DSP2

Module 2
Discrete Fourier Transform
Digital Signal Processing. Slide 2.1
Contents
Review of Fourier analysis

Fourier series Fourier transform

Discrete-time Fourier transform Discrete Fourier transform

Analysis, properties and applications Short-time Fourier transform Windowing Decimation-in-time FFT
Spectral estimation

Fast computation of DFT

Review of Fourier Analysis

Analysis of signals in terms of their components at different frequencies

What are signals?

Information bearing functions of time usually observed using transducers (microphone, antenna) often represented by analogue voltages or sampled data streams
What are frequency components?

Fourier analysis considers signals to be constructed from a sum of complex exponentials with appropriate frequencies, amplitudes and phase frequency components are the complex exponentials (sines and cosines) which, when added together, make up the signal
Fourier Series
For a periodic continuous-time signal x(t) with period T

x(t ) = A0 + Ak cos k 0t + Bk sin k 0t
k =1 k =1
This can be written in exponential form x(t ) = The ak are found from ak =
1 x(t )e jk 0t dt TT
k =
a e
k
jk 0t
x(t) is made from a d.c. term plus a weighted sum of sinusoids at integer multiples of the fundamental frequency
0 = 2 T
Fourier Transform
For an aperiodic signal, we can say that it is equivalent to a periodic signal with infinite period
hence the Fourier Series becomes the Fourier Transform
1 x(t ) = 2 X () =
X ( )e
jt
jt
x(t )e
dt
Existence of Fourier Transform

A (Dirichlet Conditions)

x(t) is absolutely integrable x(t) has a finite number of maxima and minima in any finite interval x(t) has a finite number of discontinuities within any finite interval and all discontinuities are finite
x(t ) dt <
B
If
1. X() is finite
x(t ) dt < then
2.
e(t )
dt = 0
(t ) x(t ) for e(t ) = x (t ) = IFT {X ( )} where x

Discrete-Time Signals and the Fourier Transform

So far, so good - mostly review Now consider a discrete-time signal x(n) and its ztransform X(z)

x(n) is just a sequence of numbers what does it mean for a sequence of numbers to have a spectrum?
How can the spectrum of a discrete-time signal (sequence) be found using the Fourier transform when the Fourier transform is defined for continuous-time signals?
Discrete-Time Fourier Transform

Consider a discrete-time signal x(n) and a continuous-time signal x'(t) given by *

x' (t ) =
Now
X ' (e ) = = =
j
n =
x(n) (t nT )
j t
x' (t )e
dt
j t
n =
x ( n) e
(t nT )dt
n =
x ( n )e
jnT
Hence

If x(n) is an infinite sequence, we can compute the discrete-time Fourier transform (DTFT) a sequence of numbers having a spectrum is OK if we use the trick of * the DTFT is a continuous function of frequency
Inverse DTFT
There exists and inverse DTFT

x ( n) = T 2
2 T
X ' (e
)e jnT d
where
2 T
means integrate over any 2 range of
Discrete Fourier Transform

The DTFT is defined using an infinite sum over a discretetime signal and yields a continuous function
not very useful since we dont often have real signals of infinite length
Now introduce the DFT which is defined using a sum over only N (finite) samples of a discrete-time signal The DFT yields a discrete function of frequency with points uniformly spaced around the unit circle in z
The DFT yields a sampled version of the DTFT
Im
X(2) X(3) X(1)
X ( k ) = X ( e j )
Re Unit Circle
= 2k N
X(4)
X(0)
X(5) X(6)
X(7)

DTFT yields a continuous function

X() is defined at all values of z around the unit circle X() is defined only at certain points around the unit circle
DFT yields sampled version

Applications of the DFT

Spectrum analysis of periodic discrete-time signals Short-time spectral estimation of aperiodic signals
normally in conjunction with a window
Fast implementation of convolution (filtering) via fast algorithm for DFT (FFT).
Derivation of DFT from DTFT

