稀疏正交联合约束多通道非负矩阵分解声信号分离算法

针对复杂环境下多通道声信号分离问题,提出稀疏正交联合约束多通道非负矩阵分解声信号分离方法。首先设计基于多通道扩展坂仓斋藤(Itakura-Saito,IS)散度的稀疏正交联合约束项构造代价函数,给出信号稀疏和信号正交约束辅助函数,实现代价函数最小化求解。然后通过迭代更新规则设计,得到稀疏正交优化的多通道非负矩阵分解基矩阵和系数矩阵,讨论了稀疏正交约束对基矩阵和系数矩阵稀疏性与连续性影响。最后基于多通道信号空间特性,进行了非负矩阵分解基聚类以获得多通道非负矩阵分解声信号的分离结果。双通道音频数据与四通道声学目标分离实验数据测试表明,对音频数据,所提算法在性能指标信号失真比(SDR)上提高了0.84 dB,对于直升机声源数据,所提算法在SDR上提高了4.53 dB。

Sparse orthogonal joint constrained multichannel non-negative matrix factorization algorithm for acoustic signal separation

Aiming at the problem of multichannel acoustic signal separation in complex environment, a sparse orthogonal joint constrained multichannel non-negative matrix factorization acoustic signal separation method is proposed. First, the cost function is constructed by sparse orthogonal joint constraints based on multichannel extended Itakura-Saito (IS) divergence, and auxiliary functions of signal sparse and signal orthogonal constraints are given to minimize the cost function. Then, through the design of iterative update rules, the basis matrix and coefficient matrix of multichannel non-negative matrix factorization for sparse and orthogonal optimization are obtained, and the influence of sparse orthogonal constraint on the sparsity and continuity of the basis matrix and coefficient matrix is discussed. Finally, based on the spatial characteristics of multichannel signals, non-negative matrix decomposition basis clustering is performed to obtain the separation results of multichannel Non-negative Matrix Factorization (NMF) acoustic signals. The experimental data tests of two-channel audio data and four-channel acoustic target separation show that the proposed algorithm improves the performance index Signal Distortion Ratio (SDR) by 0.84dB for audio data, and improves the SDR by 4.53dB for helicopter sound source data.