DSP-CIS
Part-IV : Filter Banks & Subband Systems
Chapter-12 : Frequency Domain Filtering
Marc Moonen
Dept. E.E./ESAT-STADIUS, KU Leuven
marc.moonen@esat.kuleuven.be
www.esat.kuleuven.be/stadius/
Part-IV : Filter Banks & Subband Systems
Chapter-10 Filter Bank Preliminaries
Chapter-11 Filter Bank Design
Chapter-12 Frequency Domain Filtering
• Frequency Domain FIR Filter Realization
• Frequency Domain Adaptive Filtering
Chapter-13 Transmultiplexers
Chapter-14 Time-Frequency Analysis & Scaling
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 2
FIR Filter Realization
FIR Filter Realization
=Construct (realize) LTI system (with delay elements,
adders and multipliers), such that I/O behavior is given by..
y[k] = b0.u[k]+ b1.u[k -1]+... + bL .u[k - L]
Several possibilities exist…
1. Direct form
2. Transposed direct form
3. Lattice realization (LPC lattice)
4. Lossless lattice realization
5. Frequency domain realization: see Part IV
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering

p. 3
Frequency Domain FIR Filter Realization
Have to know a theorem from linear algebra here:
• A `circulant’ matrix is a matrix where each row is obtained from the
previous row using a right-shift (by 1 position),
a d c b 
b a d c 
the rightmost element which spills over is




c
b
a
d
circulated back to become the leftmost element


d
c
b
a
• The eigenvalue decomposition of a circulant matrix is always given as...
(4x4 example)  a d c b 
A 0 0 0 
 A
a 
b

c

d
a
b
d
a
c
b

c 
1  0
F .
0
d


a
0
B
0
0
C
0
0
0 
.F ,
0

D
B
b 
with    F . 
C 
c 
 
 
 D
d 
with F the DFT-matrix. This means that the eigenvectors are equal to
the column-vectors of the IDFT-matrix, and that then eigenvalues are
obtained as the DFT of the first column of the circulant matrix (proof by Matlab)
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 4
Frequency Domain FIR Filter Realization
FIR Filter Realization (example filter length =L+1=4, similar for other L)
y[k] = b0.u[k]+ b1.u[k -1]+ b2.u[k - 2]+ b3.u[k - 3]
B(z) = b0 + b1z-1 + b2 z -2 + b3z-3
Consider a 'block processing’ where a block of LB (=‘block
length’) output samples are computed at once, and consider
LB=L+1:
é u[k] ù
é y[k] ù é
ê
ú ê
ê y[k -1] ú ê
ê
ú=ê
y[k
2]
ê
ú ê
êë y[k - 3] úû ê
ë
b0
b1
b2
b3
0
0
0
0
0
b0
b1
b2
b3
0
0
0
0
0
b0
b1
b2
b3
0
0
0
0
0
b0
b1
b2
b3
0
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
ê
ê
ùê
úê
úê
ú. ê
úê
úê
ûê
ê
êë
u[k -1]
u[k - 2]
u[k - 3]
u[k - 4]
u[k - 5]
u[k - 6]
u[k - 7]
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
úû
p. 5
Frequency Domain FIR Filter Realization
Now some matrix manipulation…
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 6
Frequency Domain FIR Filter Realization
• This means that a block of LB=L+1 output samples can be
computed as follows (read previous formula from right to left) :
– Compute DFT of 2(L+1) input samples, i.e. last (L+1) samples
combined (‘overlapped’) with previous (L+1) samples
– Perform component-wise multiplication with…
(=freq.domain representation of the FIR filter)
– Compute IDFT
é
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
êë
B0 ù
é 0
ú
ê
B1 ú
ê b3
ú
ê b
B2 ú
ê 2
B3 ú
ú = F. êê b1
B4 ú
ê b0
ú
B5
ê 0
ú
ê
B6 ú
ê 0
ú
êë 0
B7 úû
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
úû
– Throw away 1st half of result, select (‘save’) 2nd half
• This is referred to as an ‘overlap-save’ procedure
(and ‘frequency domain filter realization’ because of the DFT/IDFT)
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 7
Frequency Domain FIR Filter Realization
•  Overlap-save procedure is very efficient for large L :
– Computational complexity (with FFT/IFFT i.o. DFT/IDFT) is
2.[α.2(L+1).log(2(L+1))] + 2(L+1) multiplications for (L+1) output
samples, i.e. O(log(L)) per sample for large L
– Compare to computational complexity for direct form realization:
(L+1) multiplications per output sample, i.e. O(L) per sample
•  Overlap-save procedure introduces O(L) processing
delay/latency (e.g. y[k-L] only available sometime after time k)
• Conclusion: For large L, complexity reduction is large, but
latency is also large
• Will derive ‘intermediate’ realizations, with a smaller latency
at the expense of a smaller complexity reduction. This will
be based on an Nth order polyphase decomposition of B(z)…
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 8
Frequency Domain FIR Filter Realization
A compact derivation will rely on a result from filter bank theory
(return to Chapter-10…)
1
u[k]
z 1
z
2
z 3
4
4
4
4
z 3
4
4
4
4
T( z )
T(z)*u[k-3]
z 2
+
z 1
1
(…and now let B(z) take the place of T(z))
This means that a filter (specified with Nth order polyphase decomposition)
B(z) = p0 (z 4 )+ z-1 p1 (z 4 )+ z-2 p2 (z 4 )+ z-3 p3 (z 4 )
can be realized in a multirate structure, based on a
pseudo-circulant matrix
p ( z)
p ( z)
p ( z)
 p ( z)
0
 z 1. p ( z )
T( z )   1 3
 z . p2 ( z )
 1
 z . p1 ( z )
1
2
p0 ( z )
p1 ( z )
z 1. p3 ( z )
z 1. p2 ( z )
p0 ( z )
z 1. p3 ( z )
3
p2 ( z )

