Multiple Sound Source Separation via Ideal Ratio Masking by Using Probability Mixture Model

The separation method based on time-frequency masking is liable to be invalid under multi-sound source scenes, this paper presents a method of multiple sound source separation via ideal ratio masking by using probability mixture model. First, the von Mises distribution is used to fit the angle estimates at each time-frequency point and the Laplacian distribution is used to fit the vector of normalized pressure gradient signals. Thereby, this paper establishing a probabilistic mixture model. Secondly, the expectation maximization algorithm is used to estimate the model parameters, and the ideal ratio masking corresponding to each source is obtained. Finally, with the estimated ideal ratio masking, source signals are separated from the microphone signals. Experimental results reveal that the performance of the proposed separated method is better than the existing separation method based on time-frequency masking.