There are algorithms that do not rely on level differences to determine if
double-talk is present. These algorithms do not suffer from input level
differences. I would suggest you do a patent search on echo canceller
One detector I know of implements a form of the Optimal Bounding Ellipsoid by
preventing the update of the adaptive filter if no innovation is detected from
the last samle to the present.
To answer specifically your question, the echo contains the double-talk signal
as well as the echo. By applying the ALC to this signal, you are amplyfing the
echo as well as the double-talk signal. The ALC my not help since the ratio of
the 2 will not be changed. In the worst case, the echo can be 6 dB down from a
nominal input level, and the double-talk signal can be 8 dB down from a nomimal
input level. Under these conditions, you need more than level detection to
determine if double-talk is present.