DTFT: X (e ) =
n =
x ( n )e
jnT
Consider only N samples of x(n), n=0, 1, N-1 and compute only N samples of X (e j ) using
= k 0
k = 0,1,..., N 1
0 = 2 NT
Sampling frequency divided by N
giving
X (e
jk 0
) = x(n)e jkn 0T
n =0 N 1 n =0
N 1
which can be written
X (k ) = x(n)W kn
k = 0,1,..., N 1
W = e j 2
Inverse DFT
There exists an inverse relationship

x ( n) = 1 N
X (k )W
k =0
N 1
kn
n = 0 ,1,...,N-1
Notes and Summary

Notes

X(k) is the DFT of x(n) X(k) is an N-point sequence computed from x(n), another N-point sequence k and n are dimensionless variables k is the frequency index, n is the time index
Summary
X (k ) = x(n)e j 2kn N
n =0 N 1
k = 0 ,1,...,N 1 n = 0,1,..., N 1
1 x ( n) = N
X ( k )e
k =0
N 1
j 2kn N
Properties of the DFT

Linearity:
Ax1 (n) + Bx2 (n) AX 1 (k ) + BX 2 (k )
A, B, arbitrary constants
Delay:
x(n m) mod N X (k )W km
i integer
where x(q ) mod N = x(q iN )
Modulation:
W nl x(n) X (k l ) mod N
Modulation in one domain implies circular shift in the other
Convolution:
m =0
x ( m) x ( n m)
1 2
N 1
mod N
X 1 (k ) X 2 (k )
circular convolution of two sequences in the product of their DFTs
Multiplication:
1 x1 (n) x2 (n) N
X (l ) X
l =0 1
N 1
(k l ) mod N
multiplication in one domain implies circular convolution in the other
Conjugation:
x* (n) X * ( k ) mod N
Spectral Estimation using DFT

Consider a speech signal s(t):

14000 12000 10000 8000 6000 4000 2000 0 -2000 -4000 -6000 0
My name is Michael Cane
amplitude spectrum is controlled by the voice source (vocal cords) and the articulators (lips, tongue etc) s[n] is a sampled version of s(t) at sampling frequency fs Hz, find the amplitude and phase spectra at time t1. Assume the spectrum of s(t) is constant near t1.
Suppose that:
Question:
0.5
1.5
2.5 x 10
4
1. Windowing

we must select N samples near t1. use a windowing function w(n)

eg: rectangular, hanning, hamming etc N is the size of the window position the window in the region of t1 and multiply the signal by the window. The result is a new data record s1(n) of length N. N is chosen to be an integer power of 2 so that the DFT can be implemented using a fast algorithm (FFT).
create a new data record s1 (n) = s (m).w(n)

n = 0, 1, , N 1 m = m1 N , , m1 + N 1 2 2
from the properties of the DFT

2 1.5 1 0.5 amplitude 0 -0.5 -1 -1.5 -2
multiplying the signal by a window is equivalent to convolving their Fourier transforms using a rectangular window causes convolution with a sin( x ) x the errors caused by the convolution are called leakage
Signal
Spectrum of Signal 100 50 0 -50 -100 -150 -200 -250 -300
amplitude dB
500 sample index Window
1000
1500
0.1
0.2
0.3
2 1.5 1 0.5
0.4 0.5 0.6 normalised frequency Spectrum of Window
0.7
0.8
0.9
1 Spectrum after windowing 50 40 30 20 amplitude dB 10 0 -10 -20 -30 -40
50 40 30 20 amplitude dB
amplitude
10 0 -10 -20 -30
0 -0.5 -1 -1.5
-40
-2 0 500 sample index 1000 1500
-50
0.1
0.2
0.3
0.4 0.5 0.6 normalised frequency
0.7
0.8
0.9
-50
0.1
0.2
0.3
0.4 0.5 0.6 normalised frequency
0.7
0.8
0.9

the rectangular window produces relatively narrow main lobes with high amplitude side-lobes (~ -13dB) the rectangular window introduces no leakage in the special case where x(n) is periodic with period N other windows produce broader main lobes with lower amplitude side-lobes, eg:
Hamming Hanning
2 n w(n) = 0.54 0.46 cos N 1 2 n w(n) = 0.50 0.50 cos N 1
0 n N 1
2. Discete Fourier Transform

Given compute
x1(n) = x(m) . w(n)

X 1 (k ) =
N 1 n =0
x1(n) e j 2kn / N
3. Interpret the result

what is X1(k) ? X1(k) is a sampled version of the discrete-time Fourier transform of x1(n) the samples in the frequency domain are spaced by f
s 1 = 1 f = N = NT Total time of sampling

