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.

PDF file

In  NIPS

Details