Correlation

8: Correlation
Cross-Correlation
Signal Matching
Cross-corr as Convolution
Normalized Cross-corr
Autocorrelation
Autocorrelation example
Fourier Transform Variants
Scale Factors
Summary
Spectrogram
8: Correlation
E1.10 Fourier Series and Transforms (2015-5585)
Fourier Transform - Correlation: 8 1 / 11
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
The cross-correlation between two signals u(t) and v(t) is
w(t) = u(t) v(t) ,
u
( )v( + t)d
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,
u
( )v( + t)d
u ( t)v( )d
[sub: t]
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]
The complex conjugate, u ( ) makes no difference if u(t) is real-valued

but makes the definition work even if u(t) is complex-valued.
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

Correlation versus Convolution:
u(t) v(t) = u ( )v( + t)d

R
u(t) v(t) = u( )v(t )d
[correlation]
[convolution]
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

u(t) v(t) = u ( )v( + t)d

[correlation]
R
u(t) v(t) = u( )v(t )d
[convolution]
Unlike convolution, the integration variable, , has the same sign in the
arguments of u( ) and v( ) so the arguments have a constant
difference instead of a constant sum (i.e. v(t) is not time-flipped).
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

u(t) v(t) = u ( )v( + t)d

[correlation]
R
u(t) v(t) = u( )v(t )d
[convolution]
Notes: (a) The argument of w(t) is called the lag (= delay of u versus v ).
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

u(t) v(t) = u ( )v( + t)d

[correlation]
R
u(t) v(t) = u( )v(t )d
[convolution]
(b) Some people write u(t) v(t) instead of u(t) v(t).
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

u(t) v(t) = u ( )v( + t)d

[correlation]
R
u(t) v(t) = u( )v(t )d
[convolution]
(c) Some swap u and v and/or negate t in the integral.
Cross-Correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) v(t) ,

=
u
( )v( + t)d
u ( t)v( )d
[sub: t]

u(t) v(t) = u ( )v( + t)d

[correlation]
R
u(t) v(t) = u( )v(t )d
[convolution]
(c) Some swap u and v and/or negate t in the integral.
It is all rather inconsistent /.
Signal Matching
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Cross correlation is used to find where two

signals match
Signal Matching
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

signals match: u(t) is the test waveform.
1
u(t)
8: Correlation
0.5
0
0
200
400
600
Signal Matching
1
u(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
Example 1:
v(t) contains u(t) with an unknown delay
and added noise.
200
400
600
1
v(t)
8: Correlation
0.5
0
0
200
400
600
800
Signal Matching
1
u(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
Example 1:
and added noise.
200
400
600
1
v(t)
8: Correlation
0.5
0
0
200
400
600
800
w(t) = u(t) v(t)
Signal Matching
1
u(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
Example 1:
and added noise.
200
400
600
1
v(t)
8: Correlation
0.5
0
0
200
400
600
800
w(t) = u(t)
R v(t)
= u ( t)v( )dt
Signal Matching
1
u(t)

0.5
0
0
Example 1:
and added noise.
w(t) = u(t)
R v(t)
= u ( t)v( )dt
200
400
600
1
v(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
200
400
600
800
200
400
600
800
20
w(t)
8: Correlation
10
0
Signal Matching
1
u(t)

0.5
0
0
Example 1:
and added noise.
w(t) = u(t)
R v(t)
= u ( t)v( )dt gives a peak
at the time lag where u( t) best
matches v( )
200
400
600
1
v(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
200
400
600
800
200
400
600
800
20
w(t)
8: Correlation
10
0
Signal Matching
1
u(t)

0.5
0
0
Example 1:
and added noise.
w(t) = u(t)
R v(t)
matches v( ); in this case at t = 450
200
400
600
1
v(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
0.5
0
0
200
400
600
800
200
400
600
800
20
w(t)
8: Correlation
10
0
Signal Matching
1
u(t)

0.5
0
0
w(t) = u(t)
R v(t)
Example 2:
y(t) is the same as v(t) with more noise
200
400
600
1
v(t)
Example 1:
and added noise.
0.5
0
0
200
400
600
800
200
400
600
800
200
400
600
800
20
w(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
10
0
y(t)
8: Correlation
2
0
-2
Signal Matching
1
u(t)

0.5
0
0
w(t) = u(t)
R v(t)
Example 2:
z(t) = u(t) y(t) can still detect the
correct time delay (hard for humans)
200
400
600
v(t)
1
0.5
0
0
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
20
w(t)
Example 1:
and added noise.
10
0
y(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
2
0
-2
20
z(t)
8: Correlation
10
0
-10
Signal Matching
1
u(t)

