基于高斯混合密度函数估计的语音分离

Speech Separation Based on Gaussian Mixture Model Probability Density Function Estimation

基于最大熵法（Ｍａｘｉｍ ｕｍ Ｅｎｔｒｏｐｙ， ＭＥ）、最小互信息量法（Ｍｉｎｉｍ ｕｍ Ｍｕｔｕａｌ Ｉｎｆｏｒｍ ａ－ｔｉｏｎ， ＭＭＩ）和最大似然法（Ｍａｘｉｍ ｕｍ Ｌｉｋｅｌｉｈｏｏｄ， ＭＬ）是解决盲信号分离问题的常用算法，分析了ＭＥ、ＭＭＩ以及ＭＬ算法之间关系．基于高斯混合模式（Ｇａｕｓｓｉａｎ Ｍｉｘｔｕｒｅ Ｍｏｄｅｌ， ＧＭＭ）概率密度函数估计，提出了一种采用反馈结构的扩展最大熵语音分离算法．与传统ＭＥ的计算机模拟实验结果比较得知，新算法具有更好的收敛性能和语音分离效果．

The speech separation task was seen as a convolution mixture blind signal separation (BSS) problem. There are 3 kinds of main approaches to solve the BSS problem: ME(maximum entropy) algorithm, MMI(minimum mutual information) algorithm, and ML (maximum likelihood) algorithm. The relationship among the 3 kinds of algorithms was analyzed in this paper. Based on the feedback architecture and Gaussian mixture model (GMM) probability density function (pdf) estimation, a new extended ME algorithm speech separation algorithm was proposed. Based on the computer simulations of the proposed algorithm and traditional ME algorithm, it can be concluded that the proposed algorithm has better convergence performance.