Phase Constraint and Deep Neural Network for Speech Separation

摘要

The phase response of speech is an important part in speech separation. In this paper, we apply the complex mask to the speech separation. It both enhances the magnitude and phase of speech. Specifically, we use a deep neural network to estimate the complex mask of two sources. And considering the importance of the phase, we also explore a phase constraint objective function, which can ensure the phase of the sum of estimated sources that is close to the phase of the mixture. We demonstrate the efficiency of the method on the TIMIT speech corpus for single channel speech separation.