Speech Denoising and Dereverberation Using Probabilistic Models

This paper presents a unified probabilistic framework for denoising and dereverberation of speech signals. The framework transforms the denoising nd 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.