Multirate Implementation of Aliasing-free Adaptive Volterra Filters by Interpolation of Higher-Order Kernel Inputs

Abstract:
This paper discusses a novel aliasing-free implementation of adaptive Volterra filters. Since any above-Nyquist frequencies generated by a continuous-time nonlinearity are typically removed by the anti-alising filter in the A/D conversion, the nonlinear system cannot be fully identified by a discrete-time model operating at the input rate. Therefore, we propose a multirate structure of adaptive Volterra filters employing interpolated higher-order kernel inputs so as to low-pass filter the corresponding outputs and retain only baseband contributions just as in the analog domain. In particular, the implications on the Volterra kernel sizes as well as an efficient polyphase structure are discussed. The conventional and the multirate implementation are compared for a nonlinear acoustic echo cancellation scenario with regard to convergence behaviour and implementation issues.