Towards Online Maximum Kurtosis Beamforming

In prior work, the current authors investigated the use of optimization criteria for beamforming that exploit the non-Gaussianity of human speech. In particular, we examined beamforming algorithms designed to maximize the kurtosis or negentropy of the subband output of a generalized sidelobe canceller. These techniques, while effective, require making multiple passes through the data, and hence are unsuitable for online implementation. Thus, in this work, we propose an online implementation of the maximum kurtosis beamformer. In a set of distant speech recognition experiments, we compare the effectiveness of the proposed technique to several common beamformer designs. Compared to a single channel of the array, the proposed algorithm reduced word error rate from 24.0% to 10.3%, which is the best performance yet achieved on this task.