Clean speech AE-DNN PSD constraint for MCLP based reverberant speech enhancement
Srikanth Raj Chetupalli, T.V. Sreenivas · 2019
Blind inverse filtering using multi-channel linear prediction (MCLP) in short-time Fourier transform (STFT) domain is an effective means to enhance reverberant speech. Traditionally, a speech power spectral density (PSD) weighted prediction error (WPE) minimization approach is used to estimate the prediction filters, independently in each frequency bin. The method is sensitive to the estimation of desired signal PSD. In this paper, we propose an auto-encoder (AE) deep neural network (DNN) based constraint for the estimation of desired signal PSD. An auto encoder trained on clean speech STFT coefficients is used as the prior to non-linearly map the natural speech PSD. We explore two different architectures for the auto-encoder: (i) fully-connected (FC) feed-forward, and (ii) recurrent long short-term memory (LSTM) architecture. Experiments using real room impulse responses show that the LSTM-DNN based PSD estimate performs better than the traditional methods for reverberant signal enhancement.