A SI-SDR Loss Function based Monaural Source Separation

Deep neural network is very popular for performing various tasks in the field of speech signal processing nowdays. This work mainly studies the problem of source separation under single channel (monaural) conditions, which is more challenging because only a single channel of information is available, and without any constraints, an infinite number of solutions are possible. We introduce deep recurrent neural network (DRNN) as a regression model to discover the deep structure and regularity of signal reconstruction from a mixture containing two sources. To alleviate the problem of interference between different sources, we integrate scale-invariant signal-to-distortion ratio (SI-SDR) loss function and discriminative network training criterion together to generate a new training objective function. Experimental results show that the proposed method produces a superior performance compared with other approaches.