0.5
0
0
w(t) = u(t)
R v(t)
Example 2:
200
400
600
v(t)
1
0.5
0
0
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
20
w(t)
Example 1:
and added noise.
10
0
y(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
2
0
-2
20
10
z(t)
8: Correlation
0
-10
Example 3:
p(t) contains u(t)
p(t)
1
0
-1
Signal Matching
1
u(t)

0.5
0
0
w(t) = u(t)
R v(t)
Example 2:
200
400
600
v(t)
1
0.5
0
0
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
20
w(t)
Example 1:
and added noise.
10
0
y(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
2
0
-2
20
10
z(t)
8: Correlation
0
-10
Example 3:
p(t) contains u(t) so that
q(t) = u(t) p(t) has a negative peak
p(t)
1
0
-1
Signal Matching
1
u(t)

0.5
0
0
w(t) = u(t)
R v(t)
Example 2:
200
400
600
v(t)
1
0.5
0
0
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
200
400
600
800
20
w(t)
Example 1:
and added noise.
10
0
y(t)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
2
0
-2
20
10
z(t)
8: Correlation
0
-10
0
-1
0
q(t)
Example 3:
p(t) contains u(t) so that
q(t) = u(t) p(t) has a negative peak
p(t)
-10
-20
Cross-correlation as Convolution
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Correlation: w(t) = u(t) v(t) = u ( t)v( )d
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
If we define x(t) = u (t) then
x(t) v(t) ,
x(t )v( )d
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) ,
x(t )v( )d =
u
( t)v( )d
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
Fourier Transform of x(t):
X(f ) =
i2f t
x(t)e
dt
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
X(f ) =
i2f t
x(t)e
dt =
u
( t)v( )d
i2f t
u
(t)e
dt
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
X(f ) =
=
i2f t
x(t)e
dt =
u
( t)v( )d
i2f t
u
(t)e
dt
i2f t
u
(t)e
dt
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
So w(t) = x(t) v(t) W (f ) = X(f )V (f )
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
So w(t) = x(t) v(t) W (f ) = X(f )V (f ) = U (f )V (f )
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
Hence the Cross-correlation theorem:
w(t) = u(t) v(t)
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
w(t) = u(t) v(t)

= u (t) v(t)
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
w(t) = u(t) v(t)

= u (t) v(t)
W (f ) = U (f )V (f )
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
w(t) = u(t) v(t)

= u (t) v(t)
W (f ) = U (f )V (f )
Note that, unlike convolution, correlation is not associative or commutative:
v(t) u(t) = v (t) u(t)

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
w(t) = u(t) v(t)

= u (t) v(t)
W (f ) = U (f )V (f )
v(t) u(t) = v (t) u(t) = u(t) v (t)

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
x(t) v(t) , x(t )v( )d =

= u(t) v(t)
u
( t)v( )d
dt = u (t)ei2f t dt

R
R
i2f t
i2f t
dt
dt = u(t)e
= u (t)e
X(f ) =
i2f t
x(t)e
= U (f )
w(t) = u(t) v(t)

= u (t) v(t)
W (f ) = U (f )V (f )
v(t) u(t) = v (t) u(t) = u(t) v (t) = w (t)

Normalized Cross-correlation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

If we define y(t) = u(t t0 ) for some fixed t0 , then Ey = Eu :
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Ey =
|y(t)| dt
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Ey =
|y(t)| dt =
|u(t t0 )| dt
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Ey =
|u(t t0 )| dt
R
2
= |u( )| d = Eu
|y(t)| dt =
[ t + t0 ]
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

Cauchy-Schwarz inequality: y ( )v( )d Ey Ev
Ey =
|y(t)| dt =
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d
Ey =
|y(t)| dt =
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev
Ey =
|y(t)| dt =
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev = Eu Ev
Ey =
|y(t)| dt =
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev = Eu Ev
but t0 was arbitrary, so we must have |w(t)| Eu Ev for all t

Ey =
|y(t)| dt =
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev = Eu Ev

Ey =
|y(t)| dt =
We can define the normalized cross-correlation
z(t) =
u(t)v(t)
Eu Ev
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev = Eu Ev

