基于特征融合与改进GMM-UBM的方言分析

Dialect Analysis Based on Feature Fusion and Modified GMM-UBM

针对单一特征在描述方言间差异性方面存在不足和传统高斯混合通用背景模型(Gaussian Mixture Model-Universal Background Model,GMM-UBM)在训练时存在混叠的问题,在将经验模式分解引入到特征提取的基础上,融合多特征形成高维特征。进一步,通过多次训练GMMUBM,筛选出最具区分性的方言模型以提升各方言模型间的区分性。方言种类识别对比实验结果表明,基于融合特征与改进GMM-UBM的方法优于传统方法。

In view of the shortcomings of a single feature in describing the differences among dialects and the problem of aliasing in the conventional GMM-UBM(Gaussian Mixture Model-Universal Background Model)during training,on the basis of introducing empirical mode decomposition into feature extraction,multiple features are fused to form high-level features.Furthermore,the GMM-UBM is trained multiple times to screen the most discriminative dialect model to improve the discriminativeness among dialect models.The comparative experimental results of dialect type recognition indicate that the method based on the fusion feature and the modified GMM-UBM is superior to the conventional method.