Nonlinear echo cancellation using a partial adaptive time delay neural network

System identification of a nonlinear loudspeaker/microphone acoustic system is necessary to achieve high acoustic echo cancellation in the handsfree telephony environments where the loudspeaker often operates at high volumes. In this paper, a partial adaptive process consisting of a small order tapped delay line neural network (TDNN) followed by a delayed normalized least mean squares (NLMS) adaptive filter is used to model a loudspeaker/microphone acoustic system. The TDNN models the first part of the acoustic impulse response (AIR) where most of the energy is contained and the delayed NLMS filter models the remaining echo. Experimental measurements confirm that a short length TDNN is capable of improved identification in an undermodelled system and that by extending this to the partial adaptive TDNN structure, the ERLE performance improves by 5.5 dB at high loudspeaker volumes when compared to a NLMS structure