复高斯混合模型分布式语音分离方法研究

Distributed Speech Separation Based on Complex Gaussian Mixture Model

本文研究空域协方差矩阵初始化对复高斯混合模型下的分布式语音分离性能的影响。在不同节点的接收信号向量条件独立性假设前提下,推导出一种逐节点迭代更新所有接收信号向量对应的空域协方差矩阵和后验概率等参数的方法;基于此,本文提出用基于到达角度的导向矢量的相关矩阵初始化每个节点对应的空域协方差矩阵;同时,为保证不同节点能协同工作,提出了一个基于到达角度自聚类的方法,以实现在不同节点上选出同一个说话人的到达角度组合。实验结果表明,本文提出的分布式语音分离算法及其初始化方法在保证分离性能的同时,大幅度降低了集中式算法所需的计算复杂度,而且避免了排序问题。

In this paper, the impact of the spatial covariance matrix(SCM) initialization on the performance of the distributed speech separation was studied under the complex Gaussian mixture model. Based on the conditional independence assumption of the recordings of different nodes, the update of the SCM and the posterior probability corresponding to all received signals could be performed per node. Therefore, the SCM corresponding to each node was proposed to be initialized by the correlation matrix of the steering vector based on direction of arrival(DOA). Meanwhile, a DOA self-clustering method was proposed to find the combination of DOA corresponding to the same speaker from different nodes, which guaranteed that different nodes could cooperate. The proposed distributed speech separation and its initialization method has lower computational complexity than the centralized algorithm and avoids the permutation problem. Experimental results validate the effectiveness of the proposed method.