Digital Signal Processing Reference
In-Depth Information
however, degrades when near-end speech is present (and is even worse
if near-end speech cannot be detected correctly). Significant performance
degradation is also expected when echo is contaminated with background
noise.
Echo cancellers generally stop filter coefficient adaptation when near-end
speech is present. An accurate near-end speech detector is therefore necessary
to avoid divergence of the filter coefficients, which may have two drawbacks.
First, the cancellation performance strongly depends on the accuracy of the
near-end speech detector. The second drawback is related to the length of the
near-end speech presence. In cases where a near-end speech segment is long,
the echo characteristics may change considerably and if the filter coefficients
are not continually adapted during those segments, then the filter will lose
synchronization with the echo path changes, leading to a large change when
filter coefficient adaptation is resumed. This may result in temporary filter
divergence causing performance reduction.
An adaptive normalized least mean squared (ANLMS) algorithm has been
suggested by Al-Naimi [24] to overcome these problems. It is based on the
NLMS algorithm (with a 128-tap transversal adaptive filter [25]). The NLMS
of [25] differs from the general NLMS in that filter coefficients are updated
less frequently with a thinning factor,
M
,resultingin
M
−
1
e(i
+
M
−
m)y(i
+
M
−
m
−
k)
m
=
0
h
k
(i
+
1
)
=
h
k
(i)
+
β
(11.64)
σ(i)
2
The ANLMS includes a number of enhancements to the system in [25]
which are: increased robustness to noise contamination, continuous filter
coefficient adaptation, and elimination of the need for a near-end speech
detector. The ANLMS is given by,
M
−
1
e(i
+
M
−
m)y(i
+
M
−
m
−
k)
m
=
0
h
k
(i
+
1
)
=
h
k
(i)
+
w
k
(i)β
(11.65)
ψ(i)
2
ρ(i)
2
where
ψ(i)
and
ρ(i)
are given by,
α
e
)
y(i)
ψ(i)
=
α
e
ψ(i
−
1
)
+
(
1
−
(11.66)
and,
ρ(i)
=
α
e
ρ(i
−
1
)
+
(
1
−
α
e
)
|
z(i)
|
(11.67)
Search WWH ::
Custom Search