convergence

dX ( z ) = nx[n ]z n 1 dz n =

C R1 R2

1 = nx[n ]z n z n = 1 = Z (nx[n ]) z
43

for X(z) rational, can use a partial fraction expansion for finding inverse transforms
44

Partial Fraction Expansion

H (z) = b0 z M + b1z M 1 + ... + bM z N + a1z N 1 + ... + aN

Example of Partial Fractions

Find the inverse z-transform of H (z ) = A A A H (z ) z2 + z + 1 = = 0+ 1 + 2 z z (z + 1)(z + 2) z z + 1 z + 2 A0 = A2 = z2 + z + 1 1 = (z + 1)(z + 2) z =0 2 z2 + z + 1 3 = z(z + 1) z =2 2 A1 = z2 + z + 1 = 1 z(z + 2) z =1 z2 + z + 1 1 <| z |< 2 (z 2 + 3z + 2)

b z M + b1z M 1 + ... + bM = 0 ; (N M ) (z p1 )(z p2 )...(z pN ) A1 A2 AN + + ... + H (z) = z p1 z p2 z pN A0 H (z) A1 A2 AN = + + + ... + ; p0 = 0 z z p0 z p1 z p2 z pN Ai = (z pi ) H (z ) i = 0,1,..., N z z = pi

45

H (z ) =

1 z (3 / 2)z + 1 <| z |< 2 2 z +1 z + 2 1 3 h[n ] = [n ] (1)n u[n ] (2)n u[n 1] 2 2

46

Transform Properties
Linearity Shift
Exponential Weighting

Linear Weighting

Time Reversal Convolution

ax1[n]+bx2[n] x[n-n0] anx[n] n x[n] x[-n]

aX1(z)+bX2(z)
z n0 X ( z)

non-causal, need x[N0-n] to be causal for finite length sequence

x[n] * h[n] Multiplication of x[n] w[n] Sequences

X(a-1z) -z dX(z)/dz X(z-1) X(z) H(z)

1 2 j
C circular convolution in the frequency
domain

X (v )W (z / v )v

dv

DiscreteTime Fourier Transform (DTFT)

48

47

Discrete-Time Fourier Transform

X (e j ) = X (z ) |z =e j =

n =

Simple DTFTs
Impulse Delayed impulse Step function

x[n]e

jn

x[n ] = [n ],

X (e j ) = 1

x[n ] = [n n0 ], X (e j ) = e jn0 1 1 e j 1 e jN x[n ] = u[n ] u[n N ], X (e j ) = 1 e j 1 x[n ] = a n u[n ], X (e j ) = , a <1 1 ae j 1 , b >1 x[n ] = b n u[ n 1], X (e j ) = 1 be j x[n ] = u[n ], X (e j ) =

z = e j ; | z |= 1, arg(z ) = j x[ n ] = 1 2

X (e
n

)e jn d

Rectangular window

evaluation of X(z) on the unit circle in the z-plane sufficient condition for existence of Fourier transform is:
Exponential

n =

| x[n] | | z

|=

n =

| x[n] |< , since

| z |= 1
49

Backward exponential

50

DTFT Examples
x[n ] = nu[n ] = 0.9 1 X (e j ) = | |< 1 1 e j

DTFT Examples
x [n ] = cos(0 n ), X (e ) =
j k =

< n <
0

[ (

+ 2 k ) + ( + 0 + 2 k )]

Within interval < < , X (e j ) is comprised of a pair of impulses at 0

x[n ] = cos(0 n )

51

52

DTFT Examples
x[n ] = rect M [n ]

DTFT Examples

h[n ]

53

54

Fourier Transform Properties

0.25, /2, FS/ 4, FS/ 2

Periodic DT Signals
A signal is periodic with period N if x[n] = x[n+N] for all n For the complex exponential this condition becomes j (0 n +0 N ) j0 n j0 ( n + N )

periodicity in

0. 5, , FS/ 2, FS 0,0,0,0 1, 2, FS, 2 FS

X (e j ) = X (e j ( + 2 n ) ) f, , f period of 2 corresponds to once around

0.75, 3/2, 3FS/ 4, 3FS/ 2

D,

unit circle in the z -plane

Units of Frequency (Digital Domain) (Trap #2 - loss of FS) normalized frequency: f, 00.51 (independent of FS) normalized radian frequency: , 0 2 (independent of FS) digital frequency: fD= f *FS, 0FS/2FS digital radian frequency: D= *FS, 0 FS2FS
55

