Monaural speech separation based on linear regression optimized using gradient descent

Monaural speech separation (MSS) is useful for many real-world applications. In this work, we propose a novel method for MSS based on the observation that a composite speech signals can be modeled as the linear summation of each speaker with respect to participation coefficients. Hence, speech signals are separated using linear regression. Partial derivative with respect to each variable is then used to perform gradient descent in order to optimize the estimation and therefore the separation. The proposed speech separation method for is applicable to known speakers.The proposed method was assessed using metrics characterized by good correlation coefficients with subjective listening tests. Evaluation results reveal the effectiveness of the proposed approach.