X1(0) is the d.c. component X1(N/2) is the component at the Nyquist frequency X1(N-1) is the component at f s f
4. Present the result

normally we graph the result of a spectrum estimate several options:

power spectrum (spectrogram) S(k)

tells you how much power there is in the signal at each discrete frequency 1 X ( k ) 2 = 1 X * ( k ). X ( k ) S (k ) = N k = 0, 1, ..., N 1 N
Amplitude and phase spectrum

define the amplitude and phase spectra to be
X (k ) =
(Re{X (k )})2 + (Im{X (k )})2
X (k ) = tan 1 Re{X ( k )}
Im{X ( k )}
Computation for FFT

Computation for the DFT of a sequence requires N 2complex multiples
X [n] = x[n]e 2 nk / N , k = 0,1,...N 1

n =0
N 1
N complex multiples for each value of k Cooley and Tukey (1965) showed that the DFT could be computed using only N 2 log 2 N complex multiples when N is a power of 2 Their algorithm is called the radix-2 decimation-in-time Fast Fourier Transform
Number of complex multiples N 2 4 8 16 32 1024

N 1 n=0
DFT 4 16 64 256 1024 1048576
FFT 1 4 12 32 80 5120
2 point DFT
DFT :
X (k ) = x(n) WNnk
k = 0,1, , N 1
WN = e j 2 / N
For 2 points, N = 2, k = 0,1.
X (k ) = x(0) W20 + x(1) W2k

n=0 n=1
W2 = e j = 1
W20 = 1, W2k = (1) k

X (0) = x(0) + x(1) X (1) = x(0) x(1)
4 point DFT
X (k ) = x(n) W4nk
n =0 3
k = 0,1,2,3
= x(0)W40 + x(1)W4k + x(2)W42 k + x(3)W43k

Note that also
W40 k = e j 2 .0.k / 4 = e j 2 .0.k / 2 = W20 k
W42 k = e j 2 .2.k / 4 = e j 2 .1.k / 2 = W21k
Now consider only the even-indexed samples x(0) and x(2)

Call this even sequence
xe (0), xe (1)
x(0)W40 k + x(2)W42 k = xe (0)W20 k + xe (1)W21k

Thus the 1st and 3rd terms of X(k) are the 2 point DFT of the even samples
The 2nd and 4th terms can be written
x(1)W4k + x(3)W43k = W4k ( x(1)W40 k + x(3)W42 k ) = W4k ( x0 (0)W20 k + x0 (1)W2k )

since W42 k = W2k and where x0 (n) are the odd samples
Thus the 2nd and the 4th terms of X(k) are the 2point DFT of the odd samples Finally X (k ) = X e (k ) + W4k X 0 (k )
x(n) 2point DFT of
even samples
k = 0, 1, 2, 3
4 point DFT of
W4k 2point DFT of

odd samples
Recombination Equations
X (0) = X e (0) + W40 X 0 (0) = X e (0) + X 0 (0) X (1) = X e (1) + W41 X 0 (1) X (2) = X e (2) + W42 X 0 (2) = X e (2) X 0 (2) = X e (0) X 0 (0) X [3] = X e (3) + W43 X 0 (3) = X e (3) + W41W42 X 0 (3) X e (3) + W41 X 0 (3) X e (1) + W41 X 0 (1) because of periodicity of the DFT
since W4 0 = e j 2 0 k / 4 = 1
since W4 2 = e j 2 2 k / 4 = 1 because of periodicity of the DFT
Periodicity of the DFT

Complex exponentials are periodic functions and hence
X e (2m) = X e (0) m integer X e (2m + 1) = X e (1) m integer

Similarly for X o
Signal flow graph: 4point DFT

Step 1 : Split x(n) into an even sequence and an odd sequence Step 2 : perform 2 2point DFTs Step 3 : recombine
From the 2-point and 4-point FFT implementations, we can generalise the procedure : 1. Decompose the N-point DFT into two N/2-point DFTs and determine the recombination equations 2. Decompose each N/2-point DFT into two N/4-point DFTs and determine the recombination equations 3. Repeat the decomposition until only 2-point DFTs remain The computational saving comes from the fact that each butterfly requires only 1 complex multiplication (plus one addition and one subtraction) the number of butterflies is N 2 log 2 N , N a power of 2
Implementing an FFT on a computer

