Over

Overlap-Save and Overlap-Add for Real-time Processing
Overlap-Save and Overlap-Add for Real-time

Processing
Overlap-Save and Overlap-Add for Real-time
Filtering Reference:
Professor Deepa Kundur Section 7.3 of
University of Toronto John G. Proakis and Dimitris G. Manolakis, Digital Signal Processing:
Principles, Algorithms, and Applications, 4th edition, 2007.
Professor Deepa Kundur (University of Toronto)

Overlap-Save and Overlap-Add for Real-time Filtering 1 / 59 Professor Deepa Kundur (University of Toronto)
Overlap-Save and Overlap-Add for Real-time Filtering 2 / 59
Overlap-Save and Overlap-Add for Real-time Processing Circular and Linear Convolution Overlap-Save and Overlap-Add for Real-time Processing Circular and Linear Convolution
The Discrete Fourier Transform Pair Important DFT Properties
Property Time Domain Frequency Domain

I DFT and inverse-DFT (IDFT): Notation: x(n) X (k)
Periodicity: x(n) = x(n + N) X (k) = X (k + N)
N−1 Linearity: a1 x1 (n) + a2 x2 (n) a1 X1 (k) + a2 X2 (k)
n
X Time reversal x(N − n) X (N − k)
X (k) = x(n)e −j2πk N , k = 0, 1, . . . , N − 1 Circular time shift: x((n − l))N X (k)e −j2πkl/N
n=0 Circular frequency shift: x(n)e j2πln/N X ((k − l))N
N−1 Complex conjugate: x ∗ (n) X ∗ (N − k)
1 X n Circular convolution: x1 (n) ⊗ x2 (n) X1 (k)X2 (k)
x(n) = X (k)e j2πk N , n = 0, 1, . . . , N − 1 Multiplication: x1 (n)x2 (n) 1
X1 (k) ⊗ X2 (k)
N Parseval’s theorem:
P N−1 ∗
N P
1 N−1 ∗
k=0 n=0 x(n)y (n) N k=0 X (k)Y (k)

DTFT Theorems and Properties Circular and Linear Convolution
Property Time Domain Frequency Domain

Notation: x(n) X (ω)
x1 (n) X1 (ω)
x2 (n) X1 (ω) F
Linearity: a1 x1 (n) + a2 x2 (n) a1 X1 (ω) + a2 X2 (ω)
x1 (n) ⊗ x2 (n) ←→ X1 (k)X2 (k)
Time shifting: x(n − k) e −jωk X (ω) F
Time reversal x(−n) X (−ω) x1 (n) ∗ x2 (n) ←→ X1 (ω)X2 (ω)
Convolution: x1 (n) ∗ x2 (n) X1 (ω)X2 (ω)
Correlation: rx1 x2 (l) = x1 (l) ∗ x2 (−l) Sx1 x2 (ω) = X1 (ω)X2 (−ω)
= X1 (ω)X2∗ (ω) [if x2 (n) real]
Wiener-Khintchine: rxx (l) = x(l) ∗ x(−l) Sxx (ω) = |X (ω)|2 Q: What is circular convolution?

Circular Convolution Circular Convolution

Assume: x1 (n) and x2 (n) have support n = 0, 1, . . . , N − 1. Assume: x1 (n) and x2 (n) have support n = 0, 1, . . . , N − 1.
N−1
Examples: N = 10 and support: n = 0, 1, . . . , 9 X
x1 (n) ⊗ x2 (n) = x1 (k)x2 ((n − k))N
k=0
N−1
1 1
X
= x2 (k)x1 ((n − k))N
-3 -2 -1 0 1 2 3
4 5
6 7 8 9 10 11 12
n -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12
n k=0
where (n)N = n mod N = remainder of n/N.

