Digital Signal Processing Reference
In-Depth Information
let us work in the frequency domain. Then, the signal model in (1.1) becomes
Y
m
(
ω
)=
X
m
(
ω
)+
V
m
(
ω
)
(1.3)
−
ω
(
t
+
τ
m
)
+
V
m
(
ω
)
,m
=1
,
2
,...,M,
=
X
(
ω
)
e
where
ω
=2
πf
is the angular frequency,
f
denotes the temporal frequency,
=
√
−
1 represents the imaginary unit, and
Y
m
(
ω
),
X
m
(
ω
),
V
m
(
ω
), and
X
(
ω
) are the frequency-domain representations of
y
m
(
k
),
x
m
(
k
),
v
m
(
k
), and
x
(
k
), respectively.
Now, let us process the
M
signals
Y
m
(
ω
)
,m
=1
,
2
,...,M
, in order to
extract the desired signal
X
(
ω
) (up to a delay) and reduce the effect of
V
m
(
ω
).
The most straightforward and simple way of doing this is through the use
of the so-called delay-and-sum (DS) beamformer. The basic principle of this
approach is that it compensates
y
m
(
k
) with a delay
τ
m
to align all the
M
microphone signals, i.e., multiply
Y
m
(
ω
) with
e
ωτ
m
, and then average the
results together. The DS beamformer output is then
M
Z
(
ω
)=
1
M
Y
m
(
ω
)
e
ωτ
m
(1.4)
m=1
M
1
M
−
ωt
+
V
m
(
ω
)
e
ωτ
m
.
=
X
(
ω
)
e
m=1
To check whether the beamformer output is less noisy than its input, let
us compare the input and output signal-to-noise ratios (SNRs). The input
SNR of the DS beamformer, according to the signal model given in (1.3), is
defined as the SNR at the first (reference) microphone, i.e.,
iSNR(
ω
)=
φ
X
1
(
ω
)
φ
V
1
(
ω
)
(1.5)
=
φ
X
(
ω
)
φ
V
1
(
ω
)
,
,
φ
V
1
(
ω
)=
E
, and
φ
X
(
ω
)=
|X
1
(
ω
)
|
2
|V
1
(
ω
)
|
2
where
φ
X
1
(
ω
)=
E
|X
(
ω
)
|
2
E
, with
E
[
·
] denoting mathematical expectation. The output SNR
of the beamformer, from (1.4), is then
φ
X
(
ω
)
oSNR(
ω
)=
.
(1.6)
1
M
2
E
2
m=1
V
m
(
ω
)
e
ωτ
m
M
Now, we evaluate the following two cases.
•
The noise signals at the different microphones are uncorrelated.
Search WWH ::
Custom Search