PROJECT TITLE:

A Regression Approach to Speech Enhancement Based on Deep Neural Networks - 2015

ABSTRACT:

In distinction to the conventional minimum mean square error (MMSE)-primarily based noise reduction techniques, we have a tendency to propose a supervised technique to reinforce speech by means that of finding a mapping function between noisy and clean speech signals based mostly on deep Neural Networks (DNNs). In order to be ready to handle a wide selection of additive noises in real-world situations, a large coaching set that encompasses several doable mixtures of speech and noise varieties, is initial designed. A DNN architecture is then used as a nonlinear regression perform to ensure a strong modeling capability. Several techniques have also been proposed to improve the DNN-based mostly speech enhancement system, as well as world variance equalization to alleviate the over-smoothing downside of the regression model, and the dropout and noise-aware coaching strategies to any improve the generalization capability of DNNs to unseen noise conditions. Experimental results demonstrate that the proposed framework will achieve important improvements in each objective and subjective measures over the conventional MMSE based technique. It is additionally attention-grabbing to watch that the proposed DNN approach will well suppress highly nonstationary noise, that is robust to handle generally. Furthermore, the ensuing DNN model, trained with artificial synthesized data, is additionally effective in coping with noisy speech information recorded in real-world scenarios without the generation of the annoying musical artifact commonly observed in standard enhancement ways.