基于混合微粒群算法的说话人识别

Speaker recognition based on hybrid particle swarm optimization algorithm

为了解决传统高斯混合模型(GMM)对初值敏感,在实际训练中极易得到局部最优参数的问题,提出了一种采用微粒群算法优化GMM参数的新方法。该方法将最大似然估计融入到微粒群算法迭代过程中,形成了新的混合算法。它利用微粒群算法的全局优化性及最大似然估计的局部寻优性求解高斯混合模型的参数,以提高参数精度。说话人辨认实验表明,与传统的方法相比,新方法可以得到更优的模型参数,使得系统的识别率进一步提高。

The traditional training methods of Gaussian Mixture Model （GMM） are sensitive to the initial model parameters, which often leads to a local optimal parameter in practice. To resolve this problem, a new GMM optimization method was proposed based on Particle Swarm Optimization （ PSO）. It utilized Maximum Likelihood （ML） algorithm in the PSO iteration and provided a new architecture of hybrid algorithm. Because of the global optimization characteristic of the particle swarm optimizer method and the strong local searching capacity of ML, it can obtain model parameters with high precision. Experiment for text-independent speaker identification shows that this method can obtain more optimum GMM parameters and better results than the traditional method.