Marginalization of static observation parameters in a Rao-Blackwellized particle filter with application to sequential blind speech dereverberation

Enhancement of an unknown signal from distorted observations is an extremely important Engineering problem. In addition to noise, the observation space often contains a degrading filter component. A typical example is blind speech enhancement, where a reverberant channel between a stationary source and the receiver can be modeled as a static infinite impulse response component. Particle filters have become popular and versatile estimators for estimating the clean source signal and unknown model parameters by sequentially drawing a large number of samples from a hypothesis distribution. However, direct sampling of static components leads to particle impoverishment as a dynamic is implicitly enforced on the parameters. To circumvent this issue, this paper proposes a novel approach by exploiting analytically tractable substructures of the state space to marginalize static components, facilitating separate estimation of the static parameters using their optimal estimator. The approach is tested for blind dereverberation of speech. Results show that the proposed algorithm effectively removes the effects of the static reverberant channel.