声纹识别系统关键技术研究

发布时间：2018-03-11 13:33

本文选题：声纹识别　切入点：特征提取　出处：《哈尔滨理工大学》2014年硕士论文　论文类型：学位论文

【摘要】：声纹如同指纹、人脸一样，是人体特有的一种生物特征。由于其方便性和经济性，声纹识别作为生物认证技术的一种，逐渐走入大众的视线。声纹识别不同于语音识别，顾名思义，声纹识别注重的是待识别的语音信号中说话人的声纹特征，无需知道说话内容。由于每个人都有其独一无二的声纹特征，很难被模仿伪造，因此，相对于其他生物认证技术，声纹识别技术在身份认证领域更加地安全、可靠。本文主要是对与文本无关的声纹识别系统的关键技术进行研究，力图在前人研究的基础上有所创新，以提高系统的识别率。首先在宏观上分析了声纹识别的课题背景、发展现状及研究难点等，并对声纹识别系统的结构原理进行介绍，其次分析了声纹识别系统的端点检测部分，，然后重点对声纹识别的关键部分特征提取模块进行研究。主要包括分析声纹识别系统的特征提取模块中线性预测倒谱系数(Linear Prediction Cepstrum Coefficient，LPCC)与Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient，MFCC)的提取原理，并对MFCC参数的提取进行了改进，提出了基于小波变换和改进的MFCC参数组合特征的提取算法。最后用高斯混合模型通过实验的方式对特征提取部分的不同算法进行分析比较，达到了提高了系统的效率的目的。
[Abstract]:Sound pattern, like fingerprint and face, is a special biological feature of human body. Because of its convenience and economy, sound pattern recognition, as a kind of biometric authentication technology, has gradually come into the public's sight. Voice pattern recognition is different from speech recognition. As the name implies, voice-pattern recognition focuses on the speaker's voice-pattern feature in the speech signal to be recognized, without knowing what to say. Because everyone has its unique voice-pattern feature, it is difficult to imitate and forge, so, Compared with other biometric authentication technologies, voiceprint recognition technology is more secure and reliable in the field of identity authentication. This paper mainly studies the key technology of text-independent voiceprint recognition system, and tries to innovate on the basis of previous research in order to improve the recognition rate of the system. This paper introduces the structure principle of the voiceprint recognition system, and then analyzes the endpoint detection part of the voiceprint recognition system. Then the key part of feature extraction module of sound stripe recognition is studied, including the principle of linear predictive cepstrum coefficient (linear Prediction Cepstrum coefficient) and Mel frequency cepstrum coefficient (Mel Frequency Cepstrum efficient Prediction) in the feature extraction module of sound stripe recognition system, and the principle of linear predictive cepstrum coefficient (LPC) and Mel frequency cepstrum coefficient (Mel Frequency Cepstrum efficient Prediction) are analyzed. After improving the extraction of MFCC parameters, an algorithm based on wavelet transform and improved MFCC parameter combination feature extraction is proposed. Finally, the different algorithms of feature extraction part are analyzed and compared by using Gao Si mixed model through experiments. The efficiency of the system is improved.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【参考文献】