A MAP Approach to Noise Compensation of Speech

We show that estimation of parameters for the popular Gaussian model of speech in noise can be regularised in a Bayesian sense by use of simple prior distributions. For two example prior distributions, we show that the marginal distribution of the uncorrupted speech is non-Gaussian, but the parameter estimates themselves have tractable solutions. Speech recognition experiments serve to suggest values for hyper-parameters, and demonstrate that the theory is practically applicable.