Ae

= Ae

= Ae

which requires 0N = 2 k for some integer k Thus, not all DT sinusoids are periodic! Consequence: there are N distinguishable frequencies with period N
e.g., k = 2 k/N, k=0,1,,N-1
56

Periodic DT Signals
Example 1: Fs = 10, 000 Hz Is the signal x[n] = cos(2 100n / Fs ) a periodic signal? If so, what is the period.
Solution: If the signal is periodic with period N , then we have: x[ n] = x[ n + N ] cos(2 100n / Fs ) = cos(2 100( n + N ) / Fs ) 2 100 N = 2 k (k an integer ) Fs k= 100 N 100 N N = = 10, 000 100 Fs

Periodic DT Signals
Example 2: Fs = 11059 Hz; Is the signal x[n] = cos(2 100n / Fs ) periodic? If so, what is the period.

Solution: If the signal is periodic with period N , then we have: x [ n ] = x[ n + N ] cos(2 100n / Fs ) = cos(2 100( n + N ) / Fs ) 2 100 N = 2 k ( k an integer ) Fs k= 100 N 100 N = 11, 059 Fs

For k an integer we get N = 100k = 100 (for k = 1) Thus x[ n] is periodic of period 100 samples.

For k an integer we get N =

57

11059 k which is not an integer 100 Thus x[n ] is not periodic at this sampling rate.

58

Periodic DT Signals
Example 3: Fs = 10, 000 Hz Is the signal x[n] = cos(2 101n / Fs ) a periodic signal? If so, what is the period.
Solution: If the signal is periodic with period N , then we have: x[ n] = x[ n + N ] cos(2 101n / Fs ) = cos(2 101(n + N ) / Fs ) 2 101N = 2 k ( k an integer ) Fs k= 101N 101N = which is not an integer 10, 000 Fs
59

Periodic Sequences??

Thus x[ n] is not periodic at this sampling rate.

60

10

Discrete Fourier Transform

consider a periodic signal with period N (samples)

% [n ] = x %[n + N ], < n < x

The DFT Discrete Fourier Transform

61

%[n ] can be represented exactly by a discrete x

sum of sinusoids
N 1 n =0

% [k ] = x X %[n]e j 2 kn / N % [n ] = x 1 % [k ]e j 2 kn / N X N k =0
N 1

N sequence values

N DFT coefficients

exact representation of the discrete periodic sequence

62

DFT Unit Vectors (N=8)

k = 0; e j 2 k /8 = 1 k = 1; e j 2 k /8 = k = 2; e
j 2 k /8

DFT Examples

2 (1 j ) 2 =j

[ n]
1

DFT{ [n]}
1

2 k = 3; e = (1 j ) 2 j 2 k /8 k = 4; e = 1
j 2 k /8

k = 5; e j 2 k /8 = k = 6; e j 2 k /8 = j k = 7; e j 2 k /8 =

2 ( 1 + j ) 2 2 (1 + j ) 2
63

64

DFT Examples
x[n] X [k]

DFT Examples
%[n ] = (0.9)n 0 n 31 (N = 32) x

65

66

11

Circularly Shifting Sequences

x[n] x[n-2] x((n-2))

Review
DTFT of sequence { x[n ], < n < } X (e j ) = x[n ] =
n =

x [n ] e

j n

x[n-1]

x((n-1))

x[-n]

x((-n))

1 X (e j ) e jn 2 % [n], 0 n N 1} DFT of periodic sequence { x % [k ] = x X % [n] e j 2 nk /N , 0 k N 1

n =0 N 1

1 N 1 % % [n ] = X x [k ] e j 2 nk / N , 0 n N 1 N k =0
67 68

Finite Length Sequences

consider a finite length (but not periodic) sequence, x[n], that is zero outside the interval
0 n N 1

DFT for Finite Length Sequences

X (z ) = x[n ]z n
n =0

N 1

evaluate X(z) at N equally spaced points on the unit circle,

zk = e j 2 k / N , k = 0,1, ..., N 1 )= N 1

69

x[n ]e j 2 kn /N , k =0,1,...,N 1 n =0 70 --looks like DFT of periodic sequence!

