Monaural Speech Separation Using Dual-Output Deep Neural Network with Multiple Joint Constraint

Monaural speech separation is a significant research field in signal processing. To achieve better performance, we propose three novel joint-constraint loss functions and multiple function for monaural based on dual-output deep neural network (DNN). The DNN model not only restricts the ideal ratio mask (IRM) errors of two outputs, but also constrains relationship estimated IRMs magnitude spectrograms clean signals, spectrogram mixed signal. constraint strength adjusted through parameters to improve accuracy model. Furthermore, solve optimal weighting coefficients optimization idea, which further improves performance system. We conduct series experiments GRID corpus validate superiority proposed method. results show that using perceptual evaluation quality, short-time objective intelligibility, source distortion ratio, interference artifact as metrics, method out-performs conventional Taking gender into consideration, carry out among Female-Female, Male-Male Male-Female cases, our robustness system compared with some previous approaches.