On the combination of voice prompt suppression with maximum kurtosis beamforming

In earlier work, we proposed a voice prompt suppression (VPS) algorithm based on a Kalman filter, in which the temporal update or correction step is performed in information space. The advantage of this approach is that the information matrix can be diagonally loaded in order to control the magnitude of the subband filter coefficients, which provides for better robustness. In this work, we extend that earlier work by proposing a square root implementation of the information filter VPS algorithm, as well as a technique for diagonally loading the Cholesky factor of the error covariance matrix used in this implementation. We also investigate the effectiveness of cascading VPS after maximum kurtosis beamforming, which has been shown to provide performance superior to all conventional beamforming techniques. In a set of distant speech recognition experiments we demonstrate that VPS can reduce word error rate from 19.9% to 16.1% for an adult speaker, and from 44.4% to 40.0% for a child.