Reverberated speech enhancement using neural networks

In this paper, we propose to use neural networks for blind speech dereverberation. Reverberation impulse response consists of early and late reverberation components. A neural network is first trained and then used to estimate the early and late reverberation characteristics of a reverberant room from a chosen string of speech. Using these estimated values, an inverse filter of the reverberation room impulse response is constructed. The reverberated speech is then dereverberated using the constructed room inverse filter. The proposed method shows an improvement in intelligibility and quality of speech. In addition, the method can be used for online dereverberation.