Consider the 4 point example
note that the output sequence is correctly ordered the input sequence is not correctly ordered
Data input x[0] X[2] X[1] X[3]
Binary index 00 10 01 11
this turns out to be a standard binary count but with the order of the bits reversed
Consider an 8 point case
data point x[0] x[4] x[2] x[6] x[1] x[5] x[3] x[7]
binary index 000 100 010 110 001 101 011 111
Hence the array of input data must be read in bitreversed order. If the input data is the correct order (sequential), then the output of the FFT will appear in bitreversed order
Decimation in Frequency FFT

X (k ) = x(n)WNkn
n =0 N 1
= = =
N /2 - 1
n =0
x(n)W +
kn N kn N
n= N / 2
N 1
x(n)WNkn
N /2 1
N /2 - 1
n =0 n =0
x(n)W + W
Nk / 2 N
n =0
x(n +
N kn )WN 2
N /2 - 1
N kn k x(n) + (1) x(n + ) WN 2
The decimation comes from dividing X(k) into even and odd terms
giving
N kn x(n) + x(n + ) WN / 2 k = 0,1,..., N / 2 - 1 2 n=0 N /2 - 1 N X (2k + 1) = x(n) x(n + ) WNn WNkn k = 0,1,..., N / 2 - 1 /2 2 n =0 X (2k ) =
N /2 - 1
Which can be written X (2k ) =

N /2 - 1
n=0
g1 (n)WNkn /2 g 2 (n)WNkn /2
X (2k + 1) =
N /2 - 1
n =0
4 point decimation-in-frequency
Summary
So far, we have seen radix-2 algorithms

signal is divided into elementary groups of 2 no. of points integer power of 2 no. of points integer power of 4 extra efficiency compared to radix-2 uses both radix-2 and radix-4 decomposition highly efficient but loses regular processing structure
important for VLSI
Radix-4 algorithms

Split Radix algorithms


DSP2

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

DSP2

Hochgeladen von

Copyright:

Verfügbare Formate

Module 2

Discrete Fourier Transform

Digital Signal Processing. Slide 2.1

Review of Fourier analysis

Fourier series Fourier transform

Discrete-time Fourier transform Discrete Fourier transform

Fast computation of DFT

Review of Fourier Analysis

Analysis of signals in terms of their components at different frequencies

What are signals?

What are frequency components?

Digital Signal Processing. Slide 2.3

For a periodic continuous-time signal x(t) with period T

hence the Fourier Series becomes the Fourier Transform

Digital Signal Processing. Slide 2.5

Existence of Fourier Transform

x(t ) dt < then

(t ) x(t ) for e(t ) = x (t ) = IFT {X ( )} where x

Discrete-Time Signals and the Fourier Transform

Digital Signal Processing. Slide 2.7

Discrete-Time Fourier Transform

Consider a discrete-time signal x(n) and a continuous-time signal x'(t) given by *

Digital Signal Processing. Slide 2.8

Digital Signal Processing. Slide 2.9

There exists and inverse DTFT

means integrate over any 2 range of

Digital Signal Processing. Slide 2.10

Discrete Fourier Transform

The DFT yields a sampled version of the DTFT

Digital Signal Processing. Slide 2.11

DTFT yields a continuous function

DFT yields sampled version

Digital Signal Processing. Slide 2.12

Applications of the DFT

normally in conjunction with a window

Digital Signal Processing. Slide 2.13

Derivation of DFT from DTFT

which can be written

Digital Signal Processing. Slide 2.14

There exists an inverse relationship

Digital Signal Processing. Slide 2.15

Notes and Summary

Properties of the DFT

Ax1 (n) + Bx2 (n) AX 1 (k ) + BX 2 (k )

where x(q ) mod N = x(q iN )

Modulation in one domain implies circular shift in the other

Digital Signal Processing. Slide 2.17

circular convolution of two sequences in the product of their DFTs

multiplication in one domain implies circular convolution in the other

Digital Signal Processing. Slide 2.18

Spectral Estimation using DFT

Consider a speech signal s(t):

My name is Michael Cane

Digital Signal Processing. Slide 2.19

we must select N samples near t1. use a windowing function w(n)

create a new data record s1 (n) = s (m).w(n)

Digital Signal Processing. Slide 2.20

from the properties of the DFT

Spectrum of Signal 100 50 0 -50 -100 -150 -200 -250 -300

500 sample index Window

0.4 0.5 0.6 normalised frequency Spectrum of Window

1 Spectrum after windowing 50 40 30 20 amplitude dB 10 0 -10 -20 -30 -40

10 0 -10 -20 -30

0.4 0.5 0.6 normalised frequency