2012-11-28 Dan Ellis 1

ELEN E4810: Digital Signal Processing

Topic 10:
The Fast Fourier Transform
1. Calculation of the DFT
2. The Fast Fourier Transform algorithm
3. Short-Time Fourier Transform
2012-11-28 Dan Ellis 2
1. Calculation of the DFT

Filter design so far has been oriented to

time-domain processing - cheaper!

But: frequency-domain processing

makes some problems very simple:

use all of x[n], or use short-time windows

Need an efficient way to calculate DFT

Fourier domain
processing
x[n] y[n] DFT IDFT
X[k] Y[k]
2012-11-28 Dan Ellis 3
The DFT

Recall the DFT:

discrete transform of discrete sequence

Matrix form:

X[k] = x[n]W
N
kn
n=0
N1

W
N
= e
j
2
N
( )
W
N
r
has only
N distinct values
W
N
@2/N
Structure
opportunities
for
efficiency

X[0]
X[1]
X[2]
.
.
.
X[N 1]

1 1 1 1
1 W
1
N
W
2
N
W
(N1)
N
1 W
2
N
W
4
N
W
2(N1)
N
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1 W
(N1)
N
W
2(N1)
N
W
(N1)
2
N

x[0]
x[1]
x[2]
.
.
.
x[N 1]

2012-11-28 Dan Ellis 4

Computational Complexity

N complex multiplies
+ N-1 complex adds per point (k)
N points (k = 0.. N-1)

cpx mult: (a+jb)(c+jd) = ac - bd + j(ad + bc)

= 4 real mults + 2 real adds

cpx add = 2 real adds

N points: 4N
2
real mults, 4N
2
-2N real adds

X[k] = x[n]W
N
kn
n=0
N1

2012-11-28 Dan Ellis 5

Goertzels Algorithm

Now:

i.e.
where


X k
[ ]
= x
[ ]
W
N
k
=0
N1

= W
N
kN
x
[ ]

W
N
k N
( )
looks like a
convolution

X k
[ ]
= y
k
N
[ ]
y
k
n
[ ]
= x
e
n
[ ]
h
k
[n]
x
e
[n] = {
x[n] 0 n < N
0 n = N
h
k
[n] = {
W
N
-kn
n 0
0 n < 0
+
z
-1
W
N
-k
x
e
[n]
x
e
[N] = 0
y
k
[n]
y
k
[-1] = 0
y
k
[N] = X[k]
2012-11-28 Dan Ellis 6
Goertzels Algorithm

Separate filters for each X[k]

can calculate for just a few values of k

No large buffer, no coefficient table

Same complexity for full X[k]

(4N
2
mults, 4N
2
- 2N adds)

but: can halve multiplies by making the

denominator real:

H z
( )
=
1
1W
N
k
z
1
=
1W
N
k
z
1
1 2cos
2k
N
z
1
+ z
2
evaluate only
for last step
2 real mults
per step
2012-11-28 Dan Ellis 7
2. Fast Fourier Transform FFT

Reduce complexity of DFT

from O(N
2
) to O(NlogN)

grows more slowly with larger N

Works by decomposing large DFT into

several stages of smaller DFTs

Often provided as a highly optimized

library
2012-11-28 Dan Ellis 8
Decimation in Time (DIT) FFT

Can rearrange DFT formula in 2 halves:

X k
[ ]
= x n
[ ]
W
N
nk
n=0
N1

= x 2m
[ ]
W
N
2mk
+ x 2m+1
[ ]
W
N
2m+1
( )
k
( )
m=0
N
2
1

= x 2m
[ ]
W
N
2
mk
m=0
N
2
1

+W
N
k
x 2m+1
[ ]
W
N
2
mk
m=0
N
2
1

N/2 pt DFT of x for even n N/2 pt DFT of x for odd n

X
0
[<k>
N/2
] X
1
[<k>
N/2
]
Arrange
terms
in pairs...
Group terms
from each
pair
k = 0.. N-1
2012-11-28 Dan Ellis 9
Decimation in Time (DIT) FFT

We can evaluate an N-pt DFT as two

N/2-pt DFTs (plus a few mults/adds)

