Adaptive Beamformer Performance In Reverberation

Introduction. Recent applications of adaptive filtering to hearing aids [4-81 have shown that very simple two-microphone systems can provide large improvements in target-to-jammer ratio under anechoic conditions. Some of these studies also considered non-anechoic conditions and showed that the presence of reverberation has a strong effect on performance. Because of differences among the acoustic and signal-processing conditions of these studies, a more detailed summary cannot be given. The present study illustrates , for a particular adaptive beamformer, the effects of reverberation on performance and the interactions among reverberation, target-to-jammer ratio (TJR) and filter length (L). System Description and Methods. The system used here (Figure 1) is a modified, two-microphone version of the Griffiths-Jim (21 constnained adaptive beamformer. The modifications were developed to deal with the problems of misadjustnient and misalignment at high TJRs and do so by exploiting the fluctuations in speech to allow adaptation during intervals of low TJR [l]. The first method employs the input correlation p between bandpass-filtered microphone signals as a measure of TJR and inhibits adaptation when p exceeds a threshold. The second method includes output power in the normalization of the weight update: Aw&z + 1) = 2ay[n]d[n-k]/{L(P,,[n] + Pd[n])}, where Py and P d are running estimates of power in the system output and the adaptive filter' The study employed computer simulations of this beamformer with 7-cm spacing between microphones in free-space. Input signals were generated by convolving single-talker target and babble jammer sources with synthetic source-to-microphone impulse responses [3]. For all conditions, the target was located at 0", broadside to the array, the jammer was at 45", and both sources were 0.9 m from the center of the array. Output target and jammer were measured separately through use of a master and two slave processors. The master processed target and jammer summed together, while the slave systems processed the target and jammer separately using adaptive filter weights copied from the master. The performance metric, GI, is a spectrally-weighted gain in target-to-jammer ratio from input to output, measured in the steady-state [6]. Results. A sampling of rooms and source/array geometries was simulated to study the joint effects of TJR, degree of reverberation, and filter length. Condition A employed a room with dimensions 5.2 x 3.4 x 2.8 mtders and a uniform absorption coefficient of 1.0 (anechoic), 0.6, or 0.2, resulting in direct-to-reverberant energy ratios at the array of 00, 5.7,oI-2.4 dB, respectively. The …