p1 ( z ) 

p0 ( z ) 
DSP-CIS
2015 / Part
IV / for
Chapter-12:
Domain Filtering
PS:
formulas
given
N=4, Frequency
for conciseness
(but without loss of generality)
p. 9
Frequency Domain FIR Filter Realization
Now some matrix manipulation…
(compare to p.6)
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 10
Frequency Domain FIR Filter Realization
• An (8-channel) filter bank representation of this is...
1
u[k]
z 1
z 2
z 3
4
4
4
4
E( z )
Analysis bank:
H (z )
R (z )
4
4
4
4
z 3
z 2
z 1
+
y[k]
1
 I4 x4 
E( z )  F . 1
.

 z .I 4 x 4 
Subband processing:
H( z )  diag{P0 ( z ), P1 ( z ),...P7 ( z )} 
Synthesis bank:
R ( z )  0
I 4 x 4 . F 1
This is a 2N-channel filter bank, with N-fold downsampling
The analysis FB is a 2N-channel uniform DFT filter bank (see Chapter 10)
The synthesis FB is matched to the analysis bank, for PR: R ( z ).E( z )  z 1 I 4 x 4
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 11
Frequency Domain FIR Filter Realization
• This is again known as an `overlap-save’ realization :
– Analysis bank: performs 2N-point DFT (FFT) of a block of (N=4)
samples, together with the previous block of (N) samples
(hence `overlap’)
 I4 x4 
E( z )  F . 1
.

 z .I 4 x 4 
`block’
`previous block’
– Synthesis bank: performs 2N-point IDFT (IFFT), throws away the
1st half of the result, saves the 2nd half
(hence `save’)
R ( z )  0 I . F 1
4x4
`throw away’
`save’
– Subband processing corresponds to `frequency domain’ operation
• Complexity/latency?
See p.15…
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 12
Frequency Domain FIR Filter Realization
Derivation on p.10 can also be modified as follows…
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 13
Frequency Domain FIR Filter Realizationng
• This is known as an `overlap-add’ realization :
– Analysis bank: performs 2N-point DFT (FFT) of a block of (N=4)
samples, padded with N zero samples
 I4x4 
E( z )  F .
.

