基于深度关注神经网络的汉语方言辨识

发布时间：2018-04-03 16:30

本文选题：汉语方言辨识　切入点：方言语音库　出处：《江苏师范大学》2017年硕士论文

【摘要】：伴随科技进步,信息处理智能化成为社会发展的方向。语音识别是智能信息处理的核心技术,经历近百年的研究发展,已取得长足进步,但复杂多语言语音处理系统的真正实现效果不尽如人意。本文研究的汉语方言辨识是利用计算机自动辨别语音片段方言类属特性的语音处理技术,其在公共安全、信息服务、言语工程等诸多领域,均有重要的应用价值和广阔的应用前景。目前,经济有效的汉语方言辨识方法是利用机器学习算法提取方言的类属特征并设计建立方言模型,通过模型匹配完成方言类属判定。但是,由于汉语方言本身的复杂异变性,该领域仍存在语音库匮乏、数据失衡;特征的类属特性表征模糊、不稳定;模型的学习及分类能力不足等诸多问题需要解决。本文针对以上问题展开研究,具体工作如下:(1)完善汉语方言语音库。在前人已建汉语方言语音库的基础上继续收录各方言语音,同时积极寻求各方资源共享,先后增补吴、闽及官话方言近百小时。其次对方言粗语音进行参数规整、数据平衡、标注以及时长自动切分操作,使其符合语音库标准及实验要求。(2)提出基于身份认证矢量(Identify vector,I-vector)与韵律特征融合的汉语方言辨识方法。从句法结构上提升方言类属特性的全局表征效果,增强方言语音的可区分性。实验结果显示,对比高斯混合模型-通用背景模型(Gaussian Mixture Models-Universal Background Model,GMM-UBM)方法,融合全局信息的方法将汉语方言辨识等错误率(Equal Error Rate,EER)相对降低56.32%,性能提升显著。(3)提出基于深度关注神经网络的汉语方言辨识方法。深度神经网络(Deep Neural Network,DNN)可以有效解决大输入量、复杂非线性信号的分类处理问题,关注机制体现模型对分配资源进一步有效利用。本文设计不同输入特征探究深度关注神经网络对汉语方言辨识的影响。结果显示,深度关注神经网络将汉语方言辨识EER相对降低28.3%,提高辨识正确率的同时展现模型高度的稳定性。
[Abstract]:With the progress of science and technology, intelligent information processing has become the direction of social development.Speech recognition is the core technology of intelligent information processing. After nearly 100 years of research and development, it has made great progress, but the real effect of complex multilingual speech processing system is not satisfactory.The Chinese dialect recognition studied in this paper is a kind of speech processing technology which can automatically identify the characteristics of speech segment dialects by computer. It is used in many fields, such as public security, information service, speech engineering and so on.It has important application value and broad application prospect.At present, an economical and effective method for identifying Chinese dialects is to extract the generic features of dialects by using machine learning algorithm and to design and establish dialect models, and complete the classification of dialects by model matching.However, due to the complex variability of Chinese dialects, there are still many problems in this field, such as lack of phonetic corpus, unbalance of data, fuzzy and unstable representation of generic characteristics of features, inadequate learning and classification ability of models, and so on.This paper aims at the above problems, the specific work is as follows: 1) perfect the Chinese dialect phonetic corpus.On the basis of the previous Chinese dialect phonetic database, we continue to collect the pronunciation of each dialect, and actively seek the sharing of resources, and supplement Wu, Min and Guanyu dialects for nearly 100 hours.Secondly, the parameters, data balance, tagging and time length automatic segmentation of dialect coarse speech are carried out to meet the standard of phonetic corpus and experimental requirements. (2) an identification method for Chinese dialects based on the fusion of identity authentication vector identifying vector vectoring I-vectorand prosodic features is proposed.From the syntactic structure, we can improve the global representation effect of dialect generic characteristics and enhance the differentiability of dialect pronunciation.The experimental results show that the Gaussian Mixture Models-Universal Background Model GMM-UBM method is compared with Gao Si mixed model and general background model.The method of integrating global information reduces the equal Error rate of Chinese dialect recognition and equal Error error rate by 56.32, and improves the performance significantly. 3) A method of Chinese dialect identification based on the neural network of depth concern is proposed.Deep Neural Network (DNN) can effectively solve the problem of classification and processing of large input and complex nonlinear signals.This paper designs different input features to explore the effects of neural networks on Chinese dialect recognition.The results show that the neural network with deep attention reduces the EER of Chinese dialect recognition by 28.3and improves the accuracy of identification and shows the high stability of the model at the same time.
【学位授予单位】：江苏师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3;TP18

【相似文献】