1
-3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12
n

Modulo Indices and Periodic Repetition Modulo Indices and Periodic Repetition
(n)N = n mod N = remainder of n/N (n)N = n mod N = remainder of n/N
Example: N = 4 Example: N = 4
n -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 n -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
(n)4 0 1 2 3 0 1 2 3 0 1 2 3 0 (n)4 0 1 2 3 0 1 2 3 0 1 2 3 0
n nonneg integer < N x((n))4 will be periodic with period 4. The repeated pattern will be
= integer +
N N consist of: {x(0), x(1), x(2), x(3)}.
5 1 −2 2
=1+ = −1 + Thus, x((n))N is a periodic signal comprised of the following
4 4 4 4
repeating pattern: {x(0), x(1), · · · x(N − 2), x(N − 1)}.

Overlap During Periodic Repetition Overlap During Periodic Repetition

A periodic repetition makes an aperiodic signal x(n), periodic to A periodic repetition makes an aperiodic signal x(n), periodic to
produce xp (n). produce xp (n).
∞
There are two important parameters:
X
xp (n) = x(n − lN) 1. smallest support length of the signal x(n)
l=−∞ 2. period N used for repetition that determines the period of xp (n)
I smallest support length > period of repetition
I there will be overlap
I smallest support length ≤ period of repetition
I there will be no overlap
⇒ x(n) can be recovered from xp (n)

Periodic Repetition: Example N = 4 Period Repetition: Example N = 4

l=-2 l=-1 l=0 l=1 l=2 l=-2 l=-1 l=0 l=1 l=2
x (n+N) x (n-N) x (n+N) x (n-N)
x (n) x (n+2N) x (n) x (n-2N) x (n) x (n+2N) x (n) x (n-2N)
2 + + 2
+ + + + +
... ... ...
1 1
+ +
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n
no overlap overlap
support length = 4 = N xp(n) support length = 6 > N xp(n)
2 2
1 1
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n
= x (n) =x(n)

Modulo Indices and the Periodic Repetition Modulo Indices and the Periodic Repetition
x(n)
2
Assume: x(n) has support n = 0, 1, . . . , N − 1.
1
∞
n
X -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
x((n))N = x(n mod N) = xp (n) = x(n − lN)

l=−∞
Note: Because the support size and period size are the same, there is
no overlap when taking the periodic repetition x((n))N .

x(n) 2 x(n) 2 x((n))N
2
2 2
1 1 1 1
1 0 1 0
1 1
n n
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 0

x(n) 2 x((n))N
2 x(n) 2 x((n))N
2
12 12
2 8 2 8
1 1 4 1 1 4
0 0
1 0 1 0
-4 -4
15
11
15
11
-1
-1
7
3
7
3
1 1 1 1
n n
13
13
-3
-3
1
5
9
1
5
9
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 0
-2 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 0
-2
2
~ 2
6 x (n) 6
10 10
14 2 14
n
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 ...
Therefore x((n))N = xp (n).

Circular Convolution: One Interpretation Circular Convolution: One Interpretation

Assume: x1 (n) and x2 (n) have support n = 0, 1, . . . , N − 1.
PN−1 PN−1
To compute k=0 x1 (k)x2 ((n − k))N (or k=0 x2 (k)x1 ((n − k))N ):
1. Take the periodic repetition of x2 (n) with period N:
N−1 N−1
∞ X X
x1 (k) x2 ((n − k))N = x1 (k) x2p (n − k)
X
x2p (n) = x2 (n − lN)
l=−∞ k=0 k=0
2. Conduct a standard linear convolution of x1 (n) and x2p (n) for

n = 0, 1, . . . , N − 1: . . . which makes sense, since x((n))N = xp (n).
∞
X N−1
X
x1 (n) ⊗ x2 (n) = x1 (n) ∗ x2p (n) = x1 (k)x2p (n − k) = x1 (k)x2p (n − k)
k=−∞ k=0
Note: x1 (n) ⊗ x2 (n) = 0 for n < 0 and n ≥ N.