04 x 4 
`block’
`zero padding’
– Synthesis bank: performs 2N-point IDFT (IFFT), adds 2nd half
of the result to 1st half of previous IDFT
(hence `add’)
R ( z )  z 1. I
I . F 1

4x4
4x4

`overlap’
`add’
– Subband processing corresponds to `frequency domain’ operation
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 14
Frequency Domain FIR Filter Realization
• Computational complexity is (with FFT/IFFT i.o. DFT/IDFT, plus
subband processing) 2.[α.2N.log(2N)] + 2(L+1) multiplications for N
output samples, i.e. O(log(N))+O(L/N) per sample
 For large N≈L this is O(log(L)) i.e. dominated by FFT/IFFT (cheap!)
 For N<<L this is O(L), i.e. dominated by subband processing
• Processing delay/latency is O(N)
• Standard `overlap-add’ and `overlap-save’(=p.7) realizations are
derived when 0th order poly-phase components are used in the above
derivation (N=L+1, i.e. each poly-phase component has only 1 filter
coefficient). For large L, this leads to a large complexity reduction, but
also a large latency (=O(L))
• In the more general case, with higher-order polyphase components
(N<L+1, i.e. each poly-phase component has >1 filter coefficients) a
smaller complexity reduction is achieved, but the latency is also smaller
(=O(N)).
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 15
Frequency Domain Adaptive Filtering
• A similar derivation can be made for LMS-based adaptive
filtering with block processing ('Block-LMS'). The adaptive
filter then consist in a filtering operation plus an adaptation
operation, which corresponds to a correlation operation.
Both operations can be performed cheaply in the frequency
domain..
• Starting point is the LMS update equation
wk+1 = wk + m.x k .(d[k]- xTk wk )
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 16
Frequency Domain Adaptive Filtering
Consider block processing with so-called 'Block-LMS'
• Remember that LMS is a ‘stochastic gradient’ algorithm, where
instantaneous estimates of the autocorrelation matrix and crosscorrelation vector are used to compute a gradient (=steepest descent vector)
• Block-LMS uses averaged estimates, with averaging over a block of
LB (=‘block length’) samples, and hence an averaged gradient.
The update formula is then..
w n+1 = w n + m.
where n is the block index
nLB + LB
å
x k .(d[k]- xTk w n )
k = nLB +1
• Compared to LMS, Block-LMS does fewer updates (one per LB
samples), but with (presumably) better gradient estimates.
Overall,
convergence could be faster or slower (=unpredictable).
• The important thing is that Block-LMS can be realized cheaply…
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 17
Frequency Domain Adaptive Filtering
é
ê
ê
ê
ê
êë
e[4n + 4]
e[4n + 3]
e[4n + 2]
e[4n +1]
