A DNN parameter mask for the binaural reverberant speech segregation

The reverberant speech segregation is a basic problem in speech enhancement and automatic speech recognition. Based on the deep neural networks (DNN), a novel binaural speech segregation method is proposed. The binaural feature is extracted and used as the cue to train a DNN with a ideal parameter mask. The trained DNN is used to distinguish the target speech and noise, and output the estimated parameter mask. The performance of the proposed method is systematic evaluated, and it is better than the binary mask method. The proposed method has good performance on untrained locations and reverberant conditions too.