基于统计模型的人声识别优化研究

Research on Optimization of Voice Recognition Based on Statistical Models

为研究基于变分推断的高斯混合模型(Gaussian Mixture Model,GMM)在人声识别中的优化方法,首先设计人声识别系统框架,其次阐述传统GMM在人声识别系统中的基本原理和特点,再次详细介绍变分推断的基本原理及其在GMM优化中的应用,最后采用公开数据集进行实验评估。仿真结果表明,优化后的GMM在识别准确率、精确率、召回率以及F1分数等指标上均显著优于传统GMM。

In order to study the optimization method of Gaussian Mixture Model(GMM)based on variational inference in human voice recognition,the framework of human voice recognition system is first designed,and the basic principle and characteristics of traditional GMM in human voice recognition systems are described.Then,the basic principle of variational inference and its application in GMM optimization are introduced in detail.Finally,experimental evaluation is carried out with open data set.The simulation results show that the optimized GMM is significantly better than the traditional GMM in recognition accuracy,precision,recall,and F,score.