• Transforms make freq analysis and design much easier.

F.S à F.T à L.T à Z.T

jW ® s e jw ® z
a fixed time window
STFT
non-stantionary
signal analysis CWT DWT
• Stationary and non-stationary signal

(i ) stationary signal

time
• Stationary signal is constant in its statistical parameters over time
such as the mean and variance

• FT is suitable to analyze the stationary signal.

J.C. Yoo
 (ii ) non - stationary signal :different statistical characteristics from time to time

FT FT FT FT
chirp signal chirp signal

freq 2 freq 3 freq 2 freq 3

freq 1 freq 1

freq freq3

freq3 STFT freq 2

FT freq 2 freq1
freq1

time

doesn ' t give a time resolved freq analysis give a time resolved freq analysis

Three freq components appear over

the whole time at the same time.
J.C. Yoo
 CHAP.5 STFT(Short Time Fourier Transform)

For short enough windows, speech can be considered to be stationary

-Define analysis window (e.g., 30ms narrowband, 5 ms wideband)

–Define the amount of overlap between windows (e.g., 30%)
–Define a windowing function (e.g., Hanning, Gaussian)
–Generate windowed segments (multiply signal by windowing function)
–Apply the FFT to each windowed segment

( frames )

J.C. Yoo
 Each FT provides the spectral information of a separate
time-slice of the signal, providing simultaneous time and
frequency analysis

J.C. Yoo
 windowed segments
(short time sliced x(n))
A / D conversion
xraw (t ) x(t ) Continuous-to- x[n] v[n] V [k ] ( STFT )
Anti-aliasing
lowpass filter
discrete-time
conversion
Ä DFT
cutoff £ f s / 2
with f s ³ 2 f max
w(n)
A
¥
1 æ w 2p r ö
DTFT of x ( n) : °
X (e jw ) = å çè T
X j + j ÷
T r =-¥ T ø
p
° (e ) = 1
V jw ° ° (e j (w -q ) ) dq
X (e jq ) × W
2p ò
-p
fs / 2
N -1
V [ k ] = å v[ n]e - j (2p / N ) kn , k = 0,1,..., N - 1
n =0

DFT of v ( n) : V [ k ] = V (e jw ) w = 2p k / N

2p k w
Wk = Ü (W = )
NT T

J.C. Yoo
 fs / 2

= w /T

FT of x(t )

FT of x(n)
¥
° 1
jw æ w 2p r ö
X (e ) = å X ç j + j ÷
T r =-¥ è T T ø
FT of w(n)

FT of v(n) ° jw 1 p
° ° (e j (w -q ) )dq
V (e ) = òp X (e jq ) × W
2p -

¥
2p -n
sampling over w with
N
V ( z) = å v ( n) z
n =-¥

N -1
V° (k ) = V° (e )
jw
= å v(n)WNnk
w = 2p k / N
n =0
J.C. Yoo
 fs / 2

p æ 2500 ö
W
ç ÷ 1 2p
Tè T ø X a ( jW + j )
1 T T
X ( j W)
T
W

fs

Wk

If we want DW £ 2p (10) Hz
2p
i.e f s £ 2p (10) Hz
N
\ N ³ 500
J.C. Yoo To use a radix-2 FFT , we choose N = 512
 The Effect of Windowing:

<Rectangular Window>

w(n)
0 L-1
L -1
1 - e - jw L e - jw L / 2 (e jw L / 2 - e - jw L / 2 )
W (e jw ) = å e - jw n = - jw
= - jw / 2
n =0 1- e e (e jw / 2 - e - jw / 2 )

jw - jw( L -1) / 2 sin(wL / 2)

W (e ) = e
sin(w / 2)
J.C. Yoo
•Sidelobe causes leakage
•How to reduce the height of sidelobe peak? jw - jw( L -1) / 2 sin(wL / 2)
•How to reduce mainlobe width? W (e ) = e
•Dwm affects spectrum resolution sin(w / 2)

| W ( e jw ) | L=8
8

Peak sidelobe

0
- 2p -p 2p 2p p 2p
-
L L

Mainlobe
Dwm
width
J.C. Yoo
 Commonly Used Windows

ì1 0 £ n £ L - 1
L -1
w(n) = í
Rectangular î0 otherwise

Bartlett ì2 n / M
ï
w(n) = í2 - 2n / M
0£ n£ M /2
M /2< n£ M /2
(Triangular) ï0
î otherwise

ì0.5 - 0.5cos(2p n / M ) 0 £ n £ L - 1
w(n) = í
Hanning î0 otherwise

ì0.54 - 0.46 cos(2p n / M ) 0 £ n £ L - 1

w(n) = í
Hamming î0 otherwise

ì0.42 - 0.5cos(2p n / M ) + 0.8cos(4p n / M ) 0 £ n £ L - 1

w(n) = í
Blackman î0 otherwise

J.C. Yoo where M = L - 1

 Hanning

Least mainlobe width

Rectangular
Hamming

Bartlett
Blackman

J.C. Yoo
 The Effect of Windowing :