But if DFT
N
{} ~ O(N
2
)
then DFT
N/2
{} ~ O((N/2)
2
) = 1/4 O(N
2
)
Total computation ~ 2 1/4 O(N
2
)
= 1/2 the computation (+") of direct DFT

DFT
N
x n
[ ] { }
= DFT
N
2
x
0
n
[ ] { }
+W
N
k
DFT
N
2
x
1
n
[ ] { }
x[n] for even n x[n] for odd n
2012-11-28 Dan Ellis 10
One-Stage DIT Flowgraph
Classic FFT structure
Even
points
from
x[n]
Odd
points
from
x[n]
Same as
X[0..3]
except for
factors on
X
1
[]
terms
twiddle factors:
always apply to
odd-terms output
NOT mirror-image

X k
[ ]
= X
0
k
N
2
[ ]
+W
N
k
X
1
k
N
2
[ ]
x[0]
x[2]
x[4]
x[6]
x[1]
x[3]
x[5]
x[7]
X[0]
X[1]
X[2]
X[3]
X[4]
X[5]
X[6]
X[7]
X
0
[0]
X
0
[1]
X
0
[2]
X
0
[3]
X
1
[0]
X
1
[1]
X
1
[2]
X
1
[3]
DFT
N
2
DFT
N
2
W
N
0
W
N
1
W
N
2
W
N
3
W
N
4
W
N
5
W
N
6
W
N
7
2012-11-28 Dan Ellis 11

If decomposing one DFT

N
into two
smaller DFT
N/2
s speeds things up ...
Why not further divide into DFT
N/4
s ?

i.e.

make:

Similarly,
Multiple DIT Stages

X k
[ ]
= X
0
k
N
2
[ ]
+W
N
k
X
1
k
N
2
[ ]
0 k < N

X
0
k
[ ]
= X
00
k
N
4
[ ]
+W
N
2
k
X
01
k
N
4
[ ]
0 k < N/2
N/4-pt DFT of even points
in even subset of x[n]
N/4-pt DFT of odd points
from even subset

X
1
k
[ ]
= X
10
k
N
4
[ ]
+W
N
2
k
X
11
k
N
4
[ ]
2012-11-28 Dan Ellis 12
Two-Stage DIT Flowgraph
x[0]
x[4]
x[2]
x[6]
x[1]
x[5]
x[3]
x[7]
X[0]
X[1]
X[2]
X[3]
X[4]
X[5]
X[6]
X[7]
X
0
[0]
X
00
X
01
X
10
X
11
X
0
[1]
X
0
[2]
X
0
[3]
X
1
[0]
X
1
[1]
X
1
[2]
X
1
[3]
DFT
N
4
DFT
N
4
DFT
N
4
DFT
N
4
W
N
0
W
N
1
W
N
2
W
N
3
W
N
4
W
N
5
W
N
6
W
N
7
W
N/2
3
W
N/2
0
W
N/2
3
W
N/2
0
different from before same as before
2012-11-28 Dan Ellis 13
Multi-stage DIT FFT

Can keep doing this until we get down

to 2-pt DFTs:
N = 2
M
-pt DFT reduces to M stages of
twiddle factors & summation
(O(N
2
) part vanishes)
real mults < M4N , real adds < 2M2N
complexity ~ O(NM) = O(Nlog
2
N)
DFT
2
X[0] = x[0] + x[1]
X[1] = x[0] - x[1]
-1 = W
2
1
1 = W
2
0
butterfly element

2012-11-28 Dan Ellis 14

W
N
r+
N
2
= e
j
2 r+
N
2
( )
N
= e
j
2r
N
e
j
2N / 2
N
= W
N
r
FFT Implementation Details

Basic butterfly (at any stage):

Can simplify:
W
N
r
W
N
r+N/2
X
X
[r]
X
X
[r+N/2]
X
X0
[r]
X
X1
[r]

2 cpx mults
X
X
[r]
X
X
[r+N/2]
X
X0
[r]
X
X1
[r]
W
N
r
-1
just one cpx mult!
i.e. SUB rather than ADD
2012-11-28 Dan Ellis 15
8-pt DIT FFT Flowgraph

