基于改进的高斯混合模型算法的说话人识别

Speaker Recognition Based on Improved Gaussian Mixture Model

高斯混合模型(GMM)已广泛运用于文本无关的说话人识别系统中,该方法具有简单高效的特点。在使用EM算法训练GMM时,GMM模型的初始化参数必须首先确定。本文采用改进后的模糊C均值聚类(FCM)方法将特征矢量归为与混合数相等的各个类中,然后计算参数作为初始值。实验表明,此训练方法能够获得更优的模型参数且识别率有较大的提高。

Gaussian mixture model （GMM） was widely used in text-independent speaker recognition system, which had a simple and efficient features.When using EM to train GMM model, GMM model initialization parameters had to be first determined. Hereby the improved fuzzy C-means clustering （FCM） method was used to go with hybridzing feature vector to be the same number of each individual class and then the calculation parameters were taken as initial values. Experiments expatiated that this training method could obtain better model parameters and the recognition rate was significantly improved.