Circular Convolution: Another Interpretation Using DFT for Linear Convolution

Assume: x1 (n) and x2 (n) have support n = 0, 1, . . . , N − 1.
PN−1 PN−1
Therefore, circular convolution and linear convolution are related as
To compute k=0 x1 (k)x2 ((n − k))N (or k=0 x2 (k)x1 ((n − k))N ): follows:
1. Conduct a linear convolution of x1 (n) and x2 (n) for all n: ∞
X
∞ N−1 xC (n) = x1 (n) ⊗ x2 (n) = xL (n − lN)
X X
xL (n) = x1 (n) ∗ x2 (n) = x1 (k)x2 (n − k) = x1 (k)x2 (n − k) l=−∞
k=−∞ k=0
for n = 0, 1, . . . , N − 1
2. Compute the periodic repetition of xL (n) and window the result for
n = 0, 1, . . . , N − 1: Q: When can one recover xL (n) from xC (n)?
When can one use the DFT to compute linear convolution?
∞
A: When there is no overlap in the periodic repetition of xL (n).
X
x1 (n) ⊗ x2 (n) = xL (n − lN), n = 0, 1, . . . , N − 1
l=−∞ When support length of xL (n) ≤ N.

Overlap-Save and Overlap-Add for Real-time Processing Circular and Linear Convolution Overlap-Save and Overlap-Add for Real-time Processing Filtering of Long Data Sequences
Using DFT for Linear Convolution Filtering of Long Data Sequences
Let x(n) have support n = 0, 1, . . . , L − 1.

Let h(n) have support n = 0, 1, . . . , M − 1.
We can set N ≥ L + M − 1 and zero pad x(n) and h(n) to have I The input signal x(n) is often very long especially in real-time
support n = 0, 1, . . . , N − 1. signal monitoring applications.
I For linear filtering via the DFT, for example, the signal must be
1. Take N-DFT of x(n) to give X (k), k = 0, 1, . . . , N − 1.
limited size due to memory requirements.
2. Take N-DFT of h(n) to give H(k), k = 0, 1, . . . , N − 1.
3. Multiply: Y (k) = X (k) · H(k), k = 0, 1, . . . , N − 1.
4. Take N-IDFT of Y (k) to give y (n), n = 0, 1, . . . , N − 1.

Overlap-Save and Overlap-Add for Real-time Processing Filtering of Long Data Sequences Overlap-Save and Overlap-Add for Real-time Processing Filtering of Long Data Sequences
Filtering of Long Data Sequences Filtering of Long Data Sequences
I All N-input samples are required simultaneously by the DFT

operator. I Strategy:
1. Segment the input signal into fixed-size blocks prior to
I If N is too large as for long data sequences, then there is a processing.
significant delay in processing that precludes real-time 2. Compute DFT-based linear filtering of each block separately.
processing. 3. Fit the output blocks together in such a way that the overall
output is equivalent to the linear filtering of x(n) directly.
I Main advantage: samples of the output y (n) = h(n) ∗ x(n) will
signal Data Acquisition Data Processing signal
input Delay Delay output
be available real-time on a block-by-block basis.

Overlap-Save and Overlap-Add for Real-time Processing Filtering of Long Data Sequences Overlap-Save and Overlap-Add for Real-time Processing Overlap-Add Method
Filtering of Long Data Sequences Overlap-Add Method
I Goal: FIR filtering: y (n) = x(n) ∗ h(n)
I Two approaches to real-time linear filtering of long inputs:

I Overlap-Add Method
I Overlap-Save Method
Overlap-Add Method
I Assumptions:
I FIR filter h(n) length = M
I Block length = L M

Overlap-Save and Overlap-Add for Real-time Processing Overlap-Add Method Overlap-Save and Overlap-Add for Real-time Processing Overlap-Add Method
x(n)
Overlap-Add Method 1
* h(n)
-3 -2 -1 0 1 2 3
4 5
6 7 8 9
10
n
L=4 L=4 L=4 L=4 Input x(n) is divided
Deals with the following signal processing principles:
x1(n) into non-overlapping
blocks xm (n) each of
I The linear convolution of a discrete-time signal of length L and a = 1
* h(n) length L.
discrete-time signal of length M produces a discrete-time -3 -2 -1 0 1 2 3 4 5 6 7 8 9
10
n
convolved result of length L + M − 1. x2(n)
L=4
Each input block
xm (n) is individually
+ 1
* h(n)
I Addititvity: filtered as it is received
-3 -2 -1 0 1 2 3
4 5
6 7 8 9
10
n to produce the output
L=4
(x1 (n)+x2 (n)) ∗ h(n) = x1 (n) ∗ h(n)+x2 (n) ∗ h(n) x3(n) block ym (n).
+ 1
* h(n) + ...
-3 -2 -1 0 1 2 3 6 7 8 9
10
n
L=4

Overlap-Add Filtering Stage Using DFT for Linear Convolution
Let xm (n) have support n = 0, 1, . . . , L − 1.