ù é
ú ê
ú ê
ú=ê
ú ê
úû êë
é
ê
as on p.6 ê
=
ê
ê
êë
d[4n + 4]
d[4n + 3]
d[4n + 2]
d[4n +1]
d[4n + 4]
d[4n + 3]
d[4n + 2]
d[4n +1]
with Wi=…
ù é w[0]
ú ê
0
ú ê
ú-ê
0
ú ê
0
úû êë
ù
ú
ú é
ú - ë 04 x 4
ú
úû
w[1]
w[0]
0
0
I4x 4
w[2]
w[1]
w[0]
0
w[3]
w[2]
w[1]
w[0]
é W
0
ê
ê 0
ê
ê 0
ê
ù.F -1. ê 0
û
ê 0
ê 0
ê
ê 0
ê
êë 0
0
w[3]
w[2]
w[1]
0
0
w[3]
w[2]
0
0
0
w[3]
0
0
0
0
é
ê
ê
ùê
úê
úê
ú. ê
úê
úû ê
ê
ê
êë
x[4n + 4]
x[4n + 3]
x[4n + 2]
x[4n +1]
x[4n]
x[4n -1]
x[4n - 2]
x[4n - 3]
0
0
0
0
0
0
0
W1
0
0
0
0
0
0
0
W2
0
0
0
0
0
0
0
W3
0
0
0
0
0
0
0
W4
0
0
0
0
0
0
0
W5
0
0
0
0
0
0
0
W6
0
0
0
0
0
0
0
W7
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
úû
ù é
ú ê
ú ê
ú ê
ú ê
ú ê
ú.F. ê
ú ê
ú ê
ú ê
ú ê
ú ê
úû ë
x[4n + 4]
x[4n + 3]
x[4n + 2]
x[4n +1]
x[4n]
x[4n -1]
x[4n - 2]
x[4n - 3]
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
úû
=frequency domain filtering
Will consider case where block length LB = filter length = L+1
The update formulas are then given as follows
1) Compute a priori residuals (example LB=L+1=4 , similar for other L)
p. 18
Frequency Domain Adaptive Filtering
é
ê
ê
ê
ê
ê
ë
w[0]
w[1]
w[2]
w[3]
ù
é
ú
ê
ú
ê
=
ú
ê
ú
ê
ú
ûn +1 ê
ë
é
ê
ê
=ê
ê
ê
ë
é
ê
as on p.6 ê
=
ê
ê
ê
ë
w[0]
w[1]
w[2]
w[3]
w[0]
w[1]
w[2]
w[3]
w[0]
w[1]
w[2]
w[3]
ù
é
ú
ê
ú
ê
+
m
.
ú
ê
ú
ê
ú
ê
ûn
ë
x[4n + 4]
x[4n + 3]
x[4n + 2]
x[4n +1]
ù
é e[4n + 4]
ú
ê
0
ú
ê
ú + m. ê
0
ú
ê
0
ú
ê
ûn
ë
ù
ú
ú
é
ú + m. ë 0 4 x 4
ú
ú
ûn
with Ei=…
I4 x 4
x[4n + 3]
x[4n + 2]
x[4n +1]
x[4n]
e[4n + 3]
e[4n + 4]
0
0
é E
0
ê
ê 0
ê
ê 0
ê
ù.F -1. ê 0
û
ê 0
ê 0
ê
ê 0
ê
ê
ë 0
x[4n + 2]
x[4n +1]
x[4n]
x[4n -1]
e[4n + 2]
e[4n + 3]
e[4n + 4]
0
x[4n +1]
x[4n]
x[4n -1]
x[4n - 2]
e[4n +1]
e[4n + 2]
e[4n + 3]
e[4n + 4]
ùé
úê
úê
ú. ê
úê
ú
ûê
ë
e[4n + 4]
e[4n + 3]
e[4n + 2]
e[4n +1]
0
e[4n +1]
e[4n + 2]
e[4n + 3]
0
0
e[4n +1]
e[4n + 2]
0
0
0
0
0
0
0
E1
0
0
0
0
0
0
0
E2
0
0
0
0
0
0
0
E3
0
0
0
0
0
0
0
E4
0
0
0
0
0
0
0
E5
0
0
0
0
0
0
0
E6
0
0
0
0
0
0
0
E7
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
ù
ú
ú
ú
ú
ú
û
ù
é
ú
ê
ú
ê
ú
ê
ú
ê
ú
ê
ú.F. ê
ú
ê
ú
ê
ú
ê
ú
ê
ú
ê
ë
ú
û
0
0
0
e[4n +1]
x[4n + 4]
x[4n + 3]
x[4n + 2]
x[4n +1]
x[4n]
x[4n -1]
x[4n - 2]
x[4n - 3]
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û
0
0
0
0
é
ê
ê
ùê
úê
úê
ú. ê
úê
ú
ûê
ê
ê
ê
ë
x[4n + 4]
x[4n + 3]
x[4n + 2]
x[4n +1]
x[4n]
x[4n -1]
x[4n - 2]
x[4n - 3]
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û
=frequency domain correlation
Will consider case where block length LB =filter length =L+1
The update formulas are then given as follows
2) Filter update (example LB=L+1=4 , similar for other L)
p. 19
Frequency Domain Adaptive Filtering
This is referred to as FDAF (’Frequency Domain Adaptive Filtering’)
• FDAF is functionally equivalent to Block-LMS
