Single-Channel Speech Restoration using Deep Speech Features Reconstruction

Single-channel deep neural network (DNN) methods for noise suppression have shown great promise in recent years. In many applied telecommunication applications the desired speech quality is limited by processing performed either on the far end (FE) side, or by a speech processing chain which is implemented by a third-party vendor. In this work, we present a method for single-channel speech enhancement. We propose positioning a DNN following a traditional noise suppression module, aiming to restore the missing speech features lost by the traditional module, without affecting the noise profile, thereby restoring quality and audibility to the desired signal. The DNN is designed to induce a small processing delay, thereby making it an attractive addition to a classical speech processing chain. The proposed network improvement to the existing chain is presented and compared to recent speech enhancement solutions.