-1s absorbed into summation nodes

W
N
0
disappears

in-place algorithm: sequential stages

b
i
t
-
r
e
v
e
r
s
e
d

i
n
d
e
x
i
n
g
W
4
W
4
W
8
W
8
W
8
x[0]
x[4]
x[2]
x[6]
x[1]
x[5]
x[3]
x[7]
000
100
010
110
001
101
011
111
X[0]
X[1]
X[2]
X[3]
X[4]
X[5]
X[6]
X[7]
-
-
-
-
-
-
-
-
-
-
-
-
2
3
2012-11-28 Dan Ellis 16
FFT for Other Values of N

Having N = 2
M
meant we could divide
each stage into 2 halves = radix-2 FFT

Same approach works for:

N = 3
M
radix-3

N = 4
M
radix-4 - more optimized radix-2

etc...

Composite N = abcd mixed radix

(different N/r point FFTs at each stage)

.. or just zero-pad to make N = 2

M

M
2012-11-28 Dan Ellis 17
Inverse FFT

Recall IDFT:

Thus:

Hence, use FFT to calculate IFFT:

x[n] =
1
N
X[k]W
N
nk
k=0
N1

only differences
from forward DFT

Nx
*
[n] = X[k]W
N
nk
( )
*
k=0
N1

= X
*
[k]W
N
nk
k=0
N1

Forward DFT of x[n] = X

*
[k]|
k=n
i.e. time sequence made from spectrum
DFT
Re{x[n]}
Im{x[n]}
Re
Im
Re
Im
Re{X[k]}
Im{X[k]}
1/N
-1/N -1
pure real flowgraph
x n


1
N
X
*
k

W
N
nk
k0
N 1

*
2012-11-28 Dan Ellis 18

If x[n] is pure-real, DFT wastes mults

Real x[n] Conj. symm. X[k] = X

*
[-k]

Given two real sequences, x[n] and w[n]

call y[n] = jw[n] , v[n] = x[n] + y[n]

N-pt DFT V[k] = X[k] + Y[k]

but: V[k]+V
*
[-k] = X[k]+X
*
[-k]+Y[k]+Y
*
[-k]
X[k]=
1
/
2
(V[k]+V
*
[-k]) , W[k]=
-j
/
2
(V[k]-V
*
[-k])

i.e. compute DFTs of two N-pt real

sequences with a single N-pt DFT
DFT of Real Sequences
X[k]
-Y[k]
2012-11-28 Dan Ellis 19
3. Short-Time
Fourier Transform (STFT)

Fourier Transform (e.g. DTFT) gives

spectrum of an entire sequence:

How to see a time-varying spectrum?

e.g. slow AM of a sinusoid carrier:

x n
[ ]
= 1 cos
2tn
N
|
\

|
.
|
cose
0
n
0 200 400 600 800 1000
-2
-1
0
1
2
n
x[n]
2012-11-28 Dan Ellis 20
Fourier Transform of AM Sine

Spectrum of
whole sequence
indicates
modulation
indirectly...

... as
cancellation
between
closely-
tuned
sines
2cAcB
= cA+B
+cA-B
Nsin2kn
N
-Nsin2(k-1)n
N 2
-Nsin2(k+1)n
N 2
0 0.02 0.04 0.06 0.08
0
200
400
600
N
N/2
\X[k]\
WP
k/(N/2)
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1
0 128 256 384 512 640 768 896
-1
-0.5
0
0.5
1
M
2012-11-28 Dan Ellis 21
Fourier Transform of AM Sine

Sometimes wed rather separate

modulation and carrier:
x[n] = A[n]cos!
0
n

A[n] varies on a
different (slower) timescale

One approach:

chop x[n] into short sub-sequences ..

