A study on target feature activation and normalization and their impacts on the performance of DNN based speech dereverberation systems

Abstract

We adopt a linear activation function at the output layer and globally normalize the target features into zero mean and unit variance to learn the complicated mapping from reverberant to anechoic speech with a regression model based on deep neural networks (DNNs). The proposed feature activation and normalization framework was found to retain clearly observable harmonics and improve the speech quality better than a recently proposed sigmoid activation and min-max normalization scheme. It also outperforms this state-of-the-art algorithm in all objective performance metrics at all reverberation times tested. With a large training set, the proposed DNN-based dereverberation system can consistently improve the restoration of the low-frequency and intermediate-frequency contents of the estimated anechoic spectrograms, essential for human perception. As for a small training set, the proposed DNN system also exhibits a better robustness than the competing algorithm.