利用概率混合模型的理想比率掩蔽多声源分离方法

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

针对基于时频掩蔽的分离方法在多声源场景下的分离效果不佳的问题,论文提出一种利用概率混合模型的理想比率掩蔽多声源分离方法。首先,利用冯·米塞斯分布对时频点处方位角估计进行拟合以及拉普拉斯分布对归一化声压梯度信号向量进行拟合,由此建立概率混合模型。其次,利用期望最大化算法对模型参数进行求解,估计各声源对应的理想比率掩蔽。最后,利用估计出的理想比率掩蔽,从麦克风采集信号中分离得到各声源信号。实验结果表明,与现有基于时频掩蔽的多声源分离方法相比,论文所提方法在欠定场景下具有更好的分离效果。

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.