X [k ]=X (e

j 2 k / N

Relation to Periodic Sequence

% [n ], consisting of an infinite -consider a periodic sequence, x sequence of replicas of x[n]

a m p litu d e 1 0.9 0.8 0.7 0.6

Periodic and Finite Length Sequences

~
x[n]

% (n ) = x

r =

x [n + r N ]

0.5 0.4 0.3 0.2

X[k]

0.1 0

8 10 s ample number

12

14

16

18

% [k ], are then identical to the values - the Fourier coefficients, X of X (e j 2 k /N ) for the finite duration sequence => a sequence
x[n]

periodic signal => line spectrum in frequency

n
k

of length N can be exactly represented by a DFT representation

of the form:

X(ej) finite duration => continuous spectrum in frequency

N 1 j 2 nk / N X [k ] = x[n ]e , k = 0,1,..., N 1 n=0 1 N 1 j 2 kn / N , n = 0,1,..., N 1 x[ n ] = X [k ]e Nk =0

Works for both finite sequence and for periodic sequence

71

72

12

Sampling in Frequency (Time Domain Aliasing)

Consider a finite duration sequence: x[n ] 0 for 0 n L 1 i.e., an L point sequence, with discrete time Fourier transform X (e j ) = x[n ] e jn 0 2
n =0 L 1

Sampling in Frequency (Time Domain Aliasing)

S(e )
. . .
0 1 2 3 4 N-1 k =2k/N

If the duration of the finite duration signal satisfies the relation N L, then only the first term in the infinite summation affects the interval 0 n L 1 and there is no time domain aliasing, i.e., %[n ] = x[n ] 0 n L 1 x If N < L, i.e., the number of frequency samples is smaller than the duration of the finite duration signal, then there is time domain aliasing and the resulting aliased signal (over the interval 0 n L 1) satisfies the aliasing relation: %[n ] = x[n ] + x[n + N ] + x[n N ] 0 n N 1 x

Consider sampling the discrete time Fourier transform by multiplying it by a signal that is defined as: S (e ) = [ 2 k / N ]
j k =0 N 1

s[n]

. . .
-N 0 N 2N

. . .
n

with time-domain representation s[n ] =

r =

[n rN ]

Thus we form the spectral sequence % (e j ) = X ( e j ) S (e j ) X which transforms in the time domain to the convolution % [n ] = x[n ] s[n ] = x[n ] [n rN ] = x
r = r =

x[n rN ]
73 74

%[n ] = x[n ] + x[n N ] + x[n + N ] + ... x

Time Domain Aliasing Example

Consider the finite duration sequence x[n ] = (m + 1) [n m] = [n ] + 2 [n 1] + 3 [n 2] + 4 [n 3] + 5 [n 4]
m =0 4

DFT Properties
Periodic Sequence Finite Sequence

Period=N Length=N Sequence defined for all n Sequence defined for n=0,1,,N-1 DFT defined for all k DTFT defined for all

-2 -1

The discrete time Fourier transform of x[n] is computed and sampled at N frequencies around the unit circle. The resulting sampled Fourier transform is inverse %[n], transformed back to the time domain. What is the resulting time domain signal, x (over the interval 0 n L 1) for the cases N = 11, N = 5 and N = 4.
SOLUTION: % [ n ] = x[ n ] For the cases N = 11 and N = 5, we have no aliasing (since N L) and we get x over the interval 0 n L 1. For the case N = 4, the n = 0 value is aliased, giving %[0] = 6 (as opposed to 1 for x[0]) with the remaining values unchanged. x
75

when using DFT representation, all sequences behave as if they

were infinitely periodic => DFT is really the representation of the % [n ] = extended periodic function, x

r =

x[n + rN ]

alternative (equivalent) view is that all sequence indices must be interpreted modulo N % [n ] = x
r =

x[n + rN ] = x[n modulo N ] = x ([n])

76

DFT Properties for Finite Sequences

X[k], the DFT of the finite sequence x[n], can be viewed as a sampled version of the z-transform (or Fourier transform) of the finite sequence (used to design finite length filters via frequency sampling method) the DFT has properties very similar to those of the z-transform and the Fourier transform the N values of X[k] can be computed very efficiently (time proportional to N log N) using the set of FFT methods DFT used in computing spectral estimates, correlation functions, and in implementing digital filters via convolutional methods
77

