基于语种识别系统的语言距离关系研究

发布时间：2018-04-12 08:54

本文选题：语言距离关系 + 语种识别　；参考：《浙江大学》2016年硕士论文

【摘要】：本文主要的研究内容为使用语音数据来研究语言之间的距离关系。对于语言距离关系的定义,语言学家通过语言是否同源以及语言之间发音、字词、语法规则的对应关系来研究语言之间的距离关系,最终生成语言谱系树来反映语言的从属关系,这属于定性研究；经济学家将语言之间的距离定义为母语为一种语言的人学习其他语言的难易程度,这属于定量研究。不论是在文化还是经济领域,语言距离关系的研究都具有一定的应用价值。本文通过语种识别系统的输出——识别率来定义语言之间的距离关系,其中最关键的就是语种识别系统的搭建,它直接关系到语言之间距离关系的可靠程度。语种识别系统的搭建主要包括特征提取和模型训练两部分。在特征提取部分,本文先提取输入语音信号的56种声学特征以及它们的一阶、二阶差分值,然后使用39种统计函数来描述这些声学特征,最终生成6552维特征向量；在模型训练部分,本文采用区分性建模方法——Gentle AdaBoost算法。最后将语言距离矩阵映射到二维平面,生成语言关系图,直观的反映了语言之间的距离关系。本文的主要研究成果包括：(1)本文使用语音数据来研究语言之间的距离关系。通过语种识别系统的输出——识别率来定义语言之间的距离,并将本文得到的语言距离与之前语言学家和经济学家的研究成果作比较,实验表明本文的距离度量方法与之前的研究成果较为一致。(2)在语种识别系统的搭建中,采用一种全新的方法提取输入语音信号的特征,该方法与以往的方法最大的不同在于,一个输入语音样本对应一个特征向量,而不再是每一帧都对应一个特征向量。然后采用Gentle AdaBoost算法训练语言模型,在OGI数据库和NIST LRE15数据库上进行语种识别系统的性能测试,实验表明本文的语种识别系统性能要优于传统的基于GMM的语种识别系统；(3)将得到的语言距离矩阵映射到二维平面,生成语言关系图。在OGI数据库和NIST LRE15两个数据库上进行实验,验证了本文所使用的语言距离研究方法在不同的数据库上的一致性。
[Abstract]:The main research content of this paper is to use speech data to study the distance relationship between languages.As to the definition of language distance relation, linguists study the distance relationship between languages by means of the homology of language and the correspondence of pronunciation, words and grammar rules between languages.Finally, a language tree is generated to reflect the subordination of language, which is a qualitative study; the distance between languages is defined by economists as the degree of difficulty for native speakers to learn other languages, which is a quantitative study.The study of language distance relation has certain application value both in culture and economy.In this paper, the distance relationship between languages is defined by the output-recognition rate of the language recognition system, and the most important one is the construction of the language recognition system, which is directly related to the reliability of the distance relationship between languages.Language recognition system consists of feature extraction and model training.In the part of feature extraction, we first extract 56 acoustic features of the input speech signal and their first-order and second-order differences, then use 39 statistical functions to describe these acoustic features, and finally generate 6552 dimensional feature vectors.In the part of model training, this paper adopts the discriminative modeling method-Gentle AdaBoost algorithm.Finally, the language distance matrix is mapped to the two-dimensional plane, and the language relation graph is generated, which directly reflects the distance relationship between languages.The main research results of this paper include: 1) this paper uses speech data to study the distance relationship between languages.The distance between languages is defined by the output-recognition rate of the language recognition system, and the language distance obtained in this paper is compared with the previous research results of linguists and economists.The experimental results show that the distance measurement method in this paper is consistent with previous research results. In the language recognition system, a new method is used to extract the features of the input speech signal. The biggest difference between this method and the previous methods is that,An input speech sample corresponds to a feature vector instead of a feature vector for each frame.Then the language model is trained by Gentle AdaBoost algorithm, and the performance of language recognition system is tested on OGI database and NIST LRE15 database.The experimental results show that the performance of the language recognition system in this paper is better than that of the traditional language recognition system based on GMM. The language distance matrix is mapped to the two-dimensional plane and the language relation graph is generated.Experiments on two databases, OGI database and NIST LRE15 database, have been carried out to verify the consistency of the language distance research methods used in this paper on different databases.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN912.34

【相似文献】