一种新的高斯混合模型参数估计算法

该文提出了一种高斯混合模型（GMM）参数估计的改进算法．原始的特征向量先经Schmidt正交化消除各维间的相关性，再用数学形态学方法估计出各维概率分布中混合分量的真实个数，最后按真实的混合分量个数用EM算法对各维分别作标量GMM参数估计．该方法能缓解GMM传统参数估计算法引起的“不易扩展”的不便．实验结果表明，将其应用于说话人辨认，能在较大幅度提高训练速度的基础上相对传统GMM参数估计方法获得更高的识别率．

A New Parameter Estimation Algorithm of Gaussian Mixture Model

A modified algorithm for parameter estimation of Gaussian mixture model (GMM) is proposed. The original feature vector is first transformed with the Schmidt orthogonal procedure to remove correlation between pairs of dimensions. Subsequently, mathematic morphology (MM) is adopted to evaluate the actual number of mixture components in the probability distribution of each dimension approximated by the scalar GMM. The parameters of each scalar GMM are finally estimated by EM algorithm based on the corresponding number of the mixture components. This algorithm can alleviate the inconvenience caused by the traditional vector-based parameter estimation algorithm. When applied to speaker identification experiments, the results show that a higher recognition rate is achievable compared with that obtained with conventional methods, and the training time is also significantly reduced.