.. where slow modulator is ~ constant

DFT spectrum of pieces show variation

!
A[n]
!
0
2012-11-28 Dan Ellis 22
FT of Short Segments

Break up x[n] into successive, shorter

chunks of length N
FT
, then DFT each:
Shows amplitude modulation
of !
0
energy
0 128 256 384 512 640 768 896 1024 = N
-2
-1
0
1
2
0 64
0
50
100
n
k
k
x
0
[n]
x[n]
X
0
[k]
x
1
[n]
X
1
[k]
x
2
[n]
X
2
[k]
x
3
[n]
X
3
[k]
x
4
[n]
X
4
[k]
x
5
[n]
X
5
[k]
x
6
[n]
X
6
[k]
x
7
[n]
X
7
[k]
k
0
= W
0
N
FT
2P
N
FT
= N/8
2012-11-28 Dan Ellis 23
The Spectrogram

Plot successive DFTs in time-frequency:

This image is called the Spectrogram

time hopsize (between successive frames)
= 128 points
128 256 384 512 640 768 896 1024
k
n
k k
\X
i
[k]\
\X[k,n]\
0
0
5
10
15
20
0
40
60
80
100
120
X
0
[k] X
1
[k] X
2
[k] X
3
[k] X
4
[k] X
5
[k] X
6
[k] X
7
[k]
2012-11-28 Dan Ellis 24
Short-Time Fourier Transform

Spectrogram = STFT magnitude

plotted on time-frequency plane

STFT is (DFT form):

intensity as a function of time & frequency

X k,n
0
[ ]
= x n
0
+ n
[ ]
w[n] e
j
2kn
N
FT
n=0
N
FT
1

frequency
index
time
index
N
FT
points of x
starting at n
window
DFT
kernel
2012-11-28 Dan Ellis 25
STFT Window Shape

w[n] provides time localization of STFT

e.g. rectangular
selects x[n], n
0
n < n
0
+N
W

But: resulting spectrum has same

problems as windowing for FIR design:
n
w[n]

X e
j
,n
0
( )
= DTFT x n
0
+ n
[ ]
w n
[ ] { }
= e
jn
0
X e
j
( )
W e
j
( )
( )
d

spectrum of short-time window

is convolved with (twisted) parent spectrum
DTFT
form of
STFT
2012-11-28 Dan Ellis 26
STFT Window Shape

e.g. if x[n] is a pure sinusoid,

Hence, use tapered window for w[n]

blurring (mainlobe)
+ ghosting (sidelobes)
e.g. Hamming

w n
[ ]
=
0.54 + 0.46cos(2
n
2M+1
)
sidelobes
< -40 dB
W
P
W

X(e
jW
) W(e
jW
)
W
-10 -5 0 5 10
W n
w[n] W(e
jW
)
2012-11-28 Dan Ellis 27
STFT Window Length

Length of w[n] sets temporal resolution

Window length 1/(Mainlobe width)

more time detail less frequency detail

short window measures
only local properties
longer window averages
spectral character
shorter window
more blurred
spectrum
0 200 400 600 800 1000
-0.1
0
0.1
0.2
0 200 400 600 800 1000
-0.1
0
0.1
0.2
x[n] x[n]
w [n]
L S
w [n]
-100 -50 0 50 100
0
0.5
1
-P -0.5P 0 0.5P P
0
10
20
-100 -50 0 50 100
0
0.5
1
-P -0.5P 0 0.5P P
0
10
20
n
n
w
L
[n]
w
S
[n]
W
W
W
S
(e
jW
)
W
L
(e
jW
)
N
1
pts
N
2
pts
N
1
zero at 4
N
2
zero at 4
2012-11-28 Dan Ellis 28
STFT Window Length

Can illustrate time-frequency tradeoff

on the time-frequency plane:

Alternate tilings
of time-freq:
disks show blurring
due to window length;
area of disk is constant
Uncertainty principle:
ft k
half-length window half as many DFT samples
0 100 200 300
0
0.5
1
0
50
100
150
200
250
n
k
2012-11-28 Dan Ellis 29
Spectrograms of Real Sounds
individual t-f
cells merge
into continuous
image
time-domain
successive
short
DFTs
time / s
time / s
f
r
e
q

/

H
z
i
n
t
e
n
s
i
t
y

/

d
B
2.35 2.4 2.45 2.5 2.55 2.6
0
1000
2000
3000
4000
f
r
e
q

/

H
z
0
1000
2000
3000
4000
0
0.1
-50
-40
-30
-20
-10
0
10
0 0.5 1 1.5 2 2.5
2012-11-28 Dan Ellis 30
Narrowband vs. Wideband

