9.4.4 Double-Talk Effects and Solutions
Another extremely important problem of designing adaptive echo canceler is to handle
double-talking, which is the simultaneous presence of both echo and near-end speech.
An adaptive echo canceler estimates the impulse response of echo path using x(n) and
d(n) as shown in Figure 9.11. For correctly identifying the characteristics of P(z), d(n)
must originate solely from its input signal. During the double-talk periods, the error
signal e(n) described in (9.4.4) contains the residual echo, the uncorrelated noise v(n),
and the near-end speech u(n). The effect of interpreting near-end speech as an error
signal and making large corrections to the adaptive filter coefficients is a serious
problem.
As long as the far-end signal x(n) is uncorrelated with the near-end speech u(n), this
signal will not affect the asymptotic mean value of the filter coefficients. However, the
variation in the filter coefficients about this mean will be increased substantially in the
presence of the near-end talker due to the introduction of another large stochastic
component in the adaptation. Thus the adaptive filter W(z) is greatly disturbed in a
very short time, resulting in performance degradation of the echo canceler. An unpro-
tected algorithm may exhibit unacceptable behavior during double-talk periods and so
some mechanisms to avoid its effects must be included. This problem may be solved by
using a very small step size m. However, this may result in slow adaptation.
An effective approach for the solving double-talk problem is to detect the occurrence
of double-talking and to disable the adaptation of W(z) during these periods. Note
that nly the coefficient adaptation is disabled as illustrated in Figure 9.15, and the
transmission channel remains open in both directions at all times. If the echo path
does not change appreciably during the double-talk periods, the echo still can be
canceled by the previously converged coefficients of W(z) that are fixed during
double-talk periods.
As shown in Figure 9.15, the speech detection and control block is used to control the
adaptation of the adaptive filter W(z) and the residual echo suppressor. The complexity
of the double-talk detector (DTD), which detects the presence of near-end speech when
the far-end speech is present, is much higher. A DTD is a very critical element in echo
cancelers since an adaptive filter diverges quickly during double-talk unless the adapta-
tion process is inhibited.
Telephone
H
x(n)
e(n)
LMS
W(z)
d(n)
y(n)
+
-
Residual
echo
suppressor
Detection
& control
To
far-end
Figure 9.15 Adaptive echo canceler with speech detectors and residual echo suppressor
ADAPTIVE ECHO CANCELLATION
423