Let h(n) have support n = 0, 1, . . . , M − 1.
I makes use of the N-DFT and N-IDFT where: N = L + M − 1
We set N ≥ L + M − 1 (the length of the linear convolution result)
I Thus, zero-padding of x(n) and h(n) that are of length and zero pad xm (n) and h(n) to have support n = 0, 1, . . . , N − 1.
L, M < N is required.
1. Take N-DFT of xm (n) to give Xm (k), k = 0, 1, . . . , N − 1.
I The actual implementation of the DFT/IDFT will use the fast
Fourier Transform (FFT) for computational simplicity. 2. Take N-DFT of h(n) to give H(k), k = 0, 1, . . . , N − 1.
3. Multiply: Ym (k) = Xm (k) · H(k), k = 0, 1, . . . , N − 1.
4. Take N-IDFT of Ym (k) to give ym (n), n = 0, 1, . . . , N − 1.

x(n)
Linear Convolution via the DFT 1
* h(n)
Length of linear convolution result = Length of DFT
-3 -2 -1 0 1 2 3
4 5
6 7 8 9
10
n
l=-2 l=-1 l=0 l=1 l=2 L=4 L=4 L=4 L=4 Input x(n) is divided
x (n+N) x (n-N)
x (n) x (n+2N) x (n) x (n-2N) x1(n) into non-overlapping
2 + + blocks xm (n) each of
... + +
... = 1
* h(n) length L.
1
-3 -2 -1 0 1 2 3 4 5 6 7 8 9
10
n
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n L=4
x2(n) Each input block
xm (n) is individually
no overlap + 1
* h(n)
xp(n)
filtered as it is received
support length = 4 = N 4 5 10
n
-3 -2 -1 0 1 2 3 6 7 8 9 to produce the output
2 L=4
x3(n) block ym (n).
1
+ 1
* h(n) + ...
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n
-3 -2 -1 0 1 2 3 6 7 8 9
10
n
= x (n) L=4

x(n)
Overlap-Add Addition Stage
1
* h(n)
-3 -2 -1 0 1 2 3
4 5
6 7 8 9
10
n
L=4 L=4 L=4 L=4 Output blocks ym (n)
These extra M-1 samples (i.e., “tail”) must be fitted together From the Addititvity property, since:
y1(n) due to convolution expanding
the support must be added to the appropriately to gener- ∞
begininning of the next output
= 1
X
block ate: x(n) = x1 (n) + x2 (n) + x3 (n) + · · · = xm (n)
-3 -2 -1 0 1 2 3 4 5 6 7 8 9
10
n m=1
L=4 M-1=2 y (n) = x(n) ∗ h(n) x(n) ∗ h(n) = (x1 (n) + x2 (n) + x3 (n) + · · · ) ∗ h(n)
y2(n) ADD
to
next
The support overlap = x1 (n) ∗ h(n) + x2 (n) ∗ h(n) + x3 (n) ∗ h(n) + · · ·
+ 1
block
∞ ∞
4 5 6 8 9 10
n amongst the ym (n) X X
-3 -2 -1 0 1 2 3 6 7 = xm (n) ∗ h(n) = ym (n)
L=4 M-1=2 blocks must be ac-
y3(n) m=1 m=1
ADD
to
counted for.
+ 1 next
block
+ ...
-3 -2 -1 0 1 2 3 6 7 8 9
10
n
L=4 M-1=2

x(n)
Input signal:
1
* h(n)
1
4 5 10
n
=-3 -2 -1 0 1
L=4
2 3 6
L=4
7 8 9
L=4 L=4 Output blocks ym (n)