Ey =
|y(t)| dt =
z(t) =
u(t)v(t)
Eu Ev
with properties: (1) |z(t)| 1 for all t
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

|u(t t0 )| dt
R
2
= |u( )| d = Eu
[ t + t0 ]
R
2

R
2

2
|w(t0 )| = u ( t0 )v( )d Ey Ev = Eu Ev

Ey =
|y(t)| dt =
z(t) =
u(t)v(t)
Eu Ev
with properties: (1) |z(t)| 1 for all t
(2) |z(t0 )| = 1 v( ) = u( t0 ) with constant
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
The correlation of a signal with itself is its autocorrelation:
w(t) = u(t) u(t)
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
The autocorrelation at zero lag:
w(0) =
u
( 0)u( )d
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
w(0) =
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
The autocorrelation at zero lag, w(0), is the energy of the signal.
w(0) =
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
The normalized autocorrelation:
z(t) =
u(t)u(t)
Eu
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
z(t) = u(t)u(t)
Eu
satisfies z(0) = 1 and |z(t)| 1 for any t.
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
z(t) = u(t)u(t)
Eu
Wiener-Khinchin Theorem: [Cross-correlation theorem when v(t) = u(t)]
w(t) = u(t) u(t)
W (f ) = U (f )U (f )
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
z(t) = u(t)u(t)
Eu
w(t) = u(t) u(t)
W (f ) = U (f )U (f ) = |U (f )|
Autocorrelation
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
w(t) = u(t) u(t) =
u
( t)u( )d
u
( 0)u( )d
R
= u ( )u( )d
R
2
= |u( )| d = Eu
w(0) =
z(t) = u(t)u(t)
Eu
w(t) = u(t) u(t)
W (f ) = U (f )U (f ) = |U (f )|
The Fourier transform of the autocorrelation is the energy spectrum.
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Cross-correlation is used to find when two different signals are similar.
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Autocorrelation is used to find when a signal is similar to itself delayed.
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

First graph shows s(t) a segment of the microphone signal from the initial
vowel of early spoken by me.
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

vowel of early spoken by me. The waveform is quasi-periodic =
almost periodic but not quite.
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Second graph shows normalized autocorrelation, z(t) =
Norm Autocorr, z(t)
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
s(t)s(t)
.
Es
1
0.82
Lag = 6.2 ms (161 Hz)
0.5
0
10
Lag: t (ms)
15
20
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

s(t)s(t)
.
Es
z(0) = 1 for t = 0 since a signal always matches itself exactly.
Norm Autocorr, z(t)
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
1
0.82
Lag = 6.2 ms (161 Hz)
0.5
0
10
Lag: t (ms)
15
20
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

s(t)s(t)
.
Es

z(t) = 0.82 for t = 6.2 ms = one period lag (not an exact match).
Norm Autocorr, z(t)
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
1
0.82
Lag = 6.2 ms (161 Hz)
0.5
0
10
Lag: t (ms)
15
20
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

s(t)s(t)
.
Es

z(t) = 0.82 for t = 6.2 ms = one period lag (not an exact match).
z(t) = 0.53 for t = 12.4 ms = two periods lag (even worse match).
Norm Autocorr, z(t)
Speech s(t)
1
0.5
0
-0.5
-1
10
20
30
Time (ms)
40
50
1
0.82
Lag = 6.2 ms (161 Hz)
0.5
0
10
Lag: t (ms)
15
20

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
There are three different versions of the Fourier Transform in current use.

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
(1) Frequency version (we have used this in lectures)
U (f ) =
i2f t
u(t)e
dt
u(t) =
i2f t
U
(f
)e
df

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
u(t) = U (f )ei2f t df
U (f ) = u(t)e
Used in the communications/broadcasting industry and textbooks.
The formulae do not need scale factors of 2 anywhere.
,,,
i2f t

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
(2) Angular frequency version
e () =
U
it
u(t)e
dt
u(t) =
1
2
e ()eit d
U

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
Continuous spectra are unchanged: U
it
u(t)e
dt
1
2

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
2
However -function spectral components are multiplied by 2 so that
e () = 2 ( 2f0 )
U (f ) = (f f0 ) U
it
u(t)e
dt
1
2

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
2
e () = 2 ( 2f0 )
U (f ) = (f f0 ) U
Used in most signal processing and control theory textbooks.
it
u(t)e
dt
1
2

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
2
e () = 2 ( 2f0 )
U (f ) = (f f0 ) U
it
u(t)e
1
2
dt
(3) Angular frequency + symmetrical scale factor
b () =
U
1
2
it
u(t)e
dt
u(t) =
1
2
b ()eit d
U

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
2
e () = 2 ( 2f0 )
U (f ) = (f f0 ) U
it
u(t)e
1
2
dt
b () =
U
1
2
it
u(t)e
dt
b () = 1 U
e ()
In all cases U
u(t) =
1
2
b ()eit d
U

