A dual microphone speech enhancement method with a smoothing parameter mask

The dual microphone system provides high quality of speech in noisy conditions, and is always an important topic in signal processing. In this paper, a computational auditory scene analysis (CASA) based speech enhancement method with a smoothing parameter mask is proposed. With a flexible dual microphone setting, we focus on the speech enhancement between the matched and unmatched training and test conditions. Cooperated with a deep neural network (DNN), the parameter mask is estimated and smoothing with simulated and recording data. The recording data is used to smooth the estimated parameter mask trained with simulate data as a transition to real application. We use recording data to train the DNN. The various configuration recording data are used to test the proposed speech segregation system right away. The proposed system has a positive results on trained and untrained conditions and low signal to noise ratio (SNR) test conditions. It also has a good performance on an office application.