Crosstalk robust structures based on sub-band noise estimation

Abstract

This paper will focus on a new approach, which was derived from the multi-band ANC concept. The proposed structures referred as Fast Multi-Band ANCs (FMB-ANC) are immune to crosstalk, allowing higher robustness to instability and faster convergence than the CrossTalk Resistant ANC approach (CTRANC). The method is efficient for speech enhancement because it removes reasonably the correlation between the primary and reference estimated speech components. A comparison between FMB-ANCs based on the transversal filter (FIR-LMS) and on the linear combined gradient adaptive lattice filter (GAL) is presented. Noisy speech acquired in real environments and coloured synthetic input signals, a single sinusoid and a chirped sinusoid in white noise were used to evaluate and compare the two adaptive structures. The adaptive lattice filters are less sensitive to eigenvalue spread and present faster convergence than the conventional ANC. However, in the FMB-ANC the sub-band noise estimation facilitates the convergence and increases the level of sinusoid cancellation. The experimental results obtained with real time implementations on a DSP, proved the superiority of the simpler solution based on transversal filters, while the computational burden is kept low.