【作者】 王金甲；

【导师】 王成儒；

【作者基本信息】 燕山大学 ， 电路与系统， 2003， 硕士

【摘要】 说话人辨认系统的目的是提取、特征化和辨认表征说话人身份的语音信号信息，因此它在身份验证领域具有广阔的应用前景。本文对噪声环境下闭集文本自由的说话人辨认系统中的两个部分进行了研究。首先本文介绍了各种特征提取的方法，如LPCC、MFCC等，在分析了它们的缺点的基础上，本文根据人耳频率子带独立识别的特性和能量算子，提出了子带能量倒谱特征参数，并用提出的瑞利倒谱提升进行加权，突出说话人的个性信息。其次，在对说话人辨认分类器的特性进行深入分析的基础上，本文提出了概率神经网络说话人辨认方法，并就概率神经网络的模型、训练算法、实时性、噪声鲁棒性、网络结构改进等方面进行了深入研究。(1)提出了一种基于高斯核函数有不同协方差的混合模型的异方差PNN模型及其算法。(2)提出了一种基于最小分类错误准则的概率神经网络的训练算法。(3)提出了一种基于声学分类的并行异方差PNN的实时说话人辨认系统。(4)提出了一种新的基于PNN分类器的噪声自适应更新训练方案。(5)提出了一种新的类条件密度函数估计的PNN模型及其算法。最后，Matlab和Visual C＋＋说话人辨认实验证明了本文提出的参数、模型和算法提高了噪声环境下的系统辨认性能更多还原

【Abstract】 The goal of speaker identification systems is to extract, characterize and identify the information in the speech signal conveying speaker identity. So they have wide applied prospect in the identity authentication domain. The paper involves two parts of closed-set text-independent speaker identification systems in noise.Firstly, the paper introduces shortcomings of the parameters, such as LPCC, MFCC etc. So sub-band energy cepstral feature parameters based on multi-rate sub-band processing and Teager energy operator are described. While the paper discusses the use of rayleigh cepstral liftering to weigh sub-band cepstral coefficients so as to emphasize speaker personality information.Secondly, the paper proposes probabilistic neural networks (PNN) methods of speaker identification, and thoroughly researches the models, training algorithms, real-time fabric, noise robustness and network structure of PNN.(1)The heteroscedastic PNN model with training algorithms that is a mixture of Gaussian basis functions having different variances is considered. (2)An efficient training algorithm for PNN using the minimum classification error criterion is presented.(3)A real-time speaker identification system that introduces acoustic classification information into a heteroscedastic PNN model is proposed.(4)A novel noise-adaptive updating and training approach is proposed for PNN classification.(5)A novel PNN model with training algorithms is proposed for class conditional density estimation.Lastly, speaker identification experiments results using Matlab or VC++ program tool indicate that the proposed parameters, models and algorithms improve identification accuracy in noise更多还原