These 7 M-1
6extra samples (i.e., “tail”) n must be fitted together
-3 -2
y1(n)
-1 0 1 2 3 4 5 8 9 10
due to convolution expanding
the support must be added to the appropriately to gener-
= 1 begininning of the next output zeros
block ate:
10
n
-3 -2 -1 0 1 2 3 4 5 6 7 8 9
zeros
L=4 M-1=2 y (n) = x(n) ∗ h(n)
y2(n) ADD
to Output signal: zeros
+ 1 next
block
The support overlap
4 5 6 8 9 10
n amongst the ym (n)
-3 -2 -1 0 1 2 3 6 7
blocks must be ac- Add
L=4 M-1=2
y3(n) ADD
to
counted for. points Add
+ 1 next + ...
block
points
-3 -2 -1 0 1 2 3 6 7 8 9
10
n
L=4 M-1=2

Overlap-Add Method Overlap-Add Method: Cautionary Note
1. Break the input signal x(n) into non-overlapping blocks xm (n) of length L.
2. Zero pad h(n) to be of length N = L + M − 1.
3. Take N-DFT of h(n) to give H(k), k = 0, 1, . . . , N − 1. If you DO NOT overlap and add, but only append the output blocks
4. For each block m: ym (n) for m = 1, 2, . . ., then you will not get the true y (n) sequence.
4.1 Zero pad xm (n) to be of length N = L + M − 1.
4.2 Take N-DFT of xm (n) to give Xm (k), k = 0, 1, . . . , N − 1. Q: What sequence will you obtain instead?
4.3 Multiply: Ym (k) = Xm (k) · H(k), k = 0, 1, . . . , N − 1.
4.4 Take N-IDFT of Ym (k) to give ym (n), n = 0, 1, . . . , N − 1.
5. Form y (n) by overlapping the last M − 1 samples of ym (n) with the first
M − 1 samples of ym+1 (n) and adding the result.

Overlap-Add Method: Cautionary Note Overlap-Add Method: Cautionary Note

Output signal:
Add
points Add
zeros
points
zeros
zeros

Overlap-Save and Overlap-Add for Real-time Processing Overlap-Save Method Overlap-Save and Overlap-Add for Real-time Processing Overlap-Save Method
Overlap-Save Method Overlap-Save Method
Deals with the following signal processing principles:
I The N = (L + M − 1)-circular convolution of a discrete-time

signal of length N and a discrete-time signal of length M using
Overlap-Save Method an N-DFT and N-IDFT.
Overlap-[Discard] Method
I Time-Domain Aliasing:
∞
X
xC (n) = xL (n − lN) , n = 0, 1, . . . , N − 1
| {z }
l=−∞ support=M + N − 1

Overlap-Save Method
yL(n+N) yL(n-N)
yL(n) linear yL(n+2N) yL(n) yL(n-2N)
convolution I Convolution of xm (n) with support n = 0, 1, . . . , N − 1 and h(n)
2
result + + + ... with support n = 0, 1, . . . , M − 1 via the N-DFT will produce a
...
1
+ +
result yC ,m (n) such that:
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n

Add and keep points only aliasing corruption n = 0, 1, . . . , M − 2
yC ,m (n) =
at n=0, 1, ..., N-1 yL,m (n) n = M − 1, M, . . . , N − 1
overlap since
N < L+M-1 yC (n) where yL,m (n) = xm (n) ∗ h(n) is the desired output.
corruption from
circular 2 previous repetitions
convolution I The first M − 1 points of a the current filtered output block
1
result ym (n) must be discarded.
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n
I The previous filtered block ym−1 (n) must compensate by
= yL(n) at n=M-1, M, ..., N-1
providing these output samples.

