cepstrum analysis - mario bonusing the inverse z-transform integral: = ∫ − & q; ] ] g] m [ q...
TRANSCRIPT
Guido Geßl 21.01.2004
Cepstrum Analysis
Guido Geßl 21.01.2004
B.P. Bogert, M.J. R. Healy and J.W. Tukey, "The Quefrency Alanysis of Time Series forEchoes: Cepstrum, Psuedo-Autocovariance, Cross-cepstrum and Saphe Cracking", Proceedings of the Symposium on Time Series Analysis, M. Rosenblat, Ed., Wiley, NY, 1963, pp 209-243
Guido Geßl 21.01.2004
Guido Geßl 21.01.2004
original definition - real cepstrum
Inverse Fourier transfrom of the logarithm of themagnitude of the Fourier transform:
[ ] ( ) ωπ
ωπ
π
ω GHH;QF QMM
[ ∫−
= log2
1
magnitude is real and nonnegative
not invertible as the phase is missing
Guido Geßl 21.01.2004
more general definitioncomplex cepstrum
∫−
=π
π
ωω ωπ
GHH;Q[ QMM )](log[2
1][ˆ
∫−
+=π
π
ωωω ωπ
GHH;MH; QMMM ))](arg()(log[2
1
arg is the continuous (unwrapped) phase function.
The term complex cepstrum refers to the use of thecomplex logarithm, not to the sequence.
The complex cepstrum of a real sequence is also a real sequence.
Guido Geßl 21.01.2004
complex logarithm
The complex cepstrum exists if the complexlogarithm has a convergent power seriesrepresentation of the form :
∑∞
−∞=
− ===Q
Q ]]Q[];]; 1,][ˆ)](log[)(ˆ
So log[X(z)] must have the properties of the z-transform of a stable sequence.
Guido Geßl 21.01.2004
using the inverse z-transform integral:
∫ −=&
Q G]]];M
Q[ 1)](log[2
1][ˆ
π
The contour of integration C is within the region of convergence.
The region of convergence must include the unit cyclefor stability.
That‘s why we can use the inverse Fourier transform.
Guido Geßl 21.01.2004
The real cepstrum is the inverse transform of the real partof and therefore is equal to the conjugate-symmetric part of :
)(ˆ ωMH;][ˆ Q[
2
][ˆ][ˆ][
* Q[Q[QF[
−+=
Guido Geßl 21.01.2004
Alternative Expressions
We can use the logarithmic derivative to avoid thecomplex logarithm.Assuming log[X(z)] is analytic, then
)(
)(’)(’ˆ
];
];]; =
Using the z-transform formula we get
∫ −=−&
Q G]]];
];]
MQ[Q 1
)(
)(’
2
1][ˆ
π
Guido Geßl 21.01.2004
Dividing both sides by (-n) yields
∫ ≠−= −
&
Q QG]]];
];]
QMQ[ 0,
)(
)(’
2
1][ˆ 1
π
and per definition
∫−
=π
π
ω ωπ
GH;[ M )(ˆ2
1]0[ˆ
A (nonlinear) difference equation for x[n] is :
0,][][ˆ][ ≠−= ∑∞
−∞=
QNQ[N[Q
NQ[
N
(for computation with recursive algorithms)
Guido Geßl 21.01.2004
Complex cepstrum of exponential sequences
If a sequence x[n] consists of a sum of complexexponential sequences, it‘s z-transform is a rational function of z.
( ) ( )
( ) ( )∏∏
∏∏
==
−
==
−
−−
−−=
RL
RL
1
N
N
1
N
N
0
N
N
0
N
N
U
]G]F
]E]D]$
];
11
1
11
1
11
11)(
i…inside, o…outside of the unit circle
Guido Geßl 21.01.2004
then log[X(z)] is :
( ) ( )
( ) ( )∑∑
∑∑
==
−
==
−
−−−−
−+−++=
RL
RL
1
N
N
1
N
N
0
N
N
0
N
N
U
]G]F
]E]D]$];
11
1
11
1
1log1log
1log1log)log()log()(ˆ
A is real for real sequences but could be negative.
A nonzero r will cause a discontinuity in arg[X].
