摘要:
针对高斯混合模型(Gaussian Mixture Model, GMM)参数最优估计问题，常用的最大期望(Expectation- Maximization, EM)算法对初值敏感，在实际训练中极易得到局部最优参数，本文提出了一种GMM参数优化的新方法。将EM算法融入到粒子群优化(Particle Swarm Optimization, PSO)训练过程，形成了一种新的混合算法，利用PSO的全局探索和EM算法的局部深度搜索的混合策略，粒子在每次迭代中执行PSO速度位置更新和标准EM算法的混合更新操作，在训练语音矢量空间搜索最优高斯混合模型参数。从而避免传统EM算法陷入局部最优的缺点。说话人辨认实验表明，与EM算法相比，本文方法可以得到更优的模型参数，能有效提高系统的识别率。
Abstract: Expectation-Maximization (EM) algorithm is usually used to estimate parameters of Gaussian mixture model. Due to the hill-climbing characteristic of EM, any arbitrary estimation of the initial model parameters will usually lead to a sub-optimal model in practice. To resolve this problem, a hybrid training method based on Particle Swarm Optimi- zation (PSO) is proposed. It utilizes the global searching capability of PSO and combines the effectiveness of EM. The particles perform basic operations of PSO (velocity updating and position updating) and EM algorithm, which can ex- plore the training speech space to move toward the global optimum. The dependence of the final model parameters on the selection of the initial model parameters is also reduced. Experimental results have showed that this method can obtain more optimized GMM parameters and has better capability than EM in speaker recognition.