Acoustic echo cancellation using NLMS-neural network structures

One of the limitations of linear adaptive echo cancellers is nonlinearities which are generated mainly in the loudspeaker. The complete acoustic channel can be modelled as a nonlinear system convolved with a linear dispersive echo channel. Two new acoustic echo canceller models are developed to improve nonlinear performance. The first model consists of a time-delay feedforward neural network (TDNN) and the second model consists of a memoryless neural network followed by an adaptive Normalized Least Mean Square (NLMS) structure. Simuations demonstrate that both neural network based structures improve the Echo Return Loss Enhancement (ERLE) performance compared to a linear NLMS acoustic echo canceller. Experimental results using the TDNN improved the ERLE by 10 dB at low to medium loudspeaker volumes. 1.0 INTRODUCTION Limitations of echo cancellers [5][7] include (a) acoustic, thermal and DSP related noise, (b) under-modelling of the room impulse response (c) slow convergence and dynamic tracking, (d) nonlinearities in the transfer function caused mainly due to the loudspeaker, and (e) resonances and vibration in the plastic enclosure. In this paper, a tapped delay line feedforward neural network and a cascaded neural network/NLMS structure are employed in an attempt to model the system nonlinearities and acoustic path in a hands-free environment. Since there is no feedback in the network, the backpropagation algorithm [6] is used to train the networks. A typical handsfree terminal is illustrated in Figure 1 and normally consists of two Adaptive Filters (AF). The first AF is used to remove acoustic echos and the second AF is used for cancelling echoes from an imperfect hybrid as well as reflections from the line. In this paper, only the acoustic echo canceller (AEC) is considered. 1.1 Distortions in the Loudspeaker A loudspeaker has several sources of nonlinearity including non-uniform magnetic field and nonlinear suspension system [1][3]. A loudspeaker consists of an electrical part and a mechanical part. The electrical part is the voice coil and the mechanical part consists of the cone, the suspension system and the air load. The two parts interact through the magnetic field resulting in a nonlinear force deflection characteristic fM of the loudspeaker cone suspension system, usually approximated [3] by;