In practice this is avoided by determining A and r and altering the input so, that A _$_�����U� ��
Guido Geßl 21.01.2004
with the power series expansions
∑
∑∞
=
−
∞
=
−−
<−=−
<−=−
1
1
1
1
,)1log(
,)1log(
Q
Q
]]Q
]
D]]Q
]
βββ
αα
we obtain
<+
>+−
=
=
∑∑
∑∑
=
−
=
−
==
��
��
1
N
Q
N
0
N
Q
N
1
N
Q
N
0
N
Q
N
G
Q
E
F
Q
D
Q$
Q[
11
11
.0,
,0,
,0,log
][ˆ
Guido Geßl 21.01.2004
From this we can derive the following general
properties:
1) the complex cepstrum decays at least as fast as 1/|n|
2) it has infinite duration, even if x[n] has finite duration
3) it is real if x[n] is real (poles and zeros arein complex conjugate pairs)
Guido Geßl 21.01.2004
minimum-phase and maximum-phase sequences
A minimum-phase sequence is a real, causaland stable sequence whose poles and zerosare all inside the unit cycle.That means that all the singularities of log[X(z)] are inside the unit cycle. So we have the property
4) The complex cepstrum is causal (0 for all n<0) if and only if x[n] is minimum phase.
Guido Geßl 21.01.2004
similarily we conclude, that the complexcepstrum for maximum-phase (=left sided) systems is also left-sided
5) The complex cepstrum is 0 for all n>0 ifand only if x[n] is maximum phase, i.e. X(z) has all poles and zeros outside the unitcircle
Guido Geßl 21.01.2004
Hilbert Transform Relations
Causality (or anticausality) of a sequenceplaces powerful constraints on the FourierTransform.
The Fourier Transform of any real, causalsequence is almost completely determinedby either the real or the imaginary part of the Fourier Transform.
Guido Geßl 21.01.2004
Thus, if is causal (x[n] is minimum phase), wecan obtain Hilbert transform relations between thereal ( ) and imaginary ( ) partsof the complex cepstrum.
As noted earlier, the cepstrum is the even part of thecomplex cepstrum:
][ˆ Q[
)(log ωMH; )](arg[ ωMH;
2
][ˆ][ˆ][
* Q[Q[QF[
−+=
So if the complex cepstrum is known to be causal, it can berecovered from the real cepstrum by frequency-invariantlinear filtering.
][][2][
][][][ˆ
min
min
QQXQ
QQFQ[ [
δ−==
A
A
Guido Geßl 21.01.2004
Homomorphic deconvolution
• Generalized superposition
T{x1[n]+x2[n]}= T{x1[n]}+ T{x2[n]}
T{c x1[n]}=c T{x1[n]}
becomes
T{x1[n] �[2[n]}= T{x1[n]} �7^[2[n]}
T{c �[1[n]}=c �7^[1[n]}
with different input and output operations
Guido Geßl 21.01.2004
Systems satisfying this generalizedprinciple of superposition are calledhomomorphic systems since they can berepresented by algebraically linear (homomorphic) mappings between inputand output signal spaces.
Linear systems are obviously a special casewhere �DQG� �are addition and �DQG� �aremultiplication.
Guido Geßl 21.01.2004
Now suppose the input of a system is the
convolution of two signals :
x[n]=x1[n]*x2[n]
Their Fourier transforms would therefore be
multiplied :
X(z)=X1(z) X2(z)
applying the logarithm :
log[X(z)]=log[X1(z)] + log[X2(z)]
So the cepstrum is:
][ˆ][ˆ][ˆ 21 Q[Q[Q[ +=
Guido Geßl 21.01.2004
Thus, the cepstrum can be interpreted as a System that satisfies the generalizedprinciple of superposition with convolutionas the input function of superposition and addition as the output function of superposition.
This system is called the CharacteristicSystem for Convolution D*.
Guido Geßl 21.01.2004
D* L D*-1
L is a linear System in the usual sense.
Now if we choose a system L that removes the additivecomponent , then x2[n] will be removed from theconvolutional combination.
In practical applications, L is a linear frequency-invariantsystem.
Such L are called ‚lifters‘ because of their similarity to filters in the frequency domain.
][ˆ2 Q[
Guido Geßl 21.01.2004
Applications
• processing signals containing echoesseismology, measuring properties of reflecting surfaces, loudspeaker design, dereverberation, restoration of acousticrecordings
• speech processingestimating parameters of the speech model
Guido Geßl 21.01.2004
• machine diagnostics
detection of families of harmonics and sidebands, for example in gearbox and turbine vibrations
• calculating the minimum phase spectrumcorresponding to a given log amplitudespectrum with Hilbert transform
Guido Geßl 21.01.2004
example: delay time detection
Guido Geßl 21.01.2004
Guido Geßl 21.01.2004
Example: echo removal
Guido Geßl 21.01.2004