说话人识别技术 毕业论文.doc

摘　　要 说话人识别技术是一种的重要生物认证手段，也是身份鉴别学术会议中的一项重要内容。说话人识别的目的是通过话语找出或核实说话人的身份，可以被用于访问控制。它属于语音信号领域的一个模式识别问题。 本文使用交叠分帧的方法短时化语音信号，使用每帧信号的能频值区分语音信号和噪声信号。特征提取方面，本文使用线性预测倒谱系数和基音频率来表征生成语音的发音器官的差异（先天的），用差分线性预测倒谱系数和差分基音频率表征发音器官发音时动作的差异（后天的）。四种特征加权扩维得到的组合特征矢量最终表征了一个特定的说话人。分类决策方面，本文使用矢量量化的方法完成对说话人语音信号的分类和判决。 本文设计的系统是基于Java语言和SQL Server 2000数据库实现的。Java语言用于实现语音样本采集、预处理、特征提取、分类决策等说话人识别所需的各种算法。SQL Server 2000数据库用于存储已注册说话人的语音码本。 本文在实现系统的基础上，分析了组合特征中各分量对说话人识别的贡献大小。得到的结论是：用于说话人识别的参数中，线性预测倒谱系数效果最好，差分线性预测倒谱系数次之，基音频率再次之，差分基音频率效果最差。根据这一结论，系统通过调整组合特征中各分量加权系数的方式突出贡献大的分量。实验表明，调整后系统识别率显著提高。对于10名男性语音的码本库，本文实现系统的识别率可达到87%。 The research of the text-independent speaker recognition system Abstract Speaker recognition technology is one of the important biometric ways, as well as an important part in academic conferences of identification. The purpose of speaker recognition is identifying or verifying the speakers identity through the discourse, which can be used to control access. It is a pattern recognition problem on speech signals. This paper uses the way of overlapping sub-frame to short the voice signal, and uses the Energy Frequency Value of each frame to separate the voice signal from the noise signal. In the respect of feature extraction, this paper uses LPCC and pitch frequency to characterize pronunciation organs generated voice sounds (congenital), and uses differential LPCC and differential pitch frequency characterize the difference of pronunciation organ moves to pronounce(acquired). Ultimately, a component feature vector, which is obtained by weighted and united that four features, characterize a particular speaker. In the respect of classification and decision, we use method of vector quantization to complete the classification and sentencing for speakers speech signal. The system that is designed by this paper is achieved based on the Java language and SQL Server 2000 database. Java language is used to implement algorithms needed by speaker recognition, such as voice sampling, preproces