DFT Properties
N-point sequences N-point DFT

1. Linearity 2. Shift

ax1[n ] + bx2 [n ] x( [n n0 ])N

aX1[k ] + bX 2 [k ] e j 2 kn0 / N X [k ] X [k ]
N

3. Time Reversal x ([n ])N 4. Convolution 5. Multiplication

m =0

x[m] h([n m])

N 1

X [k ] H [k ] 1 N 1 X [r ]W ([k r ])N N r =0

x[n ] w [n ]

78

13

Key Transform Properties

y [n ] = x1[n ] x2 [n ] Y (e j ) = X1(e j ) X 2 (e j ) convolution
j

Sampling Function
x[n]
...
0

multiplication
j j

...

y [n ] = x1[n ] x2 [n ] Y (e ) = X1(e ) X 2 (e ) multiplication circular convolution

2M

Special Case: x2 [n ] = impulse train of period M samples x2 [ n ] =

k = M 1 n =0

[k nM ]
j 2 nk / M

X [k]
...
0 1 2 3 4 5 M-1 (M-1)2 M
80

X 2 [k ] = [n ] e

= 1, k = 0,1,..., M 1

1 M 1 1 M 1 x2 [n ] = X 2 [k ] e j 2 nk / M = e j 2 nk / M sampling function M k =0 M k =0

0 2 4 6 M M M
79

Summary of DSP-Part 1
speech signals are inherently bandlimited => must sample appropriately in time and amplitude LTI systems of most interest in speech processing; can characterize them completely by impulse response, h(n) the z-transform and Fourier transform representations enable us to efficiently process signals in both the time and frequency domains both periodic and time-limited digital signals can be represented in terms of their Discrete Fourier transforms sampling in time leads to aliasing in frequency; sampling in frequency leads to aliasing in time => when processing time-limited signals, must be careful to sample in frequency at a sufficiently high rate to avoid time-aliasing
81

Digital Filters

82

Digital Filters
digital filter is a discrete-time linear, shift invariant system with input-output relation:
y [n ] = x[n ] h[n ] = Y (z ) = X (z) H (z) H (z) is the system function with H (e j ) as the complex frequency response H (e j ) = Hr (e j ) + jHi (e j ) H (e j ) =| H (e j ) | e j arg|H ( e
j

Digital Filters
causal linear shift-invariant => h[n]=0 for n<0 stable system => every bounded input produces a bounded output => a necessary and sufficient condition for stability and for the existence of
H (e j )

m =

x[m] h[n m]

real, imaginary representation magnitude, phase representation

)|

log H (e j ) = log | H (e j ) | + j arg | H (e j ) |

j log | H (e j ) |= Re log H (e ) j j arg | H (e j ) |= Im log H (e )

n =
83

| h[n] |<
84

14

Digital Filter Implementation

input and output satisfy linear difference equation of the form:
y [n ] ak y [n k ] = br x[n r ]
k =1 r =0 N M

Digital Filters
1 H(z) is a rational function in z
O

evaluating

z-transforms of both sides gives:

Y (z) ak z kY (z ) = br z r X (z)
k =1 r =0 N M

H (z) =

A (1 cr z )
1

Y (z) (1 ak z k ) = X (z) br z r
k =1 r =0

(1 dk z 1 )
k =1

r =1 N

=> M zeros, N poles

O O O

X X

Y (z ) H (z) = = X (z)

b z
r =0 N r k =1

converges for | z |> R1 , with R1 < 1 for stability

canonic form showing poles and zeros
85

=> all poles of H (z ) inside the unit circle for a stable, causal system

1 ak z k

86

Ideal Filter Responses

if ak=0, all k, then
M r =0

FIR Systems
y [n ] = br x[n r ] = b0 x[n ] + b1 x[n 1] + ... + bM x[n M ] 1. h[n ] = bn 0 n M =0
M n =0

=>

otherwise

2. H (z ) = bn z n =>

(1 c
m =0

M 1

z 1 ) => M zeros

3. if h[n ] = h[M n ] (symmetric, antisymmetric) H(e j ) = A(e j )e jM / 2 , A(e j ) = real (symmetric), imaginary (anti-symmetric)

linear phase filter => no signal dispersion because of non-linear phase => precise time alignment of events in signal event at t0 FIR Linear Phase Filter event at t0 + fixed delay