(but cheaper, see below)
• Convergence: Instead of using one and the same stepsize
 for all ‘frequency bins’, frequency dependent stepsizes
can be applied..
– In the update formula,  is removed and Ei is replaced by i.Ei
– Stepsize i dependent on the energy in the ith frequency bin
– Leads to increased convergence speed at only a small extra cost
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 20
Frequency Domain Adaptive Filtering
This is referred to as FDAF (’Frequency Domain Adaptive Filtering’)
• Complexity ≈ 5 (I)FFT’s (size 2(L+1))
per block of L+1 output samples (check!)
Hence for large L, FDAF is very efficient/cheap, only O(log(L))
multiplications per output sample (compared to O(L) for (Block-)LMS)
cost LMS
» 20
Example: LB=L+1=1024, then cost FDAF
!
• Processing delay/latency is again O(L).
Example: LB=L+1=1024 and fs=8000Hz, then delay is 256 ms !
In cases where this is objectionable (e.g. acoustic echo cancellation),
need ‘intermediate’ algorithms with smaller latency and smaller
complexity reduction, based on a polyphase decomposition…(read on)
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 21
Frequency Domain Adaptive Filtering
For large L, a block length of LB=L+1 may lead to a too large latency
If a Nth order polyphase decomposition of the adaptive filter is
considered (hence with LP+1=(L+1)/N coefficients per
polyphase component), then a frequency domain adaptive
filtering algorithm with block length LB=N can derived as
follows... (where “N takes the place of L+1”)
Example LB=N=4, i.e. (as on p.9)
W(z) = p0 (z 4 )+ z-1 p1 (z 4 )+ z-2 p2 (z 4 )+ z-3 p3 (z 4 )
with
L
p0 (z) = p00 + p10 z -1 + p02 z -2 +... + p0 p z
- Lp
p1 (z) = etc.
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 22
Frequency Domain Adaptive Filtering
(compare to p.18-19)
The update formulas are given as follows
1) Compute a priori residuals (example LB=N=4 , similar for other N)
é
ê
ê
ê
ê
ê
ë
e[4n + 4]
e[4n + 3]
e[4n + 2]
e[4n +1]
ù
ú
ú
ú=
ú
ú
û
é
ê
ê
ê
ê
ê
ë
é
ê
as on p.6 ê
=
ê
ê
ê
ë
d[4n + 4]
d[4n + 3]
d[4n + 2]
d[4n +1]
d[4n + 4]
d[4n + 3]
d[4n + 2]
d[4n +1]
with Pij=…
ì
ï
ï
ïé p0i
ù
ú L ïê
P ïê 0
ú
å íêê
ú
0
ú i = 0ï
ïê
ú
û
ïê
ë 0
ï
ï
î
ù
ú
ú é
ú - ë 04 x 4
ú
ú
û
I4x 4
p1i
p2i
p3i
i
0
i
1
i
2
p
0
0
p
i
0
p
0
p
i
1
p
i
0
p
0
i
3
p
i
2
p
i
1
p
ìé
i
ïê P0
ïê 0
ïê
ïê 0
L ïê
P ïê 0
ù.F -1.
íê
å
û
ê 0
i = 0ï
ïê
ïê 0
ïê 0
ïê
ïê 0
îë
0
0
0
0
0
0
0
0
i
3
p
i
2
i
3
é x[4(n - i) + 4] ùü
ê
úï
ê x[4n - i) + 3] úï
ùê
úï
ú ê x[4n - i) + 2] úï
ú ê x[4n - i) +1] úï
ú. ê
úý
x[4n - i)]
úê
úï
ú ê x[4n - i) -1] úï
ú
ûê
úï
ê x[4n - i) - 2] úï
ê x[4n - i) - 3] úï
ë
ûþ
p
p
0
0
0
0
0
0
0
0
P1i
0
0
0
0
0
0
0
i
2
P
0
0
0
0
0
0
0
P3i
0
0
0
0
i
4
0
0
0
P
0
0
0
0
0
0
0
P5i
