This paper presents a unified probabilistic framework for denoising and
dereverberation of speech signals. The framework transforms the denoising
and dereverberation problems into Bayes-optimal signal estimation.
The key idea is to use a strong speech model that is pre-trained on a
large data set of clean speech. Computational efficiency is achieved by
using variational EM, working in the frequency domain, and employing
conjugate priors. The framework covers both single and multiple microphones.
We apply this approach to noisy reverberant speech signals and
get results substantially better than standard methods.