Experiments to identify and track non stationarities in audio conference rooms

The author proposes a modification of the least mean square algorithm to improve its capacities for the suppression of the acoustical feedback in audioconference rooms. Time-domain or frequency-domain algorithms are usually based on a uniform approach of the whole response of the acoustical channel. The author introduces a priori geometrical information on the response to divide it into two kinds of coefficients, fast or slow varying, with the step size of the gradient algorithm depending on the kind of coefficient. This allows a better adaptation, which can be still improved, using an algorithm of detection of motion in the room, to separate stationary and nonstationary periods