0
0
i
6
0
0
0
0
0
P
0
0
0
0
0
0
0
P7i
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
ù
é
ú
ê
ú
ê
ú
ê
ú
ê
ú
ê
ú
ú.F. ê
ê
ú
ê
ú
ê
ú
ê
ú
ê
ú
ë
ú
û
x[4(n - i) + 4]
x[4(n - i) + 3]
x[4(n - i) + 2]
x[4(n - i) +1]
x[4(n - i)]
x[4(n - i) -1]
x[4(n - i) - 2]
x[4(n - i) - 3]
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û
p. 23
ü
ï
ï
ï
ï
ï
ï
ý
ï
ï
ï
ï
ï
ï
þ
Frequency Domain Adaptive Filtering
(compare to p.18-19)
The update formulas are given as follows
2) Filter update (example LB=N=4 , similar for other N)
é
ù
é
4
ê
ê p0 (z ) ú
4
ê
ê p (z ) ú
1
ú
ê
=ê
4
ê
ê p2 (z ) ú
ê
ú
ê
4
)
(z
p
ú
ê
3
ë
ûn +1 ê
ë
é
ê
ê
=ê
ê
ê
ê
ë
ù
ì
é
p0 (z 4 ) ú
ï L
ê
4
p1 (z ) ú
ï P -4 i ê
ú + m. í å z ê
p2 (z 4 ) ú
ïi = 0
ê
ú
ï
ê
ë
p3 (z 4 ) ú
î
ûn
ù
é e[4n + 4]
p0 (z 4 ) ú
ê
4
p1 (z ) ú
0
ê
ú + m. ê
0
p2 (z 4 ) ú
ê
ú
0
ê
ë
p3 (z 4 ) ú
ûn
ù
é
4
ê p0 (z ) ú
as on p.6 ê p (z 4 ) ú
1
ú + m. é 0 4 x 4
ê
=
ë
ê p2 (z 4 ) ú
ú
ê
4
ê
ûn
ë p3 (z ) ú
I4 x4
x[4(n - i) + 4]
x[4(n - i) + 3]
x[4(n - i) + 2]
x[4(n - i) +1]
e[4n + 3]
e[4n + 4]
0
0
é E
0
ê
ê 0
ê
ê 0
ê
ù.F -1. ê 0
û
ê 0
ê 0
ê
ê 0
ê
ê
ë 0
x[4(n - i) + 3]
x[4(n - i) + 2]
x[4(n - i) +1]
x[4(n - i)]
e[4n + 2]
e[4n + 3]
e[4n + 4]
0
e[4n +1]
e[4n + 2]
e[4n + 3]
e[4n + 4]
x[4(n - i) + 2]
x[4(n - i) +1]
x[4(n - i)]
x[4(n - i) -1]
0
e[4n +1]
e[4n + 2]
e[4n + 3]
0
0
e[4n +1]
e[4n + 2]
0
0
0
0
0
0
E1
0
0
0
0
0
0
0
E2
0
0
0
0
0
0
0
E3
0
0
0
0
0
0
0
E4
0
0
0
0
0
0
0
E5
0
0
0
0
0
0
0
E6
0
0
0
0
0
0
0
E7
0
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
x[4(n - i) +1]
x[4(n - i)]
x[4(n - i) -1]
x[4(n - i) - 2]
0
0
0
e[4n +1]
ù
ì
é
ú
ï
ê
ú
ï
ê
ú
ï
ê
ú
ïL
ê
P
ú
ï
ê
ú.F. í å z -4 i ê
ú
ï i=0
ê
ú
ï
ê
ú
ï
ê
ú
ï
ê
ú
ï
ê
ë
î
ú
û
ùü é
úï ê
úï ê
úý. ê
úï ê
ï
ú
ë
ûþ ê
0
0
0
0
e[4n + 4]
e[4n + 3]
e[4n + 2]
e[4n +1]
ì
é
ï
ê
ï
ê
ï
ê
ù
ï
ê
ú L
P
ï
ú
-4i ê
ú. í å z ê
ê
ú ï i=0
ï
ê
ú
û ï
ê
ï
ê
ï
ê
ë
î
x[4(n - i) + 4]
x[4(n - i) + 3]
x[4(n - i) + 2]
x[4(n - i) +1]
x[4(n - i)]
x[4(n - i) -1]
x[4(n - i) - 2]
x[4(n - i) - 3]
ù
ú
ú
ú
ú
ú
û
x[4(n - i) + 4]
x[4(n - i) + 3]
x[4(n - i) + 2]
x[4(n - i) +1]
x[4(n - i)]
x[4(n - i) -1]
x[4(n - i) - 2]
x[4(n - i) - 3]
ùü
úï
úï
úï
úï
úï
úý
úï
úï
úï
úï
úï
ûþ
p. 24
ùü
úï
úï
úï
úï
úï
úý
úï
úï
úï
úï
úï
ûþ
Frequency Domain Adaptive Filtering
This is referred to as PB-FDAF
(’Partitioned Block Frequency Domain Adaptive Filtering’)
• PB-FDAF is functionally equivalent to Block-LMS
• Complexity ≈ 3+2(LP+1) (I)FFT’s (size 2N))
per block of N output samples (check!)
Example: L+1=1024, N=128, then
cost LMS
»6
cost PB-FDAF
!
• Processing delay/latency is O(N).
Example: L+1=1024, N=128 and fs=8000Hz, then delay is 32 ms
(used in commercial acoustic echo cancellers)
DSP-CIS 2015 / Part IV / Chapter-12: Frequency Domain Filtering
p. 25