高斯混合声纹识别模型在EM训练中初始值选择问题的研究

GMM(Gaussian Mixture Model) of voiceprint recognition can be described as linearity combination with a group of PDF(probability density function). First GMM voiceprint system classify feature vectors of an utterance, and establish initial model of the utterance, then we apply the EM(expectation maximization) algorithm to obtain the optimized model .EM algorithm is a local maximization arithmetic, the selection of initial model ,the same meaning of primary parameters, greatly influences the eventual identification effect. Traditionally primary parameters are provided by K-means algorithm, which is a local clustering arithmetic, it cannot supply optimized primary parameters. This paper proposes an ant colony algorithm combined with genetic arithmetic. This algorithm can bring better recognition rate than K-means algorithm in text-dependent GMM voiceprint recognition experiment.