混响的混合高斯概率密度建模

混合高斯模型能够有效地拟合混响背景的一维概率密度分布。常用的混合高斯概率密度模型参数估计方法是EM迭代算法,但这种算法的主要缺点是估计精度过分依赖于初始值。而GreedyEM算法通过往混合模型中不断地加入高斯分量,能很好地解决这一问题。文章将多维图象处理中的GreedyEM算法加以合理简化,并给出模型自动定阶方法,从而成功应用于水声混响的一维混合高斯模型建模中。实验结果表明:应用新算法能从混响接收数据中准确拟合其概率密度曲线,并且能适应不同的数据长度,具有很好的通用性。

Gaussian mixture model for reverberation

The probability density distribution of one-dimention under the reverberation background can be efficiently modeled using Gaussian mixtures. EM is one of the popular algorithms for parameters estimation of Gaussian mixture probability density model. However, this method highly depends on the initial parameters. The GreedyEM algorithm can solve this problem efficiently by incrementally adding Gaussian components to the mixture. It has been used in multidimensional image manipulations. We properly simplify and apply it to learn Gaussian mixtures of onedimention reverberation successfully, present a method for determining the number of mixing components. We provide experimental results illustrating that the new algorithm can approximate the probability density curve of the receiving reverberation data, and has the potential to be used for different length of data due to its good adaptability.