论文总字数：24727字

摘 要

Abstract III

第一章引言 1

1.1 研究背景及意义 1

1.2 国内外研究现状 1

1.3 本文研究内容 1

第二章语音处理基本原理 3

2.1 语音信号模型 3

2.2 语音预处理 4

2.2.1 采样 4

2.2.2 量化 4

2.2.3 预加重 4

2.2.4 加窗 5

2.3 语音特征提取 5

2.3.1 短时平均能量分析 5

2.3.2 短时平均幅度分析 6

2.3.3 短时过零率分析 7

2.3.4 短时自相关分析 7

2.3.5 LPC倒谱系数(LPCC) 7

2.3.6 Mel频率倒谱系数(MFCC) 8

第三章基于HMM的声纹识别 9

3.1 声纹识别基本过程 9

3.2 矢量量化模型(VQ) 9

3.3 隐马尔可夫模型(HMM) 10

第四章声纹密码仿真平台 12

4.1 MATLAB开发简介 12

4.2 系统界面设计 12

4.3 仿真结果分析 13

第五章总结与展望 16

致谢 17

参考文献 18

附录 19

低资源文本相关的声纹密码识别方法的研究

摘要

声纹这个词，指电声仪器可以显示的带有信息的可听见的频谱。人类说话的过程是人体语言中枢与发音器官之间发生的一个包含很多含义和巨量生物信息的过程，每个人在讲话时用的各种发声器官诸如舌头、牙齿、喉咙、肺部、鼻子因不同的人在大小和形状等方面的差异，导致不同的人发声用机器记录下来的图谱都有不同。每个人的语音声学特征有双重特性，分别是相对稳定性和变异性，而且这种特性不会永远不变。而这种变化可来自各种方面，有可能源自生理、心理，也可能因为生病导致变化，或者是人为故意的模拟、伪装，同时也可能与外部环境干扰有关比如分贝较大的杂音。即便是这样，但是因为每个人所用的发音器官都不一样，正常情况下，人们一般都能够辨认并且区分不同的人所发出的声音或判断这是不是同一个人发出的声音。

声纹密码是说话人识别模式中最经典的应用方式，它通过说话人的文本内容和语音通道信息双重渠道来保证说话人的信息安全。即使对于测试说话的人，在他下次登录应用系统时，说话内容必须与录音内容相同，然后系统才能确认，因此声纹密码非常安全，语音密码系统可以看成文本相关说话人识别。

模式识别包括三个大方向的研究课题，分别是分割、特征提取、训练和识别模块。说话人识别也不能逃脱其中，它包括说话人语音分割、说话人个性参数提取和识别说话人训练识别模型(主要是隐马尔可夫模型和高斯混合模型)本文通过提取语音特征，并采用隐马尔可夫模型(HMM(Hidden Markov Model, HMM))对文本相关性进行建模。最后基于Matlab仿真平台建立了声纹密码仿真平台，平台可以灵活地设置不同的语音口令实施有效的声纹密码识别验证。

关键词：声纹识别；MATLAB；模式识别；隐马尔可夫

Abstract

The word voiceprint refers to the audible frequency spectrum with information that can be displayed by electroacoustic instruments. The process of human speech is a process which contains a lot of meanings and huge amount of biological information between the human language center and the vocal organs. The various vocal organs used by each person in speech, such as tongue, teeth, throat, lung and nose, are different in size and shape, resulting in different maps recorded by different people. Each person's speech acoustic characteristics have two characteristics, namely, relative stability and variability, and this characteristic will not remain unchanged forever. This kind of change can come from various aspects. It may come from physiology and psychology, or it may be caused by illness, or it may be artificial simulation or disguise. At the same time, it may also be related to external environmental interference, such as the noise with large decibel. Even so, because everyone uses different voice organs, under normal circumstances, people are generally able to recognize and distinguish the voice of different people or judge whether it is the voice of the same person.

Voiceprint password is the most classic application method in speaker recognition mode. It uses the dual channels of the speaker's text content and voice channel information to ensure the speaker's information security. Even for the test speaker, the next time he logs in to the application system, the content of the speech must be the same as the recording content, and then the system can confirm it. Therefore, the voiceprint password is very safe, and the voice password system can be regarded as a text-related speaker recognition.

Pattern recognition includes three major research topics: segmentation, feature extraction, training and recognition module. Speaker recognition can not escape from it, it includes speaker speech segmentation, speaker personality parameter extraction and speaker training recognition model (mainly hidden Markov model and Gaussian mixture model). This paper focuses on these three aspects to discuss and study the text related speaker recognition system based on Hidden Markov model. In this paper, speech features are extracted and hidden Markov model (HMM) is used to model text relevance. Finally, based on MATLAB simulation platform, a voiceprint password simulation platform is established, which can flexibly set different voice passwords to implement effective voiceprint password recognition verification.

剩余内容已隐藏，请支付后下载全文，论文总字数：24727字