8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
dt
U (f ) = u(t)e
,,,
i2f t
e () =
U
e ()eit d
u(t) =
U
e () = U (f ) = U ( )
2
e () = 2 ( 2f0 )
U (f ) = (f f0 ) U
it
u(t)e
1
2
dt
b () =
U
1
2
it
u(t)e
dt
b () = 1 U
e ()
In all cases U
u(t) =
1
2
b ()eit d
U
Used in many Maths textbooks (mathematicians like symmetry)
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Fourier Transform using Angular Frequency:
e () =
U
it
u(t)e
dt
u(t) =
1
2
e ()eit d
U
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
Any formula involving df will change to 2 d
[since d = 2df ]
e () =
U
it
1
2
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
[since d = 2df ]
e () =
U
it
Parsevals Theorem:
u (t)v(t)dt =
1
2
1
2
e ()Ve ()d
U
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
[since d = 2df ]
e () =
U
it
Parsevals Theorem:
1
2
e ()Ve ()d
u (t)v(t)dt =
U
2
R
R
2
1
e () d
Eu = |u(t)| dt = 2 U
1
2
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
[since d = 2df ]
e () =
U
Parsevals Theorem:
1
2
it
e ()Ve ()d
u (t)v(t)dt =
U
2
R
R
2
1
e () d
Eu = |u(t)| dt = 2 U
1
2
Waveform Multiplication: (convolution implicitly involves integration)
f () =
w(t) = u(t)v(t) W
1 e
2 U ()
Ve ()
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
[since d = 2df ]
e () =
U
Parsevals Theorem:
1
2
it
e ()Ve ()d
u (t)v(t)dt =
U
2
R
R
2
1
e () d
Eu = |u(t)| dt = 2 U
1
2
f () =
w(t) = u(t)v(t) W
1 e
2 U ()
Ve ()
Spectrum Multiplication: (multiplication ; integration)
f () = U
e ()Ve ()
w(t) = u(t) v(t) W
Scale Factors
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
e ()eit d
U
u(t)e
dt
u(t)
=
R
R
1
[since d = 2df ]
e () =
U
Parsevals Theorem:
1
2
it
e ()Ve ()d
u (t)v(t)dt =
U
2
R
R
2
1
e () d
Eu = |u(t)| dt = 2 U
1
2
f () =
w(t) = u(t)v(t) W
1 e
2 U ()
Ve ()
Spectrum Multiplication: (multiplication ; integration)
f () = U
e ()Ve ()
w(t) = u(t) v(t) W
b (),
To obtain formulae for version (3) of the Fourier Transform, U
e () =
substitute into the above formulae: U
b ().
2 U
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Cross-Correlation: w(t) = u(t) v(t)
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Cross-Correlation: w(t) = u(t) v(t) =
u ( t)v( )d
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Cross-Correlation: w(t) = u(t) v(t) = u ( t)v( )d

Used to find similarities between v(t) and a delayed u(t)
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Cross-correlation theorem: W (f ) = U (f )V (f )
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram

Cauchy-Schwarz Inequality: |u(t) v(t)| Eu Ev
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
Normalized cross-correlation: E E 1
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u
Autocorrelation: x(t) = u(t) u(t)
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R
Autocorrelation: x(t) = u(t) u(t) = u ( t)u( )d Eu
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R
Wiener-Khinchin: X(f ) = energy spectral density, |U (f )|2
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R
Used to find periodicity in u(t)
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R
Fourier Transform using :

Continuous spectra unchanged; spectral impulses multiplied by 2
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R

R
R
1
In formulae: df 2 d ; -convolution involves an integral
Summary
8: Correlation
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram


u(t)v(t)
u v
R

R
R
1
In formulae: df 2 d ; -convolution involves an integral
For further details see RHB Chapter 13.1
Spectrogram
Spectrogram of Merry Christmas spoken by Mike Brookes ()
m e rr y ch r i s t m a
-10
8
7
6
5
4
3
2
-20
1
0
-40
-30
0.2
0.4
0.6
Time (s)
0.8
Power/Decade (dB)
Cross-Correlation
Signal Matching
Autocorrelation
Scale Factors
Summary
Spectrogram
Frequency (kHz)
8: Correlation

Correlation

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Correlation

Încărcat de

Drepturi de autor:

Formate disponibile

8: Correlation

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 1 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 2 / 11

The cross-correlation between two signals u(t) and v(t) is

w(t) = u(t) v(t) ,

The complex conjugate, u ( ) makes no difference if u(t) is real-valued

u(t) v(t) = u ( )v( + t)d

Fourier Transform - Correlation: 8 2 / 11

Cross correlation is used to find where two

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 3 / 11

Cross correlation is used to find where two

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 3 / 11

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 3 / 11

w(t) = u(t) v(t)

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 3 / 11

E1.10 Fourier Series and Transforms (2015-5585)

Fourier Transform - Correlation: 8 3 / 11