ü Leakage makes the spectral resolution worse :

§ influenced primarily by the width of the main lobe of W(e
W(ejw).
§ depends on the relative amplitude of the side lobes of W(e
W(ejw).

x(t ) = cos W0t + 0.75cos W1t - ¥ < t < ¥ if Z {w(n)} = W ( z ),

w n(Q t = nT ) ¥
Wt = t Z e{ jwo n
} åe
w(n) = jwo n
w(n) z - n
T n =-¥
x(n) = cos w0 n + 0.75cos w1n ¥
- jwo
= å w(n)(e z )- n
n =-¥
v(n) = x(n) × w(n)
= w(n) × cos w0 n + 0.75 × w(n) × cos w1 (n) = W (e - jwo z )
on the unit circle
w(n) jw0 n w(n) - jw0 n 0.75w(n) jw1n 0.75w(n) - jw1n
= e + e + e + e
2 2 2 2 W (e j (w -w0 ) )
B
1 1 0.75 0.75
\ V (e jw ) = X (e jw ) * W (e jw ) = W (e j (w -w0 ) ) + W (e j (w +w0 ) ) + W (e j (w -w1 ) ) + W (e j (w +w1 ) )
2 2 2 2
J.C. Yoo
Example of spectrum leakage: w(n): rectangular window with the length of L(64)

broadened spectral lines lead

to:à reduced resolution
2p
w0 =
14
by B
* w1 =
4p
15

leakage:
（magnitudes of two frequency
replicas of components interact each other）
1/ 2
the window spectra
2p
w0 =
0.375 14
2p
2p w1 =
w0 = 12
6
2p
w1 =
3

2p reduced resolution
w0 =
14
2p
spectrum leackage by side lobe w1 =
25
Leakage:
amplitude is affected by the neighborinbg components
(spectral smearing)
The closer the spectral lines are to, the greater the leakage is
J.C. Yoo
 EXAMPLE
before windowing:
main lobe

sidelobe

after windowing:
* -W1 -W0 W0 W1

effects after windowing：

(1) spectral lines are broadened to
(2) produce sidelobe,
its attenuation is equal to sidelobe
attenuation of window frequency
spectrum

J.C. Yoo
The shape and width(L) of window will affect on resolution.
• To reduce the spectral leakage and improve spectral resolution
(a) Use wider windows (longer signals),
sinc gets narrower as Là¥, hence less leakage.

The longer the window

the narrower the mainlobe is,
Leading to a better resolution.

Least mainlobe width

4p / L

more spectral leakage less spectral leakage

(b) Use non-rectangular windows.

J.C. Yoo L -1
 (i) Use wider windows : The width (L) of window
Sequence of L = 64 v ( n)
Sequence of L = 32 v ( n)

DFT of N = 64 V [k ] V [k ]
DFT of N = 32

After zero padding (960 zeros ) After zero padding (992 zeros )

DFT of N = 1024 DFT of N = 1024

J.C. Yoo

Zero padding : for getting more samples, not for improving its resolution.
 | V (k ) |
L=32
N=32

Zero Padding:32 Zero Padding:992

| V (k ) | | V (k ) |
L=32 L=32
N=64 N=1024

Zero Padding:96 Extended window | V ( k ) |

| V (k ) |
L=64
L=32 N=1024
N=128

J.C. Yoo
 EXAMPLE Improving the spectral resolution by wider window

N = 1024 : V (k )
window size

L=32
2p
x[n] = (cos( n) k
14 2p

4p L=42
+0.75 cos( n))
15
X (e jw )
k
2p

L=54

2p
k
w

L=64

2p k
Conclusion: the wider the window size, the better the spectral resolution
J.C. Yoo
 EXAMPLE Improving the spectral resolution
by non-rectangular window

2p
x[n] = (3.5* cos( n)
14
2p V (k )
+3.5* 0.75 cos( n)) rectangular
25
wrec [n] : red
whamming [n] : blue
hamming

L = 32
N = 1024
2p k
Conclusion: shape of window has effect on frequency resolution
J.C. Yoo
 EXAMPLE Improving the spectral resolution and reducing the spectral
leakage by using non-rectangular window and increasing the
window size
N = 1024 :

almost no spectral leakage

X (e jw )

J.C. Yoo
 DFT analysis with 32
32--point Kaiser window and zero-
zero-padding

v ( n) æ 2p ö æ 4p ö
L = 32
v(n) = wkaiser (n) × cos ç n ÷ + 0.75wkaiser (n) × cos ç n÷
è 14 ø è 15 ø

n b = 5.48

n V [k ]
V [k ]
DFT of N = 128
DFT of N = 32

k
k
32 zero padding V [k ]
V [k ] linear interplated!!
DFT of N = 64
DFT of N = 1024
64 point DFT

k
k

J.C. Yoo
 Window Length vs. Zero Padding

Zero Padding
n Get more samples
n Not improve frequency resolution

Window Length
n The resolution depends on window
length and shape.

J.C. Yoo