Overlap-Save Input Segmentation Stage Overlap-Save Input Segmentation Stage
1. All input blocks xm (n) are of length N = (L + M − 1) and Input signal blocks:
contain sequential samples from x(n).
zeros
2. Input block xm (n) for m > 1 overlaps containing the first M − 1
points of the previous block xm−1 (n) to deal with aliasing
corruption.
point
overlap
3. For m = 1, there is no previous block, so the first M − 1 points point
are zeros. overlap

Overlap-Save Input Segmentation Stage Overlap-Save Filtering Stage
I makes use of the N-DFT and N-IDFT where: N = L + M − 1

x1 (n) = {0, 0, . . . 0, x(0), x(1), . . . , x(L − 1)}
| {z }
M − 1 zeros
I Only a one-time zero-padding of h(n) of length M L < N is
x2 (n) = {x(L − M + 1), . . . x(L − 1), x(L), . . . , x(2L − 1)}
| {z } required to give it support n = 0, 1, . . . , N − 1.
last M − 1 points from x1 (n)
x3 (n) = {x(2L − M + 1), . . . x(2L − 1), x(2L), . . . , x(3L − 1)} I The input blocks xm (n) are of length N to start, so no
zero-padding is necessary.
| {z }
last M − 1 points from x2 (n)
..
. I The actual implementation of the DFT/IDFT will use the fast
Fourier Transform (FFT) for computational simplicity.
The last M − 1 points from the previous input block must be saved
for use in the current input block.

Using DFT for Circular Convolution Circular Convolution via the DFT
Length of linear convolution result > Length of DFT
N = L + M − 1. yL(n+N) yL(n-N)
yL(n) linear yL(n+2N) yL(n) yL(n-2N)
Let xm (n) have support n = 0, 1, . . . , N − 1. convolution
Let h(n) have support n = 0, 1, . . . , M − 1. 2
result + + + ...
...
1
+ +
We zero pad h(n) to have support n = 0, 1, . . . , N − 1. -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
n
Add and keep points only
at n=0, 1, ..., N-1
1. Take N-DFT of xm (n) to give Xm (k), k = 0, 1, . . . , N − 1. overlap since
N < L+M-1 yC (n)
2. Take N-DFT of h(n) to give H(k), k = 0, 1, . . . , N − 1. corruption from
circular 2 previous repetitions
3. Multiply: Ym (k) = Xm (k) · H(k), k = 0, 1, . . . , N − 1. convolution 1
result
4. Take N-IDFT of Ym (k) to give yC ,m (n), n = 0, 1, . . . , N − 1. n
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
= yL(n) at n=M-1, M, ..., N-1

Overlap-Save Output Blocks Overlap-Save [Discard] Output Blocks
y1 (n) = {y1 (0), y1 (1), . . . y1 (M − 2), y (0), . . . , y (L − 1)}

| {z }
M − 1 points corrupted from aliasing
y2 (n) = {y2 (0), y2 (1), . . . y2 (M − 2), y (L), . . . , y (2L − 1)}
| {z }
aliasing n = 0, 1, . . . , M − 2 M − 1 points corrupted from aliasing
yC ,m (n) = y3 (n) = {y3 (0), y3 (1), . . . y3 (M − 2), y (2L), . . . , y (3L − 1)}
yL,m (n) n = M − 1, M, . . . , N − 1 | {z }
M − 1 points corrupted from aliasing
where yL,m (n) = xm (n) ∗ h(n) is the desired output. where y (n) = x(n) ∗ h(n) is the desired output.
The first M − 1 points of each output block are discarded.
The remaining L points of each output block are appended to form

y (n).

Overlap-Save Output Stage Overlap-Save Method

1. Insert M − 1 zeros at the beginning of the input sequence x(n).
Output signal blocks: 2. Break the padded input signal into overlapping blocks xm (n) of length
N = L + M − 1 where the overlap length is M − 1.
Discard 3. Zero pad h(n) to be of length N = L + M − 1.
points Discard 4. Take N-DFT of h(n) to give H(k), k = 0, 1, . . . , N − 1.
points Discard 5. For each block m:
points 5.1 Take N-DFT of xm (n) to give Xm (k), k = 0, 1, . . . , N − 1.
5.2 Multiply: Ym (k) = Xm (k) · H(k), k = 0, 1, . . . , N − 1.
5.3 Take N-IDFT of Ym (k) to give ym (n), n = 0, 1, . . . , N − 1.
5.4 Discard the first M − 1 points of each output block ym (n).
6. Form y (n) by appending the remaining (i.e., last) L samples of each block
ym (n).