87

88

FIR Filters
cost of linear phase filter designs
can theoretically approximate any desired response to any degree of accuracy requires longer filters than non-linear phase designs

Window Designed Filters

Windowed impulse response h[n ] = hI [n ] w [n ] In the frequency domain we get H (e j ) = HI (e j ) W (e j )

FIR filter design methods

window design => analytical, closed form method frequency sampling => optimization method minimax error design => optimal method

89

90

15

LPF Example Using RW

LPF Example Using RW

0.2

0.4
Frequency

0.6

0.8

91

0.2

0.4
Frequency

0.6

0.8

92

LPF Example Using RW

Common Windows (Time)

1 1. Rectangular w [n ] = 0 2. Bartlett 3. Blackman w [n ] = 1 0nM otherwise 2|n M / 2| M 2 n 4 n w [n ] = 0.42 0.5cos + 0.08cos M M 2 n w [n ] = 0.54 0.46 cos M 2 n w [n ] = 0.5 0.5cos M w [n ] = I0 1 ((n M / 2) / (M / 2)) 2 I0 { }

0.25
Frequency

0.5

0.75

4. Hamming 5. Hanning

6. Kaiser
0 0.25
Frequency

}
94

0.5

0.75

93

Common Windows (Frequency)

Window LPF Example

Window

Mainlobe Width

Sidelobe Attenuation 13 dB 27 dB 32 dB 43 dB 58 dB
0 0.2 0.4
Frequency

Rectangular 4 / M Bartlett 8 /M Hanning Hamming Blackman 8 /M 8 /M 12 /M

0.6

0.8

95

0.2

0.4
Frequency

0.6

0.8

96

16

Equiripple Design Specifications

Optimal FIR Filter Design

Equiripple in each defined band (passband and stopband for lowpass filter, high and low stopband and passband for bandpass filter, etc.) Optimal in sense that the cost function 1 ( ) | H () H () | d E= 2 is minimized. Solution via well known iterative algorithm based on the alternation theorem of Chebyshev approximation.
2 d

p = normalized edge of passband frequency s = normalized edge of stopband frequency p = peak ripple in passband s = peak ripple in stopband =s p = normalized transition bandwidth
97

98

MATLAB FIR Design

1. 2. Use fdatool to design digital filters Use firpm to design FIR filters 3. 4. B=firpm(N,F,A) N+1 point linear phase, FIR design B=filter coefficients (numerator polynomial) F=ideal frequency response band edges (in pairs) (normalized to 1.0) A=ideal amplitude response values (in pairs) [H,W]=freqz(B,den,NF) H=complex frequency response W=set of radian frequencies at which FR is evaluated (0 to pi) B=numerator polynomial=set of FIR filter coefficients den=denominator polynomial=[1] for FIR filter NF=number of frequencies at which FR is evaluated plot(W/pi, 20log10(abs(H)))

Remez Lowpass Filter Design

Use freqz to convert to frequency response (complex)

N=30 F=[0 0.4 0.5 1]; A=[1 1 0 0]; B=firpm(N,F,A) NF=512; number of frequency points [H,W]=freqz(B,1,NF); 99 plot(W/pi,20log10(abs(H))); 100

Use plot to evaluate log magnitude response

Remez Bandpass Filter Design

% bandpass_filter_design N=input('Filter Length in Samples:'); F=[0 0.18 .2 .4 .42 1]; A=[0 0 1 1 0 0]; B=firpm(N,F,A); NF=1024; [H,W]=freqz(B,1,NF); figure,orient landscape; stitle=sprintf('bandpass fir design, N:%d,f: %4.2f %4.2f %4.2f %4.2f %4.2f %4.2f',N,F); n=0:N; subplot(211),plot(n,B,'r','LineWidth',2); axis tight,grid on,title(stitle); xlabel('Time in Samples'),ylabel('Amplitude'); legend('Impulse Response'); subplot(212),plot(W/pi,20*log10(abs(H)),'b','LineWidth',2); axis ([0 1 -60 0]), grid on; xlabel('Normalized Frequency'),ylabel('Log Magnitude (dB)'); legend('Frequency Response');

