基于遗传模糊高斯混合模型的训练方法

为了解决传统高斯混合模型GMM(Gaussian m ixture model)的训练方法对模型初值十分敏感、在实际训练中极易得到局部最优模型参数的问题,提出了一种GMM模型参数训练的新方法。将遗传算法与基于模糊聚类分析的GMM参数估计相结合,形成一种新的混合算法,对模型参数进行全局优化,提高了参数估计的准确性。采用自适应交叉和变异算子,同时利用模糊最小目标函数FMOF(FuzzyM inimum Objection Function)准则对模型参数进行重估,提高了算法的搜索效率,加快了算法的收敛速度。使用PKU-SRSC语音数据库进行了与文本无关的说话人辨认实验。实验表明,与传统的GMM训练方法和最大似然估计方法相比,本文方法可以得到更优的模型参数,同时识别率也有所提高。

Gaussian mixture model training method based on genetic algorithms and fuzzy approach

The traditional training method of Gaussian mixture model(GMM) is sensitive to the initial model parameters and easy to lead to a sub-optimal model in practice.To solve this problem,a hybrid algorithm was proposed to optimize GMM model parameters globally and to improve parameter precision by using genetic algorithms(GA) and fuzzy clustering estimation method.An adaptive mixture crossover rate and mutation rate of GA and fuzzy minimum objective function(FMOF) algorithm were used to form a hybrid operation for re-estimation of the model parameters,which can improve GA convergence speed.Experiment results based on PKU-SRSC database show that this algorithm can obtain more optimized GMM and better than traditional GMM training method and maximum simulation training methods.