Sequentially Trained DNNs Based Monaural Source Separation in Real Room Environments

Abstract

In recent studies, deep neural networks (DNN) have been introduced to solve monaural source separation (MSS) problem within real room environments. However, the separation performance of the existing methods is limited, especially for environments with larger RT60s. In this paper, we propose a system to train two DNNs sequentially, to mitigate the challenge and improve the separation performance. Our dereverberation mask (DM) is exploited as a training target for DNN1 and new enhanced ratio mask (ERM) is used as a training target for DNN2. The IEEE and the TIMIT corpora with real room impulse responses and noise interferences from the NOISEX dataset are used to generate speech mixtures for evaluations. The proposed method outperforms the state-of-the-art methods.