FIR Implementation
x[n]
x[ n 1]
x[n 2]

x[n 3]

x[n M ]

y[n]

linear phase filters can be implemented with half the multiplications (because of the symmetry of the coefficients)
101 102

17

IIR Systems
y [n ] = ak y [n k ] + br x[n r ]
k =1 r =0 N M

IIR Design Methods

Impulse invariant transformation match the analog impulse response by sampling; resulting frequency response is aliased version of analog frequency response Bilinear transformation use a transformation to map an analog filter to a digital filter by warping the analog frequency scale (0 to infinity) to the digital frequency scale (0 to pi); use frequency pre-warping to preserve critical frequencies of transformation (i.e., filter cutoff frequencies) 104

y [n ] depends on y [n 1], y [n 2],..., y [n N ] as well as x[n ], x[n 1],..., x[n M ] for M < N H (z ) =

b z
r =0 N r k =1 N

1 ak z

=
k =1

Ak - partial fraction expansion 1 dk z 1

h[n ] = Ak (d k )n u[n ] - for causal systems

k =1

h[n ] is an infinite duration impulse response

103

IIR Filter Design

Butterworth Design

0.2

0.4
Frequency

0.6

0.8

105

0.2

0.4
Frequency

0.6

0.8

106

Chebyshev Type I Design

Chebyshev BPF Design

0.2

0.4
Frequency

0.6

0.8

107

0.2

0.4
Frequency

0.6

0.8

108

18

Chebyshev Type II Design

Elliptic BPF Design

0.2

0.4
Frequency

0.6

0.8

0.2

0.4
Frequency

0.6

0.8

0.2

0.4
Frequency

0.6

0.8

109

0.2

0.4
Frequency

0.6

0.8

110

IIR Filters
IIR filter issues:
efficient implementations in terms of computations can approximate any desired magnitude response with arbitrarily small error non-linear phase => time dispersion of waveform

Matlab Elliptic Filter Design

use ellip to design elliptic filter
[B,A]=ellip(N,Rp,Rs,Wn) B=numerator polynomialN+1 coefficients A=denominator polynomialN+1 coefficients N=order of polynomial for both numerator and denominator Rp=maximum in-band (passband) approximation error (dB) Rs=out-of-band (stopband) ripple (dB) Wp=end of passband (normalized radian frequency)

IIR design methods

Butterworth designs-maximally flat amplitude Bessel designs-maximally flat group delay Chebyshev designs-equi-ripple in either passband or stopband Elliptic designs-equi-ripple in both passband and stopband
111

use filter to generate impulse response

y=filter(B,A,x) y=filter impulse response x=filter input (impulse)

use zplane to generate pole-zero plot

zplane(B,A)
112

Matlab Elliptic Lowpass Filter

IIR Filter Implementation

M=N=4

y [n ] = ak y [n k ] + br x[n r ]
k =1 N r =0

w [n ] = ak w [n k ] + x[n ]
k =1 M

y [n ] = br w [n r ]
[b,a]=ellip(6,0.1,40,0.45); [h,w]=freqz(b,a,512); x=[1,zeros(1,511)]; y=filter(b,a,x); zplane(b,a); appropriate plotting commands; 113
r =0

114

19

IIR Filter Implementations

H (z) = A (1 cr z 1 )
N

IIR Filter Implementations

H (z) = c0 k + c1k z 1 , parallel system 1 a2 k z 2 k =1 1 a1k z
K
c01

(1 d z
k =1 k K

r =1 N

- zeros at z = cr , poles at z = d k

- since ak and br are real, poles and zeros occur in complex conjugate pairs =>
b0k + b1k z + b2 k z N + 1 , K = 1 a1k z 1 a2k z 2 2 - cascade of second order systems H(z) = A
k =1 1 2

a21

Used in formant synthesis systems based on ABS methods

115

c02

Common form for speech synthesizer implementation

a12

c12

116

DSP in Speech Processing

filtering speech coding, post filters, pre-filters, noise reduction spectral analysis vocoding, speech synthesis, speech recognition, speaker recognition, speech enhancement implementation structures speech synthesis, analysis-synthesis systems, audio encoding/decoding for MP3 and AAC sampling rate conversion audio, speech
DAT 48 kHz CD 44.06 kHz Speech 6, 8, 10, 16 kHz Cellular TDMA, GSM, CDMA transcoding
117

