基于流形学习的汉语方言辨识

发布时间：2018-04-18 11:22

本文选题：流形学习 + 低维可视化　；参考：《江苏师范大学》2014年硕士论文

【摘要】：汉语方言辨识研究的核心问题之一是方言特征的提取,因为特征提取的好坏直接关系到系统性能的高低。传统的特征提取大多数沿袭语种识别的理论和方法,忽视了汉语方言自身的有调性等区分意义较大的特征;其次,语音数据是一种典型的流形分布数据,一些好的关于流形分布的算法没有被应用到汉语方言辨识中;传统系统采用的单一特征在数据的结构描写和信息的挖掘等方面有着很大的局限性。针对以上不足,论文将流形学习算法引入到汉语方言辨识系统中,从汉语方言语音低维可视化、流形学习算法提取汉语方言特征、流形学习算法的改进和特征融合等几个方面提升汉语方言辨识系统性能。具体成果如下:1.证明了汉语方言语音中流形结构的存在。从低维可视化的角度对汉语方言语音进行了分析研究,仿真实验结果表明相对于线性降维算法,流形学习算法在汉语方言语音低维的时候能够更好地体现不同地区方言语音之间的差异性,间接证明了汉语方言语音数据中流形结构的存在。2.利用流形学习提取了汉语方言辨识新特征。通过对低维流形结构的观察与流形学习算法的分析研究,利用其中局部线性嵌入算法对汉语方言语音进行特征提取。3.在流形学习算法的基础上对局部线性嵌入算法本身进行了改进。针对局部嵌入算法本身存在的不足,对其中的距离求取方法进行改进,旨在改善样本数据集的分布,并结合聚类算法提取汉语方言语音新特征。4.构建了一套基于流形学习的汉语方言辨识系统,证明了新特征的有效性。利用高斯混合模型和支撑矢量机作为系统后端分类器。仿真实验结果表明新特征可以有效地提高系统的性能。同时利用特征融合的方法对新特征和传统特征进行了有效地融合,进一步提升了特征的有效性。
[Abstract]:The extraction of dialect features is one of the core problems in the research of Chinese dialect recognition, because the quality of feature extraction is directly related to the performance of the system.The traditional theories and methods of feature extraction mostly follow the language recognition, ignoring the tonality of Chinese dialects and other distinguishing features. Secondly, the speech data is a kind of typical manifold distribution data.Some good algorithms on manifold distribution have not been applied to Chinese dialect recognition, and the single feature used in traditional systems has great limitations in data structure description and information mining.In this paper, manifold learning algorithm is introduced into the Chinese dialect recognition system, and the feature of Chinese dialect is extracted from the low dimensional visualization of Chinese dialect speech and manifold learning algorithm.The improvement and feature fusion of manifold learning algorithm improve the performance of Chinese dialect recognition system.The concrete results are as follows: 1.The existence of manifold structure in Chinese dialect pronunciation is proved.In this paper, the Chinese dialect speech is analyzed and studied from the view of low dimension visualization. The simulation results show that compared with the linear dimensionality reduction algorithm,Manifold learning algorithm can better reflect the difference of dialect pronunciation in different regions when the dimension of Chinese dialect speech is low, which indirectly proves the existence of manifold structure in Chinese dialect phonetic data.The new features of Chinese dialect recognition are extracted by manifold learning.Based on the observation of low dimensional manifold structure and the analysis of manifold learning algorithm, the local linear embedding algorithm is used to extract the features of Chinese dialect speech.Based on the manifold learning algorithm, the local linear embedding algorithm is improved.In order to improve the distribution of sample data set and to extract the new features of Chinese dialect speech, the distance estimation method is improved to improve the distribution of the sample data set in order to overcome the shortcomings of the local embedding algorithm.A Chinese dialect recognition system based on manifold learning is constructed, which proves the validity of the new features.Gao Si hybrid model and support vector machine are used as the back-end classifier of the system.Simulation results show that the new features can effectively improve the performance of the system.At the same time, the method of feature fusion is used to fuse the new feature and the traditional feature effectively, which further improves the effectiveness of the feature.
【学位授予单位】：江苏师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【相似文献】