Effect of varying window length:

1.4 1.6 1.8 2 2.2 2.4 2.6
f
r
e
q

/

H
z
time / s
level
/ dB
0
1000
2000
3000
4000
f
r
e
q

/

H
z
0
1000
2000
3000
4000
0
0.2
W
i
n
d
o
w

=

2
5
6

p
t

N
a
r
r
o
w
b
a
n
d

W
i
n
d
o
w

=

4
8

p
t

W
i
d
e
b
a
n
d

-50
-40
-30
-20
-10
0
10
M
2012-11-28 Dan Ellis 31
Spectrogram in Matlab
>> [d,sr]=wavread(mpgr1_sx419.wav');
>> Nw=256;
>> specgram(d,Nw,sr)
>> caxis([-80 0])
>> colorbar
(hann) window length
actual sampling rate
(to label time axis)
dB
Time
F
r
e
q
u
e
n
c
y


0.5 1 1.5 2 2.5 3
0
2000
4000
6000
8000
-80
-60
-40
-20
0
2012-11-28 Dan Ellis
-60
-40
-20
0 -1
0
1
-1
0
1
n
Re{x[n]}
I
m
{
x
[
n
]
}
32
STFT as a Filterbank

Consider one row of STFT:

where

Each STFT row is output of a filter

(subsampled by the STFT hop size)

X
k
n
0
[ ]
= x n
0
+ n
[ ]
w n
[ ]
e
j
2kn
N
n=0
N1

= h
k
m
[ ]
x n
0
m
[ ]
m=0
N1
( )

h
k
n
[ ]
= w n
[ ]
e
j
2kn
N
just one freq.
convolution
with
complex IR
2012-11-28 Dan Ellis 33
STFT as a Filterbank

If
then

Each STFT row is the same bandpass

response defined by W(e
j!
),
frequency-shifted to a given DFT bin:

h
k
n
[ ]
= w
( )
n
[ ]
e
j
2kn
N

H
k
e
j
( )
= W e

( )
j
2k
N
( )
( )
shift-in-!
A bank of identical,
frequency-shifted
bandpass filters:
filterbank
\W(e
jW
)\
\H
1
(e
jW
)\ \H
2
(e
jW
)\

W
P
2012-11-28 Dan Ellis 34
STFT Analysis-Synthesis

IDFT of STFT frames can reconstruct

(part of) original waveform

e.g. if
then

Can shift by n
0
, combine, to get x[n]:

Could divide by w[n-n

0
] to recover x[n]...

X k,n
0
[ ]
= DFT x n
0
+ n
[ ]
w n
[ ] { }
IDFT X k,n
0
[ ] { }
= x n
0
+ n
[ ]
w n
[ ]
^
n
0
n
^
x[n]
x[n]w[n-n
0
]
2012-11-28 Dan Ellis 35
STFT Analysis-Synthesis

Dividing by small values of w[n] is bad

Prefer to
overlap windows:
i.e. sample X[k,n
0
]
at n
0
= rH where H = N/2 (for example)

Then
hopsize window length

x n
[ ]
= x n
[ ]
w n rH
[ ]
r

= x n
[ ]

w n rH
[ ]
r

=1
if
n
^
x[n]
x[n]w[n-rH]
2012-11-28 Dan Ellis 36
STFT Analysis-Synthesis

Hann or Hamming windows

with 50% overlap
sum to constant

Can modify individual frames of X[k,n]

and then reconstruct

complex, time-varying modifications

tapered overlap makes things OK

0.54 + 0.46cos(2
n
N
)
( )
+ 0.54 + 0.46cos(2
n
N
2
N
)
( )
=1.08
0 20 40 60 80
0
0.2
0.4
0.6
0.8
1
n
w[n] w[n-N/2]
w[n] + w[n-N/2]
2012-11-28 Dan Ellis 37
STFT Analysis-Synthesis

e.g. Noise reduction:

STFT of
original speech
Speech corrupted
by white noise
Energy threshold
mask
M
100 200 300
20
40
60
80
100
120
r
k