Sampling of Waveforms
xa(t) Sampler and Quantizer Period, T x[n],x(nT)

x[n ] = xa (nT ), < n < T = 1/ 8000 sec =125 sec for 8kHz sampling rate T = 1/10000 sec = 100 sec for 10 kHz sampling rate T = 1/16000 sec = 67 sec for 16 kHz sampling rate T = 1/ 20000 sec = 50 sec for 20 kHz sampling rate
118

The Sampling Theorem

If a signal xa(t) has a bandlimited Fourier transform Xa(j) such that Xa(j)=0 for 2FN, then xa(t) can be uniquely reconstructed from equally spaced samples xa(nT), -<n<, if 1/T2 FN (FS2FN) (A-D or C/D converter)
xa(t)

Sampling Theorem Equations

xa (t ) X a ( j ) = x[n ] X (e j T ) =

(t ) e j t dt (nT )e j nT

xa(nT)

n =

1 X (e j T ) = T
xa(nT) = xa(t) uT(nT), where uT(nT) is a periodic pulse train of period T, with periodic spectrum of period 2/T
119

k =

( j + j 2 k / T )
120

20

Sampling Theorem Interpretation

Sampling Rates
FN = Nyquist frequency (highest frequency with significant spectral level in signal) must sample at at least twice the Nyquist frequency to prevent aliasing (frequency overlap)
telephone speech (300-3200 Hz) => FS=6400 Hz wideband speech (100-7200 Hz) => FS=14400 Hz audio signal (50-21000 Hz) => FS=42000 Hz AM broadcast (100-7500 Hz) => FS=15000 Hz

To avoid aliasing need: 2 / T N > N 2 / T > 2 N Fs = 1/ T > 2FN

case where 1 / T < 2FN , aliasing occurs

can always sample at rates higher than twice the Nyquist frequency (but that is wasteful of processing)
122

121

Recovery from Sampled Signal

If 1/T > 2 FN the Fourier transform of the sequence of samples is proportional to the Fourier transform of the original signal in the baseband, i.e.,

Decimation and Interpolation of Sampled Waveforms

CD rate (44.06 kHz) to DAT rate (48 kHz)media conversion Wideband (16 kHz) to narrowband speech rates (8kHz, 6.67 kHz)storage oversampled to correctly sampled rates--coding

X (e j T ) =

1 X a ( j ), | | < T T

can show that the original signal can be recovered from the sampled signal by interpolation using an ideal LPF of bandwidth /T, i.e.,

x a (t ) =

sin( (t nT ) / T ) xa (nT ) (t nT ) / T n =

bandlimited sample interpolationperfect at every sample point, perfect in-between samples via interpolation

x[n ] = xa (nT ), X a ( j ) = 0 for | |> 2 FN if 1/ T > 2FN (adequate sampling) then X(e j T ) =

digital-to-analog converter
123

1 X a ( j ), | |< T T
124

Decimation and Interpolation

Decimation
Standard Sampling: begin with digitized signal: x[ n] = xa (nT ) X a ( j) = 0, | | 2 FN ( a) 1 2 FN T 1 X (e ) = X a ( j), | |< T T X (e jT ) = 0, 2 FN | | 2 ( Fs FN ) Fs =
j T

Decimation, M=2 => T=2T

(b)

can achieve perfect recovery of xa (t ) from digitized samples under these conditions

Interpolation, L=2 => T=T/2

125

126

21

Decimation
want to reduce sampling rate of sampled signal by factor of M 2 want to compute new signal xd [n] with sampling rate Fs ' = 1/ T ' = 1( MT ) = Fs / M such that xd [n] = xa (nT '' ) with no aliasing one solution is to downsample x[n] = xa (nT ) by retaining one out of every M samples of x[n], giving xd [n] = x[nM ]

Decimation

need Fs ' 2 FN to avoid aliasing for M = 2 when Fs ' < 2 FN we get aliasing for M = 2 (c )

(d )

127

128

Decimation
DTFTs of x[n] and xd [n] related by aliasing relationship: 1 M 1 X (e j ( 2 k )/ M ) M k =0 or equivalently (in terms of analog frequency): X d (e j ) = X d (e
j T '