Overlap-Save and Overlap-Add for Real-time Processing Overlap-Save Method
Overlap-Save Method
Input signal blocks:
zeros
point
overlap
point
Output signal blocks: overlap
Discard
points Discard
points Discard
points



Over

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Over

Uploaded by

Copyright:

Available Formats

Overlap-Save and Overlap-Add for Real-time Processing

Overlap-Save and Overlap-Add for Real-time

Professor Deepa Kundur Section 7.3 of

Professor Deepa Kundur (University of Toronto)

The Discrete Fourier Transform Pair Important DFT Properties

Property Time Domain Frequency Domain

Professor Deepa Kundur (University of Toronto)

DTFT Theorems and Properties Circular and Linear Convolution

Property Time Domain Frequency Domain

Professor Deepa Kundur (University of Toronto)

Circular Convolution Circular Convolution

where (n)N = n mod N = remainder of n/N.

Professor Deepa Kundur (University of Toronto)

(n)N = n mod N = remainder of n/N (n)N = n mod N = remainder of n/N

Professor Deepa Kundur (University of Toronto)

Overlap During Periodic Repetition Overlap During Periodic Repetition

Professor Deepa Kundur (University of Toronto)

Periodic Repetition: Example N = 4 Period Repetition: Example N = 4

Professor Deepa Kundur (University of Toronto)

x((n))N = x(n mod N) = xp (n) = x(n − lN)

Professor Deepa Kundur (University of Toronto)

Professor Deepa Kundur (University of Toronto)

Therefore x((n))N = xp (n).

Professor Deepa Kundur (University of Toronto)

Circular Convolution: One Interpretation Circular Convolution: One Interpretation

2. Conduct a standard linear convolution of x1 (n) and x2p (n) for

Note: x1 (n) ⊗ x2 (n) = 0 for n < 0 and n ≥ N.

Professor Deepa Kundur (University of Toronto)

Circular Convolution: Another Interpretation Using DFT for Linear Convolution

Professor Deepa Kundur (University of Toronto)

Using DFT for Linear Convolution Filtering of Long Data Sequences

Let x(n) have support n = 0, 1, . . . , L − 1.

Professor Deepa Kundur (University of Toronto)

Filtering of Long Data Sequences Filtering of Long Data Sequences

I All N-input samples are required simultaneously by the DFT

Professor Deepa Kundur (University of Toronto)

Filtering of Long Data Sequences Overlap-Add Method

I Goal: FIR filtering: y (n) = x(n) ∗ h(n)

I Two approaches to real-time linear filtering of long inputs:

Professor Deepa Kundur (University of Toronto)

Professor Deepa Kundur (University of Toronto)

Overlap-Add Filtering Stage Using DFT for Linear Convolution

Let xm (n) have support n = 0, 1, . . . , L − 1.

Professor Deepa Kundur (University of Toronto)

Professor Deepa Kundur (University of Toronto)

Professor Deepa Kundur (University of Toronto)

L=4 L=4 Output blocks ym (n)

Professor Deepa Kundur (University of Toronto)

Overlap-Add Method Overlap-Add Method: Cautionary Note

Professor Deepa Kundur (University of Toronto)

Overlap-Add Method: Cautionary Note Overlap-Add Method: Cautionary Note

Professor Deepa Kundur (University of Toronto)

Overlap-Save Method Overlap-Save Method

Deals with the following signal processing principles:

I The N = (L + M − 1)-circular convolution of a discrete-time

Professor Deepa Kundur (University of Toronto)

Professor Deepa Kundur (University of Toronto)

Overlap-Save Input Segmentation Stage Overlap-Save Input Segmentation Stage

contain sequential samples from x(n).

Professor Deepa Kundur (University of Toronto)

Overlap-Save Input Segmentation Stage Overlap-Save Filtering Stage

I makes use of the N-DFT and N-IDFT where: N = L + M − 1

Professor Deepa Kundur (University of Toronto)