Decimation
to decimate by factor of M with no aliasing, need to ensure that the highest frequency in x[n] is no greater than Fs / (2M ) thus we need to filter x[n] using an ideal lowpass filter with response: | |< / M 1 H d (e j ) = 0 / M <| | using the appropriate lowpass filter, we can downsample the reuslting lowpass-filtered signal by a factor of M without aliasing

1 )= M

M 1 k =0

X (e

j ( T ' 2 k )/ M

assuming Fs ' =

1 2 FN , (i.e., no aliasing) we get: MT 1 1 1 1 j T ' jT '/ M )= X d (e ) = X (e X (e jT ) = X a ( j) M M MT 1 = X a ( j), < < T' T' T'

129

130

Decimation
using a lowpass filter gives: 1 Wd (e jT ' ) = H d (e jT ) X a ( j), < < T' T' T' if filter is used, the down-sampled signal, wd [ n], no longer represents the original analog signal, xa (t ), but instead the lowpass filtered version of xa (t ) the combined operations of lowpass filtering and downsampling are called decimation.

Interpolation
assume we have x[n] = xa ( nT ), (no aliasing) and we wish to increase the sampling rate by the integer factor of L we need to compute a new sequence of samples of xa (t ) with period T '' = T / L, i.e., xi [ n] = xa (nT '') = xa ( nT / L) It is clear that we can create the signal xi [ n] = x[ n / L] for n = 0, L, 2 L,... but we need to fill in the unknown samples by an interpolation process can readily show that what we want is: sin[ (nT '' kT ) / T ] [ ( nT '' kT ) / T ] equivalently with T '' = T / L, x[n] = xa ( nT ) we get xi [ n] = xa (nT '') =
k =

x (kT )
a

xi [ n] = xa (nT '') =
131

k =

x (k )
a

sin[ (n k ) / L] [ ( n k ) / L]
132

which relates xi [ n] to x[ n] directly

22

Interpolation
implementing the previous equation by filtering the upsampled sequence x [n / L] n = 0, L, 2 L,... xu [ n] = otherwise 0 xu [ n] has the correct samples for n = 0, L, 2 L,..., but it has zero-valued samples in between (from the upsampling operation) The Fourier transform of xu [n] is simply: X u (e j ) = X (e j L ) X u (e
j T ''

Interpolation
(a ) Plot of X ( e jT ) (b) Plot of X u (e jT '' ) showing double periodicity for L = 2, T '' = T / 2 (c ) DTFT of desired signal with | | 2 FN (2 / T ) X a ( j) X i ( e jT '' ) = 0 2 FN <| | / T '' can obtain results of ( c) by applying ideal lowpass filter with gain L (to restore amplitude) and cutoff frequency 2 FN = / T , giving: (1 / T '') X (e j L ) 0 | |< / L X i ( e j ) = 0 / L | | | |< / L L H i ( e j ) = 0 / L | |
134

) = X (e

j T '' L

) = X (e

j T

Thus X u (e jT '' ) is periodic with two periods, namely with period 2 / L, due to upsampling) and 2 due to being a digital signal

133

Interpolation
Original signal, x[n], at sampling period, T , is first upsampled to give signal xu [n] with sampling period T '' = T / L lowpass filter removes images of original spectrum giving: xi [ n] = xa (nT '') = xa ( nT / L)

SR Conversion by Non-Integer Factors

T=MT/L => convert rate by factor of M/L need to interpolate by L, then decimate by M (why cant it be done in the reverse order?)

Lowpass Filter

Interpolation LPF

Decimation LPF

need to combine specifications of both LPFs and implement in a single stage of lowpass filtering

can approximate almost any rate conversion with appropriate values of L and M

135

for large values of L, or M, or both, can implement in stages, i.e., L=1024, use L1=32 followed by L2=32

136

Summary of DSP-Part II
digital filtering provides a convenient way of processing signals in the time and frequency domains can approximate arbitrary spectral characteristics via either IIR or FIR filters, with various levels of approximation can realize digital filters with a variety of structures, including direct forms, serial and parallel forms once a digital signal has been obtain via appropriate sampling methods, its sampling rate can be changed digitally (either up or